Macroeconomic Analysis and International Finance 9781783507566, 9781783507559

Citation preview

MACROECONOMIC ANALYSIS AND INTERNATIONAL FINANCE

International Symposia in Economic Theory and Econometrics Series Editor: William A. Barnett Volume 14:

Economic Complexity: Non-Linear Dynamics, Multi-Agents Economies, and Learning Edited by William A. Barnett, Christophe Deissenberg and Gustav Feichtinger

Volume 15:

Modelling Our Future: Population Ageing, Social Security and Taxation Edited by Ann Harding and Anil Gupta

Volume 16:

Modelling Our Future: Population Ageing, Health and Aged Care Edited by Anil Gupta and Ann Harding

Volume 17:

Topics in Analytical Political Economy Edited by Melvin Hinich and William A. Barnett

Volume 18:

Functional Structure Inference Edited by William A. Barnett and Apostolos Serletis

Volume 19:

Challenges of the Muslim World: Present, Future and Past Edited by William W. Cooper and Piyu Yue

Volume 20:

Nonlinear Modeling of Economic and Financial Time-Series Edited by Fredj Jawadi and William A. Barnett

Volume 21:

The Collected Scientific Works of David Cass
Parts A
C Edited by Stephen Spear

Volume 22:

Recent Developments in Alternative Finance: Empirical Assessments and Economic Implications Edited by William A. Barnett and Fredj Jawadi

International Symposia in Economic Theory and Econometrics Volume 23

MACROECONOMIC ANALYSIS AND INTERNATIONAL FINANCE EDITED BY GEORGIOS P. KOURETAS Athens University of Economics and Business, Athens, Greece ATHANASIOS P. PAPADOPOULOS University of Crete, Rethymno, Greece

United Kingdom
North America
Japan India
Malaysia
China

Emerald Group Publishing Limited Howard House, Wagon Lane, Bingley BD16 1WA, UK First edition 2014 Copyright r 2014 Emerald Group Publishing Limited Reprints and permission service Contact: [email protected] No part of this book may be reproduced, stored in a retrieval system, transmitted in any form or by any means electronic, mechanical, photocopying, recording or otherwise without either the prior written permission of the publisher or a licence permitting restricted copying issued in the UK by The Copyright Licensing Agency and in the USA by The Copyright Clearance Center. Any opinions expressed in the chapters are those of the authors. Whilst Emerald makes every effort to ensure the quality and accuracy of its content, Emerald makes no representation implied or otherwise, as to the chapters’ suitability and application and disclaims any warranties, express or implied, to their use. British Library Cataloguing in Publication Data A catalogue record for this book is available from the British Library ISBN: 978-1-78350-755-9 ISSN: 1571-0386 (Series)

ISOQAR certified Management System, awarded to Emerald for adherence to Environmental standard ISO 14001:2004. Certificate Number 1985 ISO 14001

Contents List of Contributors . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .

vii

Editorial Advisory Board Members . . . . . . . . . . . . . . . . . . . . . . . . . . . . .

ix

Guest Editorial: An Overview of the Special Volume on Macroeconomic Analysis and International Finance . . . . . . . . . . . . . . . . . . . . . . . . . . . . .

xi

What Explains House Price Booms? History and Empirical Evidence Michael D. Bordo and John Landon-Lane

1

Economic Growth and Inequality: Evidence from the Young Democracies of South America Manoel Bittencourt

37

Operational Currency Exposure and Firm Level Performance: Evidence from India Anubha Dhasmana

59

Exchange Rates, Fundamentals, and Nonlinearities: A Review and Some Further Evidence from a Century of Data Panayiotis F. Diamandis, Anastassios A. Drakos and Georgios P. Kouretas

85

A Dynamic Gravity Model for Global Bilateral Investment Holdings Konstantinos Drakos, Ekaterini Kyriazidou and Ioannis Polycarpou

125

Does China’s International Competitiveness Fluctuate in Consistency with PPP Equilibrium? Nikolaos Giannellis and Georgios P. Kouretas

153

Linkages between the Eurozone and the South-Eastern European Countries: A VECMX* Analysis Minoas Koukouritakis, Athanasios P. Papadopoulos and Andreas Yannopoulos v

185

vi

Contents

Predicting Economic Activity with Financial Market Data in a Small Open Economy: Revisiting Stylized Facts During Economic Turbulence Petri Kuosmanen and Juuso Vataja What Drives the Bank
Firm Relationship? A Case Study of the Polish Credit Market Małgorzata Pawłowska, Krzysztof Gajewski and Wojciech Rogowski

217

235

List of Contributors Manoel Bittencourt, Department of Economics, University of Pretoria, Pretoria, South Africa (Ch. 2) Michael D. Bordo, Department of Economics, Rutgers University, New Brunswick, NJ, USA; and the National Bureau of Economic Research (NBER), Cambridge, MA, USA (Ch. 1) Anubha Dhasmana, Indian Institute of Management Bangalore, Bangalore, India (Ch. 3) Panayiotis F. Diamandis, Department of Business Administration, Athens University of Economics and Business, Athens, Greece (Ch. 4) Anastassios A. Drakos, Department of Business Administration, Athens University of Economics and Business, Athens, Greece (Ch. 4) Konstantinos Drakos, Department of Accounting and Finance, Athens University of Economics and Business, Athens, Greece (Ch. 5) Krzysztof Gajewski, The National Bank of Poland, Economic Institute, Warsaw, Poland (Ch. 9) Nikolaos Giannellis, Department of Economics, University of Crete, Rethymno, Greece (Ch. 6) Minoas Koukouritakis, Department of Economics, University of Crete, Rethymno, Greece (Ch. 7) Georgios P. Kouretas, Athens University of Economics and Business, Athens, Greece (Chs. 4, 6) Petri Kuosmanen, Department of Economics, University of Vaasa, Vaasa, Finland (Ch. 8) Ekaterini Kyriazidou, Department of Economics, Athens University of Economics and Business, Athens, Greece (Ch. 5) John Landon-Lane, Department of Economics, Rutgers University, New Brunswick, NJ, USA (Ch. 1) Athanasios P. Papadopoulos, Department of Economics, University of Crete, Rethymno, Greece (Ch. 7) vii

viii

List of Contributors

Małgorzata Pawłowska, The National Bank of Poland, Economic Institute, Warsaw, Poland; and Warsaw School of Economics, Warsaw, Poland (Ch. 9) Ioannis Polycarpou, Department of Economics, Athens University of Economics and Business, Athens, Greece (Ch. 5) Wojciech Rogowski, The National Bank of Poland, Economic Institute, Warsaw, Poland; and Warsaw School of Economics, Warsaw, Poland (Ch. 9) Juuso Vataja, Department of Economics, University of Vaasa, Vaasa, Finland (Ch. 8) Andreas Yannopoulos, Department of Economics, University of Crete, Rethymno, Greece (Ch. 7)

Editorial Advisory Board Members Scientific Committee W. A. Barnett, University of Kansas and Center for Financial Stability, USA Ma. Bellalah, University of Jules Verne, France G. Dufre´not, University of Aix-Marseille 2, France B. Dumas, INSEAD, France G. Gallais-Homono, University of Orle´ans, France E. Girardin, University of Aix-Marseille 2, France G. Gregoriou, State University of New York at Plattsburgh, USA K. Hadri, Queen’s University Belfast, UK S. Hall, Leicester University, UK F. Jawadi, University of Evry Val d’Essonne, France B. Mizrach, Rutgers University, USA G. Prat, University of Paris West Nanterre La De´fense and CNRS, France F. Quittard-Pinon, University of Lyon 1, France Ch. Rault, University of Orle´ans, France G. Talmain, University of Glasgow, UK A. Tarrazi, Univesity of Limoges, France T. Tera¨svirta, Aarhus University, Denmark ix

Guest Editorial: An Overview of the Special Volume on Macroeconomic Analysis and International Finance Introduction Since 1997, the Department of Economics of the University of Crete has organized annual international conferences on Macroeconomic Analysis and International Finance. The articles included in this special volume are refereed versions of papers presented at the 17th International Conference on Macroeconomic Analysis and International Finance held at the University Campus, Rethymno from May 29 to June 1, 2013 in collaboration with International Symposia in Economic Theory and Econometrics. The topics discussed in this volume deal with new directions in banking in the aftermath of the 2007
2009 financial crisis; macroeconomic analysis and empirics; exchange rates and nonlinearities; China’s exchange rate policy and competitiveness; economic growth and inequality; and the causes and implications of house price booms. We open this Special Volume with an overview of these papers. In “What Explains House Price Booms? History and Empirical Evidence,” Michael D. Bordo and John Landon-Lane investigate the relationship between loose monetary policy, low inflation, and easy bank credit with house price booms. They conduct their analysis using a panel of 11 OECD countries from 1920 to 2011. They estimate a panel VAR in order to identify shocks that can be interpreted as loose monetary policy shocks, low inflation shocks, bank credit shocks, and house price shocks. The authors show that loose monetary policy played an important role in housing booms along with the other shocks. Furthermore, they show that during boom periods there is a heightened impact of all three “policy” shocks with the bank credit shock playing an important role. However, when they look at individual house price boom episodes the cause of the price boom is not so clear. They argue that the house price boom that occurred in the United States during the 1990s and 2000s was not due to easy bank credit.

xi

xii

Guest Editorial

They consider financial innovation and the shadow banking system to be significant factors in explaining the house price boom. Manoel Bittencourt in “Economic Growth and Inequality: Evidence from the Young Democracies of South America,” investigates whether income growth has played any role on inequality in all nine young South American democracies during 1970
2007. Based on dynamic panel timeseries analysis, the analysis suggests that income growth has played a progressive role in reducing inequality during the period. Moreover, the results suggest that this negative relationship is stronger in the 1990s and early 2000s, a period in which the continent achieved macroeconomic stabilization, political consolidation, and much improved economic performance. On the contrary, during the 1980s (the so-called “lost decade”), the negative income growth experienced by the continent at the time has hit the poor the hardest, which has consequently led to an increase in inequality. The author puts forward the argument that consistent growth, and all that it encompasses, is an important equalizer which should be taken into consideration by policy makers in the designing of economic policies leading to a more equal income distribution. In “Operational Currency Exposure and Firm Level Performance: Evidence from India,” Anubha Dhasmana, looks at the determinants and effects of exchange rate exposure using data on 500 Indian firms over the period 1995
2011. The analysis adopts a new measure of “operational” currency exposure based on foreign currency revenues and costs of firms. Exchange rate volatility is found to be a significant determinant of average firm level exposure with the direction of relationship supporting the presence of “Moral Hazard” in firm’s risk taking behavior. Furthermore, it is shown that large “operational” exposure is associated with significantly lower output growth, profitability, and capital expenditure during episodes of large currency depreciation at the firm level. Together this indicates that the policy makers must take into account the incentive effects of their intervention in foreign exchange markets. Panayiotis F. Diamandis, Anastassios A. Drakos, and Georgios P. Kouretas in their paper “Exchange Rates, Fundamentals, and Nonlinearities: A Review and Some Further Evidence from a Century of Data,” provide a review of the theoretical and empirical literature on the development and empirical testing of the monetary model of exchange rate determination. Furthermore, the authors test the flexible price monetarist variant and the sticky price Keynesian variant of the monetary model. We conduct our analysis employing a sample of 14 advanced economies using annual data spanning the period 1880
2012. The analysis provides strong evidence of the existence of a nonlinear relationship between exchange rates and fundamentals. Therefore, a model capturing the time-varying nature of this relationship was employed by allowing for Markov regime switches for

Guest Editorial

xiii

the exchange rate regimes. The results show that linearity is rejected in favor of an MS-VECM specification, which forms statistically an adequate representation of the data. Two regimes are implied by the model; the one of the estimated regimes describes the monetary model whereas the other matches in most cases the constant coefficient model with wrong signs. Furthermore it is shown that depending on the nominal exchange rate regime in operation, the adjustment to the long run implied by the monetary model of the exchange rate determination came either from the exchange rate or from the monetary fundamentals. Moreover, based on a Regime Classification Measure, we showed that our chosen Markov-switching specification performed well in distinguishing between the two regimes for all cases. In “A Dynamic Gravity Model for Global Bilateral Investment Holdings,” Konstantinos Drakos, Ekaterini Kyriazidou and Ioannis Polycarpou, they argue that global bilateral investment holdings are characterized by a substantial number of zeroes and strong serial persistence. The authors develop a gravity framework and they consider investment behavior at the extensive (participation) margin. They employ alternative dynamic first-order Markov probit models, controlling for unobserved cross-sectional heterogeneity and serial correlation in the transitory error component, in order to explore the sources of persistence. The data support that the strong persistence is driven by true state dependence, implying that past investment experiences strongly impact on the trajectory of future investment holdings. This suggests that inward-investment stimulating policy measures could have a more pronounced effect, since they are likely to induce a permanent change to the future trajectory of inward investment. Nikolaos Giannellis and Georgios P. Kouretas in “Does China’s International Competitiveness Fluctuate in Consistency with PPP Equilibrium?” examine China’s exchange rate policy. The paper argues that there is a widely accepted view, mainly from the United States and the Eurozone, that China manipulates its currency
keeping its value artificially low
in order to boost its exports. Thus, a key question that this paper addresses is whether China’s international competitiveness fluctuates in consistency with PPP equilibrium. Following the PPP equilibrium condition and by employing linear and nonlinear unit root tests, the analysis finds that China’s price competitiveness was not constantly following a disequilibrium process. Furthermore, the two-regime threshold model shows that PPP equilibrium was confirmed in periods of relatively high
compared to the estimated threshold
rate of real yuan appreciation. Moreover, they find that the fixed exchange rate regime cannot ensure external balance since it can neither establish equilibrium in the foreign exchange market, nor confirm that China’s international competitiveness adjustment follows an equilibrium process.

xiv

Guest Editorial

In “Linkages between the Eurozone and the South-Eastern European Countries: A VECMX* Analysis,” Minoas Koukouritakis, Athanasios P. Papadopoulos, and Andreas Yannopoulos assess the impact of the Eurozone’s economic policies on specific South-Eastern European countries, namely Bulgaria, Croatia, Cyprus, Greece, Romania, Slovenia, and Turkey. Since these countries are connected to the EU or the Eurozone and the economic interdependence among them is evolving, the paper carries out the present analysis using the VECMX* framework. Our results indicate that the transition economies in our sample react in a similar manner to changes in international macroeconomic policies. Cyprus and Greece react also in a similar way, but these responses are very small in magnitude. Finally, Turkey behaves in a different way, probably due to the inflationary pressures in its economy. Petri Kuosmanen and Juuso Vataja in “Predicting Economic Activity with Financial Market Data in a Small Open Economy: Revisiting Stylized Facts During Economic Turbulence” examine the predictive content of financial variables above and beyond past GDP growth in a small open economy in the Eurozone. Moreover, the analysis aims to clarify potential differences in forecasting economic activity during periods of steady growth and economic turbulence. The econometric analysis is done with data from Finland and the main findings suggest that the proper choice of forecasting variables relates to general economic conditions. Thus, during steady economic growth, the preferable financial indicator is the short-term interest rate combined with past growth. However, during economic turbulence, the traditional term spread and stock returns are more important in forecasting GDP growth. Furthermore, the main results underline the importance of long-term interest rates in determining the level of the term spread when the central bank implements a zero interest rate policy. Finally, the authors show that during economic turbulence, stock markets are able to signal the expected effects of unconventional monetary policy on GDP growth. In the last paper of the present Special Volume Małgorzata Pawłowska, Krzysztof Gajewski, and Wojciech Rogowski in “What Drives the Bank
Firm Relationship? A Case Study of the Polish Credit Market,” examine the relationship between banks and nonfinancial corporations within Poland (which are considered relationship banking from this point onward). The analysis is conducted with the implementation of panel logit models to test the way different factors affect bank
firm relationships. Three main groups of factors have been investigated: the characteristics of the firm (i.e., size, forms of ownership, and R&D activity); the characteristics of the financial sector (i.e., competition in the banking sector); and macroeconomic conditions. The authors use a large credit database (credit register) of the National Bank of Poland and financial statements from

Guest Editorial

xv

nonfinancial companies. The results of the econometric analysis demonstrate that Polish firms readily establish single-bank relationships. Furthermore, the econometric analysis confirm the testable hypotheses with respect to what determines the relationship banking in Poland from the perspective of characteristics of firms, their crediting banks, and the macroeconomic environment. We wish to thank the discussants, referees, and all participants at the Conference, whose comments have substantially improved the papers presented in this special issue. We are also grateful to the University of Crete, the Bank of Greece, and National Bank of Greece for their generous financial support. Last but not least we would like to thank Ioanna Yotopoulou and Maria Mouzouraki for their superb secretarial assistance, and Pericles Drakos and Kostis Pigounakis for their technical support. Georgios P. Kouretas Athanasios P. Papadopoulos Editors

What Explains House Price Booms? History and Empirical Evidence Michael D. Bordoa and John Landon-Laneb a

Department of Economics, Rutgers University, New Brunswick, NJ, USA; NBER, Cambridge, MA, USA, e-mail: [email protected] b Department of Economics, Rutgers University, New Brunswick, NJ, USA

Abstract Purpose
In this paper we investigate the relationship between loose monetary policy, low inflation, and easy bank credit and house price booms. Method
Using a panel of 11 OECD countries from 1920 to 2011 we estimate a panel VAR in order to identify loose monetary policy shocks, low inflation shocks, bank credit shocks, and house price shocks. Findings
We show that during boom periods there is a heightened impact of all three “policy” shocks with the bank credit shock playing an important role. However, when we look at individual house price boom episodes the cause of the price boom is not so clear. The evidence suggests that the house price boom that occurred in the United States during the 1990s and 2000s was not due to easy bank credit. Research limitations/implications
Shocks from the shadow banking system are not separately identified. These are incorporated into the fourth “catch-all” shock. Practical implications
Our evidence on housing price booms that expansionary monetary policy is a significant trigger buttresses the case for central banks following stable monetary policies based on well understood and credible rules. International Symposia in Economic Theory and Econometrics, Vol. 23 G. P. Kouretas and A. P. Papadopoulos (Editors) Copyright r 2014 by Emerald Group Publishing Limited. All rights reserved ISSN: 1571-0386/doi:10.1108/S1571-038620140000023001

2

M. D. Bordo and J. Landon-Lane

Originality/value of paper
This paper uses historical evidence to evaluate the relative importance of three main causes of house price booms. Our results bring into question the commonly held view that loose bank credit was to blame for the U.S. house price bubble of the later 1990s. Keywords: Asset price booms, monetary policy, bank credit, historical evidence JEL Classifications: E52, N20

Introduction Does expansionary monetary policy lead to house price booms? There is an extensive theoretical, empirical, and policy literature on this topic. The traditional view sees expansionary monetary policy as raising asset prices in general as part of the transmission mechanism of monetary policy. It works through the adjustment of the community’s portfolio as agents substitute from cash to government securities to corporate securities; to equities; to real estate; old masters and commodities
eventually leading to overall inflation. Another view attributed to the Austrian economists in the 1920s and more recently to the BIS sees an environment of low inflation and accommodative monetary policy as creating an environment conducive to asset booms and consequent busts.1 Finally, Schularick and Taylor (2012), Jorda, Schularick, and Taylor (2012), and Christiano, Ilut, Motto, and Rostagno (2010) have emphasized the importance of rapid bank credit growth, possibly driven by financial innovation, in contributing to asset price booms. Asset booms (especially those leading to bubbles) are often followed by busts which can have serious economic effects. There is a long historical incidence of infamous boom busts ranging from the South Sea bubble in the early eighteenth century, many famous stock market crashes in the nineteenth century, the 1929 Wall Street Crash, the UK housing boom bust of 1973, the Nordic crises of the 1980s, the Japanese housing and equity bubble and crash of 1990, and the more recent dotcom and subprime mortgage boom busts. This history keeps repeating itself. The policy implications of asset booms are significant, especially since asset busts have often tended to lead to banking crises and serious and prolonged recessions. To the extent monetary policy is a contributing factor, the question arises whether the monetary authorities should use their policy

1

Related approaches emphasize financial liberalization and innovation accommodated by loose monetary policy as conducive to creating booms.

What Explains House Price Booms? History and Empirical Evidence

3

tools to defuse booms before they turn into busts. A vociferous debate raged in the early 2000s and until the aftermath of the recent financial crisis over the subject of preemptive policy action. Central banks were unwilling to divert much attention away from their traditional concern over price and overall macro stability. However, the tide has recently turned and the new emphasis on macro prudential monetary policy suggests that asset price booms have been elevated to the top level of interest. Finally, the issue still remains that asset price booms in addition to sometimes ending with damaging busts can be the precursors to a future run up in inflation. This then leads to the question of when central banks should tighten their policies to prevent inflation from becoming embedded in expectations. In this paper we develop a method to demarcate asset price booms. We focus on house price booms for 11 OECD countries from 1920 to the present. We then ascertain whether our set of boom events can be related to expansionary monetary policy measured by deviations from Taylor rules as well as to low inflation and bank credit growth. Finally, we use panel vector autoregression techniques to identify orthogonalized shocks and their effect on house prices on average and use historical decompositions to identify the effects of the orthogonalized shocks on individual house price booms.

The Issues Debate swirls over the causes of the subprime Mortgage Crisis of 2007
2008 and the Great Recession of 2007
2009 and the subsequent slow recovery. Two views predominate; the first is that it was caused by global imbalances: a global savings glut in Asia which financed a consumption boom, persistent budget deficits, and current account deficits in the United States and other advanced countries. The second that it reflected domestic imbalances in the United States leading to an unprecedented nationwide housing boom which burst in 2006 precipitating the crisis. This paper focuses on the second view.2

2

The possibility that monetary policy can produce asset price bubbles has also been studied extensively in equilibrium rational expectations models. In such models, poorly designed monetary policies, such as the use of interest rate rules without commitment to a steady long-run inflation rate, can lead to self-fulfilling prophecies and asset price bubbles. Such outcomes are less likely, Woodford (2003) argues, if monetary policymakers follow a clear rule in which the interest rate target is adjusted sufficiently to stabilize inflation. The theoretical literature thus suggests that consideration of the monetary policy environment may be crucial to understanding why asset booms come about.

4

M. D. Bordo and J. Landon-Lane

A key element of the domestic U.S. story is that the Federal Reserve kept monetary policy too loose from 2002 to 2006 which fueled a housing boom that had its origins in a long tradition of policies to encourage home ownership in succeeding administrations, financial innovation, lax regulatory supervision, and oversight and corporate malfeasance. John Taylor (2007, 2009) has led the indictment of the Fed for fueling the housing boom in the early 2000s. Based on the Taylor Rule (1993) he shows that the Federal Funds rate was as low as 3 percentage points below what a simple Taylor rule would generate for the period 2002
2005. Taylor then simulated the path of housing starts had the Fed followed the Taylor rule over the period 2000
2006. His calculations suggest that most of the run up in housing starts from 2002 to 2005 would not have occurred. An earlier OECD study by Ahrend, Cournede, and Price (2008) found a close relationship between negative deviations of the Taylor rule and several measures of housing market buoyancy (mortgage lending, housing investment, construction investment, and real house prices) for a number of OECD countries in the early 2000s. The principal examples are the United States (2000
2006), Canada (2001
2007), Denmark (2001
2004), and Australia (2000
2003). For the euro area as a whole, they find that ECB policy rates are not far below the Taylor rule but for a number of individual members (Portugal, Spain, Greece, Netherlands, Italy, Ireland, and Finland) they are well below it. This evidence as well as evidence in several other papers (Hott & Jokipii, 2012; Assenmacher-Wesche & Gerlach, 2008a) suggests that expansionary monetary policy had a key role to play in fostering recent housing booms, some of which led to devastating busts. Other literature finds evidence linking expansionary monetary policy to equity booms and commodity price booms (Assenmacher-Wesche & Gerlach, 2008b; Pagano, Lombardi, & Anzuini, 2010). There is an extensive earlier literature on the relationship between monetary policy and asset prices in general. Asset prices are viewed as a key link in the transmission mechanism of monetary policy. The traditional view argues that added liquidity causes asset prices to rise as a link in the transmission mechanism of monetary policy actions to the economy as a whole. Another view, the Austrian/BIS view argues that asset price booms are more likely to arise in environments of low and stable inflation and thus asset price booms can arise because monetary policy is geared to credibly stabilizing prices. The traditional view has a long history. Early Keynesian models like Metzler (1951) had central bank operations affecting the stock market directly. Friedman and Schwartz (1963) and later Tobin (1969) and Brunner and Meltzer (1973) spelled out the transmission mechanism following an expansionary Fed open market purchase. It would first affect the

What Explains House Price Booms? History and Empirical Evidence

5

prices (rate of return) on short-term government securities, then via a portfolio balance substitution mechanism, the price (rate of return) of longterm government securities then corporate securities, equities, real estate, old masters, and commodities including gold would be bid up (their returns lowered). Thus, substitution from more to less liquid assets would occur as returns on the former decline relative to the latter. Thus, the impact of expansionary monetary policy will impact securities, assets, and commodities and finally, the overall price level. This view sees asset prices as possible harbingers of future inflation. The Austrian/BIS view which goes back to Hayek, von Mises, Robbins,3 and others in the 1920s posits that an asset price boom whatever its fundamental cause, can degenerate into a bubble if accommodative monetary policy allows bank credit to rise to fuel the boom. This view argues that unless policy-makers act to defuse the boom, a crash will inevitably follow that in turn may cause a serious recession. The Austrians equated rising asset prices with a rise in the overall price level. Although the level of U.S. consumer prices was virtually unchanged between 1923 and 1929, the Austrians viewed the period as one of rapid inflation fueled by loose Federal Reserve policy and excessive growth of bank credit (Rothbard, 1983). The Austrian view has carried forward into the modern discussion of asset price booms. It has been incorporated into the BIS view of Borio and Lowe (2002), Borio and White (2004), and others. They focus on the problem of “financial imbalances” defined as rapid growth of credit in conjunction with rapid increases in asset prices and possibly investment. Borio and Lowe (2002) argue that a build-up of such imbalances can increase the risk of a financial crisis and macroeconomic instability. They construct an index of imbalances based on a credit gap (deviations of credit growth from trend), an equity gap, and an output gap, to identify incipient asset price declines that can lead to significant real output losses, and advocate its use as a guide for proactive action. In this vein Borio (2012) discusses a financial cycle based on property prices and credit growth which has much greater amplitude than the business cycle and when its peak coincides with a business cycle peak, a housing bust, banking crisis, and deep protracted recession can follow, as occurred in 2007. Borio and Lowe (2002) argue that low inflation can promote financial imbalances regardless of the cause of an asset price boom. For example, by generating optimism about the macroeconomic environment, low inflation

3

See Laidler (2003).

6

M. D. Bordo and J. Landon-Lane

might cause asset prices to rise more in response to an increase in productivity than they would otherwise would. Similarly, an increase in demand is more likely to cause asset prices to rise if the central bank is credibly committed to price stability. A commitment to price stability that is viewed as credible, Borio and Lowe (2002) argue, will make product prices less sensitive and output and profits more sensitive in the short-run to an increase in demand. At the same time, the absence of inflation may cause policy makers to delay tightening as demand pressures build up.4 Thus they contend (pp. 30
31) “these endogenous responses to credible monetary policy (can) increase the probability that the latent inflation pressures manifest themselves in the development of imbalances in the financial system, rather than immediate upward pressure in higher goods and service price inflation.”5 Christiano et al. (2010) present historical evidence showing that stock price booms in the United States and Japan often occurred in periods of low inflation. Productivity shocks which raise the natural rate of interest are accommodated by expansion in bank credit which pushes up stock prices. According to their analysis based on a DSGE model, following a Taylor type rule in the face of low inflation will lead to lower interest rates which will further fuel the asset boom. Below we present some evidence consistent with the loose monetary policy explanation for asset price booms and also the Austrian BIS view that regards monetary policy dedicated to low inflation and bank credit expansion as creating an environment conducive to an asset boom. However, the weight attributed to the different explanations differs across historical boom episodes.

4

A related issue to the impact of expansionary monetary policy on asset prices is whether the price index targeted by the central bank should include asset prices. Alchian and Klein (1973) contend that a theoretically correct measure of inflation is the change in the price of a given level of utility, which includes the present value of future consumption. An accurate estimate of inflation, they argue, requires a broader price index than one consisting only of the prices of current consumption goods and services. To capture the price of future consumption, Alchian and Klein (1973) contend that monetary authorities should target a price index that includes asset prices. Bryan, Cecchetti, and O’Sullivan (2000) concur, arguing that because it omits asset prices (especially housing prices), the CPI seriously understated inflation during the 1990s. 5 For evidence that low inflation contributed to the housing booms of the 1990s and 2000s see Frappa and Mesonnier (2010).

What Explains House Price Booms? History and Empirical Evidence

7

Historical Narrative In this section we give a brief overview of house price booms and busts from a historical perspective. For a more detailed discussion on asset price booms and busts from history see Bordo and Landon-Lane (2013).

The 1920s The most famous episode of an asset price boom during the 1920s is the Wall Street Boom beginning in 1923 and ending with the Crash in October 1929. During the boom stock prices rose by over 200%, the collapse from 1929 to 1932 had prices decline by 66%. The boom was associated with massive investment that brought the major inventions of the late nineteenth century, for example electricity and the automobile, to fruition. In addition, major innovations also profoundly changed industrial organization and the financial sector, including the increased use of equity as a financial instrument. The economy of the 1920s (following the sharp recession of 1920
1921) was characterized by rapid real growth, rapid productivity advance, and slightly declining prices, punctuated by two minor recessions. Irving Fisher and other contemporaries believed that the stock market boom reflected the fundamentals of future profits from the high growth industries that were coming on stream and that it was not a bubble. Recent work by McGrattan and Prescott (2003) concurs with that view although many others regard it as a bubble (Galbraith, 1955; Rapoport & White, 1994). Debate continues over the role of expansionary Federal Reserve policy in fueling the boom. In 1932, Adolph Miller, a member of the Federal Reserve Board blamed the New York Fed and its President Benjamin Strong for pursuing expansionary open market purchases to help Britain restore the pound to its prewar parity in 1924 and then again in 1927 to protect sterling from a speculative attack. In both occasions, the U.S. economy was in recession justifying expansionary policy (Friedman & Schwartz, 1963). Miller indicted Strong (who died in 1928) for fueling the stock market boom and the resultant crash. His views were instrumental in legislation in 1933 which prohibited Reserve banks from engaging in international monetary policy actions. As mentioned above the Austrian economists later followed by economists at the BIS saw the 1920s as a credit boom accommodated by monetary policy. Eichengreen and Mitchener (2004) present evidence for the BIS view for the 1920s as a credit boom gone wild, based on their measures of a credit boom (deviations from trend of the ratio of broad money to GDP, the investment ratio, and real stock prices) for a panel of nine countries.

8

M. D. Bordo and J. Landon-Lane

The 1920s also witnessed a major house price boom in the United States from 1923 to 1925. White (2009) argues that the boom was in part triggered by expansionary monetary policy. He finds that deviation from a Taylor rule has some explanatory power for the run up in real housing prices. He also argues that the Fed, established in 1914 to act as a lender of last resort and to reduce the seasonal instability in financial markets, created some elements of a “Greenspan Put”
the view that emerged after Chairman Greenspan engineered a massive liquidity support for the New York money center banks during the October 1987 Wall Street Crash
that the Fed would bail out the financial sector in the event of a crash. Unlike the Wall Street stock market boom, the housing boom bust in the 1920s had little impact on the economy as a whole or on the financial system.

Post-World War II The postwar period has exhibited a large number of housing boom busts. Many of these episodes occurred in an environment of loose monetary policy. We briefly discuss a number of salient episodes.

Asset Booms in the United Kingdom The United Kingdom had a massive house price and stock market boom in 1971
1974, referred to by Tim Congdon (2005) as the Heath Barber Boom after the then Prime Minister and Chancellor of the Exchequer. Congdon (2005) documents the rapid growth in broad money (M4) after the passage of the Competition and Credit Control Bill in 1971 which liberalized the UK financial system and ended the rate setting cartel of the London clearing banks. He shows both rapid growth in M4 and a shift in its composition toward balances held by the corporate and financial sectors away from the household sectors. Following the Friedman and Schwartz (1963) transmission story, the excess cash balances went into equities first and properties second, greatly pushing up their prices. The big asset price booms were soon followed by an unprecedented rise in inflation to close to 20% per year by the end of the 1970s. Congdon (2005) also shows a tight connection between expansion in broad money supply in 1986/1987 and subsequent asset price booms which he calls the Lawson boom after the Chancellor of the Exchequer. As in the 1970s boom, rapid growth in M4 and in its holdings by the corporate and financial sectors fueled a stock market boom which burst in 1987 and a housing boom which burst in 1989. Finally he attributes a big run up in financial sector real broad money holdings in

What Explains House Price Booms? History and Empirical Evidence

9

1997/1998 to an equities boom in the late 1990s and a housing boom which peaked in 2006.

Nordic Asset Booms in the 1980s The Nordic countries, Norway, Sweden, and Finland, all experienced major asset booms and busts in the 1980s. In each country the run up in asset prices followed liberalization of their financial sectors after five decades of extensive controls on lending rates and government control over the sectoral allocation of bank lending. Asset booms were accommodated by expansionary monetary policy as each country adhered to pegged exchange rates which tended to make monetary policy pro-cyclical. In the case of Norway, quantitative restrictions on bank lending were lifted in 1984 without allowing interest rates to rise. Real interest rates were low and sometimes negative. Banks used their newborn freedom to expand lending on a large scale, all of them with a firm desire to increase their market shares. This stimulated a massive real estate boom until 1986. The boom ended with tighter monetary policy in 1986. The legacy of the collapse of the real estate boom and the buildup in bad assets in the commercial banks was a banking crisis in 1991 and a recession (Steigum, 2009). Similar stories occurred in Finland and Sweden (Jonung, Kiander, & Vartia, 2009). Their crises and recessions were much worse than in Norway largely because their currencies were pegged to the DM in the EMS system and they were hard hit by tight German monetary policy in reaction to the high fiscal costs of German reunification.

Japan in the 1980s The Japanese boom-bust cycle began in the mid-1980s with a run up of real estate prices fueled by an increase in bank lending and easy monetary policy. The Bank of Japan began following a looser monetary policy after the Plaza Accord of 1985, to attempt to devalue the yen and ease the upward pressure on the dollar. The property price boom in turn led to a stock market boom as the increased value of property owned by firms raised future profits and hence stock prices (Iwaisako & Ito, 1995). Both rising land prices and stock prices in turn increased firms’ collateral encouraging further bank loans and more fuel for the boom. The bust may have been triggered by the Bank of Japan’s pursuit of a tight monetary policy in 1989 to stem the asset market boom. The subsequent asset price collapse in the next five years led to a collapse in bank lending with a decline in the collateral backing corporate loans.

10

M. D. Bordo and J. Landon-Lane

The decline in asset prices further impinged on the banking system’s capital, making many banks insolvent. This occurred because the collapse in asset prices reduced the value of their capital. Lender of last resort policy prevented a classic banking panic but regulatory forbearance propped up insolvent banks. It took over a decade to resolve the banking crisis and Japan is still mired in slow growth.

House Price Booms of the 1990s and 2000s The subprime mortgage crisis of 2007
2009 in the United States, had its origins in a massive house price boom that began in the 1990s. Its causes include: government policy to encourage housing for a broad swath of the population, loose monetary policy after the tech boom of 2001 to prevent the United States from slipping into Japan style deflation, and “global imbalances” as the newly emerging countries of Asia invested their growing international reserves in safe U.S. Treasury securities. The push to encourage housing in the United States and other countries goes back to the Great Depression of the 1930s when the Roosevelt administration set up the Federal Housing Authority and the GSEs
Fannie Mae and Freddie Mac
to encourage the development of the mortgage market and to provide housing for much of the of the population. In subsequent decades and especially in the 1990s, as argued by Rajan (2011), successive government administrations and Congress, as an attempt to reduce rising income inequality and income stagnation, pushed for affordable housing for low income families using the GSEs and allowed them to reduce their capital requirements. This led the agencies to take on more risk. Lending was encouraged and rising prices raised the GSEs profits leading them to take on more risk. The FHA in the 1990s also took on riskier mortgages, reduced the minimum down payment to 3%, and increased the size of mortgages that would be guaranteed. The housing boom came to fruition in the George W. Bush administration which urged the GSEs to increase their holding of mortgages to low income households (Rajan, 2011, p. 37). Between 1999 and 2007 national house prices doubled according to the Standard and Poor’s Case-Shiller repeat sales index. The private sector also contributed heavily to the boom in an environment of loose regulation and oversight as they recognized that the GSEs would backstop their lending. During this period, lending standards were relaxed and practices like NINJA and NODOC loans were condoned. These developments led to the growth of the subprime and Alt A mortgages which were securitized and bundled into mortgage backed securities and then given triple A ratings. Mortgage backed securities (MBSs) were

What Explains House Price Booms? History and Empirical Evidence

11

further repackaged into collateralized debt obligations (CDOs). Credit default swaps (CDSs) provided insurance on many of these new products. Financial firms ramped up leverage and avoided regulatory oversight and statutory capital requirements with special purpose vehicles (SPVs) and special investment vehicles (SIVs). These factors encouraged a lending boom. As emphasized in this paper the boom was fueled by expansionary monetary policy by the Federal Reserve after the tech boom bust of 2001. Low policy rates were kept in place until 2005 to prevent the economy from slipping into Japan style inflation. Also, as discussed above, the low interest rate environment of the Great Moderation also encouraged risky investment. An additional expansionary impulse may have come from the Asian savings glut (Bernanke, 2005). As China and other countries pegged their currencies at undervalued rates relative to the dollar to encourage export driven growth, they accumulated huge international reserves which were invested in safe U.S. Treasury securities. This imbalance allowed the United States to run a persistent current account deficit which provided fuel for the boom. Other countries had big housing booms in this period as well. The two most notable, Spain and Ireland, benefited from joining the Euro in 1999. This gave them access to massive capital flows from the core countries of Europe on the assumption that currency risk had been eliminated and that in the event of a financial crisis and sovereign debt default they would be bailed out hence reducing country risk. The booms in each case were driven by strong local fundamentals; in the case of Ireland by the development of a high tech export sector and in the Spanish case by rapid growth as Spain emerged as an advanced country. In both these cases loose monetary policy under the ECBs “one size fits” all policy also fueled the boom. Finally, the United Kingdom like the United States had a housing boom partially promoted by government housing policies, financial innovation and high leverage, and loose Bank of England policy.

Summary of Historical Narrative The wide history of house price booms displays considerable evidence of a connection between monetary expansion and booms. It also highlights the importance of bank credit expansion. However, the circumstances of the different episodes varied considerably. House price booms on some occasions reflected real shocks such as rapid immigration and financial liberalization as well as expansionary monetary policy. In the rest of the paper we provide some empirical evidence on the contribution of monetary policy, bank credit expansion, low inflation, and several other factors to a large sample of house price booms.

12

M. D. Bordo and J. Landon-Lane

Identifying House Price Booms Before outlining our econometric approach we first identify asset price booms for real house prices. Our approach to identifying boom/bust periods is a mixture of the formal and the informal. We first use a well-known dating algorithm to find turning points of our asset price series and then use our discretion to select those expansions/contraction pairs that meet our criteria. We do this to avoid some well-known problems that dating algorithms can have in identifying cycles when the underlying data is purely random (see e.g., Cogley & Nason, 1995). The first step of the process is to date the turning points of our asset price series. We do this using the method described in Harding and Pagan (2002) and Pagan and Sossounov (2003). In these two related papers the authors use the method of Bry and Boschan (1971) to date turning points of time series. The dating algorithm of Bry and Boschan (1971) was formulated to mimic the NBER dating process and is successful in dating turning points in time series. For real house prices we look for peaks (troughs) that are higher (lower) than the two nearest observations on each side of the turning point under the constraint that peaks and troughs must alternate. Note, however, that this is the first stage of our process. The second stage of our process we do the following: Once turning points are identified, we inspect each expansion (defined as the period from a trough to the next peak) to see if it fits our definition of an asset price boom. To identify asset price booms we take a “holistic” approach. That is, we first look for expansions that meet our criteria and then we visually inspect each prospective boom to check whether the dates for the boom should be corrected. For example, starting dates are moved to the point where the gradient of the asset price series first significantly picks up if the initial periods of the expansion are relatively flat. The definition of a boom that we use is that a boom is a sustained expansion in asset prices that ends in a significant correction. The expansion is such that the rate of growth is higher than what would be considered usual based on previous cycles. For an expansion to meet the definition of a sustained expansion the expansion must last at least two years and average at least 5% per year for real house and commodity prices and average at least 10% per year for real stock prices. This is similar to the criteria used in Bordo and Wheelock (2009). The second screening that we use is that the price correction that follows the expansion in prices must be greater than 25% of the expansion in price that occurred during the expansion. We believe that this definition rules out secular trends where there can be large increases in asset prices followed by small corrections followed by another large expansion. The booms that we identify all are followed by significant price corrections which suggest that the price expansion was not sustainable and hence a boom/bust period.

What Explains House Price Booms? History and Empirical Evidence

13

Table 1: Identified Real House Price Booms Booms Period Canada 1984
1989 Denmark 1982
1986 2003
2007 France 1930
1935 1971
1980 1985
1991 United Kingdom 1971
1973 1977
1980 1985
1989 Italy 1980
1981 1988
1992 Japan 1986
1991 Netherlands 1958
1964 1976
1978 Norway 1983
1986 Sweden 1974
1979 1985
1990 Switzerland 1971
1973 1983
1989 United States 1921
1925 1976
1979 1984
1989 1997
2006

Corrections

Duration

%Δ

APC

Period

Duration

%Δ

APC

5

57.52

11.5

1989
1998

9

−14.39

−1.6

4 4

53.08 53.49

13.27 13.37

1986
1990 2007
2009

4 2

−25.72 −19.24

−6.43 −9.62

5 9 6

37.69 36.74 30.84

7.54 4.08 5.14

1935
1941 1980
1984 1991
1997

6 4 6

−47.15 −16.76 −16.03

−7.86 −4.19 −2.67

2 3 4

59.27 26.18 67.18

29.64 8.73 16.8

1973
1977 1980
1982 1989
1993

4 2 4

−30.91 −10.17 −26.83

−10.30 −5.08 −6.71

1 4

24.02 49.63

24.02 12.41

1981
1985 1992
1997

4 5

−30.65 −27.58

−7.66 −5.52

5

34.16

6.83

1991
1994

3

−12.98

−4.33

6 2

51.11 36.09

8.52 18.05

1964
1966 1978
1985

2 7

−27.51 −47.75

−13.75 −6.82

3

50.29

16.76

1986
1992

6

−35.2

−5.87

5 5

22.02 36.71

4.4 7.34

1979
1985 1990
1993

6 3

−36.92 −28.58

−6.15 −9.53

2 6

21.2 43.31

10.6 7.22

1973
1976 1989
1997

3 8

−26.01 −36.61

−8.67 −4.58

4 3 5 9

19.12 14.47 18.76 79.38

4.78 4.82 3.75 8.82

1925
1932 1979
1982 1989
1993 2006
2009

7 3 4 3

−12.57 −12.74 −13.01 −33.09

−1.8 −4.25 −3.25 −11.03

The identified house prices booms are reported in Table 1.6 We have annual data on real house prices for 18 countries from 1920 to 2010.7 The approach we follow is similar to that used in IMF WEO (2003), Helbling and Terrones (2003), and Bordo and Wheelock (2009). All of these studies

6 Figures showing the identified house price booms are not reported due to space considerations but are available from the authors upon request. 7 For definitions of the data that we use see the data appendix.

14

M. D. Bordo and J. Landon-Lane

used monthly data for a smaller set of countries. Only the Bordo and Wheelock (2009) study covered the pre-World War II period.

Housing Booms With the exception of France in the 1930s and the United States in the 1920s in Table 1 we did not identify any house price booms before World War II. In the post-World War II period most countries had house price booms in the 1970s and 1980s. The literature at the time associated them with the liberalization of financial markets that occurred after the breakdown of the Bretton Woods system. Many of the boom-busts were dramatic, especially in Japan, the Scandinavian countries, Netherlands, and Switzerland. The United States only experienced mild booms and corrections in that period. Several dramatic episodes occurred in the late 1990s and early 2000s. In particular, the U.S. housing boom of 1997
2006 when real prices rose by 79% and fell by 33% really stands out. There were other significant increases in house prices during the 1990s and 2000s, for example, the United Kingdom from 1996 to 2007, but these are not included in the list of identified house booms as the subsequent correction is not large enough to meet our requirement.

Empirical Analysis In this analysis we pool data from 1920 to 2011 from across the 11 countries in our data set to investigate the impact of loose monetary policy, low inflation, and rapid bank credit growth on asset prices.8 By pooling the data across the twentieth century we are in a sense calculating the impact each of our control variables have on asset prices averaged across all the boom periods that we have identified. Low inflation could reflect the credibility for low inflation that occurred in the 1980s and 1990s and 1920s according to Borio and Lowe (2002) and Eichengreen and Mitchener (2004). In this environment, endogenous asset price booms could arise, financed by easy bank credit, accommodated by the central bank. Loose monetary policy refers to deliberately expansionary monetary policy (as evidenced in the policy rate being below the Taylor rule rate) done for

8 The countries in our sample are Canada, Denmark, France, Great Britain, Italy, Japan, Netherlands, Norway, Sweden, Switzerland, and the United States. Countries are included in our regressions if data is available.

What Explains House Price Booms? History and Empirical Evidence

15

example to prevent deflation as in the 2000s or to stimulate recovery from a recession. The asset price data that we use in the analysis are real house prices. As a measure of monetary policy we use the deviation of a short-term interest rate from the optimal Taylor rule rate.9,10 The optimal Taylor rule rate is given by the following equation: r Taylor = π t þ r  þ 0:5ðyt − yt Þ þ 0:5ðπ t − π  Þ;

ð1Þ

where the output gap term is given by the deviation in log real GDP from its long run trend (as determined by the Hodrick
Prescott filter with a smoothing parameter equal to 100 since the data are annual time series) and the inflation target is 2%. It should be noted that we do not use policy rates in this analysis and that we use for all countries a target interest rate (r*) of 2% with coefficients of 0.5 and 0.5 as in Taylor (1993). Thus, the optimal Taylor rule rate that we use is a very rough measure of the optimal policy rate for each country. The credit variable that we use is the same that is used by Schularick and Taylor (2012). This variable is bank credit as measured by total bank loans as a proportion of total GDP. It should be noted that there is some discrepancy in the literature when it comes to the discussion of credit growth. Some use a broad measure of credit including data from the formal banking sector and nonfinancial institutions. This broad measure of credit is not available for many countries before recent decades and so to be able to include as many house price booms as possible in our analysis we use Schularick and Taylor’s (2012) long-run series on bank loans as our measure of credit. One important issue is that in the recent house price booms in the United States and the United Kingdom the prevalence of credit supplied by non-bank financial institutions via the shadow banking system, has played an important role. Because of this, we have to be careful in interpreting the impact of the credit shock in our analysis. Our credit shock does not include credit innovations originating from the shadow banking system.

9

In another related paper (Bordo & Landon-Lane, 2013), as an alternative measure of monetary policy, we used deviations of the growth of monetary aggregates from Milton Friedman’s (1960) famous rule. This measure may be more relevant for earlier episodes when central banks did not use monetary aggregates as their key policy tool. 10 Using the short rate rather than the policy rate is done because of data availability issues. Using a short-term interest rate is likely to understate the looseness of monetary policy and overstate the tightness. Thus, our estimated impact of loose monetary policy on asset prices is likely to be understated.

16

M. D. Bordo and J. Landon-Lane

The three shocks that we identify in our panel, VAR, are a monetary policy shock, an inflation shock, and a bank credit growth shock. To do this we include the deviation of the short-term interest rate from the optimal rate, inflation, bank credit, and house prices. The deviation of the short-term interest rate from the optimal Taylor Rule rate is included to control for possible correlations between “loose” monetary policy and asset booms. The inflation variable is included to control for possible correlations between low inflation policy and booms and the bank credit variable is included to determine if loose or “easy” bank credit has a role in asset booms. These variables are consistent with the Austrian BIS story as well as recent papers by Schularick and Taylor (2012), Jorda et al. (2012), and Christiano et al. (2010). These are the three main alternative variables that have been argued to play a role in asset booms and the aim of this paper is to use data over the whole twentieth century to shed light on their roles. In order to do this we use a panel vector autoregression (PVAR). The PVAR that we use is yit = αi þ βi Dit þ

p X j=1

Aj yit − j þ

p X

Bj Dit × yit − j þ εit ;

ð2Þ

j=1

where the dummy variable Dit takes the value of 1 if country i is in an asset boom in period t and takes a value of 0 otherwise. This specification allows us to have a PVAR specification for “regular” periods and another specification for “boom” periods. The data is ordered with the interest rate variable first, the inflation variable second, the credit variable third, and the house price variable last. The data vector yit is therefore defined to be 0 yit = ðisit − iTR it ; Δπ it ; Δðl=yÞit ; Δ logðpit ÞÞ ;

ð3Þ

where the price vector is real house prices. Finally it is assumed that εit ∼ ð0; Σ1 Þ in regular periods and εit ∼ ð0; Σ2 Þ during “boom” periods. The PVAR is estimated with country-specific fixed effects but common slope parameters over the panel. Orthogonalized shocks are identified using the standard triangular ordering and Cholesky factor. The interpretations of the shocks are as follows: the first shock is a shock to monetary policy with a negative shock being interpreted as policy is loosening. The second shock is a shock to inflation that is orthogonal to the monetary policy shock. This shock reflects inflation pressures and negative shocks for this shock lowers inflation and lessens pressure for the monetary authority to act. This shock

What Explains House Price Booms? History and Empirical Evidence

17

plays the role of the BIS story where low inflation leads to upward pressure on asset prices because of inaction by the monetary authority. The third shock is a shock to our measure of bank credit, the ratio of bank loans to GDP, which is orthogonal to the first two shocks
the monetary policy shock and the low inflation shock. A positive shock to bank credit is interpreted as an easing of bank credit and a priori you would expect a positive bank credit shock to have a positive impact on asset prices. The last shock is the “catch-all” shock for everything not captured by the first three shocks. There is no interpretation for this shock except that it represents shocks to asset prices that are orthogonal to our monetary policy shock, our inflation shock, and our bank credit shock. Sources of this shock could include financial innovation shocks, external demand shocks, credit expansion from the shadow banking system, and “bubble” behavior shocks. We use the estimates from the PVAR to construct orthogonalized impulse response functions, forecast error variance decompositions, and historical decompositions. The first two represent average effects over the panel while the last
the historical decompositions
are an attempt to look at individual boom episodes across countries. The historical decomposition is constructed in the following way. Suppose that ε ∼ ð0; ΣÞ where Σ is a positive definite matrix. Let P0 be a lower triangular matrix such that Σ = P0 P0 0 . That is, P0 is the Cholesky factor of Σ. Then the orthogonalized shocks uit are constructed via uit = P0− 1 εit :

ð4Þ

The historical decomposition is a counterfactual series that is constructed using only one of the estimated structural shocks. For example, to construct the historical decomposition series based on only the first shock
the monetary policy shock
you would first set u~it = ðu1it ; 0; 0; 0Þ0 ; and then set ε~ it = P0 u~it . The counterfactual residual series, ε~ it , is the set of residuals that would have been created if there were only monetary policy shocks, in this example.

Real House Prices The PVAR given in Equation (2) is estimated with real house prices in the data vector. Using the Schwarz Bayesian information criterion (SBIC) it was determined that the number of lags to use was 3. The orthogonalized

18

M. D. Bordo and J. Landon-Lane

impulse response functions for both the “regular” periods and the “boom” periods are depicted in Figure 1. During “regular” periods the impact of a one standard deviation shock to the monetary policy variable
the deviation of the short-term interest rate from the Taylor rule rate
on real house prices is small. The initial impact is slightly positive and the impact takes about 7 periods to be negative. This is not what you would expect. The other three shocks do appear to impact real house prices as expected. The impact of an increase in inflation is to deflate house prices, the impact of an easing of bank credit is to increase house prices and, of course, the impact of a positive shock to house prices is indeed positive. During “boom” periods the impact of the first three shocks is heightened. The magnitudes of the initial responses are larger and the impact of a tightening of monetary policy is negative after a short period. This result suggests that the three shocks have more of an impact during “boom” periods. In order to check whether the shocks’ impacts are amplified during “boom” periods we next turn to forecast error variance decompositions. These are reported in Figure 2 with the variance decomposition for “regular” periods being represented by the solid line and the variance

Interest Rate Shock

Inflation Shock

0

0.02 0

–0.005 –0.02 –0.01

–0.04 0

5

10

15

20

25

0

5

Credit Shock

10

15

20

25

20

25

Asset Price Shock 0.1

0.04 0.02

0 0 –0.1

–0.02 0

5

10

15

Periods

20

25

0

5

10

15

Periods

Figure 1: Impulse Response Function for Real House Prices (1920
2011). The solid line represents the orthogonalized impulse response function during “regular” periods while the dashed line represents the orthogonalized impulse response function during “boom” periods.

What Explains House Price Booms? History and Empirical Evidence Interest Rate Shock

19

Inflation Shock

0.06

0.2

0.04 0.1 0.02 0

0 0

5

10

15

20

0

5

Credit Shock

10

15

20

15

20

Asset Price Shock 1

0.4

0.8 0.2 0.6 0.4

0 0

5

10 Periods

15

20

0

5

10 Periods

Figure 2: Forecast Error Variance Decomposition for Real House Prices (1920
2011). The solid line represents the FEVD for “regular” periods while the dashed line represents the FEVD for “boom” periods.

decomposition for the “boom period being represented by the dashed line. Figure 2 shows clearly that during “regular” periods the three competing shocks have little impact on real house prices compared to the “other” shock. However, for boom periods the impact of each of the three shocks increases with the largest increase for the bank credit shock. Thus, it appears that easy bank credit plays an important role in “boom” periods which reinforces the view of Schularick and Taylor (2012), Jorda et al. (2012), and Christiano et al. (2010). The impulse response functions and the forecast error decompositions represent average effects across all periods and countries in the panel. One has to be careful to use the results of panel estimates for individual countries as the results presented so far may not be appropriate for individual countries and individual boom periods. In order to check whether the results presented so far are appropriate for individual cases we now turn to a number of important house price booms that have been identified in the literature. The first house price boom we look at is the house price boom that occurred in the United Kingdom from 1985 to 1989. Congdon (2005) attributes much of this house price boom to loose monetary policy.

20

M. D. Bordo and J. Landon-Lane Interest Rate Shock

Inflation Shock

1

1

0.5

0.5

0 1985

1986

1987

1988

1989

0 1985

Credit Shock

1986

1987

1988

1989

Asset Price Shock 1

1

0.5 0.5 0 0 1985

1986

1987

1988

1989

–0.5 1985

1986

1987

1988

1989

Figure 3: Historical Decomposition for the United Kingdom (1985
1989). The solid line in the figure is the actual data while the dashed line is the counterfactual historical decomposition. The historical decomposition for this episode is reported in Figure 3. Here the solid line represents the actual house price data while the dotted line represents the counterfactual series. As you can see, the historical decomposition does accord with the historical narrative in that the monetary policy shock appears to explain a large part of the rise in prices during this episode. The bank credit shock, on the other hand, does not explain much of the increase in house prices. The next house price boom that we look at is the house price boom in Norway from 1983 to 1986. Steigum (2009) attributes much of this boom to an easing of credit restrictions. The historical decompositions for this episode can be found in Figure 4. The historical decompositions agree with the analysis of Steigum (2009) in that the house price series is pretty much wellexplained by the counterfactual series generated with only credit shocks. Another house price boom from that period occurred in Sweden from 1985 to 1990. The historical decompositions, shown in Figure 5, in this case do not allow us to make a case for only one shock being important. It does not appear, however, that credit played a role early in the boom. Two more important house price booms are the house price boom in Japan in the late 1980s and the house price boom in the United States from 1996 to 2006. The historical decompositions for these episodes are found in Figures 6 and 7 respectively. For Japan the house price boom cannot be explained by any one shock
it looks like all four shocks play equal roles.

What Explains House Price Booms? History and Empirical Evidence Interest Rate Shock

Inflation Shock

0.5

0.5

0

0

–0.5 1983

1984

21

1985

1986

–0.5 1983

1984

Credit Shock

1985

1986

Asset Price Shock 1

0.5

0.5 0 0 –0.5 1983

1984

Figure 4:

1985

1986

–0.5 1983

1984

Interest Rate Shock

Inflation Shock 0.5

0

0

1986

1987

1988

1989

1990

–0.5 1985

1986

Credit Shock 0.5

0

0

1986

1987

Figure 5:

1988

1987

1988

1989

1990

1989

1990

Asset Price Shock

0.5

–0.5 1985

1986

Historical Decomposition for Norway (1983
1986).

0.5

–0.5 1985

1985

1989

1990

–0.5 1985

1986

1987

1988

Historical Decomposition for Sweden (1985
1990).

The more interesting case is the United States. The shock that best mimics the actual data is the “other” shock. The monetary policy shock alone can only predict a small increase in house prices during this period while the inflation shock appears to do better after 2002. One thing that is

22

M. D. Bordo and J. Landon-Lane Interest Rate Shock

Inflation Shock

0.4

0.4

0.2

0.2

0 1986

1987

1988

1989

1990

1991

0 1986

1987

Credit Shock

1988

1989

1990

1991

1990

1991

Asset Price Shock

0.4

0.5

0.2 0 0 –0.2 1986

1987

1988

Figure 6:

1989

1990

1991

–0.5 1986

1987

1988

1989

Historical Decomposition for Japan (1986
1991).

Interest Rate Shock

Inflation Shock

1

1

0.5 0.5 0 –0.5 1996

1998

2000

2002

2004

2006

0 1996

1998

2000

2002

2004

2006

2004

2006

Asset Price Shock

Credit Shock 1

1

0.5 0.5 0 –0.5 1996

1998

Figure 7:

2000

2002

2004

2006

0 1996

1998

2000

2002

Historical Decomposition for the United States (1997
2006).

obvious however is that the credit shock predicts that prices should have fallen over this period. The historical decomposition appears to suggest that the house price boom in the United States during the late 1990s and early 2000s was not caused by easy bank credit. Our interpretation of the

What Explains House Price Booms? History and Empirical Evidence

23

result that the “other” shock does the best in predicting the increase in house prices for this U.S. episode is that the house price boom was mainly caused by financial innovation shocks or by bubble behavior. The historical decompositions offer sobering evidence for those who want to use the panel results to claim that bank credit shocks are important in explaining house price booms. While bank credit shocks are important on average for the important house price booms from the 1980s, 1990s, and 2000s it does not appear that easy credit play a significant role in all of them, and certainly does not appear to play a role in the U.S. house price boom of the later 1990s and early 2000s.

Summary of Empirical Analysis The results shown above show that the predictions of the PVAR does a reasonable job of matching the historical narratives of a number of important and major house price booms of the 1980s, 1990s, and 2000s. Loose monetary policy, through deviations from the Taylor Rule, is an important factor in a number of individual house price booms that we look at. Thus, there is evidence that asset price booms could be managed with regular monetary policy instruments. The low inflation
BIS/Austrian explanation
is also not discounted. The historical decompositions show for some individual episodes that low inflation shocks contributed to house price booms. We also find that easy bank credit plays a role. The aggregate results (impulse response functions and forecast error variance decompositions) suggest that bank credit shocks played an important role during boom periods but not during “regular” periods. This is in contrast to the monetary policy and inflation shocks which appear to play a role in all periods with their role magnified during boom periods. The last shock is the “other” shock that captures innovations to house prices that cannot be explained by loose monetary policy, low inflation, or easy bank credit. This is the dominant shock on average and in all individual cases. The interpretation we give to this shock is that the “other” shock is picking up innovation to the financing of houses, changes in underwriting standards not captured by shocks to loans, and other innovations including “bubble” behavior. The “other” shock would also pick up financial innovations that come via the shadow banking system since the credit variable we use only includes loans made from within the formal banking system. An interesting result that comes from the individual historical decompositions is that, while on average it plays an important role, the bank credit shock is not important for the U.S. house price boom of the 1990s and

24

M. D. Bordo and J. Landon-Lane

2000s.11 The U.S. house price boom is mainly explained by “other” shocks and somewhat by the monetary policy and inflation shocks. If there were only bank credit shocks the evidence is that there would not have been any runup of house prices at all. We explore the robustness of this result in the next section.

Robustness Checks and the Role of Bank Credit In order to check our results with respect to credit and its apparent lack of importance for the U.S. house boom of the 1990s/2000s we performed a number of robustness checks. These are reported below.

An Alternative Specification for the PVAR In this robustness check we estimated a slightly different PVAR than the one that we used to get the results reported above. In the alternative PVAR we replace the first variable
the deviation of the short rate from the Taylor rule rate
with the change in interest rates. In this alternative specification we have yit = ðΔisit ; Δπ it ; Δðl=yÞit ; Δ logðpit ÞÞ0 :

ð5Þ

The interpretations of the shocks in this specification are different. The first shock is an interest rate shock of which some component might be due to monetary policy. We cannot identify the monetary policy shocks from other interest rate shocks however. The second shock is the inflation shock orthogonal to the interest rate shocks. The third shock is a bank credit shock once interest rate changes and inflation shocks have been accounted for. Thus, the bank credit shock here represents those innovations to bank credit that are not due to changes in interest rates or inflation (i.e., not due to changes in both the nominal and the real interest rate). The impulse response functions are reported in Figure 8 and the forecast error variance decompositions are reported in Figure 9. The impact of an increase in interest rates is strongly negative and all other shocks are as what we would expect. The forecast error variance decompositions remain similar to the previous specification in that the bank credit shock accounts

11

We also show in section “Other House Price Booms of the 1990s and 2000s” that the bank credit shock did not play an important role in the house price run-ups in the United Kingdom and Canada during the 1990s and 2000s as well.

What Explains House Price Booms? History and Empirical Evidence

25

Inflation Shock

Interest Rate Shock –0.005

0

–0.01

–0.005

–0.015

–0.01 –0.015

–0.02 0

5

10

15

20

0

25

5

10

Credit Shock

15

20

25

20

25

Asset Price Shock

0.04

0.1

0.02 0 0 –0.02

–0.1 0

5

10

15

20

25

0

5

10

Periods

15

Periods

Figure 8: Impulse Response Function (1920
2011): Alternative Specification.

for

Real

Interest Rate Shock

House

Prices

Inflation Shock 0.06

0.2

0.04 0.1 0.02 0

0 0

5

10

15

0

20

Credit Shock

5

10

15

20

15

20

Asset Price Shock 1

0.4

0.8 0.2 0.6 0.4

0 0

5

10 Periods

15

20

0

5

10 Periods

Figure 9: Forecast Error Variance Decomposition for Real House Prices (1920
2011): Alternative Specification. for about a third of the overall forecast error variance. However, the historical decomposition for the U.S. house price boom of the 1990’s is quite different. This is reported in Figure 10. Using this specification it appears that

26

M. D. Bordo and J. Landon-Lane Interest Rate Shock

Inflation Shock

1

1

0.5

0.5

0 1996

1998

2000

2002

2004

2006

0 1996

1998

Credit Shock 1

1

0.5

0.5

0 1996

1998

2000

2002

2000

2002

2004

2006

2004

2006

Asset Price Shock

2004

2006

0 1996

1998

2000

2002

Figure 10: Historical Decomposition for U.S. House Price Boom (1997
2006): Alternative Specification.

about half of the actual rise in house prices can be attributed to interest rate shocks. Again the credit shock plays no role in this particular house price boom. This leads us to believe that the result that bank credit did not play an important role during the U.S. house price boom of the 1990s is quite robust.

An Alternative Ordering in the PVAR One criticism that is leveled at orthogonalized VARs is that the results are not generally robust to the order that the variables appear in the VAR. It might be claimed that the result that credit plays a small role in the U.S. house price boom of the late 1990s and early 2000s is due to the specific ordering. In this robustness check we report the historical decomposition for an alternative ordering where credit is ordered first, followed by the deviation of the interest rate from the Taylor Rule rate, the first difference of the inflation rate, and the first difference of the house price respectively. Figure 11 reports the historical decomposition of the U.S. house price boom of the later 1990s and early 2000s under the alternative ordering

What Explains House Price Booms? History and Empirical Evidence Credit Shock

27

Interest Rate Shock 1

1 0.5

0.5 0 –0.5 1996

1998

2000

2002

2004

2006

0 1996

1998

Inflation Shock 1

0.5

0.5

1998

2000

2002

2002

2004

2006

2004

2006

Asset Price Shock

1

0 1996

2000

2004

2006

0 1996

1998

2000

2002

Figure 11: Historical Decomposition for U.S. House Price Boom (1997
2006): Alternative Ordering.

specification. It is clearly apparent that the results are quantitatively and qualitatively the same as for the original ordering. We are quite confident that the result that bank credit did not play an important role in the U.S. house price boom of the 1990s is robust to ordering. Note that while not reported here the impulse response functions and the forecast error variance decomposition for this alternative ordering are almost identical to the ones reported above.

Other House Price Booms of the 1990s and 2000s During our analysis we only included house price booms that had a subsequent correction that was equivalent to at least 25% of the rise in prices during the boom. This criterion was quite strict and as a result a number of large house price increases were not treated as booms. Booms that only had a small correction or had not yet finished were not included in the boom analysis. There were a number of house price run-ups that started in the 1990s that are nonetheless interesting. Table 2 reports the house price run-ups from the 1990s and 2000s that were not included in the empirical work above. In this extension these house price run-ups are included with the identified house price booms reported in Table 1. When we add these

28

M. D. Bordo and J. Landon-Lane

Table 2: Additional House Price Booms Country Canada France Italy Netherlands Norway United Kingdom Sweden

Years

Price increase (%)

1998
2011 1997
2007 1997
2007 1991
2008 1993
2005 1995
2007 1996
2010

93.6 111.7 57.9 167.3 129.6 159.1 136.7

Interest Rate Shock

Inflation Shock

0.04

0.4

0.02

0.2

0

0 0

5

10

15

20

0

5

Credit Shock

10

15

20

15

20

Asset Price Shock

0.4

1 0.8

0.2 0.6

0

0

5

10 Periods

Figure 12:

15

20

0.4

0

5

10 Periods

FEVD for Extended House Price Boom Definition.

house price run-ups there are some changes in the results.12 The impulse response functions are relatively unchanged but the forecast error variance decompositions are somewhat different. Figure 12 reports the forecast error variance decomposition with the solid blue line depicting the FEVD during “regular” periods and the dashed red line depicting the FEVD during the extended “boom” periods. The important thing to note here is the added influence of the low-inflation

12

Note that we avoid using the word boom in these house price run-ups as the subsequent correction has either not occurred yet or was small.

What Explains House Price Booms? History and Empirical Evidence

29

shock and the lessoned impact of the bank credit shock. This result is consistent with the house price booms and/or run-ups of the 1990s and 2000s being influenced by interest rate and low-inflation shocks rather than credit shocks. To look further into the causes of the individual house price booms we turn again to historical decompositions. Figure 13 shows the historical decomposition for the 1990s house price boom in Canada. Just like the results for the United States in the 1990s bank credit growth does a poor job of explaining the house price boom. Figure 14 reports the historical decomposition for the United Kingdom. Again, just like the case for Canada and the United States, the house price boom/run-up of the 1990s and 2000s is not an easy credit story. For Canada and the United Kingdom there is strong evidence that loose monetary policy played an important role
much stronger than for the United States where deviations from the Taylor rule did not explain much of the run-up in prices. Not all booms in Table 2 are like those reported for Canada, the United Kingdom, and the United States. For France and Italy, credit and interest rate shocks play an important role whereas for the Netherlands, Norway, and Sweden, both the credit and interest rate shocks play no role. Thus, the result that credit shocks, which on average play an important role, did not necessarily cause the house price booms of the 1990s and 2000s is retained when we add in the large house price run-ups from Canada, France, Italy, the Netherlands, Norway, the United Kingdom, and Sweden. The evidence Interest Rate Shock

Inflation Shock

1

0.5

0.5 0 0 –0.5 1998

2000

2002

2004

2006

–0.5 1998

Credit Shock

2000

2002

2004

2006

Asset Price Shock

0.5

1 0.5

0 0 –0.5 1998

2000

2002

2004

2006

–0.5 1998

2000

2002

2004

2006

Figure 13: Historical Decomposition for House Price Boom in Canada (1998
2011).

30

M. D. Bordo and J. Landon-Lane Interest Rate Shock

Inflation Shock

1

1

0.5

0.5

0

0

–0.5 1995

2000

2005

2010

–0.5 1995

Credit Shock 1

0.5

0.5

0

0

2000

2005

2005

2010

Asset Price Shock

1

–0.5 1995

2000

2010

–0.5 1995

2000

2005

2010

Figure 14: Historical Decomposition of House Price Boom in the United Kingdom (1995
2007). suggests that the UK and the U.S. house price booms during the 1990s and 2000s were not caused by credit booms.

Discussion and Conclusion Using a panel VAR we show that the three main explanations, including loose monetary policy, for house price booms all have merit. Averaging across all countries and boom periods the loose monetary policy shock, low inflation shock, and easy credit shock all contribute to house prices and this is magnified during boom periods. There is evidence that loose monetary policy played an important role in some historical episodes
for example the UK house price boom of the 1980s and to a lesser extent Sweden and Japan again in the 1980s. The BIS/Austrian explanation is also not ruled out as there are some episodes where the identified low-inflation shocks did contribute to the run-up of house prices. The same is true for the credit shock explanation as well. However, there is still room for alternative explanations as the majority of the forecast error variance is explained by the “other” shock identified from the panel VAR. The “other” shock is the dominant shock on average and for a majority of individual cases the “other” shock plays the dominant role in explaining the house price boom (or run-up in some cases). The “other” shock picks up all innovations to house prices that are not

What Explains House Price Booms? History and Empirical Evidence

31

explained by deviations of interest rates from the Taylor rule, inflation shocks, or shocks to credit (as measured by total bank loans). The “other” shock could include interest rate or monetary policy shocks that are not measured by the deviation of the interest rate from the Taylor Rule, financial innovation not measured by banks loans, the impact of the shadow banking system on the housing market, or they could just be picking up bubble behavior in house prices. One interesting result we found was that while credit shocks played an important role on average, it did not play a role at all in some of the major house price booms or run-ups of the 1990s and 2000s. In particular for the United States, Canada, and the United Kingdom during this period, the rise in house prices cannot be explained by innovations to loans from the banking sector. In these individual cases the historical decomposition suggests that house prices would have remained stable if only bank credit shocks were present. Two of these countries, the United States and the United Kingdom, have significant shadow banking sectors and it could be that financial innovations or easy credit from the shadow banking system are to blame for the house price booms rather than easy credit through the formal banking system. The housing bust of 2006 in the United States and the subsequent financial crisis and Great Recession then led many policy makers to decide that financial stability should be an important goal of monetary policy along with low inflation (and real macroeconomic stability). This view emphasized the use of the tools of macroprudential regulation such as countercyclical capital requirements and liquidity ratios (Kashyap, Rajan, & Stein, 2008). This case, however, has in part been predicated on the assumption that excessive bank credit was at the heart of the recent boom. The results in this paper cast some doubt on this assumption. The results also cast doubt on the usefulness of using panel estimators in attempting to understand the causes of house price booms in general. The results, especially when we look at individual episodes, suggest that there is no single magic broad spectrum policy prescription for house price booms. The house price booms that we examined all looked different in terms of their causes. However, the evidence that loose monetary policy (along with low inflation and credit expansion) does contribute significantly to booms in house prices suggests that something should be done about it. There is evidence that raising interest rates could have prevented the house price run-ups we saw in the 1990s and 2000s. This subject received considerable attention during the tech boom of the late 1990s and again during the housing boom of the early 2000s. Economists argued both for and against using the tools of monetary policy to defuse asset price booms but little was changed (Bordo & Landon-Lane, 2013).

32

M. D. Bordo and J. Landon-Lane

Our evidence on housing price booms, using data going back nearly 100 years for a number of countries, that expansionary monetary policy is a significant trigger buttresses the case for central banks following stable monetary policies based on well understood and credible rules.

Acknowledgment The authors would like to acknowledge the excellent research assistance provided by Antonio Cusato during this project. All remaining errors are our own.

References Ahrend, R., Cournede, B., & Price, R. (2008). Monetary policy, market excesses and financial turmoil. OECD Economics Department Working Paper No. 597. doi:10.1787/244200148201. Alchian, A., & Klein, B. (1973). On a correct measure of inflation. Journal of Money, Credit, and Banking, 5(1), 173
191. Assenmacher-Wesche, K., & Gerlach, S. (2008a). Financial structure and the impact of monetary policy on asset prices. Swiss National Bank Working Paper No. 2008-16. Retrieved from http://www.snb.ch/n/mmr/reference/working_paper_ 2008_16/source Assenmacher-Wesche, K., & Gerlach, S. (2008b). Monetary policy, asset prices and macroeconomic conditions; A panel VAR study. National Bank of Belgium Working Paper No. 149. Retrieved from http://www.nbb.be/doc/oc/repec/reswpp/ wp149En.pdf Bernanke, B. S. (2005). The global savings glut and the US current account deficit. Remarks by Governor Ben S. Bernanke at the Sandridge Lecture, Virginia Association of Economics, Richmond, VA. Retrieved from http://www.federal reserve.gov/boarddocs/speeches/2005/200503102/ Bordo, M., & Landon-Lane, J. S. (2013). Does expansionary monetary policy cause asset price booms: Some historical and empirical evidence. NBER Working Paper No. 19585. Retrieved from http://www.nber.org/papers/w19585 Bordo, M., & Wheelock, D. (2009). When do stock market booms occur? The macroeconomic and policy of twentieth century booms. In J. Atack & L. Neal (Eds.), The origins and development of financial markets and institutions. New York, NY: Cambridge University Press. Borio, C. (2012). The financial cycle and macroeconomics: What have we learnt? BIS Working Paper No. 395. Retrieved from http://www.bis.org/publ/work395.htm Borio, C., & Lowe, P. (2002). Asset prices, financial and monetary stability: Exploring the nexus. BIS Working Paper No. 114. Retrieved from http://www. bis.org/publ/work114.htm

What Explains House Price Booms? History and Empirical Evidence

33

Borio, C., & White, W. (2004). Whither monetary and financial stability? The implications of evolving policy regimes. BIS Working Paper No. 147. Retrieved from http://dx.doi.org/10.2139/ssrn.901387 Brunner, K., & Meltzer, A. (1973). Mr. Hicks and the ‘Monetarists’. Economica, 40(157), 44
59. Bry, G., & Boschan, C. (1971). Cyclical analysis of time series: Selected procedures and computer programs. New York, NY: NBER. Bryan, M., Cecchetti, S. G., & O’Sullivan, R. (2000). Asset prices in the measurement of inflation. NBER Working Paper No. 8700. Retrieved from http://www.nber. org/papers/w8700 Christiano, L., Ilut, C., Motto, R., & Rostagno, M. (2010). Monetary policy and stock market booms. In Macroeconomic challenges: The decade ahead. A symposium sponsored by the Federal Reserve Bank of Kansas City. Jackson Hole, WY. Retrieved from http://www.kansascityfed.org/publicat/sympos/2010/201008-20-christiano-paper.pdf Cogley, T., & Nason, J. (1995). Effects of the Hodrick-Prescott filter on trend and difference stationary time series: Implications for business cycle research. Journal of Economic Dynamics and Control, 19(1
2), 253
278. Congdon, T. (2005). Money and asset prices in boom and bust. London: Institute of Economic Affairs. Retrieved from http://dx.doi.org/10.2139/ssrn.839866 Eichengreen, B., & Mitchener, K. (2004). The Great Depression as a credit boom gone wrong. Research in Economic History, 22, 183
237. Frappa, S., & Mesonnier, J.-S. (2010). The housing boom of the late 1990s: Did inflation targeting matter? Journal of Financial Stability, 6, 243
254. Friedman, M. (1960). A program for monetary stability. New York, NY: Fordham University Press. Friedman, M., & Schwartz, A. (1963). Money and business cycles. Review of Economics and Statistics, 45b, 32
64. Galbraith, J. K. (1955). The great crash, 1929. Boston, MA: Houghton Mifflin. Harding, D., & Pagan, A. (2002). Dissecting the cycle: A methodological investigation. Journal of Monetary Economics, 49, 405
438. Helbling, T., & Terrones, M. (2003). Asset price booms and busts – Stylized facts from the last three decades of the 20th century. Mimeo. Presented at the European Central Bank workshop “Asset Prices and Monetary Policy,” December 11–12, Frankfurt. Hott, C., & Jokipii, T. (2012). Housing bubbles and interest rates. Swiss National Bank Working Paper No. 2012-07. Retrieved from http:// www.snb.ch/n/mmr/reference/sem_2011_06_23_hott/source/sem_2011_06_23_ hott.n.pdf Iwaisako, T., & Ito, T. (1995). Explaining asset bubbles in Japan. NBER Working Paper No. 5350. Retrieved from http://www.nber.org/papers/w5350 Jonung, L., Kiander, J., & Vartia, P. (2009). The great financial crisis in Finland and Sweden. Northampton, MA: Edward Elgar publishers. Jorda, O., Schularick, M., & Taylor, A. (2012). When credit bites back: Leverage, business cycles and crises. NBER Working Paper No. 17621. Retrieved from http://www.nber.org/papers/w17621

34

M. D. Bordo and J. Landon-Lane

Kashyap, A., Rajan, R., & Stein, J. (2008). Rethinking capital regulation. In Maintaining stability in a changing financial system (pp. 431–471), Federal Reserve Bank of Kansas City. Laidler, D. (2003). The price level, relative prices, and economic stability: Aspects of the interwar debate. BIS Working Paper No. 136. Retrieved from https://www. bis.org/publ/work136.htm McGrattan, E., & Prescott, E. (2003). The 1929 stock market: Irving Fisher was right. Federal Reserve Bank of Minneapolis Research Department Staff Report No. 294. Metzler, L. (1951). Wealth, saving and the rate of interest. Journal of Political Economy, 59(2), 93
116. Pagan, A., & Sossounov, K. (2003). A simple framework for analysing bull and bear markets. Journal of Applied Econometrics, 18, 23
46. Pagano, P., Lombardi, M., & Anzuini, A. (2010). The impact of monetary policy on commodity prices. ECB Working paper No. 1232. Retrieved from http://www. ecb.europa.eu/pub/pdf/scpwps/ecbwp1232.pdf Rajan, R. G. (2011). Fault lines: How hidden fractures still threaten the world economy. Princeton University Press, May. ISBN-10 0691146837. Rapoport, P., & White, E. (1994). Was the crash of 1929 expected? American Economic Review, 84(1), 271
281. Rothbard, M. (1983). America’s Great Depression (4th ed.). New York, NY: Richardson and Snyder. Schularick, M., & Taylor, A. (2012). Credit booms gone bust: Monetary policy, leverage cycles and financial crises, 1870
2008. American Economic Review, 102(2), 1029
1961. Steigum, E. (2009). The boom and bust cycle in Norway. In L. Jonung et al. (Eds.), The great financial crisis in Finland and Sweden (pp. 202
244). Northampton, MA: Edward Elgar Publishers. Taylor, J. B. (1993). Discretion versus policy rules in practice. Carnegie-Rochester Conference Series on Public Policy, 39, 195
214. Taylor, J. B. (2007). Housing and monetary policy. In Housing, housing finance and monetary policy. Kansas City, MO: Federal Reserve Bank of Kansas City. Taylor, J. B. (2009). Getting off track. Stanford, CA: Hoover Press. Tobin, J. (1969). A general equilibrium approach to monetary theory. Journal of Money, Credit and Banking, 1(1), 15
29. White, E. (2009). Lessons from the great American real estate boom and bust of the 1920s. NBER Working Paper No. 15573. Retrieved from http://www.nber.org/ papers/w15573 Woodford, M. (2003). Interest and prices: Foundations of a theory of monetary policy. Princeton, NJ: Princeton University Press.

What Explains House Price Booms? History and Empirical Evidence

35

Appendix: Data sources Real GDP See Michael D. Bordo, Christopher M. Meissner “Does Inequality Lead to a Financial Crisis?” NBER Working Paper No. 17896 Real house price index, 2000 = 100 Detailed description: US [Robert J. Shiller, Irrational Exuberance, 2nd. Edition, Princeton University Press, 2005, 2009, Broadway Books, 2006, also Subprime Solution, 2008, as updated by author], Norway [Norges Bank; Eitrheim, Ø. og Erlandsen, S. “Monetary aggregates in Norway 1819
2003,” 349
376 Chapter 9 in Eitrheim, Ø., J. T. Klovland and J. F. Qvigstad (Eds.), Historical Monetary Statistics for Norway 1819
2003, Norges Bank Occasional Papers no. 35, Oslo, 2004], UK [Department for Communities and Local Government, Housing statistics], France [conseil ge´ne´ral de l’Environnement et du De´veloppement (CGEDD), Home Prices in France, 1200
2012: Historical French Property Price Trends, home price index of Paris], Netherlands [Piet M. A. Eichholtz, 1997, “The long run house price index: The Herengracht index, 1628
1973,” Real Estate Economics, (25), 175
192, this index is based on the transactions of the buildings on the Herengracht, one of the canals in Amsterdam; for recent data the source is OECD], Australia [Stapledon, Nigel David, “Long-term housing prices in Australia and some economic perspectives,” The University of New South Wales, Sep 2007; Australian median city house prices], Spain [before 1970
source: Prados de la Escosura; after 1970 source is OECD], Finland [Hjerppe, Riitta, Finland’s Historical National Accounts 1860
1994: Calculation Methods and Statistical Tables, Jyvaskylan Yliopisto Historian Laitos Suomen Historian Julkaisuja, 24, pp. 158
160; and OECD for recent data], Canada [Statistics Canada and OECD], Japan [The Japan Real Estate Institute, for data between 1910 and 1940 Nanjo, Takashi, “Developments in Land Prices and Bank Lending in Interwar Japan: Effects of the Real Estate Finance Problem on the Banking Industry,” IMES Discussion Paper Series, 2002-E-10, Bank of Japan, 2002]. For the cases of Denmark, Germany, Ireland, Italy, Sweden, Belgium, Switzerland, and New Zealand, the OECD house price index was used. Short term interest rate See Michael D. Bordo, Christopher M. Meissner “Does Inequality Lead to a Financial Crisis?” NBER Working Paper No. 17896 Credit We thank Alan Taylor for providing us with this data

36

M. D. Bordo and J. Landon-Lane

Loans to GDP ratio Total lending, or bank loans, is defined as the end-of-year amount of outstanding domestic currency lending by domestic banks to domestic households and nonfinancial corporations (excluding lending within the financial system). Banks are defined broadly as monetary financial institutions and include savings banks, postal banks, credit unions, mortgage associations, and building societies whenever the data are available. We excluded brokerage houses, finance companies, insurance firms, and other financial institutions. See Michael D. Bordo, Christopher M. Meissner “Does Inequality Lead to a Financial Crisis?” NBER Working Paper No. 17896

Economic Growth and Inequality: Evidence from the Young Democracies of South America Manoel Bittencourt University of Pretoria, Pretoria, South Africa, e-mail: [email protected]

Abstract Purpose
We investigate in this paper whether income growth has played any role on inequality in all nine young South American democracies during the 1970
2007 period. Methodology
Given the nature of our dataset, the methodology is based on dynamic panel time-series analysis. Findings
The results suggest that income growth has played a progressive role in reducing inequality during the period. Moreover, the results suggest that this negative relationship is stronger in the 1990s and early 2000s, a period in which the continent achieved macroeconomic stabilization, political consolidation, and much improved economic performance. On the contrary, during the 1980s (the so-called “lost decade”), the negative income growth experienced by the continent at the time has hit the poor the hardest (the poor usually are the ones to lose their jobs first in recessions), which has consequently led to an increase in inequality. Practical implications
All in all, we suggest that consistent growth, and all that it encompasses, is an important equalizer that affects the poorer progressively and it should not be discarded as a plausible option by policy makers interested in a more equal income distribution. Keywords: Growth, inequality, democracy, South America JEL Classifications: E20, O11, O15, O54 International Symposia in Economic Theory and Econometrics, Vol. 23 G. P. Kouretas and A. P. Papadopoulos (Editors) Copyright r 2014 by Emerald Group Publishing Limited. All rights reserved ISSN: 1571-0386/doi:10.1108/S1571-038620140000023002

38

M. Bittencourt

Introduction and Motivation South America has presented interesting characteristics in terms of longrun development, and particularly in the last 40 years or so the region has seen dramatic economic and political events taking place. To mention a few: erratic, negative, and sometimes only modest economic growth rates (with a slightly positive trend over the period though), relatively high (but not immutable) income inequality, political changes toward more democratic regimes, high rates of inflation and debt crises (and even hyperinflationary and default episodes in some instances), and finally macroeconomic stabilization (in the spirit of Alesina & Drazen, 1991) and political consolidation (in the vein of Przeworski & Limongi, 1997). More specifically, in the last 20 years or so the region has seen a period of unprecedented economic and political stabilization, with economic growth displaying a less erratic trend since the 1990s, a much improved macroeconomic performance (at least in terms of inflation rates and public debt management), slightly lower inequality and, as we speak, not a single reversal to less democratic regimes. Therefore, taking the above eventful economic and political background into account, and the always enriching debate about the role of economic growth in reducing, or increasing, income inequality, we investigate whether income growth has played any role on inequality in the young democracies of South America during the 1970
2007 period. Intuitively, some would argue that economic growth has the ability of raising the boats of the poor higher than others’
via stronger economic activity in sectors which tend to absorb workers situated at the lower tail of the distribution
and consequently of reducing inequality. For instance, given its abundance of natural resources, Brazil has benefited from the recent commodity boom, which tends to affect positively the earnings of workers situated mostly in the (underpaid) primary sector. Furthermore, the same country has experienced a boom in its services sector, for example, catering and hospitality, another sector which tends to absorb a number of workers situated more toward the lower tail of the income distribution (Ferreira, 2010). On the other hand, others would argue that, particularly in developing countries, growth can leave the poorest poorer because of, for instance, trade liberalization and technological changes, features which would leave those at the bottom of the distribution (who also happen to be unskilled) behind. More specifically, if growth is being driven by sophisticated sectors of the economy, which require the use of advanced technologies and therefore a certain degree human capital already in place, the unskilled poor, for lacking the necessary skills required by a modern economy, might well be left behind in the income distribution with respect to those with human capital (Eicher & Garcı´ a-Pen˜alosa, 2001).

Economic Growth and Inequality: Evidence from South America

39

Moreover, in young and rather unequal democracies, with the extension of the political franchise, the poor are able to demand for particular redistributive policies (based on public transfers), which might have an effect on inequality. On the other hand, the established elites, in principle, have their influence diluted by the democratic process, and consequently are not able to influence policy as during the political dictatorship periods, a factor which can also influence inequality. In essence, the political transition to democracy, or the extension of the franchise, tends to be redistributive toward the poor because the elites want to minimize “the threat of revolution” (Acemoglu & Robinson, 2000). All the same, both effects might play a role on how inequality behaves in such an environment and the South American context (given its rather recent political transition) offers us a rich ground for better understanding those possible relationships. The sample we use for the analysis covers the period 1970
2007 and all nine South American young democracies (most of these countries transitioned from military dictatorships to more democratic regimes in the 1980s), and the empirical strategy, since the time-series variation is longer than the cross-sectional one (T > N), is based on dynamic panel time-series methods. The main results reported robustly suggest that income growth has played a small, but statistically significant, role in reducing inequality in the continent over the whole period. In addition, we are able to report that during the so-called “lost decade” of the 1980s, in which income was stagnant and growth displayed even negative rates at times, inequality increased. On the other hand, during 1990
2007, a period in which the continent achieved macroeconomic stabilization and rather decent income growth rates, the Gini coefficient has, in fact, decreased. Therefore, we suggest that growth, and all the environment and institutional framework that it encompasses, has affected sectors which absorb workers located at the left end of the distribution rather progressively and consequently is a potential equalizer that policy makers and other stakeholders interested in a more equal income distribution should not overlook. Moreover, the 1980s long recession hit the poor the hardest, which suggests that, for the sake of equality, recessions (and the bad policies that tend to cause them) should be avoided as well. Intuitively, the poor are the first ones to suffer from higher unemployment and loss of income during recessions, a fact that tend to lead to higher inequality. The literature has provided us with interesting, and sometimes even conflicting, results regarding the role of income growth on inequality. Initially, Psacharopoulos et al. (1995) suggest that income growth has reduced inequality in a sample of Latin American countries in the 1980s. On similar vein, Li, Squire, and Zou (1998) use a sample of 49 countries (they use the then newly released Deininger and Squire (1996) dataset on income inequality) and panel data methods, to report that initial income reduces

40

M. Bittencourt

inequality. However, Easterly (1999), who also uses a panel of countries, reports that growth plays no role on inequality (his growth fixed effects estimates are not statistically significant). In addition, de Janvry and Sadoulet (2000) investigate 12 Latin American countries during the 1970
1994 period, to report that growth, as Easterly had done before, presents negative estimates, but not statistically significant, against inequality. In what is probably the most cited study on the subject, Dollar and Kraay (2002), make use of a sample of 92 developing and developed countries over four decades, and the GMM estimator to report that “growth is good for the poor.” Essentially, they suggest that the shares of the poorest quintile grow “equiproportionately” to average income. On the contrary, Lundberg and Squire (2003), make use of a larger sample than Dollar and Kraay (with 125 countries), to report that economic growth, in fact, increases the Gini coefficient in their broader sample. Moreover, Lopez (2006) makes use of decadal dummies interacted with income to better pinpoint the effect of growth on inequality during different periods of time in his panel of countries (he uses the Dollar and Kraay sample). Essentially, he reports that in the 1990s income growth is associated with higher inequality, and he suggests that the trade liberalization and particular technological changes taking place in the 1990s are behind his results. Furthermore, Foster and Sze´kely (2008) use data from 34 countries during 1976
2000 (their sample is composed mostly of Latin American countries), to report that the incomes of the poor do not increase equiproportionately with average incomes. On a slightly different strand of the literature, Kuznets (1955) suggests that during the processes of long-run economic development that particular societies go through over time, income inequality increases in the short run, just to decrease in the long run. This nonlinear process happens mostly because of the sectoral reallocation taking place in developing countries and the eventual widening access to education in the urban sector. This prediction has prompted researchers to test for a nonlinear relationship between income growth and inequality. In terms of evidence, on one hand, Spilimbergo, London˜o, and Sze´kely (1999) make use of a panel of 108 countries during the 1947
1994 period to report the absence of a Kuznets effect. On the other hand, Barro (2000) tests for the same Kuznets hypothesis and he is able to report some evidence in favor of it in his sample. All in all, this brief, and admittedly non-exhaustive, literature review, and given the importance of the subject, suggests first that there is no clear verdict about the role of income growth on inequality, and second that a better understanding of this relationship is important for policy purposes and also welfare (particularly in developing countries). The former and the latter provide us with enough motivation for a better understanding of the South American context, a continent with its own idiosyncrasies and which,

Economic Growth and Inequality: Evidence from South America

41

given its historical and present characteristics, provides us with a rich ground for a better understanding of this relationship. Ultimately, apart from the regional disaggregation we implement, which allow us to better understand the continent, and also to minimize generalizations which are not always warranted, we take advantage of dynamic panel time-series analysis which permits us to deal with interesting empirical issues
like heterogeneity, and statistical and economic endogeneity biases in dynamic thin panels
which have the potential to improve on previous estimates. The remainder of the paper is as follows: in the next section we explain the data, the methodology used, and then we report and discuss the results obtained. In the third section we provide some final observations.

Empirical Analysis A Look at the Data The dataset we use covers the period 1970
2007 and all nine South American young democracies, namely: Argentina, Bolivia, Brazil, Chile, Ecuador, Guyana, Paraguay, Peru, and Uruguay (T = 38 and N = 9). The income inequality measure that we use is the Gini coefficient, which is simultaneously consistent with the Anonymity, Population, Relative Income, and Dalton principles, and is therefore Lorenz consistent (Sen & Foster, 1997). The Gini coefficients (GINI) come from the UNU-WIDER files. Income per capita (GDP) and the economic growth rates (GROW) come from the Penn World Table (PWT) 6.3 files. The control variables used are relatively standard in the literature and they are as follows: the ratio of exports and imports to real GDP (OPEN), which is a proxy for economic openness; and the government share to GDP (GOV), our proxy for government size, and they both come from the PWT files. The proxy for democracy is the rather popular, and normalized (ranging from zero to one), polity variable (POLITY), which comes from the Polity IV files. The ratio of the liquid liabilities to GDP (M2) is our measure of financial development, inflation (INFLAT), which is given by the usual transformation log (1 + (INFLAT/100)), is our proxy for macroeconomic performance, and urbanization (URBAN), a proxy for long-run development, all come from the World Bank Development Indicators. Information on secondary education (EDUC) is provided by the Barro and Lee (2013) files. As an initial look at the data, in Figure 1 we plot the simple-averaged country time series over the period. In the first panel we plot the growth rates, and we can see not only the “lost decade” in the 1980s, with its

42

M. Bittencourt 9000 mlog_gdp

mgrowth

5 0 –5

8000 7000 6000 5000

–10 1970

1980

1990

2000

2010

1970

1980

1990

2000

2010

t

t

minequality

4 3.9 3.8 3.7 1970

1980

1990

2000

2010

t

Figure 1: Economic Growth (GROW), GDP Per Capita in Logs (GDP), and Inequality (GINI), South America, 1970
2007. Source: PWT and UNU-WIDER files. negative growth rates, but also the positive growth rates taking place after the structural reforms of the 1990s. All in all, growth in the region has been far from consistent, nevertheless it seems that apart from the negative effect of an external shock toward the end of the 1990s (the Asian crisis) and the odd Argentinean crisis (i.e., 2001), the region has experienced better macroeconomic performance from the 1990s onwards than in the 1980s. In the second panel we plot the averaged income per capita in logs over the period. Again, it is not difficult to visualize the “lost decade” and the economic stagnation associated with it, and also the recovery after the 1990s. Overall though, income per capita presents a positive long-run trend in the region, even when taking into account the stagnant 1980s. Finally, in the bottom panel we plot inequality. Over time, the trend in inequality in the region seems to be positive, with a notable fall starting from the mid1990s onwards, which coincides with the stabilization and better economic performance period. Moreover, in Table 1 we provide the correlation matrix among all variables used in the analysis. The statistical correlation that interests us mostly here is the one between the Gini coefficient of income inequality (GINI) and income per capita (GDP), both in logs. This particular correlation is negative and statistically significant at the 5% level, and it indicates

43

Economic Growth and Inequality: Evidence from South America

Table 1: The Correlation Matrix: South America, 1970
2007

GDP GINI OPEN GOV POLITY M2 INFLAT URBAN EDUC

GDP

GINI

OPEN

GOV

POLITY

M2

1 −.377* −.554* .019 .151* −.192* .122* .886* .172*

1 .069 −.196* .210* .247* −.123 −.340* .225*

1 −.183* .175* .555* −.431* −.627* .349*

1 −.159* −.108* .191* −.070 −.171*

1 .222* .041 .216* .681*

1 −.415* −.236* .459*

INFLAT URBAN EDUC

1 .207* −.137*

1 .235*

1

Source: PWT, UNU-WIDER, Polity IV, World Bank and Barro-Lee files. *Represents significance at the 5% level.

(without implying any causation at this early stage) that income growth is associated with lower inequality in the continent. Another notable correlation is the one between inequality and GOV, the proxy for government size, which is negative and significant. This correlation is perhaps indicating that governments have the potential (e.g., via investment in social infrastructure) of reducing inequality. The correlation between POLITY, our indicator for democracy, and inequality is positive and significant, which perhaps illustrates the rather tumultuous first years after redemocratization in the continent (which coincides with the “lost decade”). Moreover, M2, our proxy for financial development, which is positive and significant, indicates that finance is not benefiting the bottom of the income distribution in a progressive manner (perhaps because of informational asymmetries in terms of accessing formal financial markets). In addition, the correlation between URBAN and inequality is negative and significant as well, which suggests that the urban sector of those economies tends to be less unequal than their rural counterparts (probably because the cities offer more dynamic job markets and employment opportunities). Finally, EDUC presents a positive correlation with inequality, and this correlation is possibly capturing the wage premium that people with secondary education (who usually hold technical jobs) get with respect to those with only primary education (who tend to hold low-paid manual jobs). Furthermore, in Figure 2 we plot the OLS regression lines between income growth and inequality in the continent. In the first panel we make use of the whole sample (1970
2007) and the regression line is slightly negative, which weakly confirms the negative statistical correlation reported above and the prospective progressive role of income growth on inequality. In the second panel we plot only the 1980s data, and the line now is positive, which indicates that during the “lost decade” when income was stagnant

44

M. Bittencourt 3.96 3.94 3.92 3.90 3.88 3.86

4 3.9 3.8 3.7 –10

–5

0

5

–10

mgrowth minequality

–5

0

5

mgrowth80 Fitted values

minequality

Fitted values

4.02 4 3.98 3.96 3.94 3.92 –2

0

2

4

6

mgrowth90 minequality

Fitted values

Figure 2: OLS Regression Lines. Economic Growth (GROW) and Inequality (GINI), South America, 1970
2007. Source: PWT and UNUWIDER files. and growth erratic
growth even presented negative rates at the time
the Gini coefficient increased. In the bottom panel we make use of data covering only the 1990s, and what we observe now is that the regression line becomes negative again, and the relationship is stronger than in the first panel, which indicates that during the recovery of the 1990s income growth played a progressive role on inequality in the region. In essence, the above descriptive exercise (with all its caveats), and particularly the regression lines, suggest that there is an overall (negative) economic relationship between income growth and inequality in the continent, which coincidentally enough is stronger in the 1990s, the decade that the continent saw a number of structural reforms taking place (e.g., the import substitution model, and all that it encompasses, came to an end in most countries and particular economic policies that tend to lead to macroeconomic stability were implemented), which in turn might have played a role on income growth and consequently on inequality (by lifting the boats of the poor higher). On the contrary, during the 1980s, or the “lost decade,” income did not play the same sort of progressive role on inequality, perhaps because of the stagnant income and negative growth rates that took place at the time, which tends to hurt mostly the poor.

Economic Growth and Inequality: Evidence from South America

45

Empirical Strategy In terms of empirical strategy, since we have a T > N dataset and also assuming that inequality is a persistent variable, the strategy followed is based on dynamic panel time-series analysis. This is interesting in itself because, apart from dealing with relevant empirical issues in relatively thin panels
heterogeneity and endogeneity biases
the panel time-series analysis allows us to conduct a more disaggregated study of South America, which furthers our knowledge of the region. Basically, we are able to specifically study the South American experience, avoiding particular generalizations and without treating the region either as a dummy or as an outlier to be discarded from the sample.1 Initially though some would argue that cointegration could be an issue, however this is considered to be less of a problem here because inequality is a bounded variable, within zero and one, which by default cannot be nonstationary. This fact alone theoretically precludes us from using estimators that take cointegration between our variables of interest into account. First, some would suggest that by demeaning the data the Fixed Effects (FE) estimator is able to purge the statistical endogeneity problem, which is caused by the presence of the unobserved heterogeneity in the error term (Bond, 2002). On the other hand, Judson and Owen (1999) argue that the issue of the Nickell bias in dynamic T > N panels
of order O(1/T), and which is caused because the FE transformed error term (which purges the country-specific effect) tends to be correlated with the lagged dependent variable
can be a problem even with T = 30 (although we have T = 38 in our data set). Therefore, we implement the bias approximation provided by Bruno (2005), which extends on Bun and Kiviet (2003), and that allows for an unbalanced panel to give “corrected” FE estimates. In this case we use the Anderson and Hsiao option as our baseline consistent estimator. Hence, we use the FE estimator (with robust standard errors clustered at the country level) and the Bruno (2005) correction (LSDVC) which provide consistent estimates in dynamic models when T→∞ (Smith & Fuertes, 2010). The estimated dynamic equation is as follows: GINIit = αi þ βGDPit þ γOPENit þ δGOVit þ EPOLITYit þ εM2it þ ζINFLATit þ ηURBANit þ θEDUCit þ ϑGINIit − 1 þ υit ;

ð1Þ

where GINI is our measure of inequality in logs, GDP is income per capita in logs, OPEN is our proxy for trade openness, GOV is the proxy for

1

For instance, Barro (2000) and Dollar and Kraay (2002) make use of dummies for Latin America.

46

M. Bittencourt

government size, POLITY is our variable for democracy, M2 is a measure of financial development, INFLAT is inflation and it proxies for macroeconomic stability, URBAN is the share of the population living in urban areas and a proxy for long-run development, and EDUC accounts for education. Second, we follow Lopez (2006) and introduce in our FE regressions interaction terms between income growth and dummies covering the 1980s and 1990
2007 respectively, with zeros elsewhere. With those interaction terms we can better understand the role of the “lost decade” on inequality, and the behavior of inequality during the period in which the continent saw structural changes with the implementation of particular economic policies and institutions, like trade liberalization and central bank independence. The estimated dynamic equation is as follows: GINIit = αi þ EITHER β1 GDP80it OR β2 GDP90 − 07it þ γOPENit þ δGOVit þ EPOLITYit þ εM2it þ ζINFLATit þ ηURBANit þ θEDUCit þ ϑGINIit − 1 þ υit ;

ð2Þ

where GDP80 and GDP90
07 are our interaction terms between income growth and the respective decade (1980s) or time period (1990
2007) being studied, with zeros elsewhere. Third, although we use the variables and controls suggested by the previous literature (given data availability), it can be argued that there are some omitted variables or measurement error present. In addition, some would argue that there is reverse causality present (e.g., Persson and Tabellini (1994), Clarke (1995), Forbes (2000), Panizza (2002), and Banerjee and Duflo (2003) all suggest that inequality, in one way or another, determines income growth). We therefore use the Fixed Effects with Instrumental Variables (FE-IV) two-stage Least Squares estimator, and with the Solovian assumption in mind (k = sy)
(Solow, 1956)
we make use of investment (INV), from the PWT 6.3 files, as our external identifying instrument for contemporaneous income growth. The estimates provided by the FE-IV estimator are asymptotically consistent and efficient as T→∞, and it retains the time-series consistency even if the instrument set is only predetermined (Arellano, 2003).2 The estimated second-stage FE-IV dynamic equation is as follows: GINIit = αi þ ðβGDPit = β1 INVit Þ þ γOPENit þ δGOVit þ EPOLITYit þ εM2it þ ζINFLATit þ ηURBANit þ θEDUCit þ ϑGINIit − 1 þ υit ; 2

ð3Þ

Perhaps it is worth mentioning that Bond (2002) argues that GMM-type estimators are not an alternative under T > N because of the overfitting problem.

Economic Growth and Inequality: Evidence from South America

47

with investment in the first-stage regression serving as the identifying instrument for income growth. Essentially, although these countries experienced political transitions and shared similar poor macroeconomic characteristics in the 1980s and early 1990s (which makes the assumption of common slopes plausible), these Fixed Effects estimators account not only for important econometric issues
heterogeneity bias and endogeneity
but also for the fact that some of these countries do indeed present their own economic idiosyncrasies, such as different levels of economic development (e.g., Argentina and Brazil are known to be relatively more developed than Bolivia and Peru).

Results and Discussion In what follows we estimate baseline regressions of income growth against inequality with the most popular control variables previously used by the literature and then we insert other controls also used before in a stepwise fashion for robustness sake. In Table 2 we report the FE dynamic estimates of income growth (GDP) on inequality (GINI) using the variation during the whole period. Essentially, the GDP estimates are all negative and statistically significant against inequality during 1970
2007 (and they are similar, at least in terms of size, to the ones reported by Lopez, 2006). For instance, the GDP estimate in regression five indicates that a point increase in income has the Table 2: Dynamic FE Estimates, South America, 1970
2007 GINI GDP OPEN GOV POLITY M2 INFLAT URBAN EDUC GINI1 F test F* test R2

1 (FE)

2 (FE)

3 (FE)

4 (FE)

5 (FE)

6 (LSDVC)

−.055 (−2.42) .086 (1.59) −.073 (−0.92) −.008 (−0.75)

−.063 (−2.42) .072 (1.22) −.078 (−1.09) −.007 (−0.66)

−.060 (−1.99) .090 (1.70) −.093 (−1.38) −.011 (−0.93)

−.081 (−2.93) .131 (2.96) −.078 (−1.35) −.005 (−0.39)

−.093 (−3.28) .114 (2.73) −.052 (−0.97) −.011 (−0.94)

−.083 .107 −.045 −.012

.019 (2.39)

.024 (1.99) .017 (3.76)

.034 (3.05) .019 (4.26) −.369 (−1.95)

.485 (5.74) 35.16 6.42 0.61

.468 (6.79) 31.97 7.08 0.58

.032 (3.07) .017 (3.43) −.737 (−1.99) .097 (1.37) .461 (6.30) 27.65 8.05 0.49

.031 .018 −.753 .102 .500

.498 (6.08) 41.68 6.23 0.59

.476 (7.70) 29.99 7.46 0.51

T-ratios in parentheses. Number of observations: NT = 342. GINI are the Gini coefficients in logs, GDP is the GDP per capita in logs, OPEN is a measure for trade openness, GOV the government share to GDP, POLITY is a proxy for democracy, M2 are the liquid liabilities to GDP, INFLAT are the inflation rates, URBAN is the share of urban population, and EDUC is secondary education. FE is the Fixed Effects estimator and the LSDVC are the Bruno-corrected estimates.

48

M. Bittencourt

ability of reducing inequality in .09 points. All the same, given the characteristics of the Gini coefficient, we can say that income growth has affected poorer segments of the income distribution in a fashion that has reduced overall inequality (e.g., Dollar & Kraay, 2002; Li et al., 1998). More intuitively, perhaps income growth in South America has relied on the rather flexible services sector (which also includes the large informal sector seen in the continent) and these sectors make use mostly of people with some technical skills (e.g., catering, hospitality, sales, computing, office work, etc.) who happen to be at the lower tail of the distribution, and not so much on highly skilled people with tertiary education (e.g., de Janvry & Sadoulet, 2000). About the controls, trade openness (OPEN) is not entirely significant across the different specifications, however regressions four and five indicate that openness plays a regressive role on inequality. This regressive effect of openness on the Gini coefficient is perhaps illustrating the role of skills (or factor endowments) when processes of trade liberalization take place, or that those benefiting most from openness (which includes technological transfer) are those with tertiary education who happen to be better placed in the distribution (e.g., Barro, 2000; Spilimbergo et al., 1999). The control for macroeconomic performance, inflation (INFLAT), as one would expect in South America, has had the effect of increasing inequality in the continent. This inflation effect is because South America experienced episodes of high inflation, and even some bursts of hyperinflation in countries like Argentina, Bolivia, Brazil, Peru, and Uruguay, and the poor, for not having access to indexed financial assets and for carrying more cash than the better off end up paying the regressive inflation tax (e.g., de Janvry & Sadoulet, 2000; Foster & Sze´kely, 2008).3 Moreover, our proxy for financial development, (M2), presents positive and significant estimates against inequality, however one would expect negative ones (e.g., Li et al., 1998). These negative estimates are perhaps illustrating the fact that the poorest have less experience, and even lack information, on how to make formal financial markets work in their favor in terms of investment opportunities (Foster & Sze´kely, 2008).4

3 These results are in accordance with a parallel literature which deals explicitly with the role of inflation on inequality. For instance, Easterly and Fischer (2001) suggest that the poor from 38 countries consider inflation to be a more pressing problem than the rich, and Bittencourt (2009) reports that the high rates of inflation seen in Brazil in 1983
1994 contributed to increase earnings inequality. 4 On the contrary, a parallel literature suggests that access to finance can reduce inequality via investment in productive activities, for example, Beck, Demirgu¨c¸Kunt, and Levine (2007) and Bittencourt (2010).

Economic Growth and Inequality: Evidence from South America

49

Another interesting result is the one associated with urbanization, (URBAN), which indicates that the long-run process of migration to the cities that has taken place in South America during the 20th century has helped to reduce the Gini coefficient (de Janvry & Sadoulet, 2000 report similar results, however their static random effects estimates are not entirely statistically significant). In other words, it is perhaps easier to find employment (including jobs in the informal sector) and also to acquire education in cities than in rural areas (Kuznets, 1955). In addition, the first lag of inequality (GINI1) is positive and statistically significant (but not approaching one), which confirms the fact that inequality is not only a slow-moving variable, but also indicates that inequality is not nonstationary. Finally, in column six we report the LSDVC estimates using the complete specification and they are in line with the ones reported in column five (i.e., the Nickell bias is not of a significant size in regression five). In Table 3 we report the dynamic FE estimates, but now we use our interaction term between income growth and the decadal dummy for the 1980s (GDP80), with zeros elsewhere. All GDP80 estimates are positive and statistically significant, which indicates that the “lost decade,” or the stagnation of the 1980s, played a regressive role on inequality. All the same, these estimates are somehow expected, in times of macroeconomic instability and lack of income growth, those being affected mostly by recessions and rising unemployment are the poor and unskilled (e.g., de Janvry & Sadoulet, 2000; Psacharopoulos et al., 1995). For instance, a point reduction in income increases inequality in .002 points. Table 3: Dynamic FE Estimates, South America, 1970
2007 GINI GDP80 OPEN GOV POLITY M2 INFLAT URBAN EDUC GINI1 F test F* test R2

1 (FE)

2 (FE)

3 (FE)

4 (FE)

5 (FE)

6 (LSDVC)

.002 (3.18) .098 (2.11) −.056 (−1.23) −.007 (−0.84)

.002 (2.57) .094 (1.62) −.056 (−1.23) −.007 (−0.78)

.003 (2.92) .116 (2.45) −.074 (−1.56) −.011 (−1.15)

.002 (2.52) .144 (3.22) −.055 (−1.67) −.006 (−0.57)

.002 (2.79) .130 (3.01) −.038 (−0.99) −.009 (−0.90)

.002 .125 −.035 −.009

.004 (0.35)

.009 (0.88) .018 (3.75)

.017 (1.52) .020 (4.62) −.286 (−1.48)

.487 (6.03) 36.34 7.28 0.58

.466 (6.96) 33.45 8.13 0.52

.017 (1.46) .019 (3.84) −.469 (−1.41) .048 (0.88) .481 (7.81) 27.09 7.82 0.54

.015 .019 −.484 .053 .524

.487 (6.02) 44.01 7.38 0.56

.481 (8.23) 30.47 7.81 0.52

T-ratios in parentheses. Number of observations: NT = 342. GINI are the Gini coefficients in logs, GDP80 is the GDP per capita in logs in the 1980s with zeros elsewhere, OPEN is a measure for trade openness, GOV the government share to GDP, POLITY is a proxy for democracy, M2 are the liquid liabilities to GDP, INFLAT are the inflation rates, URBAN is the share of urban population, and EDUC is secondary education. FE is the Fixed Effects estimator and the LSDVC are the Bruno-corrected estimates.

50

M. Bittencourt

Furthermore, the estimates of trade openness are all positive and mostly significant this time, clearly suggesting that trade openness in South America benefits mostly those who are highly skilled in the distribution and who can operate incoming technologies. Inflation, given its nature in the continent in the 1980s and early 1990s, keeps its regressive and significant effect on inequality, and the lagged-dependent variable maintains its significant persistence over time. Finally, the Bruno-corrected estimates in column six are in accordance with the ones provided in column five, indicating that the Nickell bias is not so much of an issue in this context. In Table 4 we use our interaction term between income growth and the dummy for the period 1990
2007 (GDP90
07), with zeros elsewhere. This period is interesting because South America achieved macroeconomic stabilization, with the implementation of particular economic policies and institutions (which include trade liberalization and fiscal responsibility laws), and it has also managed to consolidate its democratic institutions. These GDP90
07 estimates are all negative and significant, which indicate that during this period of economic recovery, not to mention the real income growth that has taken place since then, income has played a progressive role on inequality (or that the sort of economic activity taking place at the time has had the ability of positively affecting the incomes of those more toward the bottom of the distribution or sectors which tend to employ poorer workers). For instance, a point increase in income reduces the Gini in .006 points.

Table 4: Dynamic FE Estimates, South America, 1970
2007 GINI GDP90
07 OPEN GOV POLITY M2 INFLAT URBAN EDUC GINI1 F test F* test R2

1 (FE)

2 (FE)

3 (FE)

4 (FE)

5 (FE)

6 (LSDVC)

−.006 (−2.86) .139 (2.78) −.080 (−2.28) −.000 (−0.08)

−.006 (−2.54) .136 (2.28) −.079 (−2.32) −.000 (−0.08)

−.007 (−3.06) .174 (3.80) −.105 (−3.20) −.004 (−0.45)

−.006 (−2.97) .177 (3.94) −.094 (−3.38) −.002 (−0.27)

−.006 (−3.31) .161 (3.95) −.068 (−2.31) −.008 (−0.83)

−.006 .154 −.064 −.008

.002 (0.31)

.007 (0.83) .022 (4.72)

.011 (1.71) .022 (4.94) −.118 (−0.57)

.487 (5.85) 38.17 8.02 0.45

.458 (6.39) 36.40 9.50 0.36

.009 (1.15) .021 (4.11) −.406 (−1.17) .078 (1.27) .461 (6.07) 28.83 8.81 0.47

.009 .021 −.430 .078 .509

.487 (5.82) 46.26 8.14 0.44

.468 (7.11) 31.76 8.44 0.40

T-ratios in parentheses. Number of observations: NT = 342. GINI are the Gini coefficients in logs, GDP90
07 is the GDP per capita in logs in 1990
2007 with zeros elsewhere, OPEN is a measure for trade openness, GOV the government share to GDP, POLITY is a proxy for democracy, M2 are the liquid liabilities to GDP, INFLAT are the inflation rates, URBAN is the share of urban population, and EDUC is secondary education. FE is the Fixed Effects estimator and the LSDVC are the Bruno-corrected estimates.

51

Economic Growth and Inequality: Evidence from South America

Furthermore, trade openness keeps its positive and significant estimates, confirming that trade openness in South America tends to benefit those with higher education mostly, as well as inflation which keeps its regressive effect on inequality. An interesting surprise is that the proxy for government size, (GOV), presents negative and significant estimates. This result is probably reflecting better governance and therefore better spending (e.g., in social infrastructure) of public money (Foster & Sze´kely, 2008). It must be said though, that this variable is highly aggregated and therefore it becomes difficult to draw more solid conclusions about the role of government on inequality (or what type of government participation plays a progressive role on income distribution). No doubt this is an issue that deserves more attention, as long as more disaggregated data become available. The lagged dependent variable keeps its persistent role against itself and the Bruno-corrected estimates provided in column six are in line with the ones provided in column five. In Table 5 we account for possible endogeneity and report the secondstage dynamic FE-IV estimates. All instrumented GDP estimates are negative and statistically significant against inequality. The estimates themselves are bigger in size than the ones reported before because of the external variation provided by our identifying instrument, investment. Essentially, these negative income estimates are somehow confirming the progressive role of growth in reducing the Gini coefficient, or in positively affecting the lower Table 5: Dynamic FE-IV Estimates, South America, 1970
2007 GINI GDP OPEN GOV POLITY M2 INFLAT URBAN EDUC GINI1 F test F* test R2

1

2

3

4

5

−.143 (−1.88) .101 (3.83) −.112 (−2.14) −.008 (−1.25)

−.157 (−2.01) .084 (2.99) −.120 (−2.26) −.008 (−1.15)

−.137 (−1.80) .099 (3.49) −.126 (−2.43) −.011 (−1.61)

−.172 (−2.19) .149 (4.28) −.114 (−2.24) −.004 (−0.61)

−.172 (−2.24) .126 (3.64) −.078 (−1.55) −.012 (−1.50)

.024 (1.48)

.028 (1.78) .017 (2.11)

.041 (2.46) .019 (2.43) −.434 (−2.61)

.437 (4.91) 33.92 6.43 0.50

.429 (4.94) 31.16 7.07 0.50

.037 (2.30) .017 (2.12) −.852 (−3.09) .113 (2.01) .422 (4.97) 26.74 7.65 0.44

.455 (5.21) 40.40 6.26 0.49

.433 (5.06) 28.84 7.19 0.43

T-ratios in parentheses. Number of observations: NT = 342. GINI are the Gini coefficients in logs, GDP is the GDP per capita in logs, OPEN is a measure for trade openness, GOV the government share to GDP, POLITY is a proxy for democracy, M2 are the liquid liabilities to GDP, INFLAT are the inflation rates, URBAN is the share of urban population, and EDUC is secondary education. FE-IV is the Fixed Effects with Instrumental Variables estimator and investment (INV) is the identifying instrument for GDP.

52

M. Bittencourt

tail of the income distribution in South America during the eventful period of 1970
2007. Furthermore, openness and inflation maintain their regressive roles on inequality, and GOV presents once again mostly significant negative estimates. Moreover, the positive and significant M2 estimates indicate the existence of asymmetries in terms of access to formal financial markets, and the negative URBAN estimates suggest again that inequality tends to be lower in the cities. The lagged dependent variable keeps its persistent effect against itself. Lastly, in the first-stage regressions, the identifying instrument, INV, is always statistically significant, and positive, against income growth, and the F tests for overall significance are also statistically significant in all first-stage regressions which minimize the issue of weak instruments (available on request). It is worth mentioning that in all tables above the F* tests suggest that we can reject the null of homogeneous intercepts, which validates the use of the Fixed Effects estimator. Second, the variable EDUC is positive against inequality, however far from statistically significant. A plausible explanation for this non-result for EDUC is the existence of a nonmonotonic relationship between education and inequality, for example, when a country accumulates human capital it can have rising or decreasing inequality, it all depends on the costs of education, particular externalities related to human capital accumulation and the elasticity of substitution between skilled and unskilled workers (Eicher & Garcı´ a-Pen˜alosa, 2001). Another plausible explanation for the (wrong) sign and for the lack of significance is perhaps the fact that the variables OPEN and M2 are the ones capturing the importance of education in terms of the need to have an educated workforce in open economies (e.g., when those countries opened up in the 1990s they started buying technologies which required human capital) and also the need of education for a better use of finance. Third, given that all these countries are young democracies and relatively unequal, we would expect the variable POLITY, our proxy for democratization, to play a progressive role on inequality. Essentially, without the constraints imposed by those military juntas, demand for redistribution would be higher in those young democracies, and perhaps inequality lower (Acemoglu & Robinson, 2000). On the other hand, in dictatorships the rich would be able to lobby for particular economic policies that would benefit themselves (Barro, 2000). Moreover, in some of those countries, the first years of democratization were marred by poor macroeconomic performance (Bittencourt, 2012), a factor which might be affecting the results somehow. Overall, given the nature of the estimates reported, and also that democratization took place in different countries at different points in time (but mostly in the 1980s), it is plausible that those effects are cancelling each other out in South America. All the same, the issue of democracy and inequality deserves more attention.

Economic Growth and Inequality: Evidence from South America

53

Also important to mention, the income growth estimates reported above are in line with some of the previous studies, for example, Li et al. (1998) and Dollar and Kraay (2002), at least in terms of income growth and reduced inequality. On the other hand, our estimates contrast with the ones provided by de Janvry and Sadoulet (2000) and also Lopez (2006). This is perhaps because we have more data (which includes the economic recovery of the 1990s and 2000s) and take advantage of better estimation analysis that deals with heterogeneity and endogeneity in dynamic panels than de Janvry and Sadoulet (2000). In the case of Lopez (2006), we find that, at least in South America, the period of economic recovery between 1990 and 2007 has seen a decrease in inequality instead. The latter highlights the importance of regional disaggregations that can have the effect of reducing unwarranted generalizations about the role of income growth on inequality. All in all, the role of the various changes taking place in different regions of the world in the 1990s (e.g., the end of the Cold War and the Washington consensus) is an interesting issue that deserves more attention. In a nutshell, by accounting for heterogeneity bias and endogeneity concerns in dynamic panel time-series, we find that income growth plays a robust progressive role on inequality in South America. In addition, the long economic and political instability of the 1980s, illustrated by a long and protracted recession, had the effect of increasing the Gini coefficient in the continent, which confirms the long-held view that recessions hurt the poor the hardest. Furthermore, coincidentally enough, after the reforms, stabilization and consolidation of the early 1990s, economic activity resumed and income growth has played the expected role in reducing inequality, which highlights once again the importance of consistent economic activity (and all that it encompasses) in reducing inequality.5

Final Observations We have investigated whether income growth increased, or reduced, income inequality in the young democracies of South America in 1970
2007. The

5 In addition, we run regressions with the inflation tax (INFLAT/(1 + INFLAT)) on the RHS and the results are in line with the ones provided above. We also use the Chinn and Ito (2006) index for financial openness as an alternative to the liquid liabilities and the results regarding inequality and growth are robust. Finally, we run regressions with the decile ratios, however the results, probably because the data are very fragmented, are not significant. Available on request. Furthermore, in the appendix we provide some extra results from regressions that include income per capita and the interaction terms on the RHS and they confirm the estimates reported above.

54

M. Bittencourt

results, based on dynamic panel time-series analysis, suggest that income growth has had the effect of reducing inequality in the continent. Moreover, the protracted recession and poor macroeconomic performance seen in the 1980s has hurt the poor the hardest, with inequality increasing at the time. Furthermore, after the stabilization, and the structural reforms taking place in the 1990s, income growth has played a progressive role on inequality (by affecting sectors which tend to absorb poorer workers, e.g., the primary and the services sectors). In addition, the results suggest that poor macroeconomic performance, in terms of high inflation, tends to be regressive on inequality, therefore the importance of institutions (e.g., central bank independence) and policies (e.g., fiscal rules) which are conducive to macroeconomic stability and therefore growth, and that were implemented in South America only in the 1990s. Moreover, although education per se is not entirely meaningful in the above analysis (probably because of particular nonmonotonicities), our proxies for openness and financial development indicate that education plays an indirect role on inequality, in the sense that human capital is an important safety net in open and technologically driven societies, and also because it allows people to make good and productive use of finance. Finally, the long-run process of urbanization taking place in the continent, seems to offer better prospects in terms of higher earnings to the poor, or lower inequality, than life in rural areas. Future work can be extended to other regions, for example, it would be interesting to see whether the recent income growth seen in sub-Saharan Africa has played any role on poverty, since poverty is a more pressing issue in the region. Moreover, the Brazilian case, given its historical inequality and recent economic growth, is an interesting case to investigate as well as the South African case with its more structural inequality and modest growth rates. All the same, such disaggregations can shed some light on how income, inequality, and other welfare variables behave in different regions and continents. Furthermore, with more historical data on income and inequality and by using a panel smooth transition regression model (Gonza´lez, Tera¨svirta, & van Dijk, 2005) we could test for the Kuznets hypothesis in the South American continent, which would certainly enrich our knowledge of the region. To conclude, we suggest that growth (and all the institutional framework and environment that it encompasses) is a prospective
and perhaps nonintrusive
equalizer which should not be overlooked by policy makers and other stakeholders interested in a more equal income distribution, particularly in developing countries. It is also always worth mentioning that without economic activity, or growth, it becomes difficult to fund particular, and alternative, redistributive policies like public transfers which specifically

Economic Growth and Inequality: Evidence from South America

55

target the poor. Ultimately, growth matters and it can be good for all, including the poor.

Acknowledgments I thank seminar participants at Pretoria, ERSA Public Economics Workshop in Pretoria, 17th ICMAIF in Crete, Melanie Khamis, an ERSA reviewer, and an anonymous referee for comments. Financial support from ERSA is acknowledged.

References Acemoglu, D., & Robinson, J. (2000). Why did the West extend the franchise? Democracy, inequality, and growth in historical perspective. Quarterly Journal of Economics, 115(November), 1167
1199. Alesina, A., & Drazen, A. (1991). Why are stabilisations delayed? American Economic Review, 81(5), 1170
1188. Arellano, M. (2003). Panel data econometrics. Oxford: Oxford University Press. Banerjee, A., & Duflo, E. (2003). Inequality and growth: what can the data say? Journal of Economic Growth, 8(3), 267
299. Barro, R., & Lee, J.-W. (2013). A new data set of educational attainment in the world, 1950
2010. Journal of Development Economics, 104, 184
198. Barro, R. J. (2000). Inequality and growth in a panel of countries. Journal of Economic Growth, 5(1), 5
32. Beck, T., Demirgu¨c¸-Kunt, A., & Levine, R. (2007). Finance, inequality, and the poor. Journal of Economic Growth, 12(1), 27
49. Bittencourt, M. (2009). Macroeconomic performance and inequality: Brazil, 1983
94. The Developing Economies, 47(1), 30
52. Bittencourt, M. (2010). Financial development and inequality: Brazil 1985
1994. Economic Change and Restructuring, 43(2), 113
130. Bittencourt, M. (2012). Democracy, populism and hyperinflation: Some evidence from Latin America. Economics of Governance, 13(4), 311
332. Bond, S. (2002). Dynamic panel data models: A guide to micro data methods and practice. Portuguese Economic Journal, 1(2), 141
162. Bruno, G. (2005). Approximating the bias of the LSDV estimator for dynamic unbalanced panel data models. Economic Letters, 87(3), 361
366. Bun, M. J. G., & Kiviet, J. F. (2003). On the diminishing returns of higher order terms in asymptotic expansion of bias. Economic Letters, 79(2), 145
152. Chinn, M., & Ito, H. (2006). What matters for financial development? Capital controls, institutions, and interactions. Journal of Development Economics, 81(1), 163
192.

56

M. Bittencourt

Clarke, G. R. G. (1995). More evidence on income distribution and growth. Journal of Development Economics, 47(2), 403
427. de Janvry, A., & Sadoulet, E. (2000). Growth, poverty, and inequality in Latin America: A causal analysis, 1970
94. Review of Income and Wealth, 46(3), 267
287. Deininger, K., & Squire, L. (1996). Measuring income inequality: A new data base. The World Bank Economic Review, 10(3), 565
591. Dollar, D., & Kraay, A. (2002). Growth is good for the poor. Journal of Economic Growth, 7(3), 195
225. Easterly, W. (1999). Life during growth. Journal of Economic Growth, 4(3), 239
275. Easterly, W., & Fischer, S. (2001). Inflation and the poor. Journal of Money, Credit and Banking, 33(2), 160
178. Eicher, T., & Garcı´ a-Pen˜alosa, C. (2001). Inequality and growth: The dual role of human capital in development. Journal of Development Economics, 66(1), 173
197. Ferreira, F. H. G. (2010). Distributions in motion: economic growth, inequality, and poverty dynamics. Policy Research Working Paper Series 5424, The World Bank. Forbes, K. J. (2000). A reassessment of the relationship between inequality and growth. The American Economic Review, 90(4), 869
887. Foster, J. E., & Sze´kely, M. (2008). Is economic growth good for the poor? Tracking low incomes using general means. International Economic Review, 49(4), 1143
1172. Gonza´lez, A., Tera¨svirta, T., & van Dijk, D. (2005). Panel smooth transition regression models. Working Paper Series in Economics and Finance No. 5424, Stockholm School of Economics. Judson, R. A., & Owen, A. L. (1999). Estimating dynamic panel data models: A guide for macroeconomists. Economics Letters, 65(1), 9
15. Kuznets, S. (1955). Economic growth and inequality. American Economic Review, 45(1), 1
28. Li, H., Squire, L., & Zou, H.-f. (1998). Explaining international and intertemporal variations in income inequality. The Economic Journal, 108(446), 26
43. Lopez, H. (2006). Growth and inequality: Are the 1990s different? Economic Letters, 93(1), 18
25. Lundberg, M., & Squire, L. (2003). The simultaneous evolution of growth and inequality. The Economic Journal, 113(487), 326
344. Panizza, U. (2002). Income inequality and economic growth: Evidence from American data. Journal of Economic Growth, 7(1), 25
41. Persson, T., & Tabellini, G. (1994). Is inequality harmful for growth? The American Economic Review, 84(3), 600
621. Przeworski, A., & Limongi, F. (1997). Modernisation: Theories and facts. World Politics, 49(2), 155
183. Psacharopoulos, G., Morley, S., Fiszbein, A., Lee, H., & Wood, W. C. (1995). Poverty and income inequality in Latin America during the 1980s. Review of Income and Wealth, 41(3), 245
264.

Economic Growth and Inequality: Evidence from South America

57

Sen, A., & Foster, J. E. (1997). On economic inequality. Oxford: Oxford University Press. Smith, R., & Fuertes, A.-M. (2010). Panel time-series. Unpublished manuscript, London. Solow, R. (1956). A contribution to the theory of economic growth. Quarterly Journal of Economics, 70(1), 65
94. Spilimbergo, A., London˜o, J. L., & Sze´kely, M. (1999). Income distribution, factor endowments, and trade openness. Journal of Development Economics, 59(1), 77
101.

58

M. Bittencourt

Appendix In this appendix we provide some extra results which come from the FE and LSDVC regressions with income on the right hand side combined with each interaction term in turn, and then both interaction terms on their own. The GDP estimates in columns one, two, three, and four confirm the estimates reported above, that is, that income growth has played a progressive role on inequality during the whole period. Moreover, the interaction terms in these equations confirm the regressive role of the stagnation of the 1980s on inequality and the progressive role of the recovery of the 1990s onwards. Finally, in regressions five and six, when we regress both interaction terms on their own against inequality, the 1980s interaction term is not entirely statistically significant this time, however the progressive role of the economic recovery of the 1990s and 2000s on inequality is confirmed. Table A.1:

Dynamic FE Estimates, South America, 1970
2007

GINI

1 (FE)

2 (LSDVC)

3 (FE)

4 (LSDVC)

−.077 (−3.03) .001 (2.24)

−.069 .001

−.067 (−2.92)

−.058

−.005 (−2.96) .163 (3.95) −.084 (−2.32) −.009 (−0.98)

−.005 .155 −.076 −.009

.016 (1.94) .020 (3.84) −.533 (−1.52) .091 (1.37) .434 (4.76) 26.59 9.20 .44

.015 .020 −.549 .094 .472

GDP GDP80 GDP90
07 OPEN GOV POLITY M2 INFLAT URBAN EDUC GINI1 F test F* test R2

.134 (2.97) −.059 (−1.30) −.010 (−1.03)

.129 −.053 −.010

.024 (1.98) .018 (3.49) −.617 (−1.78) .072 (1.13) .450 (5.58) 25.22 8.34 .48

.023 .019 −.627 .074 .486

5 (FE)

6 (LSDVC)

−.001 (−1.60) −.008 (−3.29) .162 (3.98) −072 (−2.40) −.008 (−0.79)

−.001 −.008 .154 −.071 −.008

.009 (1.29) .021 (4.18) −.423 (−1.20) .094 (1.49) .461 (5.97) 25.81 8.80 .47

.009 .021 −.449 .097 .515

T-ratios in parentheses. Number of observations: NT = 342. GINI are the Gini coefficients in logs, GDP is the GDP per capita in logs, GDP80 is the GDP per capita in logs in the 1980s with zeros elsewhere, GDP90
07 is the GDP per capita in logs in 1990
2007 with zeros elsewhere, OPEN is a measure for trade openness, GOV the government share to GDP, POLITY is a proxy for democracy, M2 are the liquid liabilities to GDP, INFLAT are the inflation rates, URBAN is the share of urban population, and EDUC is secondary education. FE is the Fixed Effects estimator and the LSDVC is the Bruno-corrected estimates.

Operational Currency Exposure and Firm Level Performance: Evidence from India Anubha Dhasmana Indian Institute of Management Bangalore, Bangalore, India, e-mail: [email protected]

Abstract Purpose
To study the determinants and effects of “Operational” exchange rate exposure resulting from the mismatch between cost and revenues of the firms by using data on 500 Indian firms. Design/methodology/approach
We conduct detailed empirical analysis of the determinants of firm level exposure and their impact using panel regression techniques and conduct several robustness tests to confirm the validity of these results. Findings
Among other factors, exchange rate volatility appears as a significant determinant of average firm level exposure with the direction of relationship supporting the presence of “Moral Hazard” in firm’s risktaking behavior. Further large “operational” exposure is associated with significantly lower output growth, profitability, and capital expenditure during episodes of large currency depreciation at the firm level. Research limitations/implications
This paper leaves several questions to be answered. Further research is called for to explore the nature of distortions in the production process encouraged by exchange rate volatility and their impact on firm level productivity. Looking at the relationship between the use of financial and operational hedges is another fruitful area of future research. Practical implications
Our results have important implications for policy makers worried about mitigating the impact of exogenous shocks. Implicit International Symposia in Economic Theory and Econometrics, Vol. 23 G. P. Kouretas and A. P. Papadopoulos (Editors) Copyright r 2014 by Emerald Group Publishing Limited. All rights reserved ISSN: 1571-0386/doi:10.1108/S1571-038620140000023003

60

A. Dhasmana

and explicit guarantees with regards to the value of exchange rate tend to raise the vulnerability of the economy to exchange rate shocks at same time that they encourage capital expenditures and possibly output growth during “normal” times. Our findings indicate that the policy makers must take into account the incentive effects of their intervention in foreign exchange markets. Originality/value
Unlike the existing papers in the literature, we use a measure of “operational” currency exposure based on foreign currency revenues and costs of firms. In most of the existing papers the focus is on the mismatch between the currency denomination of assets and liabilities. Little attention has been paid to the currency mismatch between costs and revenues of the firms. Such “operational” mismatches are potentially equally important and deserve attention of policy makers and academics alike. Keywords: Operational currency exposure, moral hazard, exchange rate volatility JEL Classifications: F30, F31, F43

Introduction Impact of exchange rate movements on economic performance is one of the key questions in international economics. Theoretically, exchange rate movements can affect economic performance through a number of channels, such as raising the cost of imported inputs relative to other factors of production, providing exporters with a relative cost advantage relative to foreign competitors, or generating higher borrowing costs and a contraction in lending. Which of these channels becomes the dominant one is therefore a question of empirical investigation. This paper looks at the firm level exchange rate exposure and its impact on firms’ performance during episodes of large currency depreciations using data on 500 Indian firms for the period 1995
2011. We use the measure of currency exposure suggested by Bodnar and Marston (2001)1 who present a measure of exchange rate exposure elasticity based on differences in revenues and costs of emerging market firms.2 Exchange rate exposure

1 Unpublished manuscript available at http://finance.wharton.upenn.edu/weiss/ wpapers/2000/00-3.pdf 2 Details of this measure are presented in the next section.

Operational Currency Exposure

61

elasticity is defined as the percentage change in firm’s cash flow in response to a one percent change in exchange rate.3 Two key results emerge out of our analysis. First, exchange rate volatility is inversely associated with operational exposure elasticity. In other words, periods of low exchange rate volatility are associated with higher average absolute exposure among the Indian firms and vice versa. This supports the “Moral Hazard” hypothesis of risk-taking behavior among Indian firms. Periods of low exchange rate volatility (associated with greater central bank intervention) encourage firms to take on more risk through higher operational exposure to exchange rate changes as measured by the absolute level of exposure elasticity. One would not expect to see such an association if unhedged exchange rate exposure were a result of incomplete markets. Second, high exposure elasticity has a significant adverse impact on firm level performance during episodes of “large” currency depreciations (“large” currency depreciations are defined in detail below). Using alternative measures of firm level performance such as output growth, earnings per share, and capital expenditure, we find that the firms with higher exposure elasticity perform much more poorly compared to the rest during episodes of large Rupee depreciations even though they seem to have a higher output growth and capital expenditure than the rest during “normal” times. Together these results suggest that Indian policy makers should be careful regarding the incentive effects of their intervention in the foreign exchange market. Further, there is a need to focus on “operational” mismatches arising out of mismatches in cost and revenue streams of firms apart from the usually discussed asset
liability mismatches. Our paper is related to a large body of microeconomic literature looking at the impact of exchange rate fluctuations on firm level performance. A section of this literature looks at the impact of exchange rate changes on firm’s value measured by stock returns. Examples of this literature include Adler and Dumas (1984), Jorion (1990), Bodnar and Wong (2000), Dominguez and Tesar (2006), and Parsley and Popper (2006). Papers in this group try to measure exchange rate “exposure” by regressing industry/ firm level stock returns on exchange rate changes and variables controlling for overall market returns and global and domestic economic shocks. Most studies in this literature find puzzlingly small estimates for exchange rate exposure. According to Bodnar and Marston (2001) this might be explained by the use of “operational” hedges by firms. However, no paper, to our knowledge, has tried to look at the factors affecting “operational” currency exposure among firms.

3

Exchange rate is defined as domestic currency (Rupee) per unit of foreign currency.

62

A. Dhasmana

Another strand of the same literature looks at the issue of pricing policies in response to currency fluctuations (e.g., Goldberg & Knetter, 1997). Looking at the case of US and Japanese bilateral trade they find imperfect pass through of USD
Yen exchange rate changes to the prices of Japanese imports. This could be due to a downward adjustment in mark ups by the Japanese firms in response to an appreciation in Yen and/or a reduction in the share of costs incurred in local currency (Yen). In either case, an imperfect pass through of exchange rate changes should reflect itself in small exchange rate exposure elasticity. The degree of pass through and hence the size of exchange rate exposure will, therefore, depend upon the structure of product and input markets at home and abroad. A small section of this literature looks at the impact of currency fluctuations on firm level investment (e.g., Goldberg, 1993; Goldberg & Campa, 1995, 1999; Nucci & Pozzolo, 2001). Overall, most papers in this literature find a significant impact of exchange rate change on investment. Further, they find that the impact of exchange rate on investment varies with the degree of international linkage and the size of mark up as both of these determine the impact of exchange rate change on firm’s profitability. While this paper is most closely related to this last strand of literature, most of the existing papers in this literature look at developed countries, with little attention being paid to the emerging markets such as India. One of the reasons for this gap is the lack of good-quality firm level data. In that respect our paper contributes to the existing literature by putting together a large firm level dataset for an emerging economy that can be used to answer questions regarding impact of macroeconomic variables such as exchange rates on firms. At the macroeconomic level this paper is related to the extensive literature on currency mismatch and its impact on growth in emerging markets. Key contributions in this literature include Goldstein and Turner (2004), Eichengreen, Hausmann, and Pannizza (2007), and Ranciere, Tornell, and Vamvakidis (2010). In most of these papers the focus is on the mismatch between the currency denomination of assets and liabilities. Little attention has been paid to the currency mismatch between costs and revenues of the firms. Such “operational” mismatches are equally important and deserve attention of policy makers and academics alike. Firms with same degree of mismatch in their assets and liabilities can have very different level of vulnerability to exchange rate shocks depending upon whether they produce tradable or non-tradable goods or the extent to which they depend upon imported inputs. This paper fills an important gap in the literature on currency mismatch by focusing on the “operational” mismatch between firm’s costs and revenues. Finally this paper is also linked to the literature on cost of sharp currency devaluations. While theory has been ambivalent regarding the impact

Operational Currency Exposure

63

of currency devaluations on real activity, empirical literature has also provided mixed evidence regarding the economic impact of sharp currency devaluations (see e.g., Gupta, Mishra, & Sahay, 2007; Hong & Tornell, 2005; Hutchinson & Noy, 2005). Unlike most papers in this literature however, we use firm level longitudinal data set for an emerging market that allows us to take into account firm level characteristics. The paper is organized as follows: the second section describes the data and our measure of currency exposure. The third section looks at the determinants of currency exposure while the fourth section looks at the impact of exchange rate exposure on firm level performance. The fifth section discusses the policy implications of our results and concludes.

Data We use annual data for 500 Indian firms listed under the BSE 500 Index. Most of the data comes from their Annual Financial Statements and covers the period 1995
2011. Firms included under the index represent roughly 93 percent of the total market capitalization on the BSE and cover all the major industries in the Indian economy including construction, infrastructure, as well as non-traditional services such as software and ITeS. The time period covered by our data includes three important economic crisis of the 20th century
the East Asian crisis, the 2001 dotcom bubble, and the 2007 Global financial meltdown. Key variables of interest in our model are growth in output and earnings per share. We use them along with the level of capital expenditure as indicators of firm performance. Our objective is to study the impact of currency exposure on firm level performance as measured by output growth and earnings per share. The key explanatory variable for our analysis is therefore the measure of currency exposure.

Measuring “Operational” Currency Exposure Important as it is for the firms and policy makers alike, measuring exchange rate exposure is fraught with various difficulties starting from the lack of data to the need for proper theoretical framework. Studies trying to measure currency exposure of firms often rely on stock returns data. They estimate exposure of individual firms in “excess” of the overall market exposure to exchange rate changes by regressing firm level stock returns on market level returns and exchange rate returns (see Adler & Dumas, 1984). However, very few of these studies find significant exchange rate exposure among firms. This might be a reflection of the fact that firms use “Operational” hedges to protect themselves against exchange rate risk.

64

A. Dhasmana

We therefore use the measure of exchange rate exposure suggested by Bodnar and Marston (2001) instead who develop a model of foreign exchange exposure dependent on only three variables: the percentage of the firm’s revenues and expenses denominated in foreign currency and its profit rate. We describe the construction of this measure in detail below.4 Consider a profit-maximizing monopoly firm producing and selling goods at home and abroad. Value of this firm can be expressed in terms of a stream of present and future cash flows as PV =

∞ X

Kt ; ð1 þ iÞt t=1

ð1Þ

where Kt is the expected future cash flow which is equal to the after tax profit less net investment and i is the discount rate. Assuming that the cash flows are equal from year to year and net investment is equal to zero one can rewrite Equation (1) as PV =

K ð1 − τÞ = π: i i

Elasticity of firm’s value with respect to exchange rate S is given by   dPV ð1 − τÞ dπ = : dS i dS

ð2Þ

ð3Þ

For simplicity one can divide the overall impact of exchange rate on π in two parts
impact due to an adjustment in the firm’s output and the direct effect in proportion to the initial level of net revenue denominated in foreign currency. Assuming that the monopolistic firm chooses its output optimally, the response of profits to changes in output is zero and only direct effect of exchange rate on profits remain.5 Using this insight exchange rate elasticity can be written as:

4

Estimates of Country level currency mismatch are based on two main, straightforward measures. The first is the ratio of net national debt or debt service requirements to the net exports of a country. The second is the ratio of foreign currency denominated liabilities to foreign currency denominated assets of the banking sector. Goldstein and Turner (2004) and Eichengreen et al. (2007) provide a review of the first strand of this literature while Lane and Ferretti (2007) and Ranciere et al. (2010) are the latest example of the second strand. 5 It must be emphasized that the currency mismatch between revenues and costs does not reflect change in demand condition following exchange rate changes, especially for exporting firms that practice PCP. Following a depreciation, rise in demand would rise pro.t of PCP firms even if revenues are labeled in the home currency (that is, even if h1 = 0 and therefore, δ ≤ 0). See for example Betts and Devereux (2000).

65

Operational Currency Exposure



 1 −1 ; δ = h1 þ ðh1 − h2 Þ r

ð4Þ

where δ is the exposure elasticity or percentage change in firm’s cash flow in response to a change in exchange rate; h1 is the foreign currencydenominated revenue as a percentage of total revenue; h2 is the foreign currency-denominated cost as a percentage of total costs; and r is the profit rate (i.e., profits as a percent of total revenues). Equation (1) implies that higher the share of foreign currency revenues and smaller the share of foreign currency costs the greater is the increase in firm’s value in response to a depreciation of the home currency. Further, higher profit after tax would lower the “absolute” size of exposure elasticity. It is important to note that our measure of currency exposure merely captures the first-order effects of changes in exchange rate on firm’s revenues and costs that are denominated in foreign currency. We use data on foreign currency costs, foreign currency revenues and profits from Center for Monitoring Indian Economy’s (CIME’s) PROWESS database to calculate exposure elasticity of the firms in our sample for the period 1995
2011. Figure 1 plots average (actual and absolute) exposure elasticity of Indian firms over 1996
2011. It does not reflect change in demand conditions facing firms following exchange rate changes. Top panel of Figure 1 plots the cross-sectional average of exposure elasticity between 1995 and 2011 along with annual average monthly 2 1.5 1 0.5 0 –0.5 –1 –1.5 –2

60 50 40 30 1996 1997 1998 1999 2000 2001 2002 2003 2004 2005 2006 2007 2008 2009 2010 2011 20 10 Average Exposure

Rupee/USD Exchange Rate

0

Panel 1 5 4 3 2 1 0 1996 1997 1998 1999 2000 2001 2002 2003 2004 2005 2006 2007 2008 2009 2010 2011 Abs. Avg Exposure

Exchange Rate volatility

Panel 2

Figure 1: Average Exposure and Rupee/USD Exchange Rate (Panel 1). Absolute Exposure Elasticity and Exchange Rate Volatility (Panel 2).

66

A. Dhasmana

Rupee
USD exchange rate. As we can see, average exposure elasticity for Indian firms in our sample has been positive for most years between 1995 and 2011. Bottom panel of the same figure plots the average absolute exposure elasticity across Indian firms along with annual volatility of weekly Rupee
USD log returns. It shows that periods of low exchange rate volatility are associated with higher absolute exposure elasticity. This indicates the presence of “moral hazard” type behavior among Indian firms whereby lower exchange rate volatility prompts firms to take on higher exchange rate risk. We try to explore this hypothesis further in the next section.

Industry-Wise Exposure Elasticity Average exposure elasticity can hide significant variation across industries. We therefore look at the industry-wise decomposition of exchange rate exposure in Figures 2
4. We rely on the industry classification provided by the CMIE in what follows (list of all the industries is available upon request from the authors). Figure 2 presents top 10 industries with largest negative and positive exposure in 2011. From the top panel we can see that Aluminum industry had the largest “negative” exposure elasticity followed by fertilizers, glass, and glass ware and refining industry. Figure 3 of the same figure shows the industries with largest positive exposure. These include Air transport which has the largest positive exchange rate exposure followed by refractories, sugar, gems, and jewelry and industrial construction. Overall, industries with large negative exposure are the ones with a very high share of imported inputs in their total cost relative to the share of foreign income in their total income.6 Opposite is true for firms with large positive exposure elasticity. To get a better picture of the industry-wise exposure we further club these industries into 11 broad categories and look at their average exchange rate exposure over time (email authors for detailed data). Figure 4 plots the average exposure elasticity for these eleven industries between 1995 and 2011. A careful look at this figure provides important insights into the sectorial impact of exchange rate movements in the case of Indian economy.

6 For example, imports comprised about 40 percent of the total cost in Aluminum industry while its share of foreign exchange earnings was only 30 percent of its total income in 2011. Air transport services, on the other hand, had 1.5 percent of its total costs going toward imports even though its share of foreign income in the total income was 18.5 percent.

67

Operational Currency Exposure

Aluminium

Fertilisers

Glass & glassware

Man Made filaments

Refinery

Housing Finance services

LNG Paper storage & and newsprint Distribution

Diversified machinery

Coal & Lignite

0 –2 –4 –6 –8 –10

Aluminium

Fertilisers

Glass & glassware

Refinery

Man Made filaments

Housing Finance services

Diversified machinery

Paper & newsprint

LNG storage and Distribution

Coal & Lignite

–12 –14 –16 –18 –20

Figure 2:

Industries with Largest Negative Exposure Elasticity in 2011.

200 150 100 50

Cloth

Polymers

Vegetables oils

Coffee

Wires & Cables

Industrial construction

Gems & Jewellery

Sugar

Refractories

Air transport

0 Vegeta

Cloth

Wires

bles o

ils

Gems

& Cab

Refrac

& Jew

tories

ellery

les

Figure 3: Industries with Largest Positive Exposure Elasticity in 2011.

10

5

0 er

ica

C

m he

cs

y

ls

y

rg

llu

–5 eta M

in

h ac

M

ni

ro

t ec

El

rt

an

Tr

er

en

bb

pm

o sp

–10

t

s

ile

xt

Te

i qu

E

ic

st

&

Ru

od

a Pl

od

&

s

ct

er

er

Fo

d

W

oo

&

a Le

l

–20

Industry-Wise Exposure Elasticity.

cia

n na

Fi

n-

No

–15

Figure 4:

o Pr

th

Fo

vic

du

in

f Re

es

s

y

ct

du

o Pr

r Se

68

A. Dhasmana

Non-financial services (including Business consultancy and IT & ITES) have the largest positive exposure elasticity followed by metallurgy and textiles. Services, especially non-traditional services such as IT and ITES, are a growing component of India’s economy and external trade. Similarly, textiles are one of the key traditional exports of India and an important source of manufacturing employment. Rupee depreciation clearly benefits these important sectors. At the same time, sectors such as refinery (oil) and food that are a source of key inputs for other sectors exhibit a negative exchange rate exposure thereby presenting a dilemma for the policy makers.

Operational Exposure Elasticity and Exchange Rate Regime Theory gives different explanations for the presence of currency mismatch in emerging markets which can be broadly divided into two categories
“Moral Hazard” and “Incomplete Markets.” While the former explanation looks at implicit or explicit government guarantees in the form of bank bailouts and fixed exchange rate regimes, the latter looks at market frictions resulting in inadequate provision against exchange rate risk. The former explanation implies that the degree of central bank intervention would have a direct impact on the risk-taking behavior of individual firms. In the latter case, one would not expect to see any discernible relationship between exchange rate regime and currency exposure. Based on this insight Shah and Patnaik (2010) test the “moral hazard” hypothesis in the case of India. They use Bai and Perron (2003) algorithm to identify structural breaks in the volatility of weekly Rupee
Dollar returns and test the impact of exchange rate regime on un-hedged currency exposure among a set of 100 Indian firms. They find evidence in support of the moral hazard hypothesis. In similar spirit, we try to test whether operational mismatches in costs and revenues are related to exchange rate regimes. We divide the entire sample in four time periods for our analysis. Division of the sample is based loosely on the study by Shah and Patnaik (2010). Using squared weekly returns on the Rupee
USD exchange rate between April 1993 and February 2007, they identify four distinct breaks in India’s exchange rate regime. Given that we have annual data unlike Shah and Patnaik (2010) that uses weekly data, we use their break points and match them with our annual series. Column 1 in the table below gives the four sub-periods used by us while column 2 gives the corresponding periods of exchange rate regime shifts identified in Shah and Patnaik (2010). Notice that Shah and Patnaik (2010) only cover period till February 2007 which leaves out the period since the year 2007. The period after 2007

69

Operational Currency Exposure

saw the global financial crisis unfolding. That is likely to have affected firm’s exposure elasticity through changes in exports, imports, and profit margins. We therefore use the period between 2009 and 2011 as the last period of our analysis in this section so as to avoid confounding our results due to the impact of global financial crisis. Table 1 gives the average exposure elasticity “delta” of Indian firms in the four periods along with the volatility in INR/USD weekly returns in those periods. The first thing to note is that the volatility of INR
USD weekly returns varies substantially across the four periods even though India has had a de jure “managed float” throughout this period. This result is in line with Shah and Patnaik (2010, 2011). Thus, even though India had a managed floating exchange rate regime throughout this period the extent to which the central bank authorities intervened in the foreign exchange market and tried to control rupee volatility varied over time. Again in line with the findings in Shah and Patnaik (2010), rupee volatility was higher in the first period that included the Asian financial crisis. It came down during the next five year period between 1999 and 2003 following which the volatility increased again. During the latest period (between years 2009
2011) volatility of Rupee has gone up even further (this holds true even if we exclude the year 2011). Column 3 above shows the mean exposure elasticity of the firms in our sample during different periods. Average exposure elasticity of the firms in our sample is positive for all the four periods under consideration. This indicates that overall, Indian firms have tended to benefit from Rupee depreciation on account of their operational currency mismatch. At the same time, the average exposure elasticity has been higher in periods when the exchange rate volatility was lower. Average exposure elasticity increased from 0.3 in the first period (volatility 0.53) to 0.87 (volatility 0.046) in the second period. Subsequently, as the exchange rate volatility increased the average exposure elasticity came down. On an average, Indian firms have tended to expose themselves more heavily to a Rupee appreciation risk during periods of low exchange Table 1: Summary Statistics of Delta Period

Exchange rate regime (Shah & Patnaik)

1996
1998 1995-02-17
1998-08-21 1999
2003 1998-08-21
2004-03-19 2004
2006 2004-03-19
2007-02-12 2009
2011

Average exposure elasticity

Mean

Volatility INR/USD

Reserve accumulation as percentage of net capital inflows

δ1 δ2 δ3 δ4

0.30 0.87 0.71 0.21

0.53 0.046 0.22 1.01

26.86 84.3 47.0 0.09

70

A. Dhasmana

rate volatility. This is most likely a reflection of the fact that periods of low exchange rate volatility in India have been associated with a higher net inflow of foreign capital that were sterilized by the authorities in order to prevent Rupee appreciation and used to build up reserves. The last column in Table 1 provides some evidence to this effect. It presents the ratio of average reserve accumulation to average net capital inflow for the four periods. As we can see, periods of low exchange rate volatility are associated with a higher ratio of reserve accumulation to net capital inflows which is reflective of monetary sterilization of foreign exchange inflow by the central bank.

Determinants of Firm Level Exposure We would like to study the factors that are associated with higher absolute level of currency exposure at the firm level especially those directly or indirectly related to government policies. In this regard, the key variable for interest for us in this exercise is the volatility in exchange rate. Table 1 highlighted the effect of this intervention on the direction of average exchange rate exposure. If foreign exchange intervention does have an impact on the risk taking behavior of the firms then one would expect to see a negative correlation between the absolute size of currency exposure and volatility of exchange rate. Empirical exercise in this section provides a formal test of the “Moral Hazard” hypothesis apart from identifying firm level characteristics determining the absolute size of firm level currency exposure. The model used for this exercise is given below:   δi;t  = αi þ βi Xi;t þ εi;t : ð5Þ   Our dependent variable is the absolute size of δi;t  or the exchange rate exposure elasticity calculated above. The set of explanatory variables Xi;t includes share of exports in sales, growth rate of sales, volatility in exchange rate and log of market capitalization.7 Table 2 presents the results from this exercise. Column 1 above presents the results from the entire sample while the remaining columns present the results of different sub-samples. We begin by discussing the results for exchange rate volatility which is the key variable of interest for us.

7 Hausman’s specification test between fixed and random effects estimator selected the former hence we used it for estimating Equation (1). Hausman’s Specification test: Chi sq (4) = 21.6, p-val. = 0.00.

71

Operational Currency Exposure

Table 2: Determinants of Operational Currency Exposure Dependent variable:   δi;t  Exports as a percentage of sales Growth in sales Exchange rate volatility Market capitalization R2 Total number of observations No. of groups

Entire sample

Excluding mining

Excluding services

Manufacturing

0.028*** [0.005] 0.16** [0.066] −0.11*** [0.03] −0.24*** [0.05]

0.029*** [0.005] 0.17** [0.068] −0.11*** [0.03] −0.26*** [0.05]

0.028*** [0.005] 0.18** [0.07] −0.11*** [0.03] −0.22*** [0.05]

0.032*** [0.008] 0.21** [0.10] −0.11*** [0.04] −0.27*** [0.07]

0.33 3615

0.36 3491

0.22 3138

0.25 2448

346

331

284

209

Note: “***” and “**” denotes “significant at 1% and 5%”. Terms inside the brackets are standard errors adjusted for hetrosckadasticity across clusters.

Exchange rate volatility has a negative and significant coefficient in our model. This indicates that higher exchange rate volatility as measured by the annual standard deviation of weekly log returns on Rupee/USD exchange rate is associated with lower absolute value of “operational” currency exposure on average. This can be due to a greater mismatch between their foreign exchange revenues and costs or due to a lower profit rate or both. To check which of these is true, we regress growth in profits and absolute size of revenue
cost mismatch (absolute value of the difference between h1 and h2 ) on a set of time and firm specific effects and exchange rate volatility. Results from this exercise are given in the appendix. We find that exchange rate volatility does not have a statistically significant effect on the growth of firm’s profits (in fact it has a positive coefficient), though it is significantly and positively associated with the “operational” foreign currency mismatch as measured by the absolute value of the difference between h1 and h2 .8 Our result is therefore in line with the findings of Shah and Patnaik (2010) and supports the “Moral Hazard” hypothesis. Periods of low

8 This is reassuring since exchange rate changes can cause changes in price which will in turn cause changes in demand and hence profits. That will make delta unsuitable as a measure of operational currency mismatch. One explanation for profits being insensitive to exchange rate changes could be the use of “hedges”
financial as well as “operational”.

72

A. Dhasmana

exchange rate volatility are associated with higher absolute level of exposure elasticity as compared to the periods with high exchange rate volatility indicating that whenever government tries to stabilize the exchange rate in order to keep its volatility under check, the result is an increase in the risktaking behavior of the private sector as reflected in higher “operational” currency mismatch. Share of exports in total sales is positively related with the size of exposure elasticity indicating that more export-oriented firms tend to see much higher levels of exposure elasticity. On an average, a one percent increase in the share of exports in total sales is associated with a 2.8 basis points increase in the size of exposure elasticity. This result holds across different sub-samples as shown by the remaining columns. This result is in line with the findings of Jorion (1990) who find a significant positive association between the size of foreign operations and exchange rate exposure. Growth rate of sales is positively associated with the size of exposure elasticity. Firms with better growth prospects as reflected in higher sales growth exhibit higher exposure elasticity. At the same time, exposure elasticity is negatively related to the firm size as measured by their market capitalization. Larger firms tend to have smaller exchange rate exposure elasticity. The last result indicates that “smaller” firms are more vulnerable to exchange rate changes due to lower profit margins. There is some evidence to support this view. However, since our main focus is on the relationship between exchange rate volatility and firm level exposure, we leave this hypothesis for future research.

Openness and Delta It is possible that more “open” firms, that is, firms with high export and/or import intensity, react differently to exchange rate volatility as compared to the rest. This might be on account of greater access to and reliance on financial instruments for hedging exchange rate risks by more open firms. To check our hypothesis we divide our sample into “export intensive” and “import” intensive firms and estimate the model in Equation (2) separately for each sub-sample. Firms are classified as having “high” export intensity if the share of exports in their total sales is above 90 percent.9 Similarly, firms are classified as import intensive if the ratio of imported inputs to

9 The cut-off of 90 percent represents the 95th percentile of firms in terms of their export share. Similarly, cut-off of 25 percent represents 95th percentile of firms in terms of their import intensity (imports/income).

73

Operational Currency Exposure

their total income is above 25 percent. Table 3 presents the results from this exercise. The coefficient on exchange rate volatility is no longer significant for firms with a high level of export intensity. In fact it carries a positive sign. For firms with low export intensity, however, the coefficient on exchange rate volatility remains negative and significant. This result seems to support our hypothesis that firms with a “high” degree of export intensity respond differently to exchange rate volatility possibly on account of greater reliance on financial instruments for hedging exchange rate risks. Unfortunately at this stage we do not have data on the use of exchange rate derivatives by these firms to test this hypothesis directly. Looking at the firms with high import intensity, we find that their exposure elasticity is negatively related to exchange rate volatility in line with our earlier results. In fact, the coefficient on exchange rate volatility is much larger in size for “high” import intensity firms than for “low” import intensity firms. This makes intuitive sense since firms with “high” import intensity are affected much more by volatility in exchange rate as compared to firms with “low” import intensity. Another interesting result is that share of exports in total sales is not significantly related to the size of operational currency mismatch in the case of “high” import intensity firms even though it is significantly related in the case of low import intensity firms. Coefficients on the remaining variables are unchanged in sign and significance.

Table 3: Determinants of Operational Currency Exposure
Export and Import Intensities Dependent   variable: δi;t  Exports as a percentage of sales Growth in sales Exchange rate volatility Market capitalization R2 Total number of observations No. of groups

High export intensity firms

Low export intensity firms

0.027*** [0.00]

0.022*** [0.00]

High import intensity firms −0.015 [0.01]

Low import intensity firms 0.03*** [0.00]

0.21** [0.09] 0.04 [0.07] −0.22*** [0.07]

0.05 [0.07] −0.18*** [0.04] −0.34*** [0.10]

−0.11 [0.20] −0.33** [0.13] −0.32*** [0.17]

0.17** [0.07] −0.09*** [0.03] −0.24*** [0.05]

0.28 2168

0.21 1447

0.001 226

0.39 3387

258

318

25

321

Note: “***” and “**” denotes “significant at 1% and 5%”. Terms inside the brackets are standard errors adjusted for hetrosckadasticity across clusters.

74

A. Dhasmana

Next section looks at the relationship between exposure elasticity and firm level performance measured by their output growth and earnings per share.

“Operational” Exchange Rate Exposure and Firm Level Performance Objective of this exercise is to look at the impact of “operational” currency exposure on firm level performance. We use the measure of exposure elasticity described above to do so.

Output Growth And Exposure Yi;t = αi þ βXi;t þ

p X

γ i Dt þ θ × AbsExposurei;t

t=0

þ λ × Dt × AbsExposurei;t − 1 þ εi;t :

ð6Þ

Our first specification is given above. The dependent variable in the above equation is the growth rate of output. Xi;t is a set of explanatory variables that vary across firms and time periods. These include the growth rates of employment (as measured by the number of workers) and unit labor cost (defined as Total Emoluments to Workers divided by the Total Value of Output) along with growth in market capitalization and share of exports in total sales.10 Since our focus is on the impact of exchange rate exposure of firms which would be expected to have a greater impact during episodes of large currency changes, we include a dummy for large nominal depreciation of Rupee and its interaction with the exposure dummy in our analysis. Dt is the dummy for large nominal depreciation of Rupee which takes a value of 1 whenever the annual rate of increase in the monthly Rupee/USD exchange rate is more than one standard deviation above the average annual rate of exchange rate change for this period and 0 otherwise. With this criterion, currency devaluations are defined as sharp decline in Rupee/USD exchange rate exceeding 10 percent on an annual basis.11 This definition helps us identify four episodes of large depreciations in Indian Rupee
1995, 1998, 2008, and 2011.

10

Fisher’s Unit Root Test allows us to reject the null hypothesis of a unit root for all the variables (including output growth) in our model. 11 The reason for using nominal Rupee
USD exchange rate for defining depreciation episodes is that Indian Rupee has been de-facto pegged to USD (Ref. Shah and Patnaik (2010)).

Operational Currency Exposure

75

In order to capture the effect of firm level currency exposure we multiply the currency depreciation dummy with firm’s lagged absolute currency exposure. Use of lagged currency exposure is done in order to avoid possible endogeneity between currency exposure and exchange rate changes. The fourth term in Equation (1) captures this interaction term. The sign of coefficient λ indicates the effect of currency exposure on Yi;t during episodes of large Rupee depreciations. The impact of “operational” currency exposure on Yi;t (output growth) is given by −θ × AbsExposurei;t − 1 during normal times and by ðθ þ λÞ × AbsExposurei;t − 1 during episodes of large Rupee depreciations. A negative θ þ λ implies that the output growth falls with an increase in the size of “operational” currency mismatch during episodes of large Rupee depreciations. The last term, εi;t is the random error. In addition to the above variables we also try a number of industry and firm level fixed effects to capture the impact of omitted variables. They do not, however, affect our main results. Table 4 gives the results from this exercise. We discuss the results in the following paragraphs. The second column of Table 4 gives the estimation results for the entire sample. The first two rows give coefficients on current and lagged dummy for large currency depreciations. While literature has found both positive and negative effects of currency depreciations on growth, theory is not clear regarding the direction of this relationship. Large currency depreciations can help growth by boosting exports. At the same time they can also have an adverse impact on growth through a rise in the cost of imported inputs, worsening of balance sheets and an increase in the financial fragility. In our sample, large currency depreciations are associated with a decline in average output growth as seen from the negative coefficients on the current and lagged dummies. Coefficient on lagged depreciation dummy is significant for all the samples except non-manufacturing firms where the coefficient is negative but insignificant. The third row presents the coefficient on lagged level of absolute currency exposure. The coefficient is negative for all the sub-samples but is significant for only the entire sample and the set of firms excluding services. More importantly, its coefficient is smaller in size than the coefficient on the interaction term between the crisis dummy and lagged currency exposure size. Finally, we look at the interaction term between the depreciation dummy and the absolute exposure size. Coefficient on this interaction term is negative and significant for all the subsamples in our study. An exposure elasticity of one increases the loss in output growth of firm due to large Rupee depreciation by 2.6 basis points on average. Impact of operational exposure on firm performance is therefore both statistically and economically significant.

76

A. Dhasmana

Table 4: Operational Currency Exposure and Output Growth Dependent variable: Output growth

Entire sample

Manufacturing

Depreciation dummy

−0.06 [0.06] −0.14** [0.06] −0.003** [0.00] −0.026** [0.00]

−0.05 [0.05] −0.19*** [0.05] −0.00 [0.00] −0.04** [0.02]

−0.09 [0.17] −0.01 [0.18] −0.00 [0.00] −0.025*** [0.00]

−0.07 [0.06] −0.16*** [0.07] −0.00 [0.00] −0.04*** [0.01]

−0.05 [0.06] −0.11* [0.05] −0.00** [0.00] −0.025*** [0.00]

0.23** [0.11] 0.026 [0.03] −0.02 [0.04]

0.13** [0.06] 0.04 [0.03] 0.01 [0.04]

0.56*** [0.15] 0.02 [0.05] −0.14 [0.09]

0.13** [0.06] 0.04 [0.03] −0.05 [0.04]

0.22** [0.11] 0.02 [0.03] −0.01 [0.07]

0.05 1409

0.056 1025

0.09 384

0.03 1289

0.05 1351

222

157

65

207

208

Lag depreciation dummy Lagged absolute exposure Deprecation dummy* lagged absolute exposure Employment growth Unit labor cost growth ΔMarket capitalization R2 Total number of observations No. of groups

NonExcluding manufacturing mining

Excluding services

Following Rupee depreciation, increase in demand would raise profit of PCP firms even if revenues are labeled in the home currency. It would therefore be interesting to separate purely domestic versus exporting firms. Unfortunately, in our sample there are only a handful of firms (11 to be precise after taking in to account missing observations etc.) that fit the criterion of being purely domestic or having zero exports throughout the sample period. This makes it difficult to distinguish between exporting and purely domestic firms. We therefore do not attempt to separate. Note: “***” & “**” denote significance at 1 and 5 percent level respectively. Figures inside the brackets are robust standard errors corrected for intragroup correlation.

Of the other variables used in the model employment growth is the only one which has a significant and positive coefficient. The rest do not appear to have a significant impact on output growth. Higher employment growth is associated with a faster output growth as expected. Employment elasticity of output growth is higher for non-manufacturing firms when compared to manufacturing firms.

“Large”
Sized Exposure As we saw in the second section, there has been a significant increase in large-sized operational exposure among Indian firms (both negative and

Operational Currency Exposure

77

positive) in recent years concomitant with the rise in exchange rate volatility. We therefore try to capture the impact of large exposure elasticity on firm performance by using a dummy for large operational exposure. DExposuret;t is the exposure elasticity dummy that takes a value 1 whenever the absolute value of exposure elasticity δi;t is greater than 2.512 and 0 otherwise. We include this dummy in our benchmark model. It is possible that the response of firms to exchange rate change varies at very high levels of exposure. To capture this we include an interaction term between the crisis dummy and firm level exposure dummy that captures the impact of “high” exchange rate exposure on firm’s output growth during episodes of large exchange rate depreciations. Our model therefore becomes: Yi;t = αi þ βXi;t þ

p X t=0

γ i Dt þ θ × AbsExposurei;t − 1 þ λ

× Dt × AbsExposurei;t − 1 þ φ × DExposurei;t − 1 × AbsExposurei;t − 1 þ εi;t

ð7Þ

An overall exposure elasticity of 1 changes the output growth of firms in our sample by θ during “normal” times. At the same time, an exposure elasticity of 1 changes the output growth of firms by θ þ φ for firms having “large” currency exposure.13 Table 5 presents the results from this augmented model. Overall our results remain unchanged with the inclusion of this additional interaction term. Coefficient on the interaction term between lagged absolute exposure and dummy for large exposure elasticity is positive but insignificant. Interestingly, it appears for firms with large currency exposure, currency mismatch does not affect output growth during normal times. To check if firms with large operational exposure are affected differently by large currency depreciation we multiply the term DExposurei;t − 1 × AbsExposurei;t − 1 with dummy for large exposure. Once again the coefficient on this last term is not statistically significant. We therefore do not include the terms for large currency exposure in what follows.

12

 The  cut-off value of 2.5 represents the top 2.5 percentile of the distribution of δi;t . Using alternative values of this cut-off does not change our results significantly. 13 It is quite possible that large negative and positive exposure elasticity has different impact on firm level performance. We therefore repeat our analysis with separate dummies for large negative and positive exposures. However, Wald test for coefficient restrictions showed that the coefficients on them were not significantly different from each other. We therefore continue with our original specification.

78

A. Dhasmana

Table 5: Operational Currency Exposure and Output Growth
Large Sized Exposure Dependent variable: Output growth Depreciation dummy Lag depreciation dummy Lagged absolute exposure Exposure dummy* lagged absolute exposure Deprecation dummy* lagged absolute exposure Employment growth Unit labor cost growth ΔMarket capitalization R2 Total number of observations No. of groups

Entire sample

Manufacturing

NonExcluding manufacturing mining

Excluding services

−0.06 [0.06] −0.14** [0.02] −0.03 [0.02] 0.03 [0.02]

−0.05 [0.05] −0.19*** [0.05] −0.01 [0.02] 0.01 [0.02]

−0.10 [0.17] −0.01 [0.18] −0.08 [0.06] 0.08 [0.06]

−0.07 [0.06] −0.15** [0.07] −0.03 [0.02] 0.03 [0.02]

−0.06 [0.05] −0.11* [0.05] −0.03 [0.02] 0.03 [0.02]

−0.026*** [0.00]

−0.04** [0.02]

−0.025*** [0.00]

−0.04*** [0.01]

−0.025*** [0.00]

0.23** [0.11] 0.02 [0.03] −0.02 [0.04]

0.13** [0.04] 0.04 [0.03] 0.02 [0.04]

0.56*** [0.15] 0.01 [0.05] 0.14 [0.09]

0.13** [0.06] 0.04 [0.03] −0.04 [0.04]

0.22** [0.11] 0.01 [0.04] −0.01 [0.04]

0.05 1409

0.058 1025

0.09 384

0.03 1289

0.05 1351

222

157

65

207

208

Note: “***,” “**,” and “*” denote significance at 1, 5, and 10 percent level respectively. Figures inside the brackets are robust standard errors corrected for intragroup correlation.

Profitability and Exposure The key insight from the above exercise is that higher level of exchange rate exposure elasticity is associated with a greater loss in output growth during episodes of large Rupee depreciation. We next try to do the same analysis for earnings per share which is used as a measure of firm’s profitability. Our model is the same as in Equation (1) except that the dependent variable is now earnings per share.14 Table 6 presents the results from this exercise. Once again we find that the interaction term between depreciation dummy and lagged absolute exposure elasticity has a negative and

14

We tested for the presence of unit roots in all our series using Fisher’s panel unit root test and were able to reject the null of unit root for all of them.

79

Operational Currency Exposure

Table 6: Operational Currency Exposure and Earnings Per Share Dependent variable: Earnings per share

Entire sample

Manufacturing

Depreciation dummy

2.0 [1.21] −0.15*** [0.05] −0.72*** [0.21]

1.7 [2.0] −0.21*** [0.07] −1.46*** [0.55]

2.6 [1.5] −0.08 [0.07] −0.57** [0.25]

2.0 [1.2] −0.13** [0.06] −0.64*** [0.24]

2.09 [1.5] −0.17*** [0.06] −0.61** [0.25]

−0.27 [0.85] −0.98 [0.50] 3.57*** [0.64]

1.0 [0.9] −0.82 [0.87] 3.78*** [1.0]

−1.0 [1.1] −1.0 [0.59] 3.4*** [0.84]

−0.34 [0.87] −1.0** [0.5] 3.6*** [0.87]

2.4** [1.6] −1.61 [0.87] 4.4*** [0.81]

0.14 2162

0.16 1000

0.09 1162

0.16 2056

0.13 1448

354

164

190

339

241

Lagged absolute exposure Depreciation dummy* lagged absolute exposure Employment growth Unit labor cost growth ΔMarket capitalization R2 Total number of observations No. of groups

NonExcluding manufacturing mining

Excluding services

Note: “***” & “**” denote significance at 1 and 5 percent level respectively. Figures inside the brackets are robust standard errors corrected for intragroup correlation.

significant coefficient in all our sub-samples. Higher absolute exposure elasticity implies a greater loss in earnings per share due to large depreciations. Coefficient on the dummy for large currency depreciation is not statistically significant in any of the sub-samples. Coefficient on the lagged level of absolute exposure is negative and significant for all except the sample of non-manufacturing firms indicating that firms with greater currency mismatch witness lower earnings per share even during “normal” times. Among the other variables, growth in market capitalization is significantly correlated with earnings per share across all the sub-samples. So far we have seen that higher operational exposure increases the cost of large depreciations in terms of lower output growth and lower earnings per share. In the next section we try to explore possible transmission channel from exposure elasticity to firm level performance. Capital Expenditure and Exposure Elasticity One of channels through which high exposure elasticity can lead to a reduction in output growth is through lower investment. We try to explore that channel in this section. Below we present the estimates of a model for firm level capital expenditure augmented with variables capturing firm level

80

A. Dhasmana

exposure. Unfortunately we do not have data on capital expenditure for the firms in our sample prior to year 2002 hence we have to restrict this part of our analysis to period after year 2002. Table 7 presents the results from this exercise. Our dependent variable is the growth of capital expenditure.15 Key variables of interest for us are the lagged currency exposure and the dummy for large currency depreciation along with the interaction term between depreciation dummy and lagged exchange rate exposure. Apart from these we use several firm-specific variables that can potentially affect the level of capital expenditure by the firm including the growth in market capitalization. At the same time we also include industry-specific time effects to capture industry specific omitted variables that vary over time. These omitted variables could include industry specific shocks to demand and/or productivity along with industry specific policy shocks. We use four alternative specifications for our analysis. Our benchmark model, Model 1, includes the dummy for large depreciation, lagged Table 7:

Operational Currency Exposure and Capital Expenditure

Dependent variable: Growth in capital expenditure Depreciation dummy Lag of absolute exposure Depreciation dummy* lag of absolute exposure Exposure dummy* lag of absolute exposure Growth in market capitalization Inflation Constant Total number of observations No. of groups

Model 1

Model 2

Model 3

Model 4

0.33*** 0.19** 0.3*** 0.19** [0.09] [0.08] [0.08] [0.08] 0.00 0.00 0.03 0.03 [0.00] [0.00] [0.04] [0.04] −0.005*** −0.005*** −0.005*** −0.005*** [0.001] [0.001] [0.001] [0.001] −0.02 −0.03 [0.04] [0.04] 0.29** 0.27*** 0.28*** 0.28*** [0.05] [0.05] [0.05] [0.05] 0.04*** 0.04*** [0.00] [0.00] 0.26*** −2.5*** 0.25*** −2.4*** [0.08] [0.5] [0.08] [0.5] 2549 2549 2549 2549 406 406 406 406

Note: “***” & “**” denote significance at 1 and 5 percent level respectively. Figures inside the brackets are robust standard errors corrected for intragroup corr.

15

We test for the presence of unit root in the series for capital expenditure using Fisher’s panel unit root test and are able to reject the null hypothesis of unit root. The test however does point toward persistence in the series in the form of lagged dependent variable we therefore use growth rate of capital expenditure as the dependent variable.

Operational Currency Exposure

81

absolute exposure, and an interaction between the dummy for large depreciation and lagged absolute exposure. Along with these we use growth in market capitalization to capture the impact of market size on capital expenditure. Model 2 adds WPI inflation rate to the benchmark model to capture the impact of domestic macroeconomic circumstances while Model 3 adds an interaction term between the dummy for large exposure as defined above, and lagged absolute exposure to allow for nonlinearity in the impact of operational currency mismatch. Finally, Model 4 incorporates both WPI inflation and interaction term for large operational exposure. Our key results remain unchanged across these different specifications. Looking at the first row in Table 7 we find that the coefficient on the depreciation dummy is positive and significant for all the specifications. This indicates that for firms with no currency exposure large rupee depreciation is associated with an increase in the growth of capital expenditure. This is likely a reflection of a positive impact on overall growth prospects of a more competitive exchange rate. Next we have the interaction term between the depreciation dummy and the size of lagged currency exposure. The coefficient on this term is negative and significant across different specifications. This indicates a robust relationship between the size of operational currency mismatch and the reduction in the growth of capital expenditures by firms during episodes of large Rupee depreciations. Firms with higher currency mismatch see a smaller increase in the growth of their capital expenditure in response to a rupee depreciation. It must be emphasized that these correlations do not necessarily imply a causal relationship between exposure elasticity and the growth of capital expenditure. Establishing such a causal relationship would require further analysis beyond the scope of this study. Of the remaining variables, growth in market capitalization and inflation are positively correlated with the growth of capital expenditure once we have taken into account time varying industry level fixed effects. The key insight of our analysis in this section is that, like output growth and earnings per share, higher exchange rate exposure elasticity is also associated with a lower growth of capital expenditure during episodes of large currency depreciation. This provides one potential channel through which high operational exposure leads to a lower output growth and lower earnings per share during periods of large currency depreciation.

Conclusion This paper aims at exploring the causes and effects of large “operational” currency exposure in one of the key emerging markets of the world
India. We use a firm level panel data set covering the period between 1995

82

A. Dhasmana

and 2011. The key findings of the paper can be summarized as follows
exchange rate volatility has a significant effect on the level of currency exposure of Indian firms apart from firm specific factors such as size growth. Further, higher operational exposure significantly increases the adverse impact of sharp exchange rate depreciation on output growth, earnings per share and capital expenditure growth even though it seems to encourage higher output growth during “normal” times. The results have important implications for policy makers worried about mitigating the impact of exogenous shocks. Implicit and explicit guarantees with regards to the value of exchange rate tend to raise the vulnerability of the economy to exchange rate shocks at same time that they encourage capital expenditures and possibly output growth during “normal” times. From an economic perspective our results indicate that exchange rate volatility can affect firm’s value by altering the way they organize their production. To the extent that this causes firms to deviate away from the optimal production path, this would lead to lower productivity and hence lower growth. Further research is called for to explore the nature of distortions in the production process encouraged by exchange rate volatility and their impact on firm level productivity. Looking at the relationship between the use of financial and operational hedges is another fruitful area of future research.

Acknowledgments I would like to thank the participants of 17th International Conference on Macroeconomic Analysis and International Finance, the editors, and an anonymous referee for their extremely valuable comments. All the remaining errors belong to the author.

References Adler, M., & Dumas, B. (1984). Exposure to currency risk: Definition and measurement. Financial Management, 13, 41
50. Bai, J., & Perron, P. (2003). Computation and analysis of multiple structural change models. Journal of Applied Econometrics, 18, 1
22. Bodnar, G. M., & Wong, M. H. F. (2000). Estimating exchange rate exposures: Some “weighty” issues. Technical Report, NBER. Bodnar, G. M., & Marston, R. C. (2001). A simple model of foreign exchange exposure. Working Papers, University of Pennsylvania, Wharton School, Weiss Center. Retrieved from http://finance.wharton.upenn.edu/weiss/wpapers/2000/00-3.pdf Dominguez, K. M., & Tesar, L. L. (2006). Exchange rate exposure. Journal of International Economics, 68, 188
218.

Operational Currency Exposure

83

Eichengreen, B., Hausmann, R., & Panizza, U. (2007). Currency mismatches, debt intolerance, and the original sin: Why they are not the same and why it matters. In S. Edwards (Ed.), Capital Controls and Capital Flows in Emerging Economies: Policies, Practices and Consequences. Chicago, IL: University of Chicago Press. Goldberg, L. (1993). Exchange rates and investment in United States Industry. The Review of Economics and Statistics, 75(4), 575
588. Goldberg, L., & Campa, J. M. (1995). Investment in manufacturing, exchange rates and external exposure. Journal of International Economics, 38(3
4), 297
320. Goldberg, L., & Campa, J. M. (1999). Investment, pass-through and exchange rates: A cross-country comparison. International Economic Review, 40(2), 287
314. Goldberg, P., & Knetter, M. M. (1997). Goods prices and exchange rates: What have we learned? Journal of Economic Literature, 35(3), 1243
1272. Goldstein, M., & Turner, P. (2004). Controlling currency mismatches in emerging market economies. Washington: Institute of International Economics. Gupta, P., Mishra, D., & Sahay, R. (2007). Behavior of output during currency crisis. Journal of International Economics, 72(2), 428
450. Hong, K., & Tornell, A. (2005). Recovery from a currency crisis: Some stylized facts. Journal of Development Economics, 76(1), 71
96. Hutchinson, M. M., & Noy, I. (2005). How bad are twins? Output costs of currency and banking crises. Journal of Money, Credit, and Banking, 37(4), 725
752. Jorion, P. (1990). The exchange-rate exposure of U.S. multinationals. The Journal of Business, 63, 331
345. Lane, P. R., & Milesi-Ferretti, G. M. (2007). The external wealth of nations mark II: Revised and extended estimates of foreign assets and liabilities, 1970
2004. Journal of International Economics, 3(2), 223
250. Nucci, F., & Pozzolo, A. F. (2001). Investment and the exchange rate: An analysis with firm-level panel data. European Economic Review, 45(2), 259
283. Parsley, D. C., & Popper, H. A. (2006). Exchange rate pegs and foreign exchange exposure in East and South East Asia. Journal of International Money and Finance, 25, 992
1009. Ranciere, R., Tornell, A., & Vamvakidis, A. (2010). Currency mismatch, systemic risk and growth in emerging Europe. Economic Policy, 25(64), 597
658. Shah, A., & Patnaik, I. (2010). Does the currency regime shape un-hedged currency exposure? Journal of International Money and Finance, 29(5), 760
769.

84

A. Dhasmana

Appendix Table A.1:

PAT and Exchange Rate Volatility

Dependent variable: PAT growth

Entire sample

Manufacturing

Non-manufacturing

3.2 [31.3] 0.002 1.45 [0.10] 6328

−2.2 [34.0] 0.004 1.96 [0.00] 3150

11.1 [50.5] 0.006 0.98 [0.47] 3178

Exchange rate volatility R2 F-statistic Total number of observations

Table A.2:

Currency Mismatch and Exchange Rate Volatility

Dependent variable: Abs. size of h1
h2 Exchange rate volatility R2 F-statistic Total number of observations

Entire sample

Manufacturing

Non-manufacturing

−0.012*** [0.002] 0.01 3.98 [0.00] 6506

−0.012*** [0.003] 0.04 3.77 [0.00] 3214

−0.012*** [0.003] 0.01 2.47 [0.00] 3292

Exchange Rates, Fundamentals, and Nonlinearities: A Review and Some Further Evidence from a Century of Data Panayiotis F. Diamandis, Anastassios A. Drakos and Georgios P. Kouretas Athens University of Economics and Business, Athens, Greece, e-mail: [email protected]

Abstract Purpose
The purpose of this paper is to provide an extensive review of the monetary model of exchange rate determination which is the main theoretical framework on analyzing exchange rate behavior over the last 40 years. Furthermore, we test the flexible price monetarist variant and the sticky price Keynesian variant of the monetary model. We conduct our analysis employing a sample of 14 advanced economies using annual data spanning the period 1880
2012. Design/methodology/approach
The theoretical background of the paper relies on the monetary model to the exchange rate determination. We provide a thorough econometric analysis using a battery of unit root and cointegration testing techniques. We test the price-flexible monetarist version and the sticky-price version of the model using annual data from 1880 to 2012 for a group of industrialized countries. Findings
We provide strong evidence of the existence of a nonlinear relationship between exchange rates and fundamentals. Therefore, we model the time-varying nature of this relationship by allowing for Markov regime switches for the exchange rate regimes. Modeling exchange rates within this context can be motivated by the fact that the change in regime should be considered as a random event and not predictable. These results show that International Symposia in Economic Theory and Econometrics, Vol. 23 G. P. Kouretas and A. P. Papadopoulos (Editors) Copyright r 2014 by Emerald Group Publishing Limited. All rights reserved ISSN: 1571-0386/doi:10.1108/S1571-038620140000023004

86

P. F. Diamandis et al.

linearity is rejected in favor of an MS-VECM specification which forms statistically an adequate representation of the data. Two regimes are implied by the model; the one of the estimated regimes describes the monetary model whereas the other matches in most cases the constant coefficient model with wrong signs. Furthermore it is shown that depending on the nominal exchange rate regime in operation, the adjustment to the long run implied by the monetary model of the exchange rate determination came either from the exchange rate or from the monetary fundamentals. Moreover, based on a Regime Classification Measure, we showed that our chosen Markov-switching specification performed well in distinguishing between the two regimes for all cases. Finally, it is shown that fundamentals are not only significant within each regime but are also significant for the switches between the two regimes. Practical implications
The results are of interest to practitioners and policy makers since understanding the evolution and determination of exchange rates is of crucial importance. Furthermore, our results are linked to forecasting performance of exchange rate models. Originality/value
The present analysis extends previous analyses on exchange rate determination and it provides further support in favor of the monetary model as a long-run framework to understand the evolution of exchange rates. Keywords: Monetary model, nonlinearity, fundamentals, cointegration, Markov switching model JEL Classifications: C22, C32, C53, F31, F47

Introduction A long standing puzzle in international finance is the difficulty of tying floating exchange rates to macroeconomic fundamentals such as money supplies, prices, outputs, and interest rates. Economic theory states that the exchange rate is determined by such fundamental variables, but floating exchange rates between countries with similar inflation rates are in fact well-approximated as random walk. Fundamental variables do not help predict future changes in exchange rates. Meese and Rogoff (1983) were the first to establish this result. Using data from the 1970s they examined the out-of-sample fit of alternative models of exchange rates. Their main finding was that if we use standard

Exchange Rates, Fundamentals, and Nonlinearities

87

measures of forecasting accuracy, such as the RMSE, then the forecasting accuracy increased overall when we forecast exchange rates with a simple random walk model as compared to that of the structural models of exchange rates. During the last decade, several works have managed to provide evidence for the forecasting superiority over long horizons of structural models such as the monetary model and portfolio balance models over the simple random walk model. However, these results are not robust while these models have poor forecasting performance in the short run. Overall, it seems that the recent results do not lead to a specific model/ specification and also that while one model will do well for one exchange rate may not do for another. An extensive subsequent literature shows the robustness of these results for the post-Bretton Woods floating period by using nonlinear econometric techniques, different currencies, data periodicity, and samples (Cheung, Chinn, & Pascual, 2005).1 Then, the difficult task to tackle is to model the exchange rates using fundamental economic variables and to obtain forward exchange rate fit both in-sample and out-of-sample to overcome the negative result of Meese and Rogoff (1983) that exchange rates and fundamentals are separated (Frankel & Rose, 1995, p. 1704). The development of cointegration theory in the mid-1980 has given a new lease of life to exchange rate modeling. This approach provides the framework to analyze the relationship between exchange rates and fundamentals from a long-run perspective. This is a natural way to model exchange rate behavior since it has been shown that fundamentals matter in the long run (see Chinn & Meese, 1995; Kim & Mo, 1995; MacDonald, 1999; Mark, 1995 among the very many studies). During the last 20 years a great deal of research has been done in order to examine whether the monetary model of exchange rate is a valid long-run framework to explain exchange rate determination and movements. Indeed, although these models do not have short-run predictive performance power (since they use low frequency data, their shorter forecast is one month or one-quarter ahead) they do find evidence of long-run exchange rate predictability. This literature has been exhaustive in analysis for most of bilateral exchange rates across different exchange rate arrangements and for different time periods. What have we learned from this long-run analysis? The evidence is mixed (Cheung et al., 2005) and we are not allowed to form some universal results for this important relationship even in the long run. Thus, many studies have found rather negative results for the existence of cointegration

1 Although the initial works on testing the validity of the monetary model in all its variants have led to some positive results, the results broke down when the period was extended beyond 1978 (Backus, 1984; MacDonald, 1999).

88

P. F. Diamandis et al.

between nominal exchange rates and monetary fundamentals during the post-Bretton Woods period; see for example Sarantis (1994), Berben and van Dijk (1998), Kilian (1999), Berkowitz and Giorgianni (2001), Faust, Rogers, and Wright (2003), Engel and West (2005), and Boudoukh, Richardson, and Whitelaw (2008). Other studies notably, MacDonald and Taylor (1994), Mark (1995), Kouretas (1997), and Diamandis, Georgoutsos, and Kouretas (1998) have found substantially support for the long-run validity of the monetary exchange rate model for several major bilateral exchange rates. Additionally, MacDonald and Taylor (1994) have also shown that the monetary model can outperform the naı¨ ve random walk model especially as we move from short-run horizons to long-run horizons. Diamandis, Georgoutsos, and Kouretas (2000) have further shown that considering the analysis in an I(2) cointegration model we can reach to additional positive results for a number of bilateral exchange rates of the Greek drachma vis-a`-vis major currencies. Recently, Dal Bianco, Camacho, and Perez-Quiros (2012) examine the short-run forecasting of the euro
dollar exchange rate with the development of a fundamentalsbased econometric model combining weekly exchange rate data with economic variables quoted at different frequencies. The analysis provides a very good forward exchange rate in-sample fit and, more importantly satisfactory out-of-sample results. Several explanations have been offered for this mixed evidence in finding cointegration between nominal exchange rates and monetary. A first explanation that is offered is linked with the underlying domestic and foreign demand for money functions. The instability in exchange rate modeling has been emerged as a stylized fact (see e.g., Faust et al., 2003). If these functions are unstable over time then it is difficult to find a cointegrating relation that resembles the monetary model. Berben and van Dijk (1998) and Berkowitz and Giorgianni (2001) show that the positive results of studies like Mark (1995) are based on the assumption that there exists a stable cointegration relationship among the variables of interest. Indeed, it is equally important with the finding of cointegration to test for the stability of this relationship over time. Diamandis et al. (1998, 2000) have examined the issue of stability by applying the tests of Hansen and Johansen (1993, 1999) and they further confirm the support in favor of the monetary model. In addition, these studies have also provided an economic and statistical identification of the cointegrating results using the theoretical framework developed by Johansen and Juselius (1994) and Johansen (1995) and they were able to identify one of the cointegrating relations with the monetary model. A second explanation for the lack of support the long-run monetary model is the relatively short sample of data usually employed which in most studies cover the post-Bretton Woods flexible exchange rates experience. Shiller and Perron (1985) and Hakkio and Rush (1991) have

Exchange Rates, Fundamentals, and Nonlinearities

89

documented that what matters for the power of the unit root and cointegration tests typically used is not the frequency of the data but the span of the data. Coupled with the evidence for the monetary model is the equally non-favorable evidence for the long-run purchasing power parity (PPP) for the recent float. Given that PPP is a building block of the monetary model we can argue that failure to establish a long-run relationship between exchange rates and domestic and foreign prices possible leads to a rejection of the validity of at least the monetarist variant of the monetary model. Therefore, the low power of the standard tests is considered to be a significant factor for the substantial negative evidence. In this paper we provide further evidence on the validity of the monetary model to the exchange rate using annual data that covers the period 1880
2013. Our analysis extends earlier works by Rapach and Wohar (2002) and Sarno, Valente, and Wohar (2004) since we adopt the approach of nonlinear Markov switching regime modeling and we estimate the flexible-price variant of the monetary model (Frenkel, 1976) and the Keynesian sticky-price variant (Dornbusch, 1976) for Australia, Belgium, Canada, Finland, France, Italy, Portugal, Spain, and the United Kingdom. We estimate a Markov Switching-Vector Error Correction model (MS-VECM) based on the evidence that at least one statistically significant cointegration vector exists. As Meese and Rose (1991) argue linear models can be modified by allowing nonlinear formulation of coefficients. Furthermore, we examine whether the importance of exchange rates and fundamentals in restoring the long-run equilibrium level implied by the exchange rate-monetary fundamentals model varies over time and whether is affected by the nominal exchange rate regime in force. There are several important findings that stem from our analysis. Our results show that nonlinearities in the relationship between the nominal exchange rate and the macroeconomic fundamentals variables are captured fairly well by the appropriately chosen estimated MS-VECM specification for each case. Furthermore, it is shown that for all the industrialized economies during fixed exchange rate regimes it is the monetary fundamentals that adjust to restore deviations from long-run equilibrium, whereas in the cases where a less restricted exchange rate regime has been in force the exchange rate adjusts to take the system back to long-run equilibrium. These alternative adjustment schemes are also reflected by the ex post (smoothed) transition probabilities. Finally, based on a Regime Classification Measure we show that our chosen Markov-switching specification performed well in distinguishing between the two regimes for all cases. The rest of the paper is organized as follows. The second section presents the literature review. In the third section we provide the key elements of the flexible-price monetary model and the motivation for considering time-varying fundamentals. In the fourth section the Markov switching

90

P. F. Diamandis et al.

regime methodology is presented. The fifth section reports our empirical results whereas our summary and concluding remarks are given in the sixth section.

A Review of the Literature The overall poor explanatory power of structural exchange rate models (see e.g., Frankel & Rose, 1995) provides a natural motivation to search for a model which will take into consideration some important features of nominal exchange rates and fundamentals. There are several approaches which have been developed in the last decade to provide more sophisticated models to study the relationship between the exchange rates and fundamentals. In addition there are several propositions that have been recently advanced by economists to provide more favorable evidence for the monetary model and its out-of-sample forecasting performance. One direction of analysis deals with the low power of standard unit root and cointegration tests. Levin and Lin (1992) offered the first response to this issue, since they recognized that the use of panel methodologies can improve substantially the power of unit root and cointegration tests. Subsequently, Groen (2000, 2005), Mark and Sul (2001), and Rapach and Wohar (2004) use panel data for the post-Bretton Woods era and with the application of panel cointegration tests they find strong support in favor of a stable long-run relationship between nominal exchange rates and monetary fundamentals. Moreover, these studies provide evidence that the estimated monetary model provide out-sample-forecasts that are superior to those provided by a naı¨ ve random walk model. However, Rapach and Wohar (2004) question the use of such aggregate data.2 A second response to the low power problem is the use of long spans of data, which in most cases cover more than a century. Rapach and Wohar (2002) and Sarno et al. (2004) are among the few studies that examined the long-run validity of the monetary model using annual data that covers the period 1880
2000 for 14 industrialized countries. They show that a stable cointegration relationship between nominal exchange rates and monetary fundamentals could be established for Belgium, Finland, France, Italy, the Netherlands, Portugal, Spain, and Switzerland.3

2

Frankel and Rose (1996), Papell (1997), and Taylor and Sarno (1998) are among the studies which have used panel cointegration techniques that led to strong support in favor of long-run PPP during the recent float. 3 Again Abuaf and Jorion (1990), Glen (1992), Lothian and Taylor (1996, 2000), and Taylor (2002) use long spans of data and they provide support in favor of longrun PPP.

Exchange Rates, Fundamentals, and Nonlinearities

91

A second direction to exchange rate modeling calls for the introduction of nonlinearities in the relationship between nominal exchange rates and fundamentals. Thus, Hsieh (1989), Meese (1990), Balke and Fomby (1997), Taylor and Peel (2000), Taylor, Peel, and Sarno (2001), and So (2001) have documented the existence of various nonlinearities in deviations of the spot nominal exchange rate from economic fundamentals. De Grauwe (2000) further underlines the significance of changing beliefs of the economic agents as a possible source for the existence of nonlinearities. Threshold cointegration and Markov switching regimes models are the two workhorses of this strand in the literature of estimating relationships with nonlinear features. Threshold models assume nonlinear mean reversion of exchange rates and smooth threshold dynamics. Within this framework Kilian and Taylor (2003) provide support in favor of the monetary model while their specification also exhibits superior out-of-sample performance compared to the random walk model. Markov switching models focus on the idea of changing regimes and timevarying coefficients. Engel and Hamilton’s (1990) important contribution provided evidence that a Markov-switching model of exchange rate outperforms the naı¨ ve random walk model. The intuition behind these models relies on the evidence offered by some studies that the monetary model performs well for some subperiod of the total sample but not for others (Meese, 1990) and also that there have been observed sudden regimes changes. Frydman and Goldberg (2001) show that such regime changes occur in the case of the dollar-mark exchange over the recent float. Mahavan and Wagner (1999), Marsh (2000), Bessec (2003), Clarida, Sarno, Taylor, and Valente (2003), Taylor and Peel (2000), Taylor et al. (2001), Sarno et al. (2004), and De Grauwe and Vansteenkiste (2007) are among several studies which analyze the monetary model in a Markov-switching model for a set of main bilateral exchange rates and they provide support in favor of a fundamental model. Bacchetta and Van Wincoop (2013) argued that large and frequent variations in the relationship between the exchange rate and macroeconomic fundamentals become evident when structural parameters in the economy are unknown and subject to changes. Furthermore, Frommel, MacDonald, and Menkhoff (2005a, 2005b) examine the RID variant of the monetary model using bilateral exchange rates of the Deutsche Mark/Euro against the US dollar and they show that in a two regime model, one regime accurately describes the RID version of the monetary model, whereas the other regime shows an inverted relationship between interest rates and exchange rates. However, the short-run forecasting performance of the monetary model with time-varying coefficients although provides better forecasts that the constant RID model is still outperformed by random walk forecasts. Ducker and Neely (2007) provided strong evidence that the Markov-switching regime models created

92

P. F. Diamandis et al.

ex ante trading rules in the foreign exchange market and delivered strong outof-sample portfolio returns for several major currencies. Recently, Syllignakis and Kouretas (2011) using data for the CEE economies employed a Markovswitching vector error correction model which allowed for regime shifts in the entire set of parameters and the variance
covariance matrix. The main finding of the analysis was that depending on the nominal exchange rate regime in operation, the adjustment to the long run implied by the monetary model of the exchange rate determination came either from the exchange rate or from the monetary fundamentals. Moreover, based on a Regime Classification Measure, it is shown that the chosen Markov-switching specification performed well in distinguishing between the two regimes for all cases. A third approach to deal with the poor forecasting performance of the monetary model is relied on the development of models that use very highfrequency data based on microeconomic variables linked to the structure of the market (Lyons, 2001). One of the arguments put forward in this line of research is that the failure of macroeconomic fundamentals in explaining and forecasting exchange rates can be resolved through the analysis of the microstructure of FX markets. This type of analysis focuses in the identification of certain elements of this market that will help us to understand the exchange rate. The relevant studies have concluded that order flows are a significant variable in understanding and forecasting the exchange rate (see Evans & Lyons, 2002, 2005; Gehrig & Menkoff, 2004; Goodhart, 1988; Lyons, 1995; Osler, 2006). The final approach to exchange rate modeling relies on the use of realtime macroeconomic data in the estimations and forecasts to evaluate the usefulness of monetary model in predicting with the same information which market participants have at each moment. Recently, Sarno and Valente (2009) using real-time data on a broad set of economic fundamentals for five major US dollar exchange rates over the recent float re-examined the predictive ability of exchange rate models within this framework. Their analysis leads to two key findings. First, they argue that the stylized fact of poor forecasting performance of exchange rate models may be the outcome of poor performance of model-selection criteria, rather than the lack of information content in the fundamentals. Second, they argue that the difficulty of selecting the best predictive model is largely due to frequent shifts in the set of fundamentals driving exchange rates, which can be interpreted as reflecting swings in market expectations over time. Furthermore, it is argued that the strength of the link between exchange rates and fundamentals differs across countries. Furthermore, as Frankel (1996) argues exchange rates are separated from fundamentals because of swings in expectations about future values of the exchange rate. He provides substantial evidence that support this argument. Therefore, this approach supports the argument that exchange

Exchange Rates, Fundamentals, and Nonlinearities

93

rate models exhibit poor performance not only because the information content of the fundamentals is deficient, but because volatile expectations and departures from rationality are likely to account for the failure of exchange rate models. Bacchetta and Van Wincoop (2004) developed an exchange rate model which incorporates the fact that practitioners in the foreign exchange market regularly change the weight they attach to different macroeconomic variables (see Cheung & Chinn, 2001). This fact is supported by the results of various survey studies based on the framework of a stylized rational-expectations model of exchange rate determination. The intuition in this model is that each time that an explanation is sought by rational agents for the observed exchange rate change, a macroeconomic variable is chosen for this purpose (more weight is put on it) whereas the rest of potentially significant macroeconomic variables are left out. Therefore it is argued that different observed variables may be taken as the scapegoat, so that weights attributed to economic variables change.

The Monetary Model and Nonlinear Characteristics of Fundamentals The monetary model of exchange rate determination is an extension of the quantity theory of money to the case of an open economy. It assumes that (i) real income and money supply are determined exogenously; (ii) capital and goods are perfectly mobile; (iii) foreign and domestic assets are perfect substitutes; (iv) goods’ prices are perfectly flexible; and (v) domestic (foreign) money is demanded only by domestic (foreign) residents. The early, flexible-price monetary model (Frenkel, 1976) relies on the twin assumptions of continuous PPP and the existence of stable money demand functions for the domestic and foreign economies. Recent experience with flexible exchange rates has shown, however, that real exchange rates have fluctuated substantially over the years causing shifts in international competitiveness. Stickiness in prices (Dornbusch, 1976) in conjunction with the uncovered interest parity (UIP) condition are usually invoked in order to allow for short-term deviations of both the nominal and the real exchange rates from their long-run levels as determined by the PPP. Moreover, the UIP condition is necessary for the derivation of the forward-looking version of the monetary model, under which the exchange rate depends on all the expected realizations of the forcing variables, that is, the monetary aggregates and the output variables. These two approaches are considered to be subcases of the real interest rate differential (Frankel, 1979). Under these assumptions a typical monetary reduced form equation is obtained (see Sarno & Taylor, 2002; Taylor, 1995 and MacDonald, 2007): et = β0 þ β1 ðmt − mt Þ − β2 ðyt − yt Þ þ ðit − it Þ þ ut ;

ð1Þ

94

P. F. Diamandis et al.

where et is the spot exchange rate (home currency price of foreign currency); mt denotes the domestic money supply; yt denotes domestic income; it denotes domestic interest rate; corresponding foreign magnitudes are denoted by an asterisk; ut is a disturbance error; and all variables apart from the interest rate terms, are expressed in natural logarithms. The expected signs of the coefficients in Equation (1) are: β1 > 0, β2 < 0. Different signs of the interest rate coefficients in Equation (1) will be produced under imperfect substitutability between the assets of the two countries. Associated with Equation (1) is a set of coefficients restrictions that are regularly imposed and tested. The most important restriction is whether proportionality exists between the exchange rate and relative monies (β1 = − β2 = 1). As we already mentioned several papers consider parameter instability as an explanation for the poor forecasting performance of the monetary model. This instability can be explained either by policy regime changes or instabilities in the money demand (an explanation offered in the early studies of monetary model) or PPP equations or agents’ heterogeneous beliefs (see Rossi, 2005, 2006). In addition, another source of the failure of monetary model to provide accurate exchange rate forecasts may be due to changes in way expectations are formed when a switch from fixed exchange rates to flexible exchange rates occur (Flood & Rose, 1995).4 Equation (1) implies that, if the departure from the exchange ratemonetary fundamentals relationship ut is stationary given et , mt − mt , yt − yt ∼ Ið1Þ the nominal exchange rate and the fundamentals exhibit a common stochastic trend and are cointegrated with cointegrating vector [1, −1], that is, the proportionality hypothesis holds. Then given the Granger Representation Theorem (Engle & Granger, 1987), the nominal exchange rate and the fundamentals must possess a VECM representation in which ut plays the part of the error correction term. We follow Sarno et al. (2004) and we use exactly a linear VECM representation in order to examine the relative importance of the nominal exchange rate and the fundamentals in restoring equilibrium in the long-run relationship linking exchange rate and fundamentals across different exchange rate regimes since the late 19th century. Therefore, we employ a generalization of a standard linear VECM which is capable of allowing all of the VECM

4 Uncovered interest parity (UIP) is frequently invoked to provide the rational expectation version of the monetary model. However, empirical evidence in favor of UIP is rather weak. Given that we use a more than a century long data we prefer to examine the validity of the monetary model in its reduced form as given by Equation (1).

Exchange Rates, Fundamentals, and Nonlinearities

95

parameters to change over time and to identify the various regimes that characterize the long sample periods that we examine in the present paper.

Econometric Methodology Johansen Multivariate Cointegration Technique The estimation of the proposed MS-VECM is conducted using a two-stage maximum likelihood procedure. The first stage refers to the cointegration analysis which is based on the multivariate cointegration technique developed by Johansen (1988, 1991) and extended by Johansen and Juselius (1990) which is a Full Information Maximum Likelihood (FIML) estimation method. It makes use of the information incorporated in the dynamic structure of the model and it also estimates the entire space of the longrun relationships among a set of variables, without imposing a normalization on the dependent variable a priori. Although the Johansen procedure is well known we discuss it briefly in light of some recent extensions of the methodology that are applied in this paper. Consider a p-dimensional vector time series zt with an autoregressive representation (AR) which in its error correction form is given by Δzt =

k−1 X

Γi Δzt − i þ Πzt − 1 þ γDt þ μ0 þ μ1 t þ εt ;

t = 1; …; T;

ð2Þ

i=1

where zt = [e; m; m ; y; y ; i; it ]t as defined in the third section, zk þ 1 ; …; z0 are fixed, and εt ∼ Niidp ð0; ΣÞ. The adjustment of the variables to the values implied by the steady state relationship is not immediate due to a number of reasons like imperfect information or costly arbitrage. Therefore, the correct specification of the dynamic structure of the model, as expressed by the parameters ðΓ1 ; …; Γk − 1 ; γÞ, is important in order that the equilibrium be revealed. The matrix Π = αβ0 defines the cointegrating relationships, β, and the rate of adjustment, α, of the endogenous variables to their steady state values. Dt is a vector of nonstochastic variables, such as centered seasonal dummies which sum to zero over a full year by construction and are necessary to account for short-run effects which could otherwise violate the Gaussian assumption, and/or intervention dummies; μ is a drift and T is the sample size. If we allow the parameters of the model θ = ðΓ1 ; …; Γk − 1; Π; γ; μ; ΣÞ to vary unrestrictedly, then model (2) corresponds to the I(0) model. The I(1) and I(2) models are obtained if certain restrictions are satisfied. Thus, the higher-order models are nested within the more general I(0).

96

P. F. Diamandis et al.

It has been shown (Johansen, 1991) that if zt ∼ Ið1Þ, then matrix Π has reduced rank r < p, and there exist p × r matrices α and P β such that Π = αβ0 . 0 Furthermore, Ψ = α⊥ ðΓÞβ⊥ has full rank, where Γ = I − κi = 1 Γi and α⊥ and β⊥ are p × ðp − rÞ matrices orthogonal to α and β, respectively. Following this parameterization, there are r linearly independent stationary relations given by the cointegrating vectors β and p − r linearly independent nonstationary relations. These last relations define the common stochastic trends of the system and the contribution to the various variables. By contrast the AR representation of model (2) is useful for the analysis of the long-run relations of the data.

The Markov Switching-Vector Error Correction Model The second stage amounts to the study of the dynamics of the regime switching and the stochastic processes evolved in a set of the nominal exchange rate and fundamental variables that mentioned before, we adopt the MS-VECM, introduced in Krolzig (1997), which is a multivariate generalization of the univariate Hamilton (1989, 1994) model. This model allows, in a multivariate context, for shifts in the stochastic volatility regime driving the foreign exchange markets. Thus, the change in regime should be considered as a random event and not predictable. In addition, the effect of these shifts must be considered when we investigate the stochastic properties of the foreign exchange market volatility and the possible links between the exchange rate and monies supplies, outputs. The usefulness of a timevarying coefficients approach against structural models with constant coefficients has been illustrated in several studies (see e.g., Schinasi & Swamy, 1989; Wollf, 1987).5 Later studies separate the switches in mean and variance either by using two distinctive state variables for mean and variance each (Dewachter, 1997) or by using a Markov switching model with four states differing in mean or variance (Dewachter, 2001). Consider that Δyt is a T × 1 vector containing the observations for the single stationary time series {Δyt}, and let ΔYt = ðΔy1t ; …; ΔyKt Þ0 , t = 1,…,T be the K-dimensional vector, where T is the sample size. A p-th order MS-VECM [MS-VECM(p)] model can be written as ΔYt = A0 ðst Þ þ A1 ðst ÞΔYt − 1 þ ⋯ þ Ap ðst ÞΔYt − p þ Bðst Þectt − 1 þ ut ; ut ∼ NIDð0; Σðst ÞÞ;

5

ð3Þ

See also, Engel (1994), Kim (1994), Hamilton and Susmel (1994), Hamilton and Lin (1996), Engel and Kim (2001), Lee and Chen (2006), and Kanas and Kouretas (2007).

Exchange Rates, Fundamentals, and Nonlinearities

97

where st is the unobservable regime, A0 ðst Þ; …; Ap ðst Þ are regime-dependent autoregressive parameter matrices, B(st) is the regime-dependent parameter matrix of the error correction term (ect), and ut is the innovation process with a regime-dependent variance
covariance matrix Σðst Þ. It is assumed that st follows an irreducible ergodic m-regime Markov process with the transition matrix 2 3 p11 p12 …: p1M 6 p21 p22 …: p2M 7 7 P=6 ð4Þ 4 …: …: …: …: 5 ut ∼ NIDð0; Σðst ÞÞ: pM1 pM2 …: pMM The transition probabilities pij in P are constant, and given by pij = Prðst þ 1 = j | st = iÞ;

m X

pij = 1

∀i; j ∈ f1; …; mg:

ð5Þ

j=1

Maximum likelihood estimation of the model is based on the Expectation Maximization (EM) algorithm.6 One can also calculate the unconditional probability that the system of the two currency is in regime i, i = 1,…m, at any given date, Pr(st = i). Also, the “smoothed” probabilities can be obtained, representing the ex post inference about the system being in regime i at date t. Further, one could date the regime switches. For instance, for two regimes, an observation is assigned to the first regime if Pr(st = 1|ΔYT) > 0.5, and to the second regime if Pr(st = 1|ΔYT) < 0.5.

Data and Empirical Results The data consist of annual observations for the nominal exchange rate (units of foreign currency per US dollar), the money supply, real GDP, and short-term interest rates for 14 advanced economies: Australia, Belgium, Canada, Denmark, Finland, France, Italy, the Netherlands, Norway, Portugal, Spain, Sweden, Switzerland, and the United Kingdom. The corresponding money supply and real GDP for the United States are denoted with an asterisk. The data spans from 1880 to 2012 except for Belgium, Finland, France, Italy, the Netherlands, Portugal, and Spain for which the sample runs until 1998 which marks the end of their national currency with the formation of the Eurozone on January 1, 1999, and thus covers a number of alternative international monetary arrangements, such as the gold

6

The EM algorithm was first developed by Dempster, Laird, and Rubin (1977) and was extended by Hamilton (1989) and Krolzig (1997).

98

P. F. Diamandis et al.

standard, the Bretton Woods period, and the current flexible exchange rate regime. Due to specific data availability in particular for the short-term interest rates the exact dates for each case are as follows: Australia (1900
2012); Belgium (1880
1998); Canada (1900
2012); Denmark (1910
2012); Finland (1910
1998); France (1880
1998); Italy (1900
1998); Netherlands (1880
1998); Norway (1900
2012); Portugal (1910
1998); Spain (1910
1998); Sweden (1880
2012); Switzerland (1880
2012); the United Kingdom (1880
2012); and the United States (1880
2012). The nominal exchange rates, the money supplies, and real GDP are obtained from the data set employed by Rapach and Wohar (2002) and the short-term interest rates were obtained from Homer and Sylla (2005). All series were updated using data taken from the International Financial Statistics of the International Monetary Fund.7 All variables are measured in natural logarithms. The standard practice is to subject variables to a battery of unit root tests. There are many different tests for a unit root in the autoregressive (AR) polynomial of a univariate process that have been proposed, but the most common is the augmented Dickey
Fuller (ADF) test proposed by Said and Dickey (1983). It is based on the AR approximation of a general ARIMA process and is given in Equation (1).8 Δyt = α þ ðρ − 1Þyt − 1 þ

k X

γ j Δyt − j þ et :

ð6Þ

j=1

The null hypothesis of a unit root can be tested by estimating (1) using OLS and then using a t-type test statistic to test the hypothesis (1 − ρ) = 0. The choice of the lag truncation parameter k is important for the small sample properties of the test because when the number of lags is greater than the true number of lags there is a decrease in the power of the test, while two few lags leads to under sized tests. There are some potential problems with unit root testing using Equation (6), however.

7

In turn, the nominal exchange rates series are from Taylor (2002), and the money supply and real GDP series are from Bordo and Jonung (1998), Bordo, Bergman, and Jonung (1998), and Bordo’s Financial Crises database: https://sites.google.com/ site/michaelbordo/home4. 8 An ARIMA or autoregressive integrated moving average process assumes that a time series can be modeled in the time domain as a function of lagged values of itself and current and lagged values of the innovation or error to the process. An ARIMA (p, d, q) takes the general form ϕðLÞΔd yt þ μ = θðLÞεt where the autoregressive lag polynomial ϕðLÞ = 1 þ ϕ1 L þ ϕ2 L2 þ ··· þ ϕp Lp is of order p, the order of integration is given by the differencing parameter d, and the moving average polynomial θðLÞ = 1 − θ1 L − θ2 L2 − ··· − θq Lq if of order q.

Exchange Rates, Fundamentals, and Nonlinearities

99

The first problem is low power of the test relative to local alternatives. Elliott, Rothenberg, and Stock (1996) (ERS) proposed an estimator that increases the power of the unit root test substantially by using a GLS detrending procedure. One can motivate the unit root tests using the data generating process (DGP) in Equation (7) yt = dt þ ut ;

ut = ρut − 1 þ vt ð7Þ P∞ P where vt = φðLÞet = j = 0 φj et − j ; dt = ζ 0 zt = pi= 0 ζ i ti for p = 0; 1. When estimating Equation (6) the parameters of the deterministic components are estimated via OLS and are treated as nuisance parameters in the distribution of the unit root tests. By estimating these nuisance parameters using OLS the power of the test statistics is diminished. ERS propose a weighted least squares or GLS method to estimate these parameters and then detrend the data prior to testing for a unit root. For series fxt gTt= 0 define ðxα0 ; xαt Þ = ðx0 ; ð1 − αLÞxt Þ for some value α = 1 þ c=T. The GLS detrended series is 0 then defined as y~t ≡ yt − ζ^ zt where ζ^ minimizes Sðα; ζÞ = ðya − ζ 0 zαt Þ0 ðya − ζ 0 zαt Þ. ERS suggest imposing c = − 7:0 for p = 0 and c = − 13:5 for p = 1.9 Testing for a unit root can then be done by estimating Equation (8) using OLS and calculating a t-type test statistic as in Equation (1), which is referred to as the DF-GLSμ statistic when p = 0 and DF-GLSτ when p = 1. Δy~t = ðρ − 1Þy~t − 1 þ

k X j=1

γ j y~t − j þ etk :

ð8Þ

Although low power is always a problem for unit root tests, another concern is that size distortions in the tests may be a problem because of the properties of the underlying DGP. One source of size distortion is the presence of large and negative moving average (MA) parameters in the DGP. Schwert (1987) was one of the first to point out that standard unit root tests like the ADF are severely oversized when there are large negative MA terms in the DGP. He suggests increasing the value of k, the lag truncation parameter in Equations (6) and (8), to more accurately allow the AR process in Equation (10) to approximate the MA components in the ARIMA. We estimate ARIMA models for each of the series of interest in this study in order to gauge how serious this source of size distortion may be in our application. Table 1 displays estimation results for our series.10

9 Cook (2006) finds that the power of the tests in finite samples under alternative DGPs can be increased with alternative values for c. 10 We used the Box
Jenkins procedure to identify several candidate models for each series and then chose the best fitting model based on residual serial correlation tests, significance of the parameter estimates, and R2.

100

P. F. Diamandis et al.

Table 1: Unit Root and Stationarity Tests Variable

ADF t-tests tμ

Australia e −0.471 −0.507 m − m −1.103 y − y −0.905 ι − ι Belgium e 1.032 0.056 m − m −0.312 y − y −0.225 ι − ι Canada e −0.151 1.133 m − m −0.335 y − y −1.156 ι − ι Denmark e −0.201 −1.225 m − m −0.335 y − y −1.155 ι − ι Finland e −0.151 1.133 m − m −0.335 y − y −1.156 ι − ι France e −0.344 1.209 m − m −0.147 y − y −1.304 ι − ι Italy e −1.902 −1.335 m − m −0.701 y − y −1.338 ι − ι Netherlands e −0.300 −0.892 m − m −1.208 y − y −1.156 ι − ι Norway e −1.988 −0.998 m − m −1.556 y − y −1.200 ι − ι Portugal e −0.909 −1.561 m − m

MZGLS a

DF-GLSu

MZGLS t

KPSS tests ημ

ητ

−2.288* 2.001 −2.061* −2.886*

1.921* 2.033* 1.904* 0.987*

0.262* 0.366* 0.152* 0.205*

−7.566 1.098 −9.051* −6.667

1.067 −3.233 −1.045 −1.889

1.546* 1.335* 2.155* 1.998*

0.513* 0.292* 0.441* 0.393*

−1.233 0.088 −0.177 1.089

−8.002 −5.787 −5.576 −9.206*

−1.776 1.093 0.098 0.205

0.887* 0.509* 0.443* 1.019*

0.311* 0.218* 0.191* 0.166*

2.023 0.306 −0.398 −0.721

−1.245 0.091 −3.445* 1.331

2.556 −9.776* −2.226 −4.121

−0.991 0.097 −0.998 −1.445

0.355* 0.322 0.591* 0.519*

0.120 0.303* 0.202* 0.291*

−1.603 −2.276 −3.998* −0.098

1.956 0.445 −0.487 −1.766

−1.233 0.088 −0.177 1.089

−8.002 −5.787 −5.576 −9.206*

−1.776 1.093 0.098 0.205

0.887* 0.509* 0.443* 1.019*

0.311* 0.218* 0.191* 0.166*

−1.901 −1.302 −1.335 −0.101

−1.001 −1.122 −0.609 −2.608

−1.036 −1.265 −0.305 −2.645

−2.361 −3.678 −9.305* 2.001

−1.445 −1.887 −1.023 1.103

0.503* 0.398* 0.201 0.491*

0.256* 0.181 0.301* 0.105

−1.701 −2.444 −1.299 −1.409

−2.132 −1.228 −1.187 −1.609

−2.098* 0.088 −0.662 −1.116

−6.222 1.116 1.105 −6.225

−1.332 1.220 0.122 −2.445*

0.698* 0.672* 0.433 0.672*

0.115 0.318* 0.202* 0.208*

−1.558 −3.405* −2.988* −0.098

−1.336 −1.288 −0.897 −1.333

−0.609 −0.293 −0.209 −1.307

−9.044* 1.667 −7.889 −5.332

−1.446 −1.111 −1.209 1.885

0.498* 0.508* 0.307 0.625*

0.177* 0.112 0.299* 0.191*

−1.206 −1.665 −2.003 −0.552

−0.445 −2.002 −2.113 −1.103

−2.333 −1.889 −1.233 −2.092

2.063 1.099 −4.433 −7.229*

0.898 1.693 0.909 0.805

0.661* 0.701* 0.551* 0.908*

0.133 0.155* 0.177* 0.206*

−3.243* −1.089

−1.990 −1.065

−0.998 0.902

−6.443 −5.609

−1.223 −2.001

0.323* 1.202*

0.221* 0.167*

tτ

tμ

tτ

−1.408 −1.885 −1.277 −1.566

−1.301 −1.623 −1.812 −2.651

−2.033 −2.083 −1.129 −1.356

−5.671 1.299 −3.213 −4.889

−1.075 −1.596 −1.723 −2.191

0.792 0.367 −0.385 −2.891*

−0.454 −1.103 −1.044 −3.011

−1.603 −2.276 −3.998* −0.098

1.956 0.445 −0.487 −1.766

−1.558 −4.228* −3.998* −2.001

101

Exchange Rates, Fundamentals, and Nonlinearities

Table 1: (Continued ) Variable

ADF t-tests tμ

y − y −1.612 −0.901 ι − ι Spain e −0.113 −1.215 m − m −0.909 y − y −2.021 ι − ι Sweden e −1.209 −1.335 m − m −0.335 y − y −1.334 ι − ι Switzerland e −1.309 1.133 m − m −1.613 y − y −0.992 ι − ι United Kingdom e −0.891 −0.391 m − m −0.335 y − y −1.156 ι − ι

MZGLS a

DF-GLSu

MZGLS t

KPSS tests ημ

ητ

−1.022 0.667

0.399 0.554*

0.222* 0.111

−3.223 −3.998 −10.28* −7.355

−2.077* −3.889* −1.756 −1.566

0.498* 0.737* 0.692* 1.203*

0.221* 0.299* 0.401* 0.303*

−0.882 −0.904 −0.177 −1.267*

−4.772 −2.668 −1.989 −6.332

0.995 1.701 −2.001 0.309

0.499* 0.509* 0.771* 0.899*

0.166* 0.218* 0.227* 0.321*

−1.442 0.901 −2.551 −1.334

−1.609 0.771 −2.991 1.066

−3.225 −5.628 −3.229 −5.228

0.881 −1.223 2.223 −2.281*

0.883* 0.495* 0.513* 0.819*

0.224* 0.333* 0.205* 0.332*

1.001 0.901 −0.487 −1.766

−1.281 0.209 −0.177 1.089

−3.818 −3.776 −5.576 −9.206*

−1.229 1.093 0.098 0.205

0.332* 0.509* 0.443* 1.019*

0.161* 0.218* 0.191* 0.166*

tτ

tμ

tτ

−1.345 −1.223

−1.233 −1.612

−1.361 0.803

0.999 −11.225*

−1.655 −1.099 −1.244 −1.344

−1.361 −1.108 −0.665 −1.335

−1.208 −1.100 −0.882 −0.998

−1.445 −1.003 −3.998* −1.345

−1.225 0.668 −0.487 −3.781*

−1.771 −2.276 −2.991* −0.598 −2.881* −1.981 −3.998* −0.098

Note: e, m − m , y − y , ι − ι are the nominal exchange rate, relative money supply, relative real output, and nominal interest rate differential, respectively. • tμ and tτ are the standard augmented Dickey
Fuller test statistics when the relevant auxiliary regression contains a constant and a constant and a trend respectively. The number of lagged differenced terms required for serial correlation correction in the ADF auxiliary regressions is selected on the basis of a general to specific testing strategy which is terminated when a sequence of t-ratio elimination tests on the lagged differenced terms leads to a rejection at the 10% significance level and the residuals of the resultant specification satisfy standard misspecification testing (Perron & Ng, 1996). The response surface regressions of MacKinnon (1991, 1994) are used for determining the significance of the ADF test statistics. • The DF-GLSu by Elliott (1999) is a test with an unconditional alternative hypothesis. The critical values for the DF-GLSu test at the 1%, 5%, and 10% significance level are: −3.28, −2.73, −2.46 (with constant) and
3.71, −3.17, −2.91 (with constant and trend), respectively (Elliott, 1999). • MZa and MZt are the Ng and Perron (2001) GLS versions of the Phillips
Perron tests. The critical values at 5% significance level are: −8.10 and −1.98 (with constant and with constant and trend), respectively (Ng & Perron, 2001, Table 1). • ημ and ητ are the KPSS-test statistics for level and trend stationarity respectively (Kwiatkowski et al., 1992). For the computation of these statistics a Newey and West (1994) robust kernel estimate of the “long-run” variance is used. The kernel estimator is constructed using a quadratic spectral kernel with VAR(l) pre-whitening and automatic data-dependent bandwidth selection (see, Newey & West, 1994 for details). The 5% critical values for level and trend stationarity are 0.461 and 0.148 respectively, and they are taken from Sephton (1995, Table 2). • * and ** indicate significance at the 95% and 99% confidence level respectively.

102

P. F. Diamandis et al.

Two features of many economic time series tend to affect the size and power of usual unit root tests. In particular, a large negative moving average root may induce size distortions, while a large autoregressive root may result in low power. When this is the case it is preferred to apply the MZα, MZt, MSB, and the MPT tests of Ng and Perron (2001), which are precisely designed to overcome both size distortion and low power problems when the data are characterized by these features. These tests are extensions of the M tests of Perron and Ng (1996) that use Generalized Least Squares (GLS) detrending of the data, together with a modified information criterion for the selection of the truncation lag parameter. Ng and Perron (2001) have developed a modified information criterion that chooses k in Equations (6) and (8) in a way that mitigates the size distortion in unit root tests. It is based upon an autoregressive estimate of the long-run variance of yt, denoted s2AR . This estimate is calculated as σ^ 2k ; ð9Þ [1 þ γð1Þ]2 P P where γð1Þ = ki= 1 γ i and σ^ 2k = ðT − kÞ − 1 Tt= k þ 1 e^2tk and γ i and fe^tk g. The parameters can all be estimated from Equation (8) using OLS.11 The modified information criteria (MIC) is given as s2AR =

MICðkÞ = lnð^σ 2k Þ þ

CT ðτT ðkÞ þ kÞ ; T − kmax

ð10Þ

P where τT ðkÞ = ð^σ 2k Þ − 1 ρ^ Tt= kmax þ 1 y~2t − 1 and kmax is the largest lag truncation considered. When CT = lnðT − kmax Þ, Equation (10) represents the modified Bayesian information criterion (MBIC) and when CT = 2 it is the modified Akaike information criterion (MAIC). Ng and Perron (2001) also suggest using three tests that have less size distortion in the presence of MA errors than standard tests. These tests are MZρ ; MZt , and MSB, collectively referred to as the M-tests. The tests are calculated from estimates of Equation (8) as follows: MZρ = ðT − 2 y~2t

− s2AR Þ

2T

−2

t X t=1

!−1 y~ 2t − 1

ð11Þ

and

11 Perron and Qu (2007) suggest that small sample power can be improved if the parameters used to construct the estimate of the long-run variance are estimated from Equation (6) rather than Equation (8).

Exchange Rates, Fundamentals, and Nonlinearities

" MSB =

T −2

PT

2

t = 1 y~t − 1

s2AR

103

#12 ð12Þ

and MZt = MZρ × MSB. Finally, since it has been shown that the standard DF and PP unit root tests are biased toward the acceptance of the unit root hypothesis we also apply the Kwiatkowski, Phillips, Schmidt, and Shin (1992) (KPSS) test for the null hypothesis of level or trend stationarity against the alternative of nonstationarity and these additional results will provide robust inference. The KPSS test has two components: in the first the null hypothesis is the stationarity of series in level; in the second the null hypothesis is that of trend stationarity. We implement the second test only when the first hypothesis is rejected. The results of the unit root and stationarity tests are presented in Table 1. The results unambiguously lead to the conclusion that we are unable to reject the null hypothesis of nonstationarity based on the DF-GLSu and MZGLS and a MZGLS tests and we reject the null hypothesis of stationarity with the KPSS t test for the levels of all series. However, when we take the first difference of each variable then all tests indicate that these series are I(1) processes. Given the low power of the unit root tests against alternative hypotheses it is important that we also test for structural breaks in the time series when analyzing the stochastic properties of the nominal exchange rates, the relative money supplies, and the relative real output. This seems appropriate given the variety of exchange rate regimes that our data covers. In that respect, we employ the Zivot and Andrews (1992) test with one structural break. The endogenous structural break test of Zivot and Andrews (1992) is a sequential test which utilizes the full sample and uses a different dummy variable for each possible break date. This test has several desirable properties: (a) it determines the structural breaks “endogenously” from the data, (b) its null distribution is invariant to level shifts in a variable, and (c) it is easy to interpret; by including breaks under both the null and alternative hypotheses, a rejection of the null hypothesis of a unit root implies unambiguously trend stationarity. Furthermore, for reasons of comparison and robustness we also apply the detrended GLS unit root tests, which was recently developed by Perron and Rodriguez (2003), against the alternative of stationarity around a structural break, which is an extension of the Elliott et al. (1996) detrended GLS unit root tests we used above. As in the Zivot and Andrews (1992) test the structural change is allowed to occur at an unknown point of time. The results are also shown in Table 2 and in all cases there is no evidence of one or two structural breaks in nominal exchange rate, relative money supplies, and relative real outputs. We conclude our unit root testing by applying the unit root tests developed by Kapetanios, Shin, and Snell (2003). These tests are constructed

104

P. F. Diamandis et al.

Table 2: Structural Breaks and Nonlinear Unit Root Tests Variable

Australia e m − m y − y ι − ι Belgium e m − m y − y ι − ι Canada e m − m y − y ι − ι Denmark e m − m y − y ι − ι Finland e m − m y − y ι − ι France e m − m y − y ι − ι Italy e m − m y − y ι − ι Netherlands e m − m y − y ι − ι Norway e m − m y − y ι − ι

Zivot
Andrews

MZGLS a

MZGLS t

c tNL

t tNL

tμ

tτ

−3.938 −3.405 −4.114 −1.109

−3.639 −2.223 −4.278 −1.925

−28.763 −19.012 −28.315 −27.251

−2.831 −2.651 −2.562 −1.983

−2.431 −3.001 −2.053 −2.981

−2.454 −3.445 −3.001 −3.521

−3.994 −2.226 −2.998 −1.688

−3.665 −1.453 −2.667 −3.911

−26.332 −22.331 −28.219 −27.665

−2.032 −3.051 −2.668 −2.883

−2.852 −3.032 −1.988 −2.033

−3.212 −3.578 −2.013 −1.987

−3.669 −4.002 −2.338 −3.771

−2.998 −4.233 −4.227 −2.135

−29.336 −26.665 −28.334 −27.887

−3.122 −2.881 −3.229 −3.166

−1.677 −1.979 −1.455 −2.199

−2.988 −2.556 −3.292 −2.989

−2.013 −3.504 −2.592 −3.202

−2.277 −3.609 −2.698 −2.168

−25.344 −23.434 −27.557 −30.252

−2.155 −2.681 −3.105 −3.200

−2.332 −1.808 −1.809 −2.509

−2.575 −2.001 −3.168 −1.628

−2.108 −4.112 −4.651* −3.223

−2.206 −4.233 −3.105 −2.991

−26.107 −28.105 −29.102 −28.993

−2.771 −4.065* −3.606 −3.221

−1.803 −2.661 −1.901 −2.709

−2.441 −3.701* −4.288* −2.665

−3.005 −3.661 −5.031* −4.221

−2.103 −4.557 −4.512 −2.908

−24.391 −29.883 −24.205 −23.689

−2.189 −3.477 −3.688 −2.704

−2.558 −3.051* −2.099 −2.101

−2.228 −2.705 −3.105 −2.609

−3.208 −3.789 −2.501 −2.990

−2.243 −4.198 −4.099 −2.301

−28.344 −29.228 −30.500 −29.809

−3.208 −2.667 −2.990 −3.202

−2.402 −1.109 −2.661 −2.509

−2.669 −2.109 −3.701* −3.001

−3.558 −4.002 −2.702 −3.609

−2.701 −4.233 −2.621 −4.235

−24.220 −26.665 −21.488 −30.702

−3.101 −2.881 −2.991 −3.555

−2.755 −1.979 −1.728 −2.402

−3.209 −2.556 −3.552 −2.559

−3.911 −3.766 −2.523 −3.200

−2.776 −2.244 −1.992 −2.900

−27.209 −29.883 −29.300 −30.505

−3.403 −2.454 −3.709 −3.208

−1.902 −2.023 −1.989 −2.098

−2.333 −3.005 −3.902 −2.009

Exchange Rates, Fundamentals, and Nonlinearities

105

Table 2: (Continued ) Variable

MZGLS a

MZGLS t

−3.111 −4.544 −4.305 −4.501

−28.554 −27.883 −30.922 −28.330

−3.333 −3.026 −3.441 −2.202

−2.012 −2.171 −1.988 −2.409

−3.566 −2.388 −3.609 −2.108

−2.446 −4.207 −1.825 −2.442

−29.708 −29.323 −21.201 −27.228

−3.601 −2.701 −3.662 −3.487

−1.901 −2.334 −2.233 −2.601

−2.101 −2.198 −2.779 −2.101

−2.335 −4.018 −4.232 −2.991

−30.202 −29.305 −29.445 −30.669

−3.559 −2.440 −3.338 −3.203

−1.773 −2.778 −2.988 −2.443

−2.619 −3.021 −3.115 −2.601

−2.781 −4.509 −4.908 −2.709

−26.446 −23.778 −29.406 −29.171

−3.337 −3.232 −3.454 −3.209

−2.708 −2.889 −1.991 −2.206

−3.202 −2.309 −2.404 −2.508

−2.504 −4.198 −2.661 −2.347

−29.552 −27.556 −27.558 −30.199

−3.773 −2.768 −3.102 −3.637

−1.388 −1.882 −1.669 −2.500

−2.052 −2.602 −2.882 −3.400

Zivot
Andrews tμ

Portugal e −3.928 m − m −4.013 −2.905 y − y ι − ι −4.238 Spain e −3.222 m − m −4.109 y − y −1.997 ι − ι −3.106 Sweden e −3.441 m − m −3.334 y − y −2.779 −3.556 ι − ι Switzerland e −3.258 m − m −3.801 y − y −4.555 ι − ι −2.881 United Kingdom e −3.202 m − m −3.776 y − y −3.199 ι − ι −3.192

c tNL

t tNL

tτ

Note: e, m − m , y − y , ι − ι are the nominal exchange rate, relative money supply, relative real output, and nominal interest rate differential, respectively. • tμ and tτ are Zivot and Andrews test statistics for the null hypothesis of a unit root against the alternative of stationarity with a structural break. The critical values at 5% significance level are: −4.80 and −5.08 (with constant and with constant and trend), respectively (Zivot & Andrews, 1992, Table 1). • MZGLS and MZGLS are the Perron and Rodriguez (2003) GLS versions to the case where a a t change in the trend function is allowed to occur . The critical values at 5% significance level are: −31.04 and −3.91 (with constant and with constant and trend), respectively (Perron & Rodriguez, 2003, Table 1). c t • tNL and tNL are the nonlinear unit root tests with the null of nonstationarity against the alternative of a nonlinear exponential smooth transition autoregressive (ESTAR) process, with a constant and a constant and a linear trend, respectively. The asymptotic critical values at the 5% (1%) critical values for an equation with a constant and a linear trend are −2.93 (−3.48) and −3.40 (−3.93) respectively (Kapetanios et al., 2003, Table 1). • * and ** indicate significance at the 95% and 99% confidence level respectively.

106

P. F. Diamandis et al.

within the nonlinear STAR framework and they have better properties as compared to the Dickey
Fuller test. Specifically, Kapetanios et al. (2003) analyze the implications of the existence of a specific type of nonlinear dynamics for unit root testing procedures. They develop a test for the null hypothesis of a unit root process against an alternative of a nonlinear exponential smooth transition autoregressive (ESTAR) process which is globally stationary. Furthermore, this testing procedure has been designed to have power against this alternative ESTAR process. The results of the application of this test are also reported in Table 2. We consider the case of a constant and a constant and a linear trend in the series. For each of our series we are unable to reject the null of a unit root in favor of nonlinearity at conventional levels of significance. Based on the evidence from the unit root and stationarity tests as well as form the structural break tests we conclude that for each case et ∼ Ið1Þ; ðmt − mt Þ ∼ Ið1Þ; ðyt − yt Þ ∼ Ið1Þ and ðit − it Þ ∼ Ið1Þ and therefore we can test the full version of the monetary model by applying the Johansen (1988, 1991) and Johansen and Juselius (1990) multivariate cointegration technique described in Section 3. Table 3 reports the cointegration results based on the trace test proposed by Johansen (1988, 1991). Our overall findings show that we are able to identify one stable and statistically significant cointegrating vector for each bilateral nominal exchange rate.12 To address the issue of volatility regime switching and to discriminate between low and high volatility regime in the relationship between the nominal exchange rate and fundamentals, we estimate and test for an MS-VECM given by Equation (3). In principle given that during the period under examination several exchange rate regimes have been adopted we could consider three potential regimes: regime 1 which covers flexible exchange rates (high volatility regime); regime 2 which covers managed float or peg exchange rates (the medium volatility regime); and regime 3 which covers fixed exchange rates (low volatility regime). Table 4 reports the estimated coefficients of the proposed MS-VECM along with the necessary test statistics for evaluation of the adequacy of the estimated model.13 The Likelihood Ratio test for the null hypothesis of linearity is statistically significant and this suggests that linearity is strongly rejected. This is a

12

The estimated cointegration coefficients have the correct sign and reasonable magnitude as predicted by the monetary model. Furthermore, when testing for the proportionality hypothesis this is found to hold for Australia, Belgium, France, Italy, Netherlands, Spain, and Switzerland. To save space these results are available upon request. 13 To save space we do not report all of our MS-VECM. We only report the estimated equilibrium correction coefficients for each equation of the MS-VECMs.

Exchange Rates, Fundamentals, and Nonlinearities

107

nonstandard LR test by Davies (1987). This outcome is reinforced from the AIC and HQIC criteria. The estimation of the MS-VECM specification was conducted with the adoption of the “bottom-up” procedure (Krolzig, 1997), which was designed to determine the appropriate MS(m)-VECM(p) model for each CEE country. Table 4 presents our results for the choice of the appropriate MS-VECM specification. Specifically for all cases we estimated an MSIAH (2)-VECM(p) specification. The Likelihood Ratio test for the null hypothesis of linearity (LR1) was statistically significant for all cases which suggest that linearity is strongly rejected. This is a nonstandard LR test advanced by Davies (1987).14 This outcome was reinforced by the AIC, SIC, and HQIC criteria. Furthermore, Table 4 reports the results from two Likelihood Ratio tests (LR2 and LR3) which were used to choose the appropriate model specification. Based on these two Likelihood Ratio test statistics (Krolzig, 1997, pp. 135
136) the most appropriate model within this class of MS-VECM model was the MSIAH(2)-VECM(p).15 Additional evidence for the appropriateness of the estimated model was given by the standardized residuals which reveal no evidence of serial correlation, heteroskedasticity, or substantial departures from normality.16 Based on these estimates we argue that they are in favor of a nonlinear relationship between exchange rates and macroeconomic fundamentals.17

14

The results from the LR test from Davies (1987) are reported with caution. It is argued that since the Markov regime switching model has both a problem of nuisance parameters and a problem of “zero score,” under the null hypothesis, we cannot use the χ 2 distribution to determine the significance of the LR test (Garcia, 1998). Therefore, Ang and Bekaert (2002a, 2002b) have suggested alternative LR tests for the case in which the regularity conditions of the Davies (1987) test are not met. However, given the support obtained by the AIC, SIC, and HQIC information criteria, we argue that the rejection of the linear model in favor of an MS specification is robust. 15 The number of regimes is 2 since the estimation of models with 3 regimes is not feasible due to large number of parameters. The ΜSIH(2)
VECM(p) specification P is given  by the following expression: ΔXt = vðzt Þ þ pi =−11 Γi ΔXt − i þ Πðzt ÞXt − 1 þ ut where ut zt ∼ NIDð0; Σ[zt ]Þ and zt ∈ f1; 2g. 16 Following Sarno et al. (2004), we further evaluated the goodness-of-fit of the 2 appropriate MSIAH-VECM specification by calculating the ratio of the R and the residual variance from each estimated MSIAH-VECM to the corresponding measure for its best linear VECM counterpart. In all cases we found that the estimated MSIAH-VECM outperforms the best alternative linear VECM as this is measured 2 by the improvement of the R and the reduction in the residual variance. 17 Dacco and Satchell, (1999), Neely and Sarno (2002), and Rapach and Wohar (2006) provide evidence which are in line with our arguments, although they also show that the forecasting performance of the Markov switching models do not produce good forecasts.

108

P. F. Diamandis et al.

Table 3: Johansen
Juselius Cointegration Trace Tests Results et = β0 þ β1 ðmt − mt Þ þ β2 ðyt − yt Þ þ β3 ðit − it Þ Country

(n − r)

r

Australia

3 2 1 3 2 1 3 2 1 3 2 1 3 2 1 3 2 1 3 2 1 3 2 1 3 2 1 3 2 1 3 2 1 3 2 1 3 2 1 3 2 1

58.93* 18.88 7.42 67.33* 17.34 8.81 61.12* 16.23 6.33 45.21* 22.44 7.86 46.38* 19.02 7.13 39.12* 17.90 5.18 40.02* 17.22 4.14 41.13* 18.02 7.63 51.15* 17.96 7.15 55.16* 18.35 6.98 44.12* 17.65 7.23 42.16* 18.33 8.19 53.16* 18.32 8.35 50.12* 18.86 7.87

Belgium

Canada

Denmark

Finland

France

Italy

Netherlands

Norway

Portugal

Spain

Sweden

Switzerland

United Kingdom

Note: r denotes the number of eigenvectors and (n − r) is the number of common trends. Trace is the Johansen Trace likelihood ratio statistic. A structure of four lags was chosen according to a likelihood ratio test, corrected for the degrees of freedom (Sims, 1980) and the Ljung
Box Q statistic for detecting serial correlation in the residuals of the equations of the VAR. A model with a constant restricted in the cointegrating vector is chosen according the Johansen (1992) testing strategy. (*) denotes statistical significance at the 5% critical level. The 5% critical values are 34.91, 19.96, and 9.24 respectively and they are taken from MacKinnon, Haug, and Michelis (1999, Table IV).

109

Exchange Rates, Fundamentals, and Nonlinearities

Table 4: “Bottom-Up” Procedure for Model Specification Model Australia Belgium Canada Denmark Finland France Italy Netherlands Norway Portugal Spain Sweden Switzerland United Kingdom

LR1
p-value*

LR2
x2ðgÞ

LR3
x2ðgÞ

AIC ratio

HQ ratio

SIC ratio

441.56 (0.0000) 333.220 (0.0000) 541.776 (0.0000) 357.111 (0.0000) 507.566 (0.0000) 246.212 (0.0000) 271.158 (0.0000) 528.022 (0.0000) 293.793 (0.0000) 317.332 (0.0000) 521.335 (0.0000) 489.221 (0.0000) 502.113 (0.0000) 494.223 (0.0000)

168.22 ðx2ð10Þ = 18:3Þ 155.12 ðx2ð10Þ = 18:3Þ 180.13 ðx2ð10Þ = 18:3Þ 201.56 ðx2ð10Þ = 18:3Þ 887.34 ðx2ð10Þ = 18:3Þ 367.28 ðx2ð10Þ = 18:3Þ 303.66 ðx2ð10Þ = 18:3Þ 121.90 (x2ð10Þ = 18.3) 180.67 ðx2ð10Þ = 18:3Þ 233.55 ðx2ð10Þ = 18:3Þ 356.13 ðx2ð10Þ = 18:3Þ 602.51 ðx2ð10Þ = 18:3Þ 589.23 ðx2ð10Þ = 18:3Þ 433.28 ðx2ð10Þ = 18:3Þ

891.23 x2ð196Þ = 229:6 334.22 ðx2ð36Þ = 50:99Þ 609.22 ðx2ð126Þ = 153:1Þ 443.28 ðx2ð84Þ = 106:3Þ 77.89 ðx2ð36Þ = 50:99Þ 272.06 ðx2ð84Þ = 106:3Þ 256.28 ðx2ð84Þ = 106:3Þ 298.23 (x2ð84Þ = 106.3) 129.65 ðx2ð52Þ = 69:83Þ 168.33 ðx2ð68Þ = 88:25Þ 157.12 ðx2ð52Þ = 69:83Þ 150.13 ðx2ð68Þ = 88:25Þ 373.28 ðx2ð84Þ = 106:3Þ 201.17 ðx2ð68Þ = 88:25Þ

1.1126

1.0459

0.9286

1.0531

1.0390

1.0174

1.1314

1.0961

1.0390

1.0626

1.0210

0.9530

1.0980

1.0489

0.9667

1.0088

0.9797

0.9337

1.0002

0.9708

0.9243

1.1858

1.1382

1.0580

1.0453

1.0208

0.9826

1.0463

1.0154

0.9664

1.0566

1.0233

0.9732

1.0998

1.0245

0.9891

1.1467

1.0325

0.9624

1.0858

1.1258

0.9765

Note: The p-values are given based on the likelihood ratio test (LR) for the null of a linear VECM. The value in parentheses next to LR is the marginal significance level of this test, based on Davies (1987). LR1 tests the null hypothesis that there is no regime switching. LR2 tests the null hypothesis that there is no regime switching in the autoregressive parameters and in the variance
covariance matrix (i.e., MSI(2)-VECM(p) against MSIH(2)-VECM(p)). LR3 tests the null hypothesis that there is no regime switching in the autoregressive parameters (i.e., MSIH(2)-VECM(p) against MSIΑH(2)VECM(p)). The statistical criteria LR2 and LR3 are distributed as a x2 with g degrees of freedom, where g is the number of restrictions. AIC, HQ, and SIC denote the Akaike information criterion, Hannan Quinn, and Schwartz information criterion respectively, and they provide the ratios between the chosen MS-VECM model and the respective linear VECM model.

In Table 5 we report the regime-dependent equilibrium correction coefficients. For Australia, in Regime 1 we observe that the coefficients of the relative money and of interest rate differential were statistically significant which implies that, in Regime 1, these monetary fundamentals contributed most to the adjustment in restoring any deviations from the long-run equilibrium. This is consistent with the fact that during the interwar period the estimated transition probability is near or equal to unity. In addition it is

110

P. F. Diamandis et al.

Table 5: Regime Dependent Error Correction Terms αΔst ðz = 1Þ αΔst ðz = 2Þ αΔmt ðz = 1Þ

αΔmt ðz = 2Þ αΔyt ðz = 1Þ αΔyt ðz = 2Þ αΔit ðz = 1Þ αΔit ðz = 2Þ

Australia
MSIAH(2)
VECM(5) 0.0010 0.221* −0.535* −0.106* (0.007) (0.058) (0.265) (0.051) Belgium
MSIΑH(2)
VECM(2) −0.063* −0.071* −0.004 −0.056* (0.004) (0.005) (0.002) (0.002) Canada
MSIΑH(2)
VECM(7) −0.002 −0.017* −0.104* −0.006 (0.001) (0.001) (0.011) (0.012) Denmark
MSIΑH(2)
VECM(5) −0.054* −0.012 0.007 0.006 (0.012) (0.033) (0.010) (0.021) Finland
MSIΑH(2)
VECM(5) −0.022 0.091* −0.903* −0.021 (0.030) (0.022) (0.191) (0.063) France
MSIΑH(2)
VECM(6) 0.088* 0.093* −0.189* −0.022 (0.025) (0.012) (0.033) (0.081) Italy
MSIΑH(2)
VECM(5) −0.022 −0.021 0.057* 0.007 (0.071) (0.044) (0.010) (0.021) Netherlands
MSIΑH(2)
VECM(5) −0.098* −0.012 0.041 −0.202* (0.015) (0.025) (0.097) (0.051) Norway
MSIΑH(2)
VECM(3) −0.008 −0.077* −0.066* 0.011 (0.021) (0.013) (0.018) (0.023) Portugal
MSIH(2)
VECM(4) 0.003 −0.028* 0.067* 0.065 (0.012) (0.005) (0.013) (0.051) Spain
MSIH(2)
VECM(3) 0.702* −0.201* 0.008 0.076* (0.187) (0.045) (0.013) (0.019) Sweden
MSIH(2)
VECM(4) 0.021 −0.072* 0.053* 0.017 (0.081) (0.017) (0.009) (0.021) Switzerland
MSIH(2)
VECM(7) 0.008 −0.021* 0.171* 0.027 (0.027) (0.006) (0.041) (0.031) United Kingdom
MSIH(2)
VECM(5) 0.015 −0.206* 0.071* 0.059 (0.031) (0.044) (0.016) (0.078)

0.441 (0.597)

−0.033 (0.076)

0.155* (0.046)

0.112 (0.215)

0.007 (0.008)

0.000 (0.025)

0.010 (0.006)

−0.022* (0.001)

−0.010 (0.024)

0.324* (0.058)

−0.021* (0.004)

0.003 (0.007)

−0.045 (0.056)

−0.071* (0.022)

−0.022* (0.004)

−0.058* (0.012)

0.298* (0.074)

0.168 (0.065)

0.031 (0.019)

0.101 (0.093)

0.303* (0.098)

−0.405* (0.111)

0.196* (0.034)

0.008 (0.023)

−0.091* (0.029)

−0.004 (0.012)

0.103* (0.018)

−0.005 (0.023)

−0.052* (0.023)

−0.109* (0.032)

−0.219* (0.035)

0.065* (0.019)

0.151* (0.032)

−0.037* (0.08)

0.007 (0.006)

0.053* (0.009)

0.014 (0.010)

0.005 (0.010)

−0.031* (0.006)

−0.105* (0.033)

0.009 (0.021)

0.007 (0.019)

−0.002 (0.007)

−0.201* (0.065)

0.008 (0.016)

0.541* (0.105)

−0.443* (0.098)

−0.023 (0.054)

0.202* (0.034)

0.092* (0.021)

−0.031 (0.054)

−0.281* (0.077)

0.019 (0.021)

0.011 (0.039)

−0.077* (0.013)

−0.322* (0.055)

Note: αΔst , αΔmt , αΔyt , and αΔit denote the estimated coefficients of the error correction terms for the exchange rate, the money supply differential, the output differential, and the interest rates differential in both regimes 1 and 2. Figures in parentheses are asymptotic standard errors and (*) denotes statistical significance at the 5% level.

Exchange Rates, Fundamentals, and Nonlinearities

111

clear that the exchange rate arrangements were not stable during that period. This evidence is also consistent with the adoption of fixed exchange rates under the Bretton Woods system. In Regime 2 the estimated coefficients of the exchange rate and relative money were statistically significant and thus the adjustment of equilibrium was achieved through the changes in these two variables. Again, the estimated probability of being in Regime 2 was near or equal to unity for the period during the recent period of flexible exchange rates.18 For the case of Belgium we estimated an MSIΑH(2)
VECM(2) model. The analysis of the estimated error correction terms shows that for the case of Regime 1 the only statistically significant coefficient is that of the exchange rate. Therefore, we argue that it is the exchange rate that adjusts to any deviations from the long-run equilibrium. This is consistent with the flexible exchange rate system that was in force up to the beginning of the interwar period. Certainly in this case we also observe a large number of switches in transition probabilities given the adoption of several alternative exchange rate arrangements. In Regime 2 we observe that the error correction coefficients of the nominal exchange rate, the relative money, and interest rate differential were statistically significant and therefore all three variables adjust to bring the system to its long-run equilibrium. The probability of being in regime 1 during the post-Bretton Woods floating rate period is close to unity for the post-1979 period which also marks the establishment of the European Monetary System. When we consider the fixed exchange rate period 1944
1979 the probability of being in Regime 2 is near or equal to unity. For the case of Canada, based on the estimation of the appropriate switching regime model we found that the relative money and the real interest rate differential adjust to restore any deviations from the long-run equilibrium since their error correction terms were statistically significant during Regime 1. This is consistent with the gold standard period up to 1914 and the interwar period of the gold standard exchange regime from 1926 to 1933. This also holds for part of the Bretton Woods period since the early 1960s. Thus, the derived transition probabilities of being in Regime 1 are almost always equal to unity for these periods. For Regime 2 the error correction terms of exchange rate and real GDP were statistically significant which implies that they both contributed to the adjustment to the long-run equilibrium after any departure from it. This result is consistent with the fact that during the period 1920
1926 as well as during the 1950s Canada adopted a flexible exchange rate regime. Canadian dollar is also freely float

18

To save space the transition probabilities diagrams are available upon request.

112

P. F. Diamandis et al.

currency since 1973. The estimated transition probabilities are almost always equal to unity in the case of Regime 2 for the corresponding period. In Denmark the error correction coefficient of the exchange rate and the interest rate differential were the only statistically significant coefficients during Regime 1 and therefore these variables adjusted to restore any deviations from the long-run equilibrium. This finding is consistent with the exchange rate arrangements adopted in Denmark during the interwar period. The probability of being in Regime 1 was almost always unity during the period of flexible exchange rates. In contrast during Regime 2 the real output differential and the interest rate differential had statistically significant error correction terms and therefore these variables were the ones that provided the adjustment. This result is again consistent with the transition probabilities since it is consistent with the adoption of a fixed exchange and a target zone exchange rate regime. Overall we also noted that there were frequent changes in the regimes. In Finland we also identified two regimes. In the case of Regime 2 the error correction coefficient of the exchange rate was statistically significant and therefore it was the variable that adjusted to restore deviations from the long-run equilibrium. This is consistent with the transition probabilities for Finland since the probability of being in Regime 2 was almost always unity during the early years of the sample when Finland’s currency was under a flexible exchange rate system. In contrast, during Regime 1 it was the monetary fundamentals that adjusted since the error correction coefficient of the relative money supply and of the real output differential were statistically significant. This is consistent with the fact that Regime 1 coincided with the adoption of a fixed exchange rate regime under the Bretton Woods agreement by the Finish monetary authorities. For France we found that in Regime 1 the error correction coefficients were statistically significant for the exchange rate and the monetary fundamentals and therefore both the exchange rate and the monetary fundamentals adjusted to deviations from long-run equilibrium. This is consistent with the gold standard that prevailed until 1914 and the subsequent adoption of flexible exchange rate from 1919 to 1926 and the return to the gold standard at the end of 1926 until the end of WWII. It is also consistent with the Bretton Woods fixed exchange rate system. We also noted that during this period the exchange rate exhibited some volatility which ceased to exist after the adoption of a fixed exchange rate system. In Regime 2 the adjustment to long-run equilibrium came from the exchange rate and the relative real output differential which had the statistically significant error correction term. This is consistent with the adoption of the post-Bretton Woods flexible exchange rate regime since 1979. For Italy the estimation of the two regimes led to the conclusion that during Regime 1 the adjustment to the long-run equilibrium came only from

Exchange Rates, Fundamentals, and Nonlinearities

113

the monetary fundamentals. This is what one would expect during fixed exchange rates periods and is related to the estimated transition probabilities. The derived transition probabilities show that being in Regime 1 was near or equal to unity during the period of fixed exchanges rates adopted at Bretton Woods. Similar evidence is drawn for the period up until the interwar period and once again we observe a large number of regime switching. For Regime 2 it was the exchange rate and the interest rate differential that adjusted to any departures from the long-run equilibrium, which was consistent with the post-Bretton Woods floating exchange rates. This was confirmed by the transition probabilities since it is shown that during the recent flexible exchange regime these were almost always unity. For the Netherlands we also found that during Regime 1 the exchange rate along with the real output and interest rate differential adjusted to any departures from the long-run equilibrium. The transition probabilities reveal that the probability of the Netherlands being in Regime 1 was near or equal to unity for the interwar period and the post-Bretton Woods flexible exchange rate period. As expected, the transition probabilities for the interwar period again exhibit a large number of switches. In Regime 2 we found that only the monetary fundamental variables adjust to restore deviations from long-run equilibrium in this regime. This is consistent with the estimated transition probabilities which are near or equal to unity within this regime. For Norway we found that for Regime 1 both the relative money supplies and the real output differential adjusted to the long-run equilibrium. It is shown from the estimated transition probabilities that the probability of Norway being in the high volatility (flexible exchange rate regime) was near or equal to unity until up to the interwar period and during the Bretton Woods exchange rates system. During these periods the adjustment came from the monetary fundamentals. In Regime 2, the error correction coefficients of the exchange rate, the real output, and the interest rate differential were statistically significant. Therefore, these variables adjusted to restore deviations from the long-run equilibrium under the flexible exchange rate regime that was implemented since 1973 and in addition for a short period of flexible exchange rates in the 1920s. The corresponding transition probabilities show that the probability of being in Regime 2 was near or equal to unity. For Portugal the estimation of the two regimes led to the conclusion that during Regime 1 the adjustment to the long-run equilibrium came only from the relative money supplies and the interest rate differential. Therefore, these variables adjusted to restore deviations from the long-run equilibrium during the gold standard period and the interwar period, and the fixed exchange rates system prevailed from 1944 to 1973. The derived transition probabilities show that being in Regime 1 was near or equal to

114

P. F. Diamandis et al.

unity during these periods of international monetary arrangements. For the period up until the interwar period and once again we observe a large number of regime switching. For Regime 2 it was the exchange rate and the interest rate differential that adjusted to any departures from the long-run equilibrium, which was consistent with the post-Bretton Woods floating exchange rates. This was confirmed by the transition probabilities since it is shown that during the recent flexible exchange regime these were almost always unity. For the case of Spain we estimated an MSIΑH(2)
VECM(3) model. The analysis of the estimated error correction terms shows that for the case of Regime 1 the only statistically significant coefficient is that of the exchange rate. Therefore, we argue that it is the exchange rate that adjusts to any deviations from the long-run equilibrium. This is consistent with the flexible exchange rate system that was in force up to the beginning of the interwar period. Certainly in this case we also observe a large number of switches in transition probabilities given the adoption of several alternative exchange rate arrangements. In Regime 2 we observe that the error correction coefficients of the nominal exchange rate, the relative money, and interest rate differential were statistically significant and therefore all three variables adjust to bring the system to its long-run equilibrium. The probability of being in regime 1 during the post-Bretton Woods floating rate period is close to unity for the post-1979 period which also marks the establishment of the European Monetary System. When we consider the fixed exchange rate period 1944
1979 the probability of being in Regime 2 is near or equal to unity. For the case of Sweden based on the estimation of the switching regime model we found that the relative money and the interest rate differential adjust to restore any deviations from the long-run equilibrium since their error correction terms were statistically significant during Regime 1. This is consistent with the gold standard period up to 1914 and the interwar period of the gold standard exchange regime from 1926 to 1933. This also holds for the Bretton Woods period since the early 1960s. Thus, the derived transition probabilities of being in Regime 1 are almost always equal to unity for these periods. For Regime 2 the error correction terms of exchange rate and real GDP were statistically significant which implies that they both contributed to the adjustment to the long-run equilibrium after any departure from it. The estimated transition probabilities are almost always equal to unity in the case of Regime 2 for the corresponding period. For Switzerland in Regime 1 the error correction coefficients of relative money, real output differential, and interest rate differential were statistically significant and therefore these were the variables that adjusted most to any deviations from the long-run equilibrium. This clearly indicated that the transition probabilities in Regime 1 during the period up

Exchange Rates, Fundamentals, and Nonlinearities

115

until the interwar period and during the Bretton Woods period (both fixed exchange rate periods) were almost always unity. In the case of Regime 2, the coefficients of the exchange rate as well as the real output and interest rate differential were statistically significant. This result is consistent with the current flexible exchange rates period and it is observed in the transition probabilities since they take a value of unity during the period of flexible exchange rates. Finally, in the case of the United Kingdom the estimation of the two regimes led to the conclusion that during Regime 1 the adjustment to the long-run equilibrium came only from the relative money supplies and the interest rate differential. Therefore, these variables adjusted to restore deviations from the long-run equilibrium during the gold standard period and the interwar period, and the fixed exchange rates system prevailed from 1944 to 1973. The derived transition probabilities show that being in Regime 1 was near or equal to unity during these periods of international monetary arrangements. For the period up until the interwar period, we observe a large number of regime switching. For Regime 2 it was the exchange rate and the interest rate differential that adjusted to any departures from the long-run equilibrium, which was consistent with the post-Bretton Woods floating exchange rates. This was confirmed by the transition probabilities since it is shown that during the recent flexible exchange regime these were almost always unity. In order to assess the regime qualification performance of the chosen Markov-switching models, we calculated the Regime Classification Statistic (RCM) developed by Ang and Bekaert (2002a, 2002b). This measure is based on the fact that the ex post (smoothed) probabilities pt are close either to one or zero and therefore a good regime-switching model should classify regimes sharply. The RCM for a model with two regimes may be calculated as follows: RCMð2Þ = 400

T 1X pt ð1 − pt Þ; T i=1

ð13Þ

where T is the sample size, p1 and ð1 − p1 Þ is the smoothed probability of being in regime j = 1; 2 at time t, and RCM takes values between 0 and 100. In general the lower the value of RCM the better the performance of the model is. The ideal model will have an RCM with a value close to zero. Weak regime inference implies that regime-switching models cannot distinguish among regimes based on the behavior of data and this may be due to misspecification. A model which poorly distinguishes between regimes will have an RCM with a value close to 100. Table 6 reports the calculated RCM statistic for the full sample which is close to zero for all countries, implying a very satisfactory regime classification.

116

P. F. Diamandis et al.

Table 6: Regime Classification Measure Country

RCM

Australia Belgium Canada Denmark Finland France Italy Netherlands Norway Portugal Spain Sweden Switzerland United Kingdom

3.98 5.33 1.10 6.38 7.01 13.78 3.48 1.15 11.28 6.92 7.02 0.13 0.28 2.69

P Note: The statistics for a model with two regimes are calculated as RCMð2Þ = 400 T1 Ti= 1 pt ð1 − pt Þ where T is the sample size, p1 and ð1 − p1 Þ is the smoothed probability to be in regime j = 1; 2 at time t, and RCM takes values between 0 and 100. A value close to zero implies a very good discrimination between the two regimes whereas a value close to one hundred implies a model that poorly distinguishes between the two regimes.

Summary and Concluding Remarks In this paper we provide an extended review of the monetary model of exchange rate over the last 40 years. Furthermore, we provide further study of the monetary model under regime switching. We analyzed the case of the link between exchange rates and monetary fundamentals for 14 industrialized countries. We considered the presence of nonlinearities in the relationship between the nominal bilateral exchange rate and macroeconomic fundamentals and we estimated the appropriate Markov-switching vector error correction model. We used annual data spanning from the late 19th century or early 20th century to the late 20th century for the bilateral nominal exchange rates against the US dollar and the respective macroeconomic fundamentals. Furthermore, from a methodological point of view it was important to examine the adjustment mechanisms to the long-run equilibrium. Since the period under examination covers a number of alternative international monetary arrangements, such the gold standard, the Bretton Woods period, and the current flexible exchange rate regime, our analysis focused on demonstrating whether the exchange rate or the macroeconomic variables were the main vehicle in achieving their target. Therefore, it was important

Exchange Rates, Fundamentals, and Nonlinearities

117

to reveal whether the adjustment back to equilibrium took place primarily through the nominal exchange rate during periods of floating exchange rates and through monetary fundamentals during periods in which some variant of a fixed exchange rate system was in force. To this end the monetary model provided the appropriate framework to study the behavior of exchange rate movements in periods of transition. There are several important findings that stem from the present analysis. First for each bilateral exchange rate and the respective macroeconomic variables we were able to capture nonlinearities with the estimation of the appropriate Markov switching regime model with two regimes. The fitted model was quite general since it allowed for regime shifts in the intercept and the complete set of parameters, as well as the variance
covariance matrix. In addition, for all cases the null hypothesis of linearity was rejected when tested against the alternative of an MSVECM specification. Second, our analysis has clearly shown that during the period when some variant of fixed exchange rates was adopted in each country, the monetary fundamentals adjust to restore deviations from the long-run equilibrium. In contrast during periods with less restricted exchange rate regimes, it was the exchange rate that adjusted to restore any disequilibrium. Finally, the application of the Regime Classification Measure developed by Ang and Bekaert (2002a, 2002b) showed that our estimated Markov-switching models distinguished very well between the two regimes.

Acknowledgments An earlier version of this paper was presented at the 17th International Conference on Macroeconomic Analysis and International Finance, Rethymno, May 30
June 1, 2013 and thanks are due to conference participants for many helpful comments and discussions. The third author acknowledges financial support by a Marie Curie Transfer of Knowledge Fellowship of the European Community’s Sixth Framework Programme under contract number MTKD-CT-014288 as well as from the Research Committee of the University of Crete under research grant #2257. We thank Philippe Bacchetta, Richard Baillie, Michael Bordo, Dimitris Georgoutsos, Paul De Grauwe, Katarina Juselius, Menelaos Karanasos, Jim Lothian, Mike Melvin, Lucio Sarno, Dimitris Thomakos, and Mark Wohar for many helpful comments and discussions. We also thank Michael Bordo and Mark Wohar for generously providing the data used in this paper. Theodoros Bratis, Ioannis Polykarpou, and Evangelos Salachas provided superb research assistance. The usual caveat applies.

118

P. F. Diamandis et al.

References Abuaf, N., & Jorion, P. (1990). Purchasing power parity in the long run. Journal of Finance, 45, 157
1714. Ang, A., & Bekaert, G. (2002a). International asset allocation with regime shifts. Review of Financial Studies, 15, 1137
1187. Ang, A., & Bekaert, G. (2002b). Regime switches in interest rates. Journal of Business and Economic Statistics, 20, 163
182. Bacchetta, P., & Van Wincoop, E. (2004). A scapegoat model of exchange-rate fluctuations. American Economic Review Papers and Proceedings, 94, 114
118. Bacchetta, P., & Van Wincoop, E. (2013). On the unstable relationship between exchange rates and fundamentals. Journal of International Economics, 91, 18
26. Backus, D. (1984). Empirical models of the exchange rate: Separating the wheat from the chaff. Canadian Journal of Economics, 17, 824
846. Balke, N. S., & Fomby, T. B. (1997). Threshold cointegration. International Economic Review, 38, 627
645. Berben, R.-P., & van Dijk, D. (1998). Does the absence of cointegration explain the typical findings in long-horizon regressions? Econometrics Institute, Erasmus University, Report 9814. Berkowitz, J., & Giorgianni, L. (2001). Long-horizon exchange rate predictability? Review of Economics and Statistics, 83, 81
91. Bessec, M. (2003). Mean-reversion vs. adjustment to PPP: The two regimes of exchange rate dynamics under the EMS, 1979
1998. Economic Modelling, 20, 141
164. Bordo, M. D., Bergman, M., & Jonung, L. (1998). Historical evidence on business cycles: The international perspective. In J. C. Fuhrer & S. Schuh (Eds.), Beyond shocks: What causes business cycles (pp. 65
113). Conference Series No. 42, Federal Reserve Bank of Boston, Boston, MA. Bordo, M. D., & Jonung, L. (1998). A return to the convertibility principle? Monetary and fiscal regimes in historical perspective. In A. Leijonhuv (Ed.), Monetary theory as a basis for monetary policy. London: MacMillan. Boudoukh, J., Richardson, M., & Whitelaw, R. F. (2008). The myth of longhorizon predictability. Review of Financial Studies, 21, 1577
1605. Cheung, Y.-W., Chinn, M. D., & Pascual, A. G. (2005). Empirical exchange rate models of the nineties: Are any fit to survive? Journal of International Money and Finance, 24, 1150
1175. Cheung, Y.-W., & Menzie, M. D. (2001). Currency traders and exchange rate dynamics: A survey of the US market. Journal of International Money and Finance, 20, 439
471. Chinn, M. D., & Meese, R. A. (1995). Banking on currency forecasts: How predictable is change in money? Journal of International Economics, 38, 161
178. Clarida, R. H., Sarno, L., Taylor, M. P., & Valente, G. (2003). The out-of-sample success of term structure models as exchange rate predictors: A step beyond. Journal of International Economics, 60, 61
83. Cook, S. (2006). The robustness of modified unit root tests in the presence of GARCH. Quantitative Finance, 6, 359
363.

Exchange Rates, Fundamentals, and Nonlinearities

119

Dacco, R., & Satchell, S. (1999). Why do regime-switching models forecast so badly? Journal of Forecasting, 18, 1
16. Dal Bianco, M., Camacho, M., & Perez-Quiros, G. (2012). Short-run forecasting of the euro-dollar exchange rate with economic fundamentals. Journal of International Money and Finance, 31, 377
396. Davies, R. B. (1987). Hypothesis testing when the nuisance parameter is present only under the alternative. Biometrika, 74, 33
43. De Grauwe, P. (2000). Exchange rates in search of fundamentals: The case of the euro-dollar rate. International Finance, 3, 329
356. De Grauwe, P., & Vansteenkiste, I. (2007). Exchange rates and fundamentals: A nonlinear relationship? International Journal of Finance and Economics, 12, 37
54. Dempster, A. P., Laird, N. M., & Rubin, D. B. (1977). Maximum likelihood estimation from incomplete data via the EM algorithm. Journal of the Royal Statistical Society, 39(Series B), 1
38. Dewachter, H. (1997). Sign predictions of exchange rate changes: Charts as proxies for Bayesian inferences. Weltwirtschaftliches Archiv, 133, 39
55. Dewachter, H. (2001). Can Markov switching models replicate chartist profits in the foreign exchange market? Journal of International Money and Finance, 20, 25
41. Diamandis, P. F., Georgoutsos, D. A., & Kouretas., G. P. (1998). The monetary approach to the exchange rate: Long-run relationships, identification and temporal stability. Journal of Macroeconomics, 20, 741
766. Diamandis, P. F., Georgoutsos, D. A., & Kouretas, G. P. (2000). The monetary model in the presence of I(2) components: Long-run relationships, short-run dynamics and forecasting of the Greek Drachma. Journal of International Money and Finance, 19, 917
941. Dornbusch, R. (1976). Expectations and exchange rate dynamics. Journal of Political Economy, 84, 1161
1176. Ducker, M., & Neely, C. J. (2007). Can Markov switching models predict excess foreign exchange returns? Journal of Banking and Finance, 31, 270
296. Elliott, G. (1999). Efficient tests for a unit root when the initial observation is drawn from its unconditional distribution. International Economic Review, 44, 767
784. Elliott, G., Rothenberg, J. T., & Stock, J. H. (1996). Efficient tests for an autoregressive unit root. Econometrica, 64, 813
836. Engel, C. (1994). Can the Markov switching model forecast exchange rates? Journal of International Economics, 36, 151
165. Engel, C., & Hamilton, J. D. (1990). Long swings in the dollar: Are they in the data and do markets know it? American Economic Review, 80, 689
713. Engel, C., & Kim, C. J. (2001). The long-run US/UK real exchange rate. Journal of Money Credit and Banking, 31, 335
356. Engel, C., & West, K. D. (2005). Exchange rates and fundamentals. Journal of Political Economy, 113, 485
517. Engle, R. E., & Granger, C. W. J. (1987). Co-integration and error correction representation, estimation and testing. Econometrica, 55, 251
276. Evans, M. D. D., & Lyons, R. K. (2002). Order flow and exchange rate dynamics. Journal of Political Economy, 110, 170
180.

120

P. F. Diamandis et al.

Evans, M. D. D., & Lyons, R. K. (2005). Meese-Rogoff redux: Micro based exchange-rate forecasting. American Economic Review Papers and Proceedings, 95, 405
414. Faust, J., Rogers, J. H., & Wright, J. H. (2003). Exchange rate forecasting: The errors we’ve really made. Journal of International Economics, 60, 35
59. Flood, R. P., & Rose, A. K. (1995). Fixing exchange rates: A virtual quest for fundamentals. Journal of Monetary Economics, 36, 3
37. Frenkel, J. A. (1976). A monetary approach to the exchange rate: Doctrinal aspects and empirical evidence. Scandinavian Journal of Economics, 78, 200
224. Frankel, J. A. (1979). On the mark: A theory of floating exchange rate based on real interest differentials. American Economic Review, 69, 610
622. Frankel, J. A. (1996). How well do foreign exchange markets function: Might a Tobin tax help? In M. ul Haq, I. Kaul, & I. Grunberg (Eds.), The Tobin tax: Coping with financial volatility. Oxford: Oxford University Press. Frankel, J. A., & Rose, A. K. (1995). Empirical research on nominal exchange rates. In K. Rogoff & G. Grossman (Eds.), Handbook of international economics (Vol. 3, pp. 1689
1729). Amsterdam: North Holland. Frankel, J. A., & Rose, A. K. (1996). A panel project on purchasing power parity: Mean reversion within and between countries. Journal of International Economics, 40, 209
224. Frommel, M., MacDonald, R., & Menkhoff, L. (2005a). Markov switching regimes in a monetary exchange rate model. Economic Modelling, 22, 485
502. Frommel, M., MacDonald, R., & Menkhoff, L. (2005b). Do fundamentals matter for the D-mark/euro-dollar? A regime switching approach. Global Finance Journal, 15, 321
335. Frydman, R., & Goldberg, M. D. (2001). Macroeconomic fundamentals and the DM/$ exchange rate: Temporal stability and the monetary model. International Journal of Finance and Economics, 6, 421
435. Garcia, R. (1998). Asymptotic null distribution of the likelihood ratio test in Markov switching models. International Economic Review, 39, 763
788. Gehrig, T., & Menkoff, L. (2004). The use of flow analysis in foreign explanatory evidence. Journal of International Money and Finance, 23, 573
594. Glen, J. D. (1992). Real exchange rates in the short, medium, and long run. Journal of International Economics, 33, 147
166. Goodhart, C. (1988). The foreign exchange market: A random walk with a dragging anchor. Economica, 55, 437
460. Groen, J. J. J. (2000). The monetary exchange rate model as a long-run phenomenon. Journal of International Economics, 52, 299
319. Groen, J. J. J. (2005). Exchange rate predictability and monetary fundamentals in a small multi-country panel. Journal of Money, Credit and Banking, 37, 495
516. Hakkio, C. S., & Rush, M. (1991). Cointegration: How short is the long run? Journal of International Money and Finance, 10, 571
581. Hamilton, J. (1989). A new approach to the economic analysis of nonstationary time series and the business cycle. Econometrica, 57, 357
384. Hamilton, J. D. (1994). Time series analysis. Princeton, NJ: Princeton University Press.

Exchange Rates, Fundamentals, and Nonlinearities

121

Hamilton, J. D., & Lin, G. (1996). Stock market volatility and the business cycle. Journal of Applied Econometrics, 11, 573
593. Hamilton, J. D., & Susmel, R. (1994). Autoregressive conditional heteroskedasticity and changes in regime. Journal of Econometrics, 64, 307
333. Hansen, H., & Johansen, S. (1993). Recursive estimation in cointegrated VAR-models. Working Paper, Institute of Mathematical Statistics, University of Copenhagen. Hansen, H., & Johansen, S. (1999). Some tests for parameter constancy in cointegrated VAR-models. Econometrics Journal, 2, 306
333. Homer, S., & Sylla, R. (2005). A history of interest rates. New York, NY: Wiley. Hsieh, D. A. (1989). Testing for nonlinear dependence in daily foreign exchange rates. Journal of Business, 62, 339
368. Johansen, S. (1988). Statistical analysis of cointegrating vectors. Journal of Economics Dynamics and Control, 12, 231
254. Johansen, S. (1991). Estimation and hypothesis testing of cointegration vectors in Gaussian vector autoregressive models. Econometrica, 59, 1551
1580. Johansen, S. (1992). Determination of cointegration rank in the presence of a linear trend. Oxford Bulletin of Economics and Statistics, 54, 383
397. Johansen, S. (1995). Identifying restrictions of linear equations with applications to simultaneous equations and cointegration. Journal of Econometrics, 69, 111
132. Johansen, S., & Juselius, K. (1990). Maximum likelihood estimation and inference on cointegration with applications to the demand for money. Oxford Bulletin of Economics and Statistics, 52, 169
210. Johansen, S., & Juselius, K. (1994). Identification of the long-run and the short-run structure: An application to the ISLM model. Journal of Econometrics, 63, 7
36. Kanas, A., & Kouretas, G. P. (2007). Regime dependence between the official and parallel foreign currency markets for dollars in Greece. Journal of Macroeconomics, 29, 431
449. Kapetanios, G., Shin, Y., & Snell, A. (2003). Testing for a unit root in the nonlinear STAR framework. Journal of Econometrics, 112, 359
379. Kilian, L. (1999). Exchange rates and monetary fundamentals: What do we learn from long-horizon regressions? Journal of Applied Econometrics, 14, 491
510. Kilian, L., & Taylor, M. P. (2003). Why is it so difficult to beat the random walk forecast of exchange rates? Journal of International Economics, 60, 85
107. Kim, C.-J. (1994). Dynamic linear models with Markov-switching. Journal of Econometrics, 60, 1
22. Kim, B. J. C., & Mo, S. (1995). Cointegration and the long-run forecast of exchange rates. Economics Letters, 48, 491
510. Kouretas, G. P. (1997). Identifying linear restrictions on the monetary exchange rate model and the uncovered interest parity: Cointegration evidence for the Canadian-U.S. Dollar. Canadian Journal of Economics, 30, 875
890. Krolzig, H. M. (1997). Markov switching vector autoregressions: Modelling, statistical inference and application to business cycles analysis. Berlin: Springer. Kwiatkowski, D., Phillips, P. C. B., Schmidt., P., & Shin, Y. (1992). Testing the null hypothesis of stationarity against the alternative of a unit root: How sure are we that economic time series have a unit root? Journal of Econometrics, 54, 159
178.

122

P. F. Diamandis et al.

Lee, H.-Y., & Chen, S.-L. (2006). Why use Markov-switching models in exchange rate prediction? Economic Modelling, 23, 662
668. Levin, A., & Lin, C.-F. (1992). Unit root tests in panel data: Asymptotic and finite sample properties. Discussion Paper No. 92-93, Department of Economics, University of California, San Diego, CA. Lothian, J. R., & Taylor, M. P. (1996). Real exchange rate behaviour: The recent float from the perspective of the last two centuries. Journal of Political Economy, 104, 488
510. Lothian, J. R., & Taylor, M. P. (2000). Purchasing power parity over two centuries: Strengthening the case for real exchange rate stability: A reply to Coddington and Liang. Journal of International Money and Finance, 19, 759
764. Lyons, R. (1995). Tests of microstructural hypotheses in the foreign exchange market. Journal of Financial Economics, 39, 321
351. Lyons, R. (2001). The microstructure approach to exchange rats. Cambridge, MA: MIT Press. MacDonald, R. (1999). Exchange rate behaviour: Are fundamentals important? Economic Journal, 109, 673
691. MacDonald, R. (2007). Exchange rate economics: Theories and evidence. Oxford: Routledge. MacDonald, R., & Taylor, M. P. (1994). The monetary model of the exchange rate: Long-run relationships, short-run dynamics and how to beat a random walk. Journal of International Money and Finance, 13, 276
290. MacKinnon, J. G. (1991). Critical values for co-integration tests. In R. F. Engle & C. W. J. Granger (Eds.), Long-run economic relationships. Oxford: Oxford University Press. MacKinnon, J. G. (1994). Approximate asymptotic distribution functions for unit-root and cointegration tests. Journal of Business & Economic Statistics, 12, 167–176. MacKinnon, J. G., Haug, A. A., & Michelis, L. (1999). Numerical distribution functions of likelihood ratio test for cointegration. Journal of Applied Econometrics, 14, 563
577. Mahavan, A., & Wagner, A. I. (1999). Nonlinear dynamics in foreign exchange rates. Global Finance Journal, 10, 1
23. Mark, N. (1995). Exchange rates and fundamentals: Evidence on long-horizon predictability. American Economic Review, 85, 201
219. Mark, N., & Sul, D. (2001). Nominal exchange rates and monetary fundamentals: Evidence from a small Bretton-Woods panel. Journal of International Economics, 53, 29
52. Marsh, I. W. (2000). High-frequency Markov switching models in the foreign exchange market. Journal of Forecasting, 19, 123
134. Meese, R. A. (1990). Currency fluctuations in the post-Bretton Woods era. Journal of Economic Perspectives, 4, 117
134. Meese, R. A., & Rogoff, K. (1983). Empirical exchange rate models of the seventies: Do they fit out of sample? Journal of International Economics, 14, 3
24.

Exchange Rates, Fundamentals, and Nonlinearities

123

Meese, R. A., & Rose, A. K. (1991). An empirical assessment of non-linearities in models of exchange rate determination. The Review of Economics and Statistics, 58, 603
619. Neely, C. J., & Sarno, L. (2002). How well do monetary fundamentals forecast exchange rates? Federal Reserve Bank of St. Louis Review, 84, 51
74. Newey, W., & West, K. (1994). Automatic lag selection in covariance matrix estimation. Review of Economic Studies, 61, 631
653. Ng, S., & Perron, P. (2001). Lag length selection and the construction of unit root test with good size and power. Econometrica, 69, 1519
1554. Osler, C. L. (2006). Macro lessons from microstructure. International Journal of Finance and Economics, 11, 55
80. Papell, D. H. (1997). Cointegration and exchange rate dynamics. Journal of International Money and Finance, 16, 445
460. Perron, P., & Ng, S. (1996). Useful modifications to some unit root tests with dependent errors and their local asymptotic properties. Review of Economic Studies, 63, 435
463. Perron, P., & Qu, Z. (2007). Estimating and testing structural changes in multivariate regressions. Econometrica, 75, 459
502. Perron, P., & Rodriguez, G. (2003). GLS detrending, efficient unit root tests and structural change. Journal of Econometrics, 115, 1
27. Rapach, D. E., & Wohar, M. E. (2002). Testing the monetary model of exchange rate determination: New evidence from a century of data. Journal of International Economics, 58, 359
385. Rapach, D. E., & Wohar, M. E. (2004). Testing the monetary model of exchange rate determination: A closer look at panels. Journal of International Money and Finance, 23, 867
895. Rapach, D. E., & Wohar, M. E. (2006). The out-of-sample forecasting performance of the nonlinear models of real exchange rate models. International Journal of Forecasting, 22, 241
261. Rossi, B. (2005). Testing long-horizon predictive ability with high persistence, and the Meese-Rogoff puzzle. International Economic Review, 46, 61
92. Rossi, B. (2006). Are exchange rates really random walks? Some evidence robust to parameter instability. Macroeconomic Dynamics, 10, 20
38. Said, S. E., & Dickey, D. A. (1983). Testing for unit roots in autoregressive moving average models of unknown order. Biometrica, 71, 599
607. Sarantis, N. (1994). The monetary exchange rate model in the long run: An empirical investigation. Weltwirtschaftliches Archiv, 130, 698
711. Sarno, L., & Taylor, M. P. (2002). The economics of exchange rates. Cambridge: Cambridge University Press. Sarno, L., & Valente, G. (2009). Exchange rates and fundamentals: Footloose or evolving relationship? Journal of the European Economic Association, 7, 786
830. Sarno, L., Valente, G., & Wohar, M. E. (2004). Monetary fundamentals and exchange rate dynamics under different nominal regimes. Economic Inquiry, 42, 179
193.

124

P. F. Diamandis et al.

Schinasi, G. J., & Swamy, P. A. V. B. (1989). The out-of-sample forecasting performance of exchange rate models when coefficients are allowed to change. Journal of International Money and Finance, 8, 375
390. Schwert, G. W. (1987). Effects of model specification on tests for unit roots in macroeconomic data. Journal of Monetary Economics, 20, 73
103. Sephton, P. S. (1995). Response surface estimates of the KPSS stationarity test. Economics Letters, 47, 255
261. Shiller, R. J., & Perron, P. (1985). Testing the random walk hypothesis: Power versus frequency of observation. Economics Letters, 18, 381
386. Sims, C. A. (1980). Macroeconomics and reality. Econometrica, 48, 1
48. So, R. W. (2001). Price and volatility spillovers between interest rate and exchange value of the US dollar. Global Finance Journal, 12, 95
107. Syllignakis, M. N., & Kouretas, G. P. (2011). Markov switching regimes and the monetary model of exchange rate determination: Evidence from the Central and Eastern European markets. Journal of International Financial Markets, Institutions and Money, 21, 707
723. Taylor, A. M. (2002). A century of purchasing power parity. Review of Economics and Statistics, 84, 139
150. Taylor, M. P. (1995). The economics of exchange rates. Journal of Economic Literature, 33, 13
47. Taylor, M. P., & Peel, D. A. (2000). Nonlinear adjustment, long-run equilibrium and exchange rate fundamentals. Journal of International Money and Finance, 19, 33
53. Taylor, M. P., Peel, D. A., & Sarno, L. (2001). Nonlinear mean-reversion in real exchange rates: Towards a solution to the purchasing power parity puzzles. International Economic Review, 42, 1015
1042. Taylor, M. P., & Sarno, L. (1998). The behavior of real exchange rates during the post-Bretton Woods period. Journal of International Economics, 46, 281
312. Wollf, C. (1987). Time-varying parameters and the out-of-sample forecasting performance of structural exchange rate models. Journal of Business and Economic Statistics, 5, 87
97. Zivot, E., & Andrews, D. W. K. (1992). Further evidence on the great crash, the oil price shock and the unit root hypothesis. Journal of Business and Economics Statistics, 10, 251
270.

A Dynamic Gravity Model for Global Bilateral Investment Holdings Konstantinos Drakosa, Ekaterini Kyriazidoub and Ioannis Polycarpoub a

Department of Accounting and Finance, Athens University of Economics and Business, Athens 10434, Greece, e-mail: [email protected] b Department of Economics, Athens University of Economics and Business, Athens 10434, Greece

Abstract Purpose
This paper seeks to explain the serial persistence as well as the substantial number of zeros characterizing global bilateral investment holdings. We explore the different sources of serial persistence in the data (unobserved country pair effects, genuine state dependence, and transitory shocks) and examine the crucial factors affecting the decision to invest in a host country. Methodology
Based on a gravity setup, we consider investment behavior at the extensive (participation) margin and employ dynamic first-order Markov probit models, controlling for unobserved cross-sectional heterogeneity and serial correlation in the transitory error component, in order to explore the sources of persistence. Within this modeling framework we explore the importance of institutional quality of the host country in attracting foreign investment. Findings
The data support that the strong persistence is driven by true state dependence, implying that past investment experiences strongly impact on the trajectory of future investment holdings. Institutional quality appears to play a significant role to attract foreign investment. Research implications
The empirical findings suggest that due to the existence of genuine state dependence, inward-investment stimulating policy International Symposia in Economic Theory and Econometrics, Vol. 23 G. P. Kouretas and A. P. Papadopoulos (Editors) Copyright r 2014 by Emerald Group Publishing Limited. All rights reserved ISSN: 1571-0386/doi:10.1108/S1571-038620140000023005

126

K. Drakos et al.

measures could have a more pronounced effect since they are likely to induce a permanent change to the future trajectory of inward investment. Originality
Both the substantial number of zeros and the salient persistence characterizing bilateral investment holdings decision have been previously overlooked in the literature. A study modeling jointly the levels and the selection mechanism could prove a fruitful direction for future research. Keywords: Bilateral investment holdings, gravity model, initial conditions, state dependence JEL Classifications: F21, F34, G11, C25, C35

Introduction The accelerating process of financial globalization in the last decades provides countries with an important opportunity for wider portfolio diversification by investing in a large variety of financial assets available in capital markets worldwide (Lane & Milesi-Feretti, 2008). This process is an integral part of capital mobility, providing vital access to international capital for developing countries. From this point of view, international asset trading is of great importance for efficient risk sharing and economic development. It is, therefore, of great interest to investigate the determinants of the global financial capital geography, and especially the factors behind the potential of a country issuing securities (host country) to attract international capital by selling financial assets to foreign investors. The unveiling of the crucial factors rendering a country attractive for international financial capital is a prerequisite for designing effective policy plans that aim to facilitate capital inflows and consequently instigate economic development through enhanced access to external financing for investment projects in the host country. Moreover, examining the dynamics characterizing the source country’s decision to invest abroad provides grounds for evaluating the effects of policy in time
that is, we can ask whether the policy followed today to attract foreign investment could have persistent effects on the future capital inflows the host country will experience. There is now a body of empirical work employing gravity equations, formerly only used to model bilateral trade in goods, to explain bilateral trade in financial assets (equity and bonds). The empirical success and analytical tractability of gravity models has established them as a standard reference model in international finance (for an influential paper on the

A Dynamic Gravity Model for Global Bilateral Investment Holdings

127

empirical estimation of financial gravity equations see Portes & Rey, 2005). Theoretical foundations for financial gravity equations can stem from a variety of modeling assumptions. Obstfeld and Rogoff (2000) demonstrate how frictions in product markets can explain home bias in equity positions, even when global financial markets are complete. Martin and Rey (2004) focus on transactional frictions in asset markets, developing a model of incomplete financial markets reaching a gravity equation for bilateral investment holdings. Recently, Okawa and Van Wincoop (2012) established in a generalized framework the theoretical foundations necessary to generate a gravity model equation in financial transactions. The gravity equation framework roughly states that under certain modeling assumptions, the level of bilateral investment is positively affected by some measure of host and source country sizes, and is negatively related to bilateral trading costs between host and source countries. Trading costs are interpreted in the literature as informational costs entailed in asset trading, reflecting uncertainty, informational asymmetries and (un-)familiarity, cultural and trust factors (Portes & Rey, 2005). Empirical studies have also established that relevant host country factors affecting its appeal for international capital inflows are: (1) the level of host country’s institutional quality (Papaioannou, 2009), and (2) the level of host country’s (financial) market development (Alfaro, Kalemli-Ozcan, & Volosovych, 2008; Aviat & Coeurdacier, 2007; Buch, 2003; Daude & Fratzscher, 2008; Egger & Merlo, 2007; Fratzscher & Imbs, 2009; Gelos & Wei, 2005; Guiso, Sapienza, & Zingales, 2009; Lane & Milesi-Ferretti, 2008; Portes, Rey, & Oh, 2001; Rose & Spiegel, 2004; Portes & Rey, 2005; Stein & Daude, 2007; Wei & Shleifer, 2000). The standard gravity literature examines variations of the gravity equation for (the levels of) bilateral investment holdings in country pairs for a sample period, and thus is occupied with examining only the observations where positive (i.e., nonzero) level of investment has been observed. This empirical strategy, simplifying and straightforward as it may be, ignores the underlying selection mechanism which determines whether investment actually takes place or not in a given time period, and thus entails a danger of endogeneity-like bias in the estimates of the coefficients for the gravity factors affecting the level of investment from sender to host country. Moreover, no light is shed on the selection equation itself, which is of significance in its own right for policy making. Another unexplored feature of the data so far has been the serial persistence characterizing the discrete zero-one investment decision. Zeroes tend to be followed by zeroes, and ones tend to be followed by ones. An important question from a policy perspective is whether we can attribute the serial persistence to a causal mechanism from previous investment decision to the current, or if it is an artifact originating from the presence of

128

K. Drakos et al.

unobserved pair heterogeneity or even transitory factors correlated between time periods. This paper aims to fill both gaps. To this ends we employ dynamic random effects discrete choice models to investigate the factors affecting the decision to enter a foreign asset market. The discrete panel setting is appropriate since it allows us to estimate consistently both time varying and invariant factors and also incorporate unobserved pair-specific heterogeneity. Furthermore, it allows us to utilize a substantial part of the data for countries which for confidentiality reasons do not report levels of investment but only whether investment has taken place or not. In addition, the models we employ allow for both the presence of genuine state dependence and unobserved heterogeneity, as well as, serial correlation in the unobserved transitory errors, enabling us to distinguish among different sources of serial dependence and conclude on the nature of this persistence with a view on implications on the impact of investment-enhancing policies. We use data from the Coordinated Portfolio Investment Survey (CPIS hereafter), a comprehensive dataset compiled by the IMF, providing information about the investment holdings of a large number of source countries over several years, broken down by host country. Our dataset is a panel consisting of country pairs (source-host) observed from 2001 to 2007, where we observe whether the source country has invested in the host for each period (a discrete zero-one decision) and a variety of source and host specific gravity-like covariates potentially affecting this discrete choice. The countries involved in the CPIS report investment holdings by end-investors and custodians of assets issued by foreign countries broken down in bonds (short and long term) and equity holdings. The CPIS features a broad coverage of countries with 73 reporting source countries in 2007, and, contrary to the also widely used Bank of International Settlements (BIS) dataset, is not restricted to the banking sector. More importantly, the seven year panel structure of the data allows us to perform the analysis on both crosscountry and time dimensions to control for variation across country pairs and within country pairs through time. Dynamics of cross-border investment holdings can be introduced in a natural fashion in this panel data framework, and well-known dynamic discrete choice methodology is used to distinguish among conceptually very different (albeit observationally similar) sources of serial persistence in the data. Our main findings may be summarized as follows. Unobserved pair heterogeneity is shown to capture a significant part of the unobserved variation, pointing to the many unobservable factors governing investment decisions. There is also strong evidence supporting the existence of genuine state dependence, so that a successful policy that enhances the probability for positive foreign capital inflows today is even more beneficial in view of the fact that positive investment induced today will enhance the probability

A Dynamic Gravity Model for Global Bilateral Investment Holdings

129

of positive investment in the future as well. The dynamic analysis of the discrete decision taken in this paper makes a contribution to the literature not only by stressing the factors relevant for policy designing, but also empirically investigating the intertemporal value of such policy measures. Among the various gravity factors that have been examined in the literature, we find very significant negative effects for distance which are robust under various specifications, validating in the discrete choice framework a recurring gravity literature finding that informational frictions play a crucial role in explaining bilateral asset trade. Familiarity effects and cultural affinities are also shown to be positive determinants of investment in our results. Moreover, institutional quality of the issuing country emerges as a significant factor in attracting foreign capital, again in full accordance with the existing literature on international financial flows determinants. The rest of the paper is organized as follows: The second section describes the data used, and provides motivation for the main analysis. We explain the econometric methodology in the third section, and the fourth section presents the results. The fifth section concludes.

The Data and Motivation Bilateral Investment Decision and State Dependence Our dataset comes from the CPIS, and covers the time period 2001
2007. For this time span we have observations consisting of 54 source countries (i.e., countries investing abroad and buying foreign securities) and 166 host countries (i.e., countries selling the securities). After eliminating observations with missing values for any of our variables, we are left with the same group of 54 sender countries, but with less receiver countries (94 remain) and an unbalanced panel in which each pair of host-receiver is observed for only those periods where all variables are observed. This unbalanced panel constitutes a sample of 17.178 observations. The variable we examine in this paper is an indicator variable assuming the value one if a source country holds positive investment (combined bonds and equity) in a host country at time t and zero if it holds no investment. Table 1 shows that the percentage of zeros in the bilateral investment decision for both the whole sample and individually for each and every time period is strikingly high (over 60% in some sample years). This makes apparent the need to analyze the binary decision to invest or not in any given time period, if one is to avoid potentially serious bias in gravity equation coefficient estimates. In this paper, we have the opportunity to analyze this binary choice using all available data (that is including the observations in which bilateral investment is zero in levels) in a panel data setting for

130

K. Drakos et al.

Table 1: Regime-Specific and Transition Probabilities of Bilateral Investment Holdings 2001

2002

2003

2004

2005

2006

2007

All years

Panel A: Unconditional probabilities 0.368 0.398 PrðHi;t > 0Þ 0.631 0.601 PrðHi;t = 0Þ

0.395 0.604

0.393 0.606

0.430 0.569

0.417 0.582

0.452 0.547

0.409 0.590

Panel B: Conditional probabilities
0.047 PrðHi;t > 0 | Hi;t − 1 = 0Þ PrðHi;t = 0 | Hi;t − 1 = 0Þ
0.953 PrðHi;t > 0 | Hi;t − 1 > 0Þ
0.935 PrðHi;t = 0 | Hi;t − 1 > 0Þ
0.065

0.045 0.954 0.950 0.049

0.044 0.955 0.954 0.045

0.068 0.931 0.960 0.039

0.072 0.927 0.963 0.036

0.067 0.932 0.950 0.049

0.057 0.942 0.952 0.047

Note: Probabilities may not sum to one due to rounding errors.

a dataset consisting of many country pairs with varying host and sender political and financial conditions allowing us to study the impact of these factors in the attractiveness of a host country over time to attract investment from abroad. While the CPIS dataset is rich enough, there are a number of problems associated with it that need to be mentioned. For instance, there is the possibility of underreporting of assets which can be due to incomplete institutional coverage of the survey. However, this does not pose a problem to our analysis since we do not model the level of investment, but rather we treat investment holdings as a dichotomous variable. In addition, there are several instances where investment holdings data for certain country pairs are confidential. The dichotomous nature of our variable again surpasses this problem. Even though the exact amount is undisclosed for the purposes of our analysis we know it clears the zero threshold. These shortfalls notwithstanding, the CPIS provides a unique perspective on cross-country investment positions that warrants a detailed analysis. The strong persistence characterizing cross-country investment holdings becomes apparent in Panel B. In particular, while the unconditional average probability of positive investment is about 41%, it increases to 95%, when it is conditioned on positive investment holdings in the previous year. Similarly, while the unconditional probability of zero investment holdings is about 59%, when conditioning on zero investment holdings in the previous year it becomes 94%. These figures indicate that the cross-border investment process exhibits strong persistence, a property that has important policy implications for host countries aiming to attract foreign capital to fund their investment projects. If this persistence can be shown to be true state dependence, meaning that positive investment holdings at one period in time affect the probability of positive investment in later periods, so that

A Dynamic Gravity Model for Global Bilateral Investment Holdings

131

current positive investment is taking place because of past positive investment, this points to a specific mechanism that translates a positive investment in the previous period into a higher probability of positive investment in the present period as well. From a theory perspective, the large number of zeros as well as the strong persistence can be explained by the fact that entry in a foreign financial market involves a sunk cost that has to be incurred by the sender country investor, and this cost makes the investment decision partially irreversible (see Daveri, 1995). This cost includes explicit taxes, authorization, and registration procedures required, but more importantly the cost of acquiring the information needed to assess the attractiveness of a foreign market in order to decide on entry or not, such as host country legislation, the quality of institutions, level of investor protection, and political stability. This cost can be seen as a one-off incurred cost, which is no longer present in subsequent periods. This means previous investment in a host country makes future investment there more probable: it is now less costly to reinvest in a given host than incur costs to find a new host country. In our estimations, state dependence is strongly present and robust under several specifications, pointing to a mechanism with the description given above.

Empirical Strategy To model the bilateral cross-country investment holdings decision we employ a dynamic binary choice panel data model of the form: 0

Invit = 1fγ  Invi;t − 1 þ Xit β þ uit > 0g i = 1; …; N;

ð1Þ

t = 1; …; T:

Our setup is a panel data model with two dimensions: The crosssectional dimension is a pair i of source
host countries, and the time dimension where we observe investment and other covariates characterizing the host and source countries within each pair for every time period t. The indicator function 1{} takes the value one if the event within the braces has occurred, and zero otherwise. The dependent variable Invit is a dichotomous variable assuming the value one if positive investment holdings are observed by the source country in host country’s assets for pair i in period t, and zero otherwise. We thus model the binary investment choice for each time period employing a threshold-crossing binary choice setup. We use an array of covariates in vector Xit to model the discrete choice to enter in a foreign financial market. The subscripts in variables denote in order the pair, country (source or host) and time period of the observation. Thus, a variable xi,h,t in Xit is a covariate for pair i, characterizing the

132

K. Drakos et al.

host-country (h), observed for time period t. Accordingly, xi,s,t is a covariate of pair i, characterizing source-country (s) and is observed for time-period t. Variables that describe pair-specific characteristics, that is, a relation between source
host countries within a pair (like common language or common legal origin) and distance between source
host countries (which characterizes the pair participating in a transaction rather than each one participating in it) are denoted just by xi, meaning pair i specific characteristics. Note that these characteristics are also invariant in time dimension, so no subscript t is needed. We use the per capita GDP of the host (GDPi;h;t ) and source (GDPi;s;t ) countries measured in year 2000 dollars to control for country size. To proxy informational costs between source and host country in a country pair, we use the logarithm of distance between host and sender (log disti ), a dummy variable for whether the two countries share a common official language (comlang off i ), a dummy for three legal origins (the United Kingdom, French, and German), and finally a dummy for the two countries sharing a common legal origin (commonlawi ). These variables account for cultural differences and affinities which play an important role in foreign investment decisions (see e.g., Guiso et al., 2009). Openness of the host economy in trade in general can also be a factor of attractiveness for equity investment (see Aviat & Coeurdacier, 2007 for the role of goods trade in financial trade), so we also include a variable defined as the percentage of the value of goods trade (exports + imports) in host country GDP. It is now well-documented in the international finance literature that institutional quality of the host country plays an essential role in facilitating foreign capital investment in the country’s financial assets (see Papaioannou, 2009). Institutional quality consists of a variety of factors, more prominently legislation regarding investor protection (e.g., protection from expropriation), political stability, rule of law, and general socioeconomic conditions. The role of investor protection in particular and more generally of better legislation regarding foreign investment is elegantly analyzed in the illuminating model by Shleifer and Wolfenzon (2002) where it is shown that countries with better investor protection laws have more valuable firms with lower share concentration, a bigger diversity in investment opportunities, and also have higher interest rates. The same paper also provides a possible explanation for the Lucas paradox of capital not flowing from the rich to poorer countries: Better investor protection leads to higher interest rates and eliminating the incentive for capital to flow to a country with worse investor protection. Also, Daveri (1995) uses a simple model to conclude that better investor protection and political stability is consistent with more capital inflows from other countries. We control for host country institutional quality via a composite index (polriski;h;t ) (described in the appendix) from ICRG where countries are

A Dynamic Gravity Model for Global Bilateral Investment Holdings

133

graded from 1 to 100, with a larger grade meaning a smaller risk, and better conditions for investment. Moreover, an index again from ICRG measuring host country’s financial sector quality (finriski;h;t ), is included in estimations. Finally, domestic credit as percentage of GDP (domcredi;h;t ) and stock turnover ratio (stockturni;h;t ) are used to describe the host country’s financial sector development and sophistication. Descriptive statistics of all the above covariates are given in the appendix. We use time-period specific effects to control for any circumstances that are special for the years included in the data, and affect the global investment conditions. Moreover, we need to control for the “multilateral resistance” term (Anderson & Van Wincoop, 2004; Baldwin & Taglioni, 2006). As in the trade literature, this term can be interpreted as a price index of all financial assets competing with an imported asset (see Coeurdacier & Martin, 2009). Omission of this term could lead to biases in the estimated coefficients for our transaction costs variables. We employ two alternative specifications to deal with this empirically: we use either regional dummies for the continent of source and host country (Europe, Asia, Oceania, America), or a full set of source and host-specific dummies. The former methodology is not as inclusive as the latter, but nevertheless allows us to keep a reasonable number of parameters to estimate. The latter, albeit being more in accordance with the theory, poses many estimation problems in a nonlinear maximum likelihood estimation framework used here due to the large number of parameters needed (as many as the sum of host and source countries). In our estimations, we use both when possible and report both sets of estimates in the tables.

Econometric Methodology In this section we present the econometric methodology and specific model assumptions we make to estimate Equation (2). The model is the discrete choice panel model first proposed and analyzed by Heckman (1981a, 1981b) and fits the purpose of this paper as it allows dynamics through the inclusion of the lagged investment decision as an explanatory variable, thus introducing genuine state dependence as a structural feature of the model. In addition, it allows the decomposition of the error term into a country pair-specific random effect component ai and a transitory error term, in the following form uit = ai þ εit :

ð2Þ

It thus allows an additional source of serial persistence through the presence of a random time invariant unobservable component in every country

134

K. Drakos et al.

pair. Serial persistence stemming from this feature of the model is described as spurious state dependence, as it produces the same data features that genuine structural state dependence would, only the underlying cause is rooted in heterogeneity and unobservable factors rather than a structural causal effect. Qualitatively, such a distinction makes a significant difference for policy, as the scope of measures encouraging capital inflows makes sense in the presence of true state dependence, where policy taken can take advantage of a mechanism translating positive investment today into higher probability of experiencing foreign capital inflows in the future as well. On the other hand, policy can hardly play any part if serial persistence is caused by unobservable idiosyncratic effects. We will also use in the most general model specification the assumption that the transitory errors are serially correlated, thus adding a third competing explanation of serial persistence. Estimation of nonlinear panel data models with unobserved heterogeneity is highly dependent on the assumptions we are willing to make. A fixed effects approach makes no assumption about the distribution of the heterogeneity and its statistical relation to the covariates, and thus is more attractive by ensuring that the conditional distribution of the effects does not play a role in the identification of the parameters of interest.1 However, fixed effects methods have stringent requirements on the covariates while they do not deliver estimates of coefficients of time-invariant variables nor predictions, and hence are less used in practice. On the other hand, random effects methods that fully specify the distributional properties of heterogeneity lead to standard maximum likelihood estimation and any computational burden impeding the estimation of the parameters has been lifted considerably by the use and development of simulation methods. We take a fully parametric random effects approach in this paper, in the spirit of Hyslop (1999) and Heckman (1981a) and we specify the distribution of the random effects and the transitory error term. Random effects require that the distributional properties of ai and εit as well as their statistical relationship to the covariates be specified, along with the initial conditions of the dynamic process (see Hsiao, 2003). In all specifications we will assume that the transitory error term εit is independent of the gravity covariates and of the ai and normally distributed. To model the pair-specific unobserved heterogeneity, ai, we will make two alternative hypotheses. In the first case ai is considered independent of all observed covariates, while in the latter case, a more flexible assumption for the conditional mean of the random effects is assumed. Following Mundlak (1978)

1

For a comprehensive review of available estimation methods of this type see Hsiao (2003), Arellano and Carasco (2003), Honore´ and Kyriazidou (2000).

A Dynamic Gravity Model for Global Bilateral Investment Holdings

135

and Chamberlain (1984), its conditional mean is assumed to be a linear function of the longitudinal averages of some of the gravity covariates, X i , and an independent normally distributed error term vi. The latter correlated random effects assumption is specified as ai = X i δ þ v i 0

vi |X; ε ∼ Νð0; σ 2v Þ:

ð3Þ

The intuition behind the latter specification is that cross-sectional differences in longitudinal averaged characteristics carry information for the permanent unobserved country pair-specific characteristics. More specifically, we could suspect that the unobserved pair heterogeneity is statistically correlated with some of the observable host characteristics we use as covariates. These pair-specific unobserved characteristics could possibly be behind the realizations of institutional quality and financial market development of the host country. Omitted (and possibly difficult to quantify) factors, like attitude of a country pair toward obeying the law, or even disposition toward liberalization of the economy, that are present in the unobserved pair-effect, are very likely to be correlated with political and financial risk indices, as well as with variables reflecting the host’s financial markets development. Thus, we allow pair-specific means of the host’s political and financial risk variables (polriski;h;t ; finriski;h;t ) and financial markets development (domcredi;h;t ; stockturni;h;t ) to be included in X i . In both cases, the distribution of the unobservable composite error, ai þ εit in the uncorrelated case and vi þ εit in the correlated case, is therefore normal and independent of X it . In the absence of state dependence (γ = 0) and of serial correlation in the transitory error component, εit, that is in a static model, parameters of interest are estimated via maximum likelihood using Gaussian
quadrature to compute the univariate integral involved in the evaluation of the likelihood function. A simple test for the presence of correlation between the individual effect and the observed covariates can be carried out by testing the null hypothesis that δ = 0. The presence of the lagged investment decision (Invi;t − 1 ) in our dynamic specifications brings us to the initial conditions problem: we need to specify the statistical relationship between the initial investment decision, Invi0 (in our setting initial period is 2001) and the unobserved heterogeneity ai. A simple approach would be to assume that Invi0 is exogenous and can therefore be treated as fixed, as, for example, might be the case if the process were observed from its initialization. Clearly this assumption is not realistic and unlikely to hold in our context. We use two standard sets of assumptions to tackle the initial conditions problem. The first one is Heckman’s approach (see also Arulampalam & Stewart, 2009;

136

K. Drakos et al.

Stewart, 2006) which specifies a flexible reduced form approximation to the initial conditions: Invi0 = 1fZ0i0 ζ þ ui0 > 0g ui0 = θai þ wi0 ;

ð4Þ

where Zi0 includes members of Xi0 and wi0 is assumed to be uncorrelated with εit, for t = 1,…, T and to follow the normal distribution. This method essentially tackles the issue of initial conditions by assuming a distribution for the initial condition conditional on the random effects and the covariates in the initial period. Under this specification, uit is equicorrelated with ui0 and a test of exogeneity of the initial conditions can then be conducted by testing whether this correlation is zero, that is, testing the null hypothesis θ = 0. An alternative approach to the initial conditions problem is proposed by Wooldridge (2005). Instead of specifying a model for the initial conditions given the observed covariates and the unobserved effect, a model is specified for the unobserved effect given observed covariates and the initial conditions. In particular it is assumed that ai = ξ0 Invi0 þ Xi ξ1 þ vi : 0

ð5Þ

The error term vi is independent of everything else and normally distributed. Substituting back into Equation (1) gives 0

0

Invit = 1fX it β þ γInvi;t − 1 þ ξ0 Invi0 þ X i ξ1 þ vi þ εit > 0g;

ð6Þ

which again becomes a two factor probit model that can be easily estimated my ML using Gaussian quadrature procedures. The essential difference is that Equation (5) allows us to form a likelihood for ðInvi1 ;…; InviT Þ conditional this time not only on ai but also on Invi0 as well. Using Wooldridge’s method, the exogeneity of the initial condition is tested by the significance of the coefficient ξ0 . In the most general specification, serial persistence in Invit may be due not only to the presence of the lagged dependent variable Invit − 1 and/or the presence of permanent unobserved heterogeneity ai in the model, but also to the fact that the transitory error term εit may be serially correlated. To allow for this possibility, we specify a first-order autoregressive (AR (1)) model: εit = ρεit − 1 þ ηit ;

ð7Þ

where ηit is an independent normal error term. Estimation now becomes computationally cumbersome, since observations across time for a given

A Dynamic Gravity Model for Global Bilateral Investment Holdings

137

pair are no longer independent conditional on the unobserved heterogeneity, thus the probability of a string of observations for a pair is now a rectangle of T-dimensional normal distribution. The GHK (Geweke
Hajivassileiou
Keane) simulator is the best way to simulate this probability, and we perform maximum simulated likelihood maximizing the simulated likelihood function for the parameters. For more details on MSL see, for example, Hajivassiliou and Ruud (1994).

Empirical Analysis Static Random Effects Models of Bilateral Investment Holdings First we focus on static random effects probit models, that is, we do not allow for sources of dynamics to enter the model. We assume no state dependence and no serial correlation in the transitory errors. In Table 2 we employ both correlated and uncorrelated random effects assumptions, and estimate this model by Gaussian quadrature with 24 points of integration. Standard errors are obtained by the inverse of the numerically approximated Hessian (using finite differences) of the likelihood function, and when this is not possible we use the last BFGS step approximation of the hessian to obtain standard errors. The sample size is 17,178 observations (after dropping observations containing missing values for any of our covariates), and the time span varies for each pair observation (i.e., we have an unbalanced panel with time span from 2001 to 2007). Note that the static nature of this model allows us to use an unbalanced panel since gaps in time observations do not pose a problem for estimation. The 3821 pairs are consisted of 54 sender and 94 host countries. Table 2 reports results from static random effects probit models, with gravity covariates of host and source characteristics. No correlation is allowed between unobserved pair heterogeneity and explanatory variables. The static framework allows us to focus solely on the roles of institutional quality, informational costs and financial development of the host country, and how their evolution over time in a panel framework affects the probability of positive investment. Time-specific fixed effects are included to capture the particular to the time period global investment conditions affecting both the source and host countries. Dummy variables for the continent of host and source are used to control for multilateral resistance terms, as well as a dummy for the host and source residing in a common continent. The random effects heterogeneity assumption captures unobserved pair-specific social linkages and trust factors that cannot be captured by other covariates.

138

K. Drakos et al.

Table 2: Static Random Effects Models of Bilateral Investment Covariates

RE probit (1)

RE probit (2)

CRE probit (3)

logdisti

−0.634*** (−6.859) 1.960*** (7.694) −0.028 (−0.209) −0.424*** (−8.000) 0.937*** (11.254) 1.369*** (22.396) 0.274*** (4.022) 0.017 (0.452) 0.552*** (9.938) 0.326*** (8.929) 17.178 54 94 Yes No No Yes Yes

−1.203*** (−13.646) 0.784*** (3.193) 0.235** (1.928) −0.460*** (−3.379) 0.086 (0.354) 0.069 (0.226) 0.286*** (3.549) −0.143*** (−2.171) 0.104* (1.130) 0.011 (0.178) 17.178 54 94 Yes Yes Yes No Yes

−0.840*** (−8.225) 2.176*** (8.470) −0.018 (−0.153) −0.109*** (−2.118) 0.536*** (6.141) 1.685*** (24.154) 0.289*** (2.732) −0.116*** (−2.284) 0.106* (1.301) 0.034 (0.701) 17.178 54 94 Yes No No Yes Yes

comlang off i commonlawi tradegdph;i;t gdpcap00ush;i;t gdpcap00uss;i;t polriskh;i;t finriskh;i;t domcredh;i;t stockturnh;i;t Observations Sender countries Host countries Time fixed effects Sender country fixed effects Host country fixed effects Geographical dummies Country pair random effects

Notes: The dependent variable is Invi;t , a binary variable assuming the value 1 for positive investment in pair i between source and host country h in time period t. The pairs contain 54 sender and 94 host countries. Time fixed effects are included in all models. Model (1) contains also geographical dummy variables for the continent of sender and host and a dummy for common continent of host and source. Model (2) contains also sender and host-specific fixed effects dummies. Model (3) assumes Mundlack
Chamberlain correlated Random effects, and is estimated with additional regressors for the means of polriski;h;t ; finriski;h;t ; domcredi;h;t ; stockturni;h;t for every pair i. Z-scores are in parentheses. *, **, *** denote significance at the 10%, 5%, and 1% level, respectively. Models were estimated in MATLAB by Gaussian quadrature with 24 integration points. Standard errors were obtained in models (1) and (3) by the inverse of the approximated Hessian of the Likelihood function at the optimum, using finite differences. In Model (2), the last BFGS approximation of the Hessian was used. A detailed description of all the covariates used is available in the Appendix.

A Dynamic Gravity Model for Global Bilateral Investment Holdings

139

The results are consistent with the findings of the financial gravity literature (Portes & Rey, 2005; Aviat & Coeurdacier, 2007; Lane & MilesiFerretti, 2008). The standard result of distance affecting negatively the levels of bilateral investment is also present in the discrete choice framework: distance enters with a negative coefficient, thus implying a negative effect on the probability of investing on a distant host country. Distance is a proxy for informational asymmetries, reflecting nonstandard costs (Papaioannou, 2009; Portes et al., 2001) negatively affecting the probability of positive investment in a host country. Cultural linkages and familiarity effects (see Guiso et al., 2009; Portes & Rey, 2005) also appear very significant, and the common language dummy enters with a large positively signed coefficient. We thus find that in a static model, informational costs (mirroring transportation costs in goods trade) are a significant barrier to asset trading over time. Host country size appears, not surprisingly to have a positive effect on the probability of attracting foreign investment, which is in accordance with the Lucas paradox of richer countries attracting foreign capital rather than poorer ones. The institutional quality of the host country, captured in the political risk index, is positively signed and statistically significant, confirming the finding in the literature on the enhancing effect of investor-friendly legislation and political stability on the attractiveness for capital inflows. The development of financial markets also appears to be a positive influence on attracting capital, as the positive domestic credit as percentage of GDP and stock turnover ratio coefficients suggest. In model (3), correlation between the unobserved pair random effect and some of the covariates is allowed. The results are roughly the same as model (1) qualitatively, but with financial risk now appearing significant at 5% level of confidence. In column (2) we maintain the same static framework, but instead of accounting for host and source-specific fixed effects by geographical dummies, we include a full set of dummies for each of the 54 source countries and 94 host countries. The full set of dummies for source countries captures the multilateral resistance term (Anderson & Van Wincoop, 2004; Baldwin & Taglioni, 2006; Coeurdacier & Martin, 2009). With host dummies we control for unobservable country factors that affect international asset holdings. This modeling approach somewhat alters the obtained results. The distance coefficient is now amplified by more than two times its previous value, reinforcing its importance as a proxy of informational costs when we can properly control for multilateral resistance, as well as pairspecific heterogeneity. The presence of a common language is now less stressed in magnitude, but its coefficient remains positive and strongly significant. An important difference is that the presence of a common legal origin turns positive and statistically significant
which is more intuitively appealing since similarities in the legislative environment between host and source country should encourage investment, through reduction of

140

K. Drakos et al.

informational costs and uncertainty. The coefficient of trade openness of the host country is negative and statistically significant, unchanged from column (1) results. This is a counter intuitive result since one would expect more open countries, in terms of trade, to attract more investment. The host institutional quality as measured by the Political risk index, maintains a positive virtually unchanged in magnitude coefficient, which is strongly statistically significant even at a 1% confidence level. Financial development variables coefficients appear insignificant in this set of estimates, overturning the previous results obtained with geographical dummies. Table 3 presents specification tests on the estimated models of Table 2. We perform joint Wald tests for joint significance of the covariates categorized in groups suggested by the content of those variables. Specifically, we include distance and common language in the category of variables expressing informational frictions. The dummies regarding legal origins and the presence of a common legal origin, alongside the indices of political and financial risk are put on a different group under the general description of institutional quality. Lastly, domestic credit and stock turnover ratios are included under the general category of variables describing financial development. We test the joint significance of the variables in each of the three groups. Informational frictions variables are strongly jointly significant across all three specifications, with a p-value of practically zero. The same is true for institutional quality variables, where the hypothesis of joint insignificance of political risk, financial risk, legal origins dummies and Table 3: Diagnostic Tests for Static Random Effects Models Values of Test-Statistics (p-values in parentheses) Test for informational frictions: logdisti ; comlang off i jointly insignificant Test for host institutional Quality: polriski;h;t ; finriski;h;t ; legorðuk; fr; geÞ; commonlawi jointly insignificant Test for host financial market development: domcredi;h;t ; stockturni;h;t jointly insignificant Test for Correlated Random Effects % of random effects variance in total error variance Log-likelihood

RE (1)

RE (2)

CRE (3)

80.25 (0.000) 89.82 (0.000)

223.78 (0.000) 43.12 (0.000)

126.00 (0.000) 132.2 (0.000)

189.20 (0.000) Does not apply 89% −5083.2

1.31 (0.519) Does not apply 65.6% −3709.7

2.28 (0.319) 269.9 (0.000) 87.8% −4956.8

Notes: All models were fitted in MATLAB using Gaussian quadrature with 24 integration points. Asymptotic Wald tests were conducted using the variance
covariance matrix as computed by the inverse of the (numerically approximated by finite differences) Hessian of the likelihood function at the optimum. For Model (2) we used standard errors from the last QuasiNewton BFGS algorithm step.

A Dynamic Gravity Model for Global Bilateral Investment Holdings

141

common legal origin is strongly rejected across all three models. Financial markets development appears to be significant only in model (1). The joint tests reaffirm our findings that informational frictions and host country institutional quality are the factors mainly behind the attractiveness for international capital inflows, restating the familiar findings of the empirical gravity research in a discrete choice framework. An advantage of random effects specification is that we can relax the assumption of the independence of unobserved heterogeneity and observed covariates in a simple tractable manner following Chamberlain’s correlated random effects assumption.2 We do so in model (3): In our fully parametric approach we assume that the conditional mean of the unobserved paireffects is a linear function of the means of political risk, financial risk, domestic credit, and stock turnover ratio variables. The assumption that the coefficients of the variables in means are jointly zero is strongly rejected in this static model pointing to a more complex statistical relationship between unobserved effects and covariates. However, due to the incomplete way in which the geographical dummies cover host- and source-specific fixed effects (i.e., unobserved factors of the source and host and multilateral resistance of the source country) could be behind this result since the random effect is not properly “cleaned” of these fixed effects and thus exhibits correlation with the observed covariates. In support of this explanation in a regression not reported in the table we allowed for a full set of dummies alongside correlated random effects and the CRE assumption was rejected. Moreover, the variance of the unobserved effects captures a large portion of total variance, roughly 89% in model (1) and 88% in model (3). The inclusion of host and sender-specific fixed effects in model (2) is naturally decreasing this percentage to a modest 65%, which is expected given that less space is left for pair heterogeneity when we proliferate in fixed effects.

Dynamic Random Effects Models of Bilateral Investment Holdings Serial persistence in the data raises the issue of investigating the possible underlying explanations in the framework provided by our econometric specification. Disentangling the effects of unobserved heterogeneity and serial correlation from genuine state dependence is of increased interest for policy makers. Primarily, we are interested in the presence of genuine state dependence which would imply a mechanism through which lagged investment decision affects the probability of positive one occurring in the

2

In a semi-parametric framework Arellano and Carrasco (2003) also relax this assumption but at the cost of a much more complicated estimation method.

142

K. Drakos et al.

present in a given pair of source
host countries. On a different layer, we examine the robustness of the results obtained by the previous section’s static models in the presence of dynamic effects. We proceed in this section with an analysis of the dynamic specification. In Table 4 we present estimation results. In the first column (1), we allow for state dependence through the introduction of the lagged investment decision Invi;t − 1 as an explanatory variable in addition to pair random effects, time-specific effects, and geographical fixed effects for source and host country. The coefficients on distance and linguistic ties are statistically significant even at 1% confidence level. The coefficient on distance is somewhat reduced and negative, and the same holds for the effect of common language which is positive but also a bit dampened. In view of the above, we conclude that the dynamic specification leaves unchanged the previously documented significance of informational frictions. The important difference to note is that political risk index in the presence of state dependence alongside pair-specific heterogeneity is now entering with a lower coefficient that is not statistically significant, and the same applies for the financial risk index. Moreover, we document that the state-dependence parameter γ is strongly positive, statistically significant, and large in magnitude, pointing to a genuine state dependence effect, that is, a structural causal effect from positive past investment decision. This finding is of significance for policy making: it suggests that increasing the attractiveness of a host country via reforms that increase investor protection and create a safer environment for foreign capital investment is not only helping to open the door to facilitate entry for foreign capital inflows, but also helps keep these inflows coming in the future. The fixed costs involved in foreign asset investment and in general costs of entering and leaving foreign financial markets have been recognized in the literature as factors that render investment partially irreversible (see Daveri, 1995). If we think of fixed costs as the resources (both pecuniary and non-pecuniary) needed for the foreign investor to familiarize herself with the conditions, both political and financial which are intertwined, that characterize a given investment opportunity in an unknown territory, once these costs are incurred it becomes easier to invest again in the same place, if the environment has not changed dramatically. The usual gravity covariates regarding country sizes also appear statistically significant. In column (2) of Table 4 we repeat the same estimation but include a full set of source
host fixed effects alongside time-specific effects.3 The

3 We could not get meaningful standard errors when all dummies were introduced and the GDP of the source was included. We report instead a reliable set of results without this variable in the table.

A Dynamic Gravity Model for Global Bilateral Investment Holdings

143

Table 4: Dynamic Models of Bilateral Investment Covariates Coefficient and z
scores

logdisti comlang off i commonlawi tradegdph;i;t gdpcap00ush;i;t gdpcap00uss;i;t polriskh;i;t finriskh;i;t domcredh;i;t stockturnh;i;t State dependence (γ) AR(1) parameter (ρ) Initial conditions parameter (θ or Invi0 ) Observations Sender countries Host countries Initial conditions specification AR(1) in transitory error term State dependence SD(1) Time fixed effects Sender country fixed effects Host country fixed effects Geographical dummies

CRE SD(1) Heckman (1)

CRE SD(1) Heckman (2)

CRE SD(1) Wooldridge (3)

CRE + AR(1)+SD(1) Heckman (4)

−0.445*** (−3.578) 1.241*** (4.147) 0.075 (0.505) 0.009 (0.134) 0.212** (1.882) 0.558*** (7.686) 0.152 (0.830) −0.108* (−1.337) 0.049 (0.354) −0.065 (−0.857) 1.651*** (15.380) Does not apply

−0.651*** (−3.158) 0.468 (0.514) 0.034 (0.085) −0.555*** (−1.801) 0.603* (1.505)
0.200 (0.325) −0.119 (−0.242) −0.072 (−0.159) −0.054 (−0.145) 1.599*** (3.445) Does not apply

−0.211*** (−1.896) 0.489*** (2.219) 0.122 (0.975) 0.004 (0.060) 0.192*** (1.694) 0.557*** (7.757) 0.144 (0.778) −0.112* (−1.378) 0.052 (0.377) −0.064 (−0.835) 1.619*** (14.540) Does not apply

2.040*** (4.507) 10.276 49 81 Heckman No Yes Yes No No Yes

0.116 (0.987) 10.276 49 81 Heckman No Yes Yes Yes Yes No

1.979*** (8.535) 10.276 49 81 Wooldridge No Yes Yes No No Yes

−0.443*** (−3.240) 1.003*** (3.391) 0.006 (0.036) −0.113* (−1.357) 0.460*** (3.087) 0.580*** (8.091) 0.364** (1.757) −0.061 (−0.592) −0.076 (−0.283) −0.090 (−1.096) 1.820*** (13.566) −0.130 −1.540* 1.958*** (8.091) 8029 44 57 Heckman Yes Yes Yes No No Yes

Notes: The dependent variable is Invi;t , a binary variable assuming the value 1 for positive investment in pair i between source and host country h in time period t. For models (1), (2), and (3) estimates were obtained with 49 sender and 81 host countries, using an unbalanced panel with a common initial period at 2001 and no gaps, and a varying endpoint year. For Model (4), 44 sender and 57 host countries were used in a balanced Panel, from 2001 to 2007 with no gaps. Gaussian quadrature with 24 integration points was used for models (1)
(3). Model (4) was estimated via the GHK simulator, with Maximum Simulated Likelihood, at 100 replications. Time fixed effects are included in all models. Models (1), (3), and (4) also contain geographical dummy variables for the continent of sender and host and a dummy for common continent of host and source. Model (2) also contains sender and host specific fixed effects dummies. All models assume Mundlack
Chamberlain correlated random effects, and include additional regressors for the means of polriski;h;t ; finriski;h;t ; domcredi;h;t ; stockturni;h;t for every pair i. Z-scores are in parentheses. *, **, *** denote significance at the 10%, 5%, and 1% level, respectively. Standard errors were obtained in models (1), (3), and (4) by the inverse of the approximated Hessian of the Likelihood function at the optimum, using finite differences. In Model (2), the last BFGS approximation of the Hessian was used. A detailed description of all the covariates used is available in the appendix.

144

K. Drakos et al.

results are qualitatively unchanged for the distance coefficient, which remains negative and significant, and is now larger in magnitude. The common language effect is insignificant in the presence of the full set of fixed effects. State dependence is a slightly lower but again statistically significant and large in magnitude. Notice that in both models (1) and (2) we employed the Heckman approach to initial conditions, specifying a distribution for the initial period investment given the unobserved heterogeneity random effect. To further investigate the robustness of our results we repeat the same analysis employing Wooldridge’s approach to the initial conditions problem, where a distribution is specified instead for the pair-heterogeneity random effect given the initial period and covariates. As illustrated in model (3), the results are qualitatively identical, excluding the reemergence of common language as statistically significant. We therefore document insensitivity of the qualitative findings to the assumption undertaken on the initial conditions. Two sources of serial dependence have been introduced so far: genuine state dependence and unobserved heterogeneity. As discussed before, a third remains to be explored: serial correlation in the transitory error component. This is introduced in model (4) alongside the other two sources and the model is estimated using simulation. Initial conditions are endogenous Heckman-type. The distance coefficient is highly significant, large in magnitude and negative, whereas in this most general specification common language retains its significance even at 1% level. State dependence appears larger in magnitude and highly significant, even as we allow for all three sources of dynamics to be present. The political risk variable retains its statistical significance at 5% level, stressing the importance of institutions for attracting investment. The correlation coefficient for the transitory error component is negative and statistically insignificant at 5%, thus it appears that the dynamics specified are enough to fit the data. Table 5 provides specification tests for the dynamic models of Table 4. The same strategy of joint significance testing for three groups of variables is employed here as well. Informational frictions covariates are jointly significant across all four specifications. Institutional quality variables are jointly significant in models (1) and (3) where no serial correlation in the error term is allowed, and turn insignificant in the full dynamic specification in model (4). However, we stress that the institutional quality as expressed through the political risk variable is significant when taken on its own, pointing to the importance of political stability and confirming in the discrete choice framework the findings of the literature (see e.g., Papaioannou, 2009, on the institutional quality importance for investment). Financial market development is not significant at 5% in any specification. Correlation between unobserved heterogeneity and covariates (i.e., Chamberlain’s CRE assumption) is rejected in our most general specification when all dynamics are

145

A Dynamic Gravity Model for Global Bilateral Investment Holdings

Table 5: Diagnostic Tests for Dynamic Random Effects Models Values of test-statistics

(P-values in parentheses)

Test for informational Frictions: logdisti ; comlang off i jointly insignificant Test for host institutional Quality: polriski;h;t ; finriski;h;t ; legorðuk; fr; geÞ; commonlawi jointly insignificant Test for host financial market development: domcredi;h;t ; stockturni;h;t jointly insignificant Test for correlated random effects Test for endogeneity of initial conditions*: parameter θ or coefficient on Invi0 equal to zero % of random effects variance in total error variance log-likelihood

CRE + AR(1) + SD(1) Wooldridge Heckman (3) (4)

CRE SD(1)

CRE SD(1)

Heckman (1)

Heckman (2)

27.25 (0.000) 19.44 (0.003)

10.188 (0.006) 1.327 (0.970)

9.046 (0.010) 21.19 (0.001)

27.20 (0.000) 6.848 (0.335)

0.85 (0.65) 14.31 (0.006) 2.040 (0.020) 54.8%

0.061 (0.970) 3.965 (0.410) 0.116 (0.453) 53.27%

0.833 (0.659) 14.12 (0.006) 1.979 (0.023) 49.60 %

1.145 (0.563) 6.932 (0.139) 1.958 (0.05) 68.3%

2258.3

1935.7

CRE SD(1)

1249.3

1514

Notes: Models (1), (2), and (3) were fitted in MATLAB using Gaussian quadrature with 24 integration points. Model (4) was estimated using the GHK Simulator with 100 replications and Maximum Simulated Likelihood was then employed to obtain coefficients. Asymptotic Wald tests were conducted using the variance
covariance matrix as computed by the inverse of the (numerically approximated by finite differences) Hessian of the likelihood function at the optimum. For Model (2) we used standard errors from the last Quasi-Newton BFGS algorithm step. (*) endogeneity of initial conditions tests are t-tests and the number in parenthesis is the p-value.

allowed for, however it is not rejected in models (1) and (3), supporting the explanation that geographical dummies do not properly control for fixed effects and multilateral resistance terms
the CRE assumption is rejected in the model equipped with a full set of dummies (model (2)). Initial conditions are found to be indeed endogenous in the most general of our specifications and 68.3% of the heterogeneity is captured by random effects.

Conclusions This paper’s contribution to the literature is twofold. First, departing from the usual analysis of the levels of cross border investment between countries, we investigate empirically the determinants of international asset trading focusing on the discrete decision to enter a foreign asset market. We use an extended panel dataset from the CPIS for seven time periods from 2001 to 2007 to estimate several random effects probit models. Drawing from the recent gravity models in the asset trading literature, and also research investigating the impact of institutional quality on the attractiveness of a

146

K. Drakos et al.

country for foreign investment, we control for relevant determinants of investment flows including country size, informational asymmetries, and the development and quality of financial and political institutions in the host country. Second, we look at the data from a dynamic perspective and analyze the serial persistence evident in the decision to invest in a host country over time. We consider a general specification allowing for genuine state dependence alongside unobserved pair heterogeneity and serial correlation in order to assess the source of this persistence and consequently evaluate policies that encourage capital inflows in the light of this analysis. Our results largely agree with the literature in identifying the significant determinants of bilateral asset trading. In a static model framework the informational costs, as proxied by distance, are shown to be negatively related to bilateral investment, and this finding is unaltered as we examine many specifications. Moreover, evidence that cultural links play an enhancing role in encouraging bilateral asset trade is found. More specifically, the presence of a common official language between host and source countries is found statistically significant and positive across all specifications. Institutional quality and development is confirmed in static specifications to play an important role in encouraging investment as shown by the positive coefficient of the relevant indices for political as well as financial institutions development indices. The usual gravity factors appear to have the sign and effect predicted by the theory. Unobserved pair heterogeneity is found to occupy large portion of the variance in the error term as well, implying that there is some space for the effect of factors like trust or other difficult to observe social linkages in explaining bilateral asset trade. A close look at the data implies that there is significant persistence in the decision to invest in a foreign financial market. We thus allow for dynamics to be present through three statistically and conceptually different routes, that is besides pair-specific random effects we allow for true state dependence and serially correlated errors. Genuine state dependence is statistically significant across all dynamic specifications, suggesting that policies to encourage foreign investment could have a lasting effect, given the positive effect of past investment on future investment as well. Informational costs and country sizes are as predicted by the gravity model theory still significant in the presence of dynamics although their effect is dampened compared to the static specifications.

References Alfaro, L., Kalemli-Ozcan, S., & Volosovych, V. (2008). Why doesn’t capital flow from rich to poor countries? An empirical investigation. Review of Economics and Statistics, 90(2), 347
368.

A Dynamic Gravity Model for Global Bilateral Investment Holdings

147

Anderson, J., & Van Wincoop, E. (2004). Trade costs. Journal of Economic Literature, 42, 691
751. Arellano, M., & Carrasco, R. (2003). Binary choice panel data models with predetermined variables. Journal of Econometrics, 115(1), 125–157. Arulampalam W., & Stewart M. (2009). Simplified implementation of the Heckman estimator of the dynamic probit model and a comparison with alternative estimators. Oxford Bulletin of Economics and Statistics, 71, 659–681. Aviat, A., & Coeurdacier, N. (2007). The geography of trade in goods and asset holdings. Journal of International Economics, 71(1), 22
51. Baldwin, R., & Taglioni, D. (2006). Gravity for dummies and dummies for gravity equations. NBER Working Papers No. 12516. National Bureau of Economic Research, Inc. Retrieved from http://www.nber.org/papers/w12516 Buch, C. (2003). Information or regulation: What drives the international activities of commercial banks? Journal of Money, Credit, and Banking, 35(6), 851
869. Chamberlain, G. (1984). Panel data. In S. Griliches & M. Intriligator (Eds.), Handbook of econometrics (pp. 1247
1318). Amsterdam: North Holland. Coeurdacier, N., & Martin, P. (2009). The geography of asset trade and the euro: Insiders and outsiders. Journal of the Japanese and International Economies, 23(2), 90
113. Daude, C., & Fratzscher, M. (2008). The pecking order of cross-border investment. Journal of International Economics, 74(1), 94
119. Daveri, F. (1995). Costs of entry and exit from financial markets and capital flows to developing countries. World Development, 23(8), 1375
1385. Egger, P., & Merlo, V. (2007). The impact of bilateral investment treaties on FDI dynamics. World Economy, 30(10), 1536
1549. Fratzscher, M., & Imbs, J. (2009). Risk sharing, finance, and institutions in international portfolios. Journal of Financial Economics, 94(3), 428
447. Gelos, G., & Wei, S. (2005). Transparency and international investor behavior. Journal of Finance, 60(6), 2987
3020. Guiso, L., Sapienza, P., & Zingales, L. (2009). Cultural biases in economic exchange? The Quarterly Journal of Economics, 124(3), 1095
1131. Hajivassiliou, V., & Ruud, P. (1994). Classical estimation methods for LDV models using simulation. In R. F. Engle & D. L. McFadden (Eds.), Handbook of econometrics (pp. 2383–2441). Ansterdam, Netherlands: Elsevier. Heckman, J. (1981a). Statistical models for discrete panel data. In C. Manski & D. McFadden (Eds.), Structural analysis of discrete data with econometric applications (pp. 114
178). Cambridge, MA: MIT Press. Heckman, J. (1981b). The incidental parameters problem and the problem of initial conditions in estimating a discrete time
discrete data stochastic process. In C. Manski & D. McFadden (Eds.), Structural analysis of discrete data with econometric applications (pp. 179
195). Cambridge, MA: MIT Press. Honore´, B. E., & Kyriazidou, E. (2000). Panel data discrete choice models with lagged dependent variables. Econometrica, 68, 839
874. Hsiao, C. (2003). Analysis of panel data. Cambridge: Cambridge University Press.

148

K. Drakos et al.

Hyslop, D. R. (1999). State dependence, serial correlation and heterogeneity in intertemporal labor force participation of married women. Econometrica, 67(6), 1255
1294. Lane, P. R., & Milesi-Ferretti, M. (2008). International investment patterns. The Review of Economics and Statistics, 90(3), 538
549. La Porta, R., Lo´pez de Silanes, F., Shleifer, A., & Vishny, R. (1998). Law and finance. Journal of Political Economy, 106, 1113–1155. Martin, P., & Rey, H. (2004). Financial super-markets: Size matters for asset trade. Journal of International Economics, 64(2), 335
361. Mundlack, Y. (1978). On the pooling of time series and cross-section data. Econometrica, 46(1), 69
85. Obstfeld, M., & Rogoff, K. (2000). The six major puzzles in international macroeconomics: Is there a common cause? In B. Bernanke & K. Rogoff (Eds.), NBER Macroeconomics Annual (pp. 339
412). Okawa, Y., & Van Wincoop, E. (2012). Gravity in international finance. Journal of International Economics, 87(2), 205
215. Papaioannou, E. (2009). What drives international financial flows? Politics, institutions and other determinants. Journal of Development Economics, 88(2), 269
281. Portes, R., Rey, H., & Oh, Y. (2001). Information and capital flows: The determinants of transactions in financial assets. European Economic Review, 45(4/6), 783
796. Portes, R., & Rey, H. (2005). The determinants of cross-border equity flows. Journal of International Economics, 65(2), 269
296. Rose, A., & Spiegel, M. (2004). A gravity model of sovereign lending: Trade, default and credit. IMF Staff Papers, 51, 64
74. Shleifer, A., & Wolfenzon, D. (2002). Investor protection and equity markets. Journal of Financial Economics, 66(1), 3
27. Stein, E., & Daude, C. (2007). Longitude matters: Time zones and the location of foreign direct investment. Journal of International Economics, 71(1), 96
112. Stewart, M. (2006). Maximum simulated likelihood estimation of random-effects dynamic probit models with autocorrelated errors. Stata Journal, 6(2), 256–272. Wei, S., & Shleifer, A. (2000). Local corruption and global capital flows. Brookings Papers in Economic Activity, 2, 303
354. Wooldridge, J. (2005). Simple solutions to the initial conditions problem in dynamic, nonlinear panel data models with unobserved heterogeneity. Journal of Applied Econometrics, 20(1), 39
54.

A Dynamic Gravity Model for Global Bilateral Investment Holdings

149

Appendix A: Countries List of host countries in sample: Angola, Albania, Argentina, Armenia, Antigua and Barbuda, Austria, Australia, Azerbaijan, Bangladesh, Barbados, Belarus, Belize, Belgium, Benin, Bhutan, Bolivia, Bosnia and Herzegovina, Botswana, Brazil, Bulgaria, Burkina Faso, Burundi, Brunei, Cambodia, Cameroon, Canada, Cape Verde, Chad, Chile, China, Central African Republic, Cote d’ Ivoire, Congo Dem. Rep. Congo, Rep. Colombia, Comoros, Costa Rica, Croatia, Czech Republic, Djibouti, Dominica, Denmark, Dominican Rep., Algeria, Ecuador, Egypt, El Salvador, Eritrea, Estonia, Ethiopia, Finland, France, FYROM, Gabon, Germany, Georgia, Ghana, Guinea, Gambia, Guinea-Bissau, Equatorial Guinea, Greece, Grenada, Guatemala, Guyana, Haiti, Honduras, Hong Kong, Hungary, Iceland, India, Indonesia, Iran, Ireland, Israel, Italy, Jamaica, Japan, Jordan, Kazakhstan, Kenya, Kiribati, Korea Rep., Kyrgyz Rep., Kuwait, Laos, Liberia, Libya, Lesotho, Lithuania, Latvia, Madagascar, Maldives, Mexico, Marshall Islands, Malawi, Malaysia, Mali, Mauritania, Micronesia, Moldova Mongolia, Morocco, Mozambique, Namibia, Netherlands, Nepal, New Zealand, Niger, Nigeria, Nicaragua, Norway, Oman, Pakistan, Papua New Guinea, Paraguay, Peru, Philippines, Poland, Portugal, Qatar, Romania, Russia, Rwanda, Samoa, Saudi Arabia, Senegal, Seychelles, Sierra Leone, Singapore, Slovakia, Slovenia, Solomon Islands, South Africa, Spain, Sri Lanka, St. Kitts and Nevis, St. Lucia, St. Vincent and the Grenadines, Sudan, Suriname, Swaziland, Sweden, Switzerland, Syria, Tanzania, Thailand, Tajikistan, Togo, Tonga, Trinidad and Tobago, Tunisia, Turkey, Turkmenistan, Uganda, Ukraine, United Arab Emirates, United Kingdom, United States of America, Uruguay, Uzbekistan, Venezuela, Vietnam, Yemen, Zambia, Zimbabwe. List of source countries in sample: Argentina, Australia, Belgium, Brazil, Bulgaria, Canada, Chile, Colombia, Costa Rica, Czech Republic, Denmark, Egypt, Estonia, Finland, France, Germany, Greece, Hong Kong, Hungary, Iceland, India, Indonesia, Ireland, Israel, Italy, Japan, Kazakhstan, Korea, Kuwait, Latvia, Malaysia, Mexico, Netherlands, New Zealand, Norway, Pakistan, Philippines, Poland, Portugal, Romania, Russia, Singapore, Slovakia, South Africa, Spain, Sweden, Switzerland, Thailand, Turkey, Ukraine, United Kingdom, United States of America, Uruguay, Venezuela. List of countries excluded: (Off shore and small financial centers): Aruba, Bahamas, Bahrain, Bermuda, Cayman Islands, Cyprus, Guernsey, Isle of Man, Jersey, Lebanon, Luxemburg, Macao, Malta, Mauritius, Netherlands Antilles, Panama.

150

K. Drakos et al.

(Data unavailability): Afghanistan, American Samoa, Andorra, Anguilla, British Indian Ocean Territory, Christmas Island, Cocos (Keeling) Islands, Cook Islands, Cuba, Falkland Islands, Faroe Islands, Fiji, French Guyana, French Polynesia, French Southern Territory, Gibraltar, Greenland, Guadeloupe, Guam, Iraq, Korea, Democratic Rep., Lichtenstein, Martinique, Mayotte, Monaco, Montserrat, Myanmar, Nauru, New Caledonia, Niue, Norfolk Island, Palau, Pitcairn, Puerto Rico, Reunion, San Marino, Sao Tome and Principe, Serbia and Montenegro, Somalia, St. Helena, St Pierre and Miquel, Taiwan, Timor Leste, Tokelau, Turks and Caicos islands, Tuvalu, United States Minor Outlying Islands, Vanuatu, Vatican City, Virgin Islands, Wallis and Fortuna Islands, West Bank, Western Sahara.

A Dynamic Gravity Model for Global Bilateral Investment Holdings

151

Appendix B: Description of control variables Variables used in baseline model Proxies for Informational Frictions • (Log-) distance between the capital cities of the host and source country (source CEPII): calculated using latitudes and longitudes of the geographic coordinates. • Common language (source CEPII): a dummy variable attaining the value of unity if the source and host countries have the same official language. Proxies for Institutional Quality • Political Risk (source: ICRG): an index ranging from 0 to 100, with higher values denoting lower political risk. It is calculated as the sum of the following components: (a) government stability, (b) socioeconomic conditions, (c) investment profile, (d) internal conflict, (e) external conflict, (f) corruption, (g) military in politics, (h) religion in politics, (i) law and order, (j) ethnic tensions, (k) democratic accountability, (l) bureaucracy quality. • Financial Risk (source: ICRG): an index ranging from 0 to 50 with higher values denoting lower financial risk. It is calculated as the sum of the following components: (a) foreign debt as a percentage of GDP, (b) foreign debt service as a percentage of exports, (c) current account as a percentage of exports, (d) net liquidity as months of import cover, (e) exchange rate stability. • Type of Legal Origin (source La Porta, Lo´pez de Silanes, Shleifer, & Vishny, 1998): a set of dummy variables identifying English, French, or German legal origin. Proxies for Financial Market Development • Domestic credit to private sector as a percentage of GDP (Source: World Bank Development Indicators). • Stocks traded turnover ratio (Source: World Bank Development Indicators): total value of shares traded during the period divided by the average stock market capitalization during the period.

152

K. Drakos et al.

Appendix C: Descriptive statistics of control variables Table C.1: Variable

Mean

Std. Dev.

Min.

Max.

Obs.

In baseline model Log (distance) Common language Investment treaty (inverse) Political risk (inverse) Financial risk English legal origin French legal origin German legal origin Scandinavian legal origin Domestic credit (as % of GDP) Stocks traded (turnover ratio)

8.720 0.110 0.271 67.930 36.797 0.326 0.528 0.116 0.030 46.109 53.325

0.801 0.313 0.444 12.621 5.996 0.468 0.499 0.320 0.171 44.488 63.845

4.088 0.000 0.000 34.291 11.500 0.000 0.000 0.000 0.000 0.682 0.000

9.892 1.000 1.000 96.083 50.000 1.000 1.000 1.000 1.000 319.721 497.380

62370 62370 62370 47250 47250 61236 61236 61236 61236 59568 33061

Does China’s International Competitiveness Fluctuate in Consistency with PPP Equilibrium? Nikolaos Giannellisa and Georgios P. Kouretasb a

Department of Economics, University of Crete, Rethymno, Greece, e-mail: [email protected] b Athens University of Economics and Business, Athens, Greece

Abstract Purpose
The aim of this study is to examine whether China’s exchange rate follows an equilibrium process and consequently to answer the question of whether or not China’s international competitiveness fluctuates in consistency with equilibrium. Design/methodology/approach
The theoretical background of the paper relies on the Purchasing Power Parity (PPP) hypothesis, while the econometric methodology is mainly based on a nonlinear two-regime Threshold Autoregressive (TAR) unit root test. Findings
The main finding is that China’s price competitiveness was not constantly following a disequilibrium process. The two-regime threshold model shows that PPP equilibrium was confirmed in periods of relatively high
compared to the estimated threshold
rate of real yuan appreciation. Moreover, it is implied that the fixed exchange rate regime cannot ensure external balance since it can neither establish equilibrium in the foreign exchange market, nor confirm that China’s international competitiveness adjustment follows an equilibrium process. Practical implications
The results do not imply that China acts as a currency manipulator. However, a main policy implication of the paper is that China should continue appreciating the yuan to establish external balance. International Symposia in Economic Theory and Econometrics, Vol. 23 G. P. Kouretas and A. P. Papadopoulos (Editors) Copyright r 2014 by Emerald Group Publishing Limited. All rights reserved ISSN: 1571-0386/doi:10.1108/S1571-038620140000023006

154

N. Giannellis and G. P. Kouretas

Originality/value
This paper is the first which accounts for a nonlinear two-regime process toward a threshold, which is defined to be the rate of change in China’s international competitiveness. Consequently, the paper draws attention to the role of China’s international competiveness in accepting the PPP hypothesis. Keywords: China, international competitiveness, PPP, threshold JEL Classifications: C22, C24, E31, F31

Introduction Nowadays, there is a growing interest from academics and policy makers in currency manipulation, which is usually considered as a practice of “currency war.” A currency is said to be manipulated if a country intervenes systematically in the foreign exchange market to keep the value of its currency low so that to boost its exports. However, to characterize a country as a currency manipulator is not a simple task. Not all interventions in foreign exchange markets constitute actions of currency manipulation. For example, a flexible exchange rate regime may not be consistent with monetary policy objectives. In such a case, the Central Bank has to intervene to prevent exchange rate fluctuation. Moreover, if a currency is overvalued, the intervention in foreign exchange markets to prevent its appreciation does not violate any international agreement.1 However, most of the countries which prevent the appreciation of their currency are already undervalued. Cline and Williamson (2010) argue that countries intervene to prevent the appreciation of their currency, but they are not eager to intervene to prevent the depreciation of an undervalued currency. One possible explanation may be that these countries do not have adequate exchange rate reserves. But, this explanation does not seem to be valid if we consider that countries with adequate reserves keep preventing the appreciation of undervalued currencies. Thus, it could be argued that countries which prevent the appreciation of undervalued currencies have a specific and clear target. According to this view, these countries hold technically the value of their currency low in

1 Bergsten (2010) states that currency manipulation violates: (1) the international monetary rules of the International Monetary Fund (IMF) articles of agreement, and (2) the global trading rules of the World Trade Organization (WTO) charter.

China’s International Competitiveness

155

order to increase their international competitiveness and increase their exports. This policy leads to large trade surpluses in these countries and to trade deficits in their trade partners.2 Obstfeld (2012) argues that such global imbalances are not innocent and may cause crises in the future. This is actually why this aggressive policy triggers the academic debate and the concern of governments on currency manipulation. Deficit countries, which face large current account deficits and high unemployment rates, lay the blame on the artificially low value of other currencies. An implicit risk, as a result of the “currency war,” is that deficit countries may impose restrictions on imports from currency manipulator countries. A well-known case of a country, which has been recently blamed by the United States and countries of the Euro area for currency manipulation, is China. China’s exchange rate policy is mainly driven by the value of the yuan (renminbi) against the U.S. dollar, while since 1995 the employed exchange rate regimes switch between a pegged regime and an appreciating crawling pegged. This exchange rate policy has been strongly criticized by U.S. politicians and economists. The main argument is that China fixes the value of the yuan to a desired level so as to increase its exports. By contrast, McKinnon (2006) and Corden (2009) argue that the aim of fixing the exchange rate was not to increase China’s exports. Instead, they argue the main objective was to maintain internal stability. Furthermore, McKinnon and Schnabl (2009) and McKinnon, Lee, and Wang (2009) present evidence that China had to fix its exchange rate in order to regain its monetary policy control. In line with the recent exchange rate developments, there is rich evidence in the literature that the yuan was undervalued (see e.g., Benassy-Quere, Lahreche-Revil, & Mignov, 2011; Coudert & Couharde, 2007; Funke & Rahn, 2005; Goldstein & Lardy, 2006; Guo, 2010). However, there is a number of empirical works which provide somewhat different results. Cheung, Chinn, and Fujii (2007, 2009) argue that the Chinese currency appears to be undervalued but the undervaluation rate is not statistically significant. Moreover, Gregory and Shelley (2011) provide evidence that the market of the Chinese yuan exchange rate against the U.S. dollar, was deviating from the long-run equilibrium value and thus incompatible with macroeconomic fundamentals but evidence in favor of equilibrium when the effective exchange rate of the yuan was under consideration. Similarly, Wang, Hui, and Soofi (2007) argue that the yuan real effective exchange rate was not considerably undervalued.

2 Subramanian (2010) notes that this kind of policy is considered as highly protectionist trade policy since it is a combination of an import tariff and an export subsidy.

156

N. Giannellis and G. P. Kouretas

The present paper provides further investigation on the issue whether the Chinese yuan exchange rate vis-a`-vis major foreign currencies,3 follows an equilibrium process toward the Purchasing Power Parity (PPP) hypothesis. Namely, independently of the economic policy objectives (i.e., current account surplus or internal stability), we investigate whether the adopted exchange rate policy can lead to the achievement of equilibrium in the foreign exchange market. Furthermore, since the latter condition can be considered as a measure of an economy’s price competitiveness, we seek to find whether China’s price competitiveness fluctuates in consistency with PPP equilibrium. In respect with the current debate on currency manipulation, the evidence in favor of PPP hypothesis
when the yuan depreciates in real terms
can provide information that China’s international competitiveness improvement is an equilibrium phenomenon. In contrast, if the real exchange rate is not mean-reverting, thereby implying that PPP is invalid, price competitiveness adjustment is not consistent with equilibrium.4 The vast majority of the empirical studies, which have examined the aforementioned issue, assumed that real exchange rates follow a linear process. However, not surprisingly, interventions in foreign exchange markets may cause nonlinearities in real exchange rate behavior. In the presence of nonlinearities, linear models are biased against the evidence of PPP equilibrium (Taylor, Peel, & Sarno, 2001). Previous studies dealt with the Chinese exchange rate behavior have not underlined the fact that the evidence of PPP may depend on the rate of change of China’s international competitiveness.5 To fill this gap in the literature, we employ a nonlinear two-regime

3 As foreign currencies we have used the U.S. dollar, the euro, and the Japanese yen. The selection of the currencies was based on the fact that United States, Eurozone, and Japan are China’s major trade partners and their currencies have the highest weight on the yuan effective exchange rate. Specifically, the U.S. dollar is weighted by 21%, while the weights on the euro and the Japanese yen are 18.4% and 16.8%, respectively. 4 It should be clear-cut that the evidence in favor of PPP does not necessarily imply absence of short-run foreign exchange (forex) market interventions. Instead, it may indicate that interventions are not persistent and systematic. The systematic use of interventions may affect the exchange rate not only in the short run, but also in the medium run and the long run. Hence, the evidence of long-run equilibrium allows us to consider interventions as temporary and nonsystematic. On the other hand, PPP rejections are not exclusively charged to forex market interventions. However, a focus on the recent history of China’s exchange rate policy indicates that forex market interventions may be the primary source of a possible rejection of the PPP hypothesis. 5 To be precise, there are an adequate number of studies which have not ignored the presence of nonlinearities in the yuan real exchange rate. However, they have focused on one only source of nonlinearity, which is the transaction cost (see among others, Ahmad & Rashid, 2008; Fan & Wei, 2006).

China’s International Competitiveness

157

Threshold Autoregressive (TAR) unit root test, originally presented by Caner and Hansen (2001). A significant advantage of this test is that it allows us to discriminate between pure and partial nonstationarity. Pure nonstationarity exists when the real exchange rate is nonstationary across both regimes. Partial nonstationarity holds when the real exchange rate behaves like a unit root process in one regime and like a stationary process in the other regime. In other words, PPP may be valid in one regime, but not in the other. To the best of our knowledge, the present paper is the first which accounts for a nonlinear two-regime process toward a threshold, which is defined to be the rate of change in China’s international competitiveness.6 This paper contributes to the PPP literature by highlighting the role of foreign exchange market interventions as a source of nonlinearity, which we argue that applies better to China’s case. Specifically, the present paper departs from previous papers, in which nonlinearity is a result of transaction costs, by considering a different type of threshold variable, which is the rate of change in China’s international competitiveness. Intuitively, this type of threshold is not related to transactions costs, but instead to foreign exchange market intervention. An accompanying contribution is that the paper draws attention to the role of China’s international competiveness in accepting the PPP hypothesis. To preview our results, we have found evidence of nonlinearity in two out of the three real exchange rates under consideration. Under the presence of nonlinear behavior, several interesting implications stem from this analysis. First, we found that PPP equilibrium was confirmed only in periods of relatively high (compared to the estimated threshold) rate of real yuan appreciation (i.e., high competitiveness loss). In contrast, PPP equilibrium could not be established in periods of low competitiveness loss. Second, these periods were directly related to China’s exchange rate policy. Specifically, low loss in China’s price competitiveness was observed at the time of employing the fixed exchange rate regime, while higher loss occurred during the adoption of the appreciating crawl regime. Third, the fixed exchange rate regime can neither establish equilibrium in the foreign exchange market, nor confirm that China’s international competitiveness adjustment follows an equilibrium process. Fourth, Chinese monetary authorities should take into

6 It is important to note that this is not the first time that the test is employed in PPP literature. Although this test has not been previously employed for the real exchange rates under consideration, researchers have already applied this test to other exchange rates (see e.g., Alba & Park, 2005; Ho, 2005). However, the emphasis given to international competitiveness as a threshold variable is shown for the first time here in this study.

158

N. Giannellis and G. P. Kouretas

account the above limitation in forming China’s exchange rate policy. Finally, an interesting finding is that the yuan exchange rate was not continuously deviating away from PPP equilibrium. Thus, there is no strong evidence that China follows a coherent manipulation rule, under which the yuan is constantly kept at an artificially low level. The structure of the paper is organized as follows. The following section presents an overview of China’s exchange rate policy, its objectives, and its impact on global imbalances. The subsequent sections illustrate the econometric methodology and the dataset, respectively. Next, the empirical results are presented and finally, a concluding section summarizes.

China’s Exchange Rate Policy and Global Imbalances A prevalent view in several countries, principally in the United States, is that current account imbalances, that is, the Chinese current account surplus and the U.S. current account deficit, are attributed to the Chinese exchange rate policy. Regardless of the source of the imbalance, Obstfeld (2012) argues that large and persistent current account imbalances may be an indicator of a crisis in the future. He argues that current account surpluses can cause the decline of the world interest rate, which in turn may lead to higher consumption and investment in deficit countries and then to higher global imbalances. Hence, the existence of global imbalances has been set at the center of economic policy debate. Policy makers should be aware of what provokes these imbalances and how they can be eliminated. In relation with our main empirical analysis, a critical question to be answered is whether the Chinese current account surplus is attributed to the Chinese exchange rate policy. China’s exchange rate policy has been mainly focused on the value of the yuan against the U.S. dollar. In Figure 1, we present the yuan nominal exchange rate vis-a`-vis the U.S. dollar, and it is shown that the exchange rate policy has changed several times during the period under examination and can be decomposed into five stages. At subperiod 1, which lasted until the end of 1995, China introduced multiple exchange rates and controls on exchange rate transactions which made the currency inconvertible before 1994. The exchange rate unification was accompanied with the rapid depreciation of the yuan at the end of 1993, while controls on current account transactions were abolished at the same time. At subperiod 2, from December 1995 to July 2005, the yuan was pegged to the U.S. dollar. This policy has been strongly criticized by a large part of economists and politicians. China has been accused for keeping its currency undervalued in order to boost its exports and run high current account

159

China’s International Competitiveness 10 9 8 7

Stage 1

Stage 2

Stage 3

Stage 5

5

Stage 4

6

4 1993 1995 1997 1999 2001 2003 2005 2007 2009 2011

Figure 1:

Nominal Yuan Exchange Rate Against the U.S. Dollar.

U.S. dollars (Millions)

400000 300000 200000 100000 0 –100000 1993 1995 1997 1999 2001 2003 2005 2007 2009 Goods and Services

Figure 2:

Goods only

Chinese Current Account.

surpluses (see e.g., Goldstein & Lardy, 2006). The pressure for letting the yuan to appreciate has been significantly strengthened since the high rise of the Chinese balance of payment surplus in 2004. McKinnon and Schnabl (2009, 2012) argue that the dramatic increase of China’s balance of payments surplus was a result of the unexpected net saving surplus and the large inflows of foreign direct investment. The former can be seen as an increased current account surplus (Figure 2), while the latter represents a decline in the capital account deficit. Similarly, Anderson (2008) argues that China’s current account surplus rose as a result of the high increase of the net saving surplus and the decline in the growth of imports. Corden (2009) adds one

160

N. Giannellis and G. P. Kouretas

more factor, which explains the increase of the Chinese current account surplus. This is considered to be the steady productivity improvements in export and import-competing industries, which resulted in an increase in the growth of exports and decreased the growth of imports (Figure 3). Moreover, Corden (2009) points out that the aim of fixing the exchange rate was irrelevant to any current account objective. He states that the fixed exchange rate regime was mainly chosen to maintain internal balance. Similarly, McKinnon and Schnabl (2009) argue that the aim of fixing the exchange rate was to anchor the domestic price level and stabilize the rate of growth, while McKinnon (2006) shows that this policy helped end the “roller coaster” ride in China’s inflation rate and growth. In addition, McKinnon and Schnabl (2012) point out that this policy was a stabilizing instrument not only for China, but also for East Asian economies and the world economy. In subperiod 3, China announced on July 21, 2005 the appreciation of the yuan against the U.S. dollar by 2.1% and a number of exchange rate policy reforms. Among them was the abandonment of the fixed regime and a move to a predictable appreciating crawl of the yuan against the U.S. dollar. Furthermore, the yuan were expected to be managed with reference to a basket of currencies instead of being fixed to the U.S. dollar (Goldstein & Lardy, 2006). During subperiod 3 (i.e., from July 2005 to August 2008), the yuan appreciated against the U.S. dollar by 21%. However, this policy was interrupted and replaced again by a pegged regime against the U.S. dollar until June 2010 (subperiod 4).

1600000

U.S. dollars

1200000

800000

400000

0 1993 1995 1997 1999 2001 2003 2005 2007 2009 Exports

Figure 3:

Imports

China’s Trade Balance.

China’s International Competitiveness

161

An interesting puzzling issue arises here. What can explain the reestablishment of the pegged regime, in subperiod 4, given that China’s exports and the current account surplus were increasing? If we assume that China utilized its exchange rate as a device to boost its exports, it would have reasons to repeg its currency if exports were cutting down during the appreciation era (subperiod 3). However, Figure 3 illustrates that the growth rate of exports increased at subperiod 3, while Figure 2 provides a similar implication for the current account surplus. A possible explanation is provided by McKinnon and Schnabl (2009) and McKinnon et al. (2009). According to these works, China had to reinstall the fixed exchange regime to regain its control on the conduct of the monetary policy. Specifically, after July 2005, the continuous depreciation (appreciation) of the U.S. dollar (yuan) prevented private capital outflows from financing China’ trade surplus. Domestic investors and financial institutions preferred Chinese assets rather than U.S. dollar denominated assets because the U.S. dollar was expected to depreciate. The degree of inadequacy of capital outflow was even higher because of the status of China as an immature creditor country.7 Namely, as a result of the currency mismatch that was held in China, domestic private investors and financial firms faced an enormous currency risk preventing them from buying U.S. dollar denominated assets. The unwillingness of the domestic private sector to invest in foreign assets (U.S. dollar assets) led to high balance of payments surpluses in China. To avoid excess appreciation of the yuan, Chinese monetary authorities intervened in the forex market by selling yuan and buying U.S. dollars. The massive purchase of U.S. dollars increased dramatically China’s official exchange reserves (Figure 4), which in turn caused an unwanted rise in domestic monetary base. Although China’s Central Bank put an effort to sterilize the inflationary effects of foreign reserves accumulation, China lost the control of the monetary policy and the domestic inflation increased.8 Furthermore, Mckinnon et al. (2009) argue that China had to fix its currency to protect it from the U.S. dollar appreciation after the U.S. credit

7

A country is said to be immature creditor country if it cannot lend abroad in its own currency. These countries continually accumulate claims on foreigners in an internationally acceptable currency. 8 China faced limitations and difficulties in sterilizing the monetary base by selling Central Bank bonds. The bonds sell caused monetary tightening which in turn led interest rates upward. The higher domestic interest rates combined with the low U.S. interest rates caused higher capital inflow, which caused even higher capital account surpluses. Thus, Chinese Central Bank had to accumulate even higher amount of foreign reserves.

162

N. Giannellis and G. P. Kouretas

Special Drawing Rights (Millions)

2000000

1600000

1200000

800000

400000

0 1993 1995 1997 1999 2001 2003 2005 2007 2009

Figure 4: China’s Official Foreign Exchange Reserves. crisis in 2008.9 They state that the appreciation of the U.S. dollar moved the value of the Chinese yuan upward with it. To prevent higher appreciation of the yuan against the other currencies of the world, China abandoned the appreciating crawl of the yuan and refixed it to the U.S. dollar. During subperiod 4 (i.e., fixed regime), China regained the control of the monetary policy. Finally, under economic and political pressure, China entered, in June 2010, a new period of appreciating crawl against the U.S. dollar (stage 5). Summing up, the overview of the Chinese exchange rate policy does not provide indications that China used the exchange rate as a device to undervalue its currency so that to boost its exports. Neither can we support the view that China’s current account surplus was an exchange rate policy objective. Instead, it can be argued that China fixed the exchange rate to regain the monetary policy control and achieve internal stability. An additional reason for keeping the exchange rate stable was the status of China as an immature creditor country. Namely, China pegged the yuan against the U.S. dollar to protect domestic holders of U.S. dollar assets from the currency risk. However, no matter what the objective of the exchange rate

9 The U.S. dollar appreciation was a result of the credit crisis of 2008 and the dollar carry trade before 2008. Due to low U.S. interest rates, investors were borrowing in U.S. dollars and investing in other economies with higher interest rates. However, during the financial crisis, U.S. banks claimed the repayment of the loans in order to manage their liquidity problem. So, investors had to sell foreign currencies and buy U.S. dollars to repay their loans.

China’s International Competitiveness

163

policy, it is still unclear whether this policy was consistent with equilibrium in the foreign exchange market.10 In the following section we provide further insights into this issue.

Econometric Methodology The evidence in favor of the PPP hypothesis implies that the real exchange rate follows a mean-reverting process. In other words, the real exchange rate should follow a stationary process. Thus, to test for the validity of the PPP hypothesis we begin by analyzing the stochastic properties of the real exchange rate. As a preliminary empirical procedure, we employ a battery of linear unit root tests, such as the Elliot, Rothenberg, and Stock (1996) and Elliott (1999) GLS augmented Dickey
Fuller and the Ng and Perron (2001) GLS versions of the modified Phillips-Perron (1988) unit root tests. For robustness we also apply the Kwiatkowski, Phillips, Schmidt, and Shin (1992) (KPSS) stationarity test.11 Given the likelihood for the presence of possible nonlinear characteristics of the real exchange rate, we employ a nonlinear two-regime unit root test, originally presented by Caner and Hansen (2001), which is described below.

Two-Regime TAR Model The two-regime unit root test, which tests the hypothesis that the real exchange rate at levels contains a unit root, is based on the following TAR model: Δqt = θ01 xt − 1 ℓðZt − 1 < λÞ þ θ02 xt − 1 ℓðZt − 1 ≥ λÞ þ et ;

ð1Þ

where, t = 1,…, T, q is the real exchange rate xt − 1 = ðqt − 1 rt0 Δqt − 1 …Δqt − k Þ0 , ℓð⋅Þ is the indicator function, et is an independent and identically distributed error term, rt is a vector of deterministic components (intercept and linear time trend), Zt − 1 is the threshold variable, and λ is the threshold parameter. The latter is treated as unknown and it is assumed to take values in the interval λ ∈ Λ = ½λ1 ; λ2  where PðZt − 1 ≤ λ1 Þ > 0 and PðZt − 1 ≤ λ2 Þ < 1. 10

McKinnon et al. (2009) show that the forward exchange rate was misaligned, during the yuan appreciation crawl period, as a result of the one-way bet appreciation of the yuan and the extremely low U.S. interest rate. 11 As these tests are very well-known and widely used tests, the reader is referred to the original papers cited above.

164

N. Giannellis and G. P. Kouretas

A critical point of analysis is the endogenous selection of the threshold variable, which should be predetermined, strictly stationary, and ergodic with a continuous distribution function. Following Caner and Hansen (2001), we choose the threshold variable of the form Zt − 1 = qt − 1 − qt − d − 1 , for the delay parameter d ≥ 1 because it combines theoretical as well as technical advantages. Specifically, this type of the threshold variable ensures stationarity for itself under the assumption that the inflation rate differential follows a unit root or a stationary process. Moreover, the theoretical advantage stands for the ability to split our sample to two regimes according to the dynamic behavior of the real exchange rate, namely the rate of change in China’s international competitiveness. The vectors θ1 and θ2 are as follows 0 1 0 1 ρ1 ρ2 θ 1 = @ β 1 A; θ 2 = @ β 2 A; α1 α2 where ρ1 and ρ2 are the slope coefficients on qt − 1 in the two regimes, β1 and β2 are the slopes on the deterministic components in the two regimes, and α1, α2 are the slope coefficients on ðΔqt − 1 ; …; Δqt − k Þ in the two regimes as well. For λ ∈ Λ, the above TAR model is estimated by ordinary least squares (OLS).12 For fixed λ, Eq. (1) is written as Δqt = θ^ 1 ðλÞ0 xt − 1 ℓðZt − 1 < λÞ þ θ^ 2 ðλÞ0 xt − 1 ℓðZt − 1 ≥ λÞ þ e^t ðλÞ; ð2Þ where the OLS estimate of the residual variance is given by P σ^ 2 ðλÞ = T − 1 Tt= 1 e^t ðλÞ2 . The OLS estimator of λ is that which minimizes the ^ the estiresidual variance, that is, λ^ = arg min σ^ 2 ðλÞ. For a given value of λ, λ∈Λ

mated TAR model is as follows 0 ^ þ θ^ 0 xt − 1 ℓðZt − 1 ≥ λÞ ^ þ e^t Δqt = θ^ 1 xt − 1 ℓðZt − 1 < λÞ 2

^ θ^ 2 = θ^ 2 ðλÞ, ^ and residual variance σ^ 2 = T − 1 with θ^ 1 = θ^ 1 ðλÞ,

ð3Þ PT

2 t = 1 e^t .

Testing for the Linearity Hypothesis The linearity hypothesis (i.e., no threshold effect) is described by the following null hypothesis: H0 : θ 1 = θ 2 ;

ð4Þ

12 Hansen (1996, 1997) has shown that, under the assumption that the error term is normally and identically distributed with zero mean and variance σ2, OLS is equivalent to maximum likelihood estimation (MLE).

China’s International Competitiveness

165

which is tested against the alternative that the estimated parameters in θ1 and θ2 are different across regimes. The null hypothesis can be tested using a standard Wald statistic,  2  σ^ 0 WT = T 2 − 1 ; σ^

ð5Þ

where σ^ 20 is the OLS estimator of the residual variance of the linear model and σ^ 2 is the OLS estimator of the residual variance of the TAR model, as it is presented in Eq. (2). Then, given the value of λ (i.e., fixed λ), the Wald statistic of hypothesis (4) is as follows, 

 σ^ 20 WT ðλÞ = T 2 − 1 : σ^ ðλÞ

ð6Þ

Since expression (6) is a decreasing function of σ^ 2 ðλÞ, we get that ^ = supλ ∈ Λ WT ðλÞ. Hence, the Wald statistic for the null hypothWT = WT ðλÞ esis (4) is also called as the “sup-Wald” statistic. The Wald test, as described in Equations (5) and (6), has a nonstandard asymptotic distribution due to the presence of nuisance parameters under the null hypothesis (Davies, 1977).13 In addition, Caner and Hansen (2001) argue that the distribution may be nonstandard due to the assumption of a unit root process.14 For this reason, Caner and Hansen (2001) introduce two bootstrap approximations to the asymptotic distribution of WT, one based on the unrestricted estimates (unrestricted bootstrap procedure) and the other based on the restriction of a unit root (restricted bootstrap procedure).15 The former is appropriate only when the series is stationary. If the series contains a unit root, the correct asymptotic distribution and robust p-values are achieved by the restricted bootstrap procedure. Although, it seems that both bootstrap procedures have near identical size, Caner and Hansen (2001) suggest conducting both bootstrap procedures and selecting the larger p-value if the true order of integration of the series is unknown.

The nuisance parameter is the threshold parameter λ, which is not identified under the null hypothesis of no threshold effect. 14 In contrast to previous TAR models that have assumed that the data are stationary, ergodic, and have no unit roots, Caner and Hansen (2001) introduce the TAR model with an autoregressive unit root. 15 For a technical and detailed description of both bootstrap methods, see Caner and Hansen (2001, pp. 1563
1565). 13

166

N. Giannellis and G. P. Kouretas

Testing for the Unit Root Hypothesis The null hypothesis of a unit root is described by the following expression: H0 : ρ1 = ρ2 = 0;

ð7Þ

which means that the real exchange rate is integrated of order one, that is, I(1). On the other hand, the series is said to be stationary autoregressive if ρ1 < 0; ρ2 < 0 and ð1 þ ρ1 Þð1 þ ρ2 Þ < 1. Thus, the alternative to the null hypothesis is as follows: H 1 : ρ1 < 0

and

ρ2 < 0:

ð8Þ

While the null hypothesis states that the real exchange rate has unit roots in both regimes, the alternative hypothesis states that it is stationary in both regimes. However, it is possible for a series to behave like a stationary process in one regime and like a random walk process in the other regime. In other words, the real exchange rate may have a unit root in one regime and may be stationary in the other regime. This partial nonstationarity is expressed by the alternative H2, 8 < ρ1 < 0; and ρ2 = 0 or : ð9Þ H2 : : ρ1 = 0; and ρ2 < 0 Since both alternative hypotheses are one-sided the null is tested against the alternative (ρ1 < 0 and ρ2 < 0) using the following one-sided Wald test statistic R1T = t12 ℓf^ρ1 < 0g þ t22 ℓf^ρ2 < 0g;

ð10Þ

where t1 and t2 are the t-ratios for OLS estimates ρ^ 1 and ρ^ 2 from TAR model (6).16 Caner and Hansen (2001) suggest examining the individual t statistics (t1 and t2) to discriminate between the two alternative hypotheses, that is, stationarity (H1) and partial nonstationarity (H2). If only one of the t-statistics is statistically significant, we should accept the alternative H2. Finally, robust p-values are computed using a bootstrap distribution.17

16 The two-sided Wald test statistic for testing the null against the alternative (ρ1 ≠ 0 and ρ2 ≠ 0), which is given by R2T = t12 þ t22 , is misleading and inappropriate. Moreover, Caner and Hansen (2001) have shown that the one-sided Wald test R1T has more power than the two-sided Wald test R2T. 17 Caner and Hansen (2001) construct two bootstrap distributions, one that imposes an identified threshold effect (identified threshold bootstrap) and another that imposes an unidentified threshold effect (unidentified threshold bootstrap). Based on a Monte Carlo analysis they suggest calculating p-values using the unidentified threshold bootstrap. For a detailed description of both bootstrap procedures, see Caner and Hansen (2001, p. 1573).

China’s International Competitiveness

167

Data and Preliminary Empirical Results The dataset consists of monthly observations from 1993:01 to 2011:08 on nominal Chinese yuan exchange rates against the U.S. dollar, the euro, and the Japanese yen as well as national Consumer Price Indices (CPI) of China, the United States, Eurozone, and Japan.18 All data were retrieved from the International Financial Statistics of the IMF database. Real (CPI-based) exchange rates have been calculated based on the following formula: qt = st þ pt − pt ;

ð11Þ

where st denotes the logarithm of the nominal yuan exchange rate against the foreign currency, pt denotes the logarithm of the Chinese CPI, and pt is the logarithm of the foreign country’s CPI. Eq. (11) is an identity which describes the absolute version of the PPP hypothesis. Hence, the real exchange rate (qt ) measures the deviation of the nominal exchange rate from PPP equilibrium. Moreover, the structure of Eq. (11) implies that an increase in the real exchange rate stands for depreciation of the domestic currency (i.e., yuan) in real terms and increase in domestic (i.e., Chinese) competitiveness in international trade.

Evidence from Linear Unit Root Tests Since the real exchange rate
given in Eq. (11)
measures the deviation of the nominal exchange rate from PPP equilibrium, our concern is focused on the stationary nature of the real exchange rate. The evidence of nonstationarity implies that deviations from PPP are expected to be persistent. If we are unable to reject the hypothesis of stationary, the real exchange rate is mean-reverting and the nominal exchange rate is expected to be driven to PPP equilibrium. To this end, we employ the unit root and stationarity tests given in the econometrics section on the bilateral real exchange rates of yuan under consideration.

18

The data span is subject to data availability. Namely, the estimated period runs from 1993:01 to 2011:08 for the yuan exchange rates against the U.S. dollar and the Japanese yen, while the estimated period is restricted to 1999:01
2011:08 for the yuan exchange rate against the euro.

168

N. Giannellis and G. P. Kouretas

Table 1: Linear Unit Root Tests Yuan per U.S. dollar (RER)

KPSS test Exogenous term Constant & trend Bandwidth 11 LM-statistic 0.311* DF
GLS test Exogenous term Constant Lags 1 t-statistic −1.452 Ng
Perron test with constant Lags 1 MZa −6.433 MZt −1.471 Ng
Perron test with constant and trend Lags 1 MZa −8.884 MZt −1.965

Yuan per euro (RER)

Yuan per Japanese yen (RER)

Constant & trend 10 0.268*

Constant 11 1.248*

Constant 1 −1.604

Constant & trend 1 −2.567

1 −5.149 −1.595

1 −5.233 −1.533

1 −5.698 −1.641

1 −13.155 −2.561

Notes: (1) RER refers to real exchange rate. (2) KPSS test is the Kwiatkowski et al. (1992) unit root test. Statistics are computed based on Newey and West (1994) robust kernel estimator of the variance, while the Bartlett kernel estimator is constructed via an automatic datadependent bandwidth selection. Asymptotic critical values are taken from Kwiatkowski et al. (1992, Table 1, p. 166). (3) DF
GLS test stands for the Elliot et al. (1996) GLS augmented Dickey
Fuller test. The lag length is automatically selected by the Schwarz criterion, while critical values are taken from Elliot et al. (1996, Table 1, p. 825). (4) MZa and MZt test statistics are the modified versions of the Phillips and Perron (1988) Za and Zt test statistics. Test statistics are calculated by Generalized Least Squares (GLS) detrended data methodology, while the lag length is selected by the Schwarz criterion. Asymptotic critical values are taken from Ng and Perron (2001, Table 1). (5) (*) implies rejection of the null hypothesis at the 5% level of significance.

The overall results given in Table 1 imply strong evidence against the PPP hypothesis under the assumption of linearity.19 These tests unanimously reveal that all real exchange rates contain a unit root, thereby implying that

19

To determine whether a determinist trend has to be included in the KPSS and GLS
ADF unit root tests, we first estimate the unrestricted model (constant and trend) and only if the trend is found to be statistically insignificant we exclude it and estimate the restricted model (only constant). In a similar way, we test the exclusion of the constant term. However, the same testing procedure is not applicable to the Ng
Perron test. In this case, to ensure robustness, we estimate both models (i.e., restricted and unrestricted).

China’s International Competitiveness

169

deviations from PPP are persistent and that China’s international competitiveness does not fluctuate in consistency with PPP equilibrium.20

Testing the Linearity Hypothesis Given the negative evidence in favor of stationarity of the three bilateral exchange rates under the ad hoc assumption of linearity we further investigate the stochastic properties of the three alternative exchange rates. Thus, if real exchange rates exhibit nonlinear behavior, standard linear unit root tests are biased against rejecting nonstationarity. Moreover, even if nonstationarity is rejected, the estimated autoregressive parameters are biased upward, thereby implying slower mean reversion than the actual one (see among others, Giannellis & Papadopoulos, 2010; Sarno, Taylor, & Chowdhury, 2004; Taylor et al., 2001). Therefore, we test the null hypothesis of linearity against the alternative of a nonlinear feature in real exchange rates. Specifically, we test the hypothesis that real exchange rates are not characterized by a threshold effect. If the null hypothesis cannot be rejected, then a series is linear and the above results seem to be robust. By contrast, if the null hypothesis is rejected, then the respective bilateral real exchange rate is characterized by a two-regime threshold process, which implies that this variable may behave non-monotonically across the two regimes. Within this framework, we test the hypothesis of no threshold effect along the lines of the two-regime TAR model. This test is undertaken by computing a sup-Wald test statistic (WT) of the form of Equation (4) and the relevant bootstrap p-values for the threshold variable (Zt − 1).21 In order to identify the threshold variable, we let the delay parameter (d) be endogenously determined given that the minimum delay parameter is equal to one and the maximum delay order is set equal to 12. The OLS estimate of d is the value that minimizes the residual variance. As the WT statistic is a monotonic function of the residual variance, equivalently, the selected value of d maximizes WT. The OLS estimates of d and λ along with the sup-Wald

20 The GLS
ADF and the Ng
Perron unit root test results show that the null hypothesis that the real exchange rate contains a unit root cannot be rejected. Moreover, KPSS test results illustrate that the null hypothesis that the real exchange rate is stationary is rejected. 21 Bootstrap p-values are calculated on the basis of both the unrestricted and restricted bootstrap procedures and by conducting 10,000 replications.

170

N. Giannellis and G. P. Kouretas

test statistics and the corresponding p-values are shown in the upper part of Table 2.22 Based on the reported estimates we conclude that we were unable to reject the linearity hypothesis for the case of the yuan
euro real exchange rate but we were able to reject it for the other two real exchange rates. In particular, for the real exchange rate of the yuan against the euro, the sup-Wald test statistic is estimated 14.7 with bootstrap p-value 0.298. Therefore, we argue that this real exchange rate series follow a linear process implying that the results of the standard unit root tests are still valid for this real exchange rate. This implies that the evidence in favor or against PPP is not regime dependent. Thus, given the results from linear Table 2: Nonlinear TAR Unit Root Test

TAR specification Exogenous term Delay parameter (d) Threshold parameter (λ) Linearity test Wald test statistic Bootstrap p-value ρ coefficient Regime 1 Regime 2 Unit root test R1T test statistic Bootstrap p-value t1 test statistic Bootstrap p-value t2 test statistic Bootstrap p-value

Yuan per U.S. dollar (RER)

Yuan per euro (RER)

Yuan per Japanese yen (RER)

Constant 10 −0.057

Constant 11 −0.116

Constant 10 −0.056

14.7 0.298

33.1** 0.058

212.0* 0.00 −0.046 −0.006 8.27 0.17 2.76** 0.07 0.79 0.63

NA NA

−0.042 0.017

NA

10.0 0.13 3.16* 0.04 0.11 0.77

NA NA

Notes: (1) RER refers to real exchange rate. (2) Bootstrap p-value stands for the p-value based on the bootstrap distribution. (3) ρ is the estimated autoregressive parameter of the nonlinear TAR model. (4) R1T stands for the one-sided unit root test in both regimes. (5) t1 stand for the unit root test in Regime 1. (6) t2 stands for the unit root test in Regime 2. (7) (*) (**) implies rejection of the null hypothesis at 5% (10%) level of significance, respectively. (8) NA stands for non-applicable.

22

The inclusion of deterministic components in the TAR model has been determined following the same testing procedure as in the KPSS and GLS
ADF models (see footnote 19).

China’s International Competitiveness

171

unit root tests, there is evidence that the corresponding nominal exchange rate is not consistent with PPP equilibrium. The evidence of linear exchange rate behavior reveals that the rejection of the PPP hypothesis might not be attributed to China’s exchange rate policy.23 However, why this exchange rate is permanently away from PPP equilibrium? One possible reason is that relative prices cannot alone determine the value of the nominal exchange rate. Indeed, by plotting the first difference of the nominal exchange rate with the first difference of relative prices (see Figure 7), we observe that the exchange rate is much more volatile, implying that relative prices may not be the exclusive determinant of the exchange rate. In line with the above, MacDonald (2000) has argued that an exchange rate may be away from its equilibrium value due to nonzero interest rate differentials. A second possible explanation is related to the difficulties in accepting the PPP hypothesis arising from the fact that Eurozone is not a single country, but instead a monetary union of 17 independent European countries. These difficulties come up from the absence of full national markets integration since price convergence among EMU members has been slow. Fan and Wei (2006) have argued, for the case of China, that although domestic authorities have removed barriers in international trade, the achievement of international market integration depends significantly on the existence of intranational market integration. Consequently, the limited market integration among EMU members may have resulted in rejection of the PPP hypothesis for the yuan exchange rate against euro. Turning our attention to the other two real exchange rates we note that the yuan real exchange rate against the U.S. dollar is found to be nonlinear at 5% level of significance, while the yuan real exchange rate against the Japanese yen is found to be nonlinear at 10% level of significance. As a consequence, we estimate a two-regime TAR model for these two nonlinear real exchange rates.

Estimated TAR Unit Root Tests The results of the two-regime TAR unit root test are shown in Table 2, while the corresponding regime classification of the series is shown in Figures 5 and 6. The specification of the corresponding TAR model is

23

Unlike the other two real exchange rates, in which the exchange rate regime does matter, the real yuan exchange rate against euro exhibits nonstationarity across both exchange rate regimes implemented by Chinese authorities, that is, the fixed exchange rate regime and the appreciating crawl one.

172

N. Giannellis and G. P. Kouretas 2.3 2.2 2.1 2.0 1.9 1.8 1.7 1993 1995 1997 1999 2001 2003 2005 2007 2009 2011 Regime 1

Regime 2

Figure 5: Regime Classification of the Real Yuan Exchange Rate Against the U.S. Dollar. –2.0

–2.2

–2.4

–2.6

–2.8

–3.0 1993 1995 1997 1999 2001 2003 2005 2007 2009 2011 Regime 1

Regime 2

Figure 6: Regime Classification of the Real Yuan Exchange Rate Against the Japanese Yen. shown in the upper part of Table 2, while unit root test results are shown in the lower part of the same table.24

24

Following Andrews (1993), we have assumed 15% minimum percentage of observations per regime.

China’s International Competitiveness

173

Real Exchange Rate of Yuan Against the U.S. Dollar For the case of the real exchange rate of yuan against the U.S. dollar, it is shown in Figure 5 that apart from the period 1997
2005, the exchange rate exhibits a general decreasing trend. The decline of the real exchange rate is equivalent to the appreciation of the yuan in real terms, and thus implies loss of international competitiveness of the Chinese economy. By contrast, the rising trend of the real exchange rate, which coincides in time with the fixed exchange rate regime of subperiod 2, implies that China gains in terms of price competitiveness. In line with the decreasing path of the real exchange rate, the estimated threshold parameter is found to be negative. Specifically, for d = 10 and λ = −0.057, the regime classification is described as follows: The first regime occurs when the real exchange rate decreases by more than j − 0:057j over a 10-month period. By contrast, the second regime occurs when the real exchange rate decreases by less than j − 0:057j, remains constant, or increases during the same period. In other words, China’s international competitiveness decreases by more than 5.7% in Regime 1, while it decreases by less than 5.7%, remains constant, or increases in Regime 2. A graphical illustration of the above regime classification is shown in Figure 5. Not surprisingly, the time periods of relatively high loss of price competitiveness (Regime 1) coincide with periods in which the yuan, in nominal terms, was following an appreciating crawl against the U.S. dollar. Instead, periods of relatively low loss or gain of price competitiveness (Regime 2) correspond chronologically to periods in which the yuan was pegged to the U.S. dollar. Specifically, the Regime 1 captures the periods during 1995; from late 2006 to late 2008 (subperiod 3); and from late 2010 until the end of the estimated period (subperiod 5), while the Regime 2 captures the periods from 1996 to late 2006 (subperiod 2) and from 2009 to late 2010 (subperiod 4). However, regime classification of the periods during 1994 and from mid-2005 to late 2006 does not fit with the applied exchange rate policy. Although during these periods the yuan was appreciating against the U.S. dollar, the real exchange rate fell in Regime 2. This mismatch can be explained by the very low rate of appreciation of the yuan during both periods. Specifically, the yuan was appreciating, in nominal terms, against the U.S. dollar by no more than 0.5% on a monthly basis. Moreover, concentrating on the subperiod from mid-2005 to late 2006, the loss in China’s competitiveness resulting from the small nominal appreciation of the yuan was offset by the deflationary pressures in the Chinese economy. We further address the question whether changes in China’s international competitiveness are consistent with PPP equilibrium. Thus, we implement the TAR unit root test by computing the test statistics R1T, t1, and t2

174

N. Giannellis and G. P. Kouretas

given that the delay parameter equals to 10. R1T tests the null hypothesis that the real exchange rate has unit roots in both regimes, against the alternative which states that it is covariance stationary in both regimes. To find whether pure or partial nonstationarity is the case, we compute t1 and t2 test statistics. The results, which are shown in the lower part of Table 2, imply that there is no common evidence against PPP hypothesis across both regimes. The null hypothesis cannot be rejected via the R1T and t2 test statistics (the p-values are 0.174 and 0.632, respectively), but it can be rejected, at 10% level of significance, when employing the t1 test statistic (p-value = 0.07). This means that the real exchange rate behaves like a stationary process in Regime 1 and it follows a unit root process in Regime 2. Thus, under the presence of nonlinear behavior, China’s competitiveness
in bilateral trade with the United States
fluctuates in consistency with PPP equilibrium only when the yuan appreciates, in real terms, against the U.S. dollar by more than 5.7% in a 10-month period. By combining the unit root test results with the regime classification according to the TAR model (Figure 5) and the pathway of the nominal exchange rate as shown in Figure 1, we argue that fixing the exchange rate was not consistent with foreign exchange market equilibrium. Instead, the exchange rate was following an equilibrium process only when the yuan was appreciating against the U.S. dollar. In this case (i.e., Regime 1), the estimated half-life (hl = 14.72) implies fast reverting process toward PPP equilibrium. This means that when China’s price competitiveness declines by more than 5.7% over a 10-month period, deviations from PPP equilibrium are expected to decrease by 50% in less than 15 months.25 In summary, two important findings emerged with respect to the bilateral yuan
dollar real exchange rate. First, China’s international competitiveness fluctuation was not permanently inconsistent with PPP equilibrium. The degree of nominal appreciation of the yuan during the appreciation crawl period, apart from two small subperiods (i.e., 1994 and mid-2005 to late 2006), was in correspondence with relative price movements. Thus, there is no sufficient evidence to state that China has followed a constant currency manipulator rule.26 Second, the adoption of a fixed exchange rate regime

The half-life is estimated based on the following formula: lnð0:5Þ=lnð^ρ þ 1Þ, where ρ^ is the estimated autoregressive parameter of the TAR model in Regime 1. 26 This argument does not imply that Chinese monetary authorities did not intervene, in the forex market, preventing the appreciation of the yuan. What this statement argues is that there is no evidence that all interventions were in contradiction with PPP equilibrium. Thus, there is no strong evidence of the presence of a consistent currency manipulation policy. 25

China’s International Competitiveness

175

was a necessary but not a sufficient condition to establish equilibrium in the foreign exchange market. Although the aim of fixing the exchange rate was to obtain monetary policy control and internal stability, instead of increasing the current account surplus (Corden, 2009; McKinnon & Schnabl, 2009; McKinnon et al., 2009), foreign exchange market equilibrium requires the appreciation of the yuan.

Real Exchange Rate of Yuan
The Japanese Yen With respect to the bilateral real exchange rate of yuan against the Japanese yen we note that apart from two periods (i.e., from mid-1998 to 2000; from 2002 to 2004 and from 2009 to 2011), the real exchange rate seems to be decreasing. An increase in the real exchange rate implies real depreciation of the yuan and improvement of China’s price competitiveness in trade with Japan. In contrast, a decrease in the real exchange rate reveals that the yuan appreciates in real terms and thus, China’s international competitiveness deteriorates. As in the case of the exchange rate of yuan against the U.S. dollar, the threshold parameter is found to be negative and the delay parameter is equal to 10. With d = 10 and λ = −0.056, the real exchange rate observations are divided into two regimes according to the following regime classification. In Regime 1, the real exchange rate (i.e., China’s price competitiveness) decreases by more than j − 0:056 j (i.e., 5.6%) over a 10-month period. While in Regime 2, the real exchange rate (i.e., China’s price competitiveness) declines by less than 5.6%, remains stable, or rises during the same period. The two classification regimes of the yuan vis-a`-vis yen real exchange rate are shown in Figure 6. A large number of observations fall into Regime 1, while Regime 2 is present from 1998 to 2000; from 2002 to 2004, and from 2008 to 2010. Studying further Figure 6 in relation with Figure 7, it is clearly shown that China exhibits a relatively smaller loss in terms of competitiveness in its trade with Japan (Regime 2) when the nominal exchange rate of yuan against the Japanese yen depreciates. By contrast, China’s trade competitiveness exhibits greater loss (Regime 1) when the yuan appreciates, in nominal terms, against the Japanese yen.27 It turns out that gains in international competitiveness are driven from the nominal

27

Figure 8 shows that the yuan exhibited nominal appreciation against the Japanese yen from late 1995 to mid-1998; from 2000 to 2002, and from 2005 to mid-2007. Moreover, the yuan depreciated, in nominal terms, from 1993 to mid1995; from late 1998 to 2000; from 2002 to 2005; and from late 2007 to 2011.

176

N. Giannellis and G. P. Kouretas .08

.04

.00

–.04

–.08 1999

2001

2003

2005

2007

2009

2011

First log difference of the nominal exchange rate First log difference of relative prices

Figure 7: Nominal Yuan Exchange Rate Against Euro and Relative Prices.

depreciation of the yuan.28 However, there is one exception to this finding. In the last subperiod (i.e., 2010
2011), observations are distributed into Regime 1 despite the fact that the yuan was clearly depreciating, in nominal terms, against the Japanese yen. One possible explanation could be that the rate of depreciation of the yuan was small. But, a more plausible explanation could be that any gains in competitiveness, resulting from the yuan depreciation, were offset by the greater change in the Chinese CPI compared to the Japanese one.29 Test statistics, R1T, t1, and t2, are calculated as before and the results are presented in the bottom part of Table 2. R1T test statistic is 10.0, while the

28

Unlike the nominal exchange rate against the U.S. dollar, which was either fixed or decreasing, the nominal exchange rate against the Japanese yen exhibits both increasing and decreasing trends, thereby implying yuan depreciation and appreciation, respectively. Given the pressure for appreciating the yuan, what can explain the depreciation of the yuan against the Japanese yen? Given that the value of the yuan was fixed to the U.S. dollar, the depreciation of the U.S. dollar against the Japanese yen led to the depreciation of the yuan as well. 29 Since October 2005, China’s CPI has been permanently higher than Japan’s CPI. At the end of the estimated period, the CPI differential was at the highest level as a result of the increasing trend of Chinese prices and the declining, or at least stable, trend of Japanese prices.

China’s International Competitiveness

177

bootstrap p-value of accepting the null hypothesis is 0.13. This means that there are signs that the real exchange rate is nonstationary in both regimes. However, t1 and t2 test statistics provide quite interesting implications. Test statistic t1 is equal to 3.16 with p-value 0.04, but test statistic t2 equals 0.11 with p-value 0.77. These estimates reveal that the real exchange rate behaves as a stationary series in Regime 1 (at 5% level of significance) and as a nonstationary series in Regime 2. In terms of the PPP hypothesis, this evidence implies that PPP is established when China’s international competitiveness (bilaterally against Japan) decreases by more than 5.6%. On the contrary, PPP cannot be valid when China’s international competitiveness decreases by less than 5.6%, remains constant, or increases. As a consequence, only the loss of China’s price competitiveness by more than this rate can be considered as an equilibrium phenomenon. If this is the case (i.e., Regime 1), the estimated half-life (hl = 16.15) implies very fast reverting process toward PPP equilibrium. Namely, when China’s competitiveness declines by more than 5.6% over a 10-month period, deviations from PPP equilibrium are expected to decrease by 50% in about 16 months. The overall evidence from the estimated TAR unit root tests implies that the real exchange rate is stationary in Regime 1, but it is nonstationary in Regime 2. Regarding the main empirical issue of the present paper, an interesting fact that stems from this analysis is that this bilateral exchange rate was not continuously deviating from PPP equilibrium. Thus, there are an adequate number of periods in which China’s price competitiveness has fluctuated in consistency with PPP equilibrium. This would imply that .12

.10

.08

.06

.04

.02 1993 1995 1997 1999 2001 2003 2005 2007 2009 2011

Figure 8:

Nominal Yuan Exchange Rate Against the Japanese Yen.

178

N. Giannellis and G. P. Kouretas

there is no strong evidence that China has implemented a manipulation rule, under which the yuan is constantly kept at an artificially low level. Turning now our attention to Figure 8, we provide evidence that the stationarity of the real exchange rate of yuan
yen (Regime 1) coincides with periods of nominal appreciation of the yuan, while nonstationarity (Regime 2) matches with periods of nominal yuan depreciation. A direct implication of this finding is that the depreciation of the yuan cannot restore the equilibrium in the foreign exchange market. Instead, the nominal appreciation of the yuan can be considered as an equilibrium phenomenon.

Summary and Concluding Remarks This paper investigated whether the Chinese yuan exchange rate against the U.S. dollar, the euro, and the Japanese yen follows an equilibrium process toward the PPP hypothesis. Special attention has been paid to the implications underlying the PPP condition. Namely, apart from the equilibrium process of the nominal exchange rate, we examined if China’s international competitiveness fluctuates in consistency with PPP equilibrium. Our study was motivated by the growing academic and policy makers’ debate about the role of China’s exchange rate policy, the low value of the yuan, and consequently, the focus on the question of whether China acts as a currency manipulator. Utilizing the nonlinear characteristics of real exchange rates, this paper brings new and interesting findings to light. First, the yuan exchange rate follows an equilibrium process in Regime 1 (high competitiveness loss regime), but it is found to be away from PPP equilibrium in Regime 2 (low competitiveness loss regime). Second, the evidence in favor or against PPP hypothesis depends on the fluctuation of the nominal exchange rate. Periods of high real yuan appreciation (Regime 1) coincide with periods in which the yuan, in nominal terms, was following an appreciating crawl against the U.S. dollar. In contrast, low real appreciation or depreciation (Regime 2) corresponds to periods in which the yuan was pegged to the U.S. dollar. Third, the pegged regime prevents China’s price competitiveness from equilibrium adjustment, while the appreciating crawl of the yuan seems to be more appropriate. Thus, China should continue appreciating the yuan to establish external equilibrium. Moreover, an important implication is that, apart from the yuan exchange rate against euro case, the exchange rate regime does matter in accepting or not the PPP hypothesis. This condition could not be identified in periods in which the yuan was pegged to the U.S. dollar, but it

China’s International Competitiveness

179

was found to be valid when the yuan was appreciating against the U.S. dollar.30 This means that the pegged regime is not consistent with equilibrium in the foreign exchange market. In addition, this exchange rate regime prevents China’s price competitiveness from equilibrium adjustment. An alternative, but similar, interpretation of this result is that real exchange rate adjustment is also regime-dependent and that periods of fixed exchange rates display delayed adjustment. As a consequence, the appreciating crawl of the yuan seems to be more appropriate. At the end of the estimated period (subperiod 5), in which the yuan follows an appreciating trend (see Figure 1), both yuan real exchange rates (i.e., against the U.S. dollar and the Japanese yen) belong to the stationary regime (see Table 2 and Figures 5 and 6), thereby implying that PPP is valid. Thus, China should continue appreciating the yuan to establish external equilibrium. However, any policy suggestion to Chinese monetary authorities should take into account a number of policy objectives. For example, we should be aware of China’s aim to establish internal stability and maintain monetary policy control. McKinnon and Schnabl (2009, 2012) and McKinnon et al. (2009) have argued that by fixing the value of the yuan against the U.S. dollar, China restored the control of the monetary policy and stabilized its domestic economy. Nonetheless, we have found in this study that the fixed exchange rate regime was not appropriate for establishing equilibrium in the forex market. By combining these findings, we may argue that the fixed exchange rate regime helps to maintain internal stability, but undermines the achievement of external balance. Therefore, our findings imply that China has to form its exchange rate policy within a complex economic environment constrained by internal and external objectives. Finally, in terms of the question of whether China can be characterized as a currency manipulator, we did not find entire periods in which China’s international competitiveness fluctuates inconsistently with PPP equilibrium. Apart from the exchange rate against the euro, the rest of the

30

The exchange rate policy against the U.S. dollar was dominant for the fluctuation of the yuan exchange rate against the Japanese yen. Given that China’s exchange rate policy was formed based on the value of the yuan against the U.S. dollar, the yuan exchange rate against the Japanese yen was following the trend of the U.S. dollar exchange rate against the Japanese yen. For example, with the yuan pegged to the U.S. dollar, the depreciation of the U.S. dollar against the Japanese yen implies the depreciation of the yuan against the Japanese yen as well.

180

N. Giannellis and G. P. Kouretas

exchange rates were not constantly away from PPP equilibrium.31 Therefore, China’s price competitiveness was not permanently following a disequilibrium process. Our results reveal that the magnitude of appreciation of the yuan, during the appreciating crawl period, was in correspondence with relative price movements, and consequently was not manipulated by Chinese authorities. However, it is true that China has intervened a lot of times in the foreign exchange market during the pegged regime period. Although Chinese authorities have periodically intervened in forex markets to prevent exchange rate fluctuation, we conclude that we did not find strong evidence confirming that China has applied an explicit and continual currency manipulation rule.32

Acknowledgments An earlier version of this paper was presented at the 17th International Conference on Macroeconomic Analysis and International Finance, University of Crete, Rethymno, May 30
June 1, 2013 and the 10th Annual Conference of the Hellenic Finance and Accounting Association, University of Piraeus, December 16
17, 2011 and thanks are due to seminar participants for many helpful comments and discussions. Kouretas acknowledges financial support from a Marie Curie Transfer of Knowledge Fellowship of the European Community’s Sixth Framework Programme

31

This finding also implies that China’s economic reforms were successful in transforming the economy from a centralized to a market economy. A number of influential papers have investigated the effectiveness of China’s economic reform. Young (2000) has argued that the economic reform resulted in a fragmented Chinese domestic market. In contrast, Fan and Wei (2006) have found strong evidence of intra-national price convergence in China, which implies the presence of regional market integration. Since the absence of regional trade barriers is a prerequisite for the effective abolishment of trade barriers in international trade, our findings are in line with those of Fan and Wei (2006). 32 In an interview on March 12, 2012 Zhou Xiaochuan, People’s Bank of China Governor mentioned that market forces were playing a bigger role in determining the exchange rate, in keeping with the central bank’s long-term policy objective (see, Oliver, 2012). Furthermore, he added that as China’s approaches an equilibrium exchange rate, the central bank will gradually reduce substantially its intervention in the foreign exchange market. In line with this statement, since April 16, 2012 the daily floating band of the yuan against the U.S. dollar has been increased from 0.5% to 1% (PBC Announcement, 2012 No. 4). Coupled with this statement it was also documented that China’s trade balance went to a deficit of $31.48 billion, after a surplus of $27.28 billion in January (see, Back, 2012). Both these events are compatible with our main findings regarding China’s exchange rate policy.

China’s International Competitiveness

181

under contract number MTKD-CT-014288, from the Research Committee of the University of Crete under research grant #2257, as well as from the Research Committee of Athens University and Business under research grant EP-1836-01/00-01. We would like to thank without implicating Joscha Beckmann, Harris Dellas, Jerry Dwyer, Bill Gavin, Dimitris Georgoutsos, Bob King, Evgenia Passari, Jean-Sebastien Pentecote, and George Tavlas for valuable comments on an earlier draft.

References Ahmad, S., & Rashid, A. (2008). Non-linear PPP in South Asia and China. Economics Bulletin, 6, 1
6. Alba, J. D., & Park, D. (2005). An empirical investigation of purchasing power parity (PPP) for Turkey. Journal of Policy Modeling, 27, 989
1000. Anderson, J. (2008). China’s industrial investment boom and the renminbi. In M. Goldstein & N. Lardy (Eds.), Debating China’s exchange rate policy (pp. 61
69). Washington, DC: Peterson Institute for International Economics. Andrews, D. (1993). Tests for parameter instability and structural change with unknown change point. Econometrica, 61, 821
856. Back, A. (2012). China swings to $31.48 billion trade deficit. Market Watch: The Wall Street Journal. Retrieved from http://www.marketwatch.com/story/chinaswings-to-3148-billion-trade-deficit-2012-03-10 Benassy-Quere, A., Lahreche-Revil, A., & Mignov, V. (2011). World consistent equilibrium exchange rates. Journal of the Japanese and International Economies, 25(2), 12
32. Bergsten, F. (2010). Correcting the Chinese exchange rate. Congressial testimony: Testimony before the hearing of China’s exchange rate policy. Committee on Ways and Means, US House of Representatives. Caner, H., & Hansen, B. (2001). Threshold autoregression with a unit root. Econometrica, 69, 1555
1596. Cheung, Y.-W., Chinn, M. D., & Fujii, E. (2007). The overvaluation of renminbi undervaluation. Journal of International Money and Finance, 26, 762
785. Cheung, Y.-W., Chinn, M. D., & Fujii, E. (2009). Pitfalls in measuring exchange rate misalignment. The yuan and other currencies. Open Economies Review, 20, 183
206. Cline, W. R., & Williamson, J. (2010). Currency wars? Peterson Institute for International Economics, Policy Brief No. PB10-26. Corden, W. M. (2009). China’s exchange rate policy, its current account surplus and the global imbalances. The Economic Journal, 119, F430
F441. Coudert, V., & Couharde, C. (2007). Real equilibrium exchange rate in China. Is the renminbi undervalued? Journal of Asian Economics, 18(4), 568
594. Davies, R. B. (1977). Hypothesis testing when a nuisance parameter is present only under the alternative. Biometrica, 64(2), 247
254. Elliott, G. (1999). Efficient tests for a unit root when the initial observation is drawn from its unconditional distribution. International Economic Review, 40, 767
783.

182

N. Giannellis and G. P. Kouretas

Elliot, G., Rothenberg, T. J., & Stock, J. H. (1996). Efficient tests for an autoregressive unit root. Econometrica, 64, 813
836. Fan, S., & Wei, X. (2006). The law of one price: Evidence from the transitional economy of China. The Review of Economics and Statistics, 88(4), 682
697. Funke, M., & Rahn, J. (2005). Just how undervalued is the Chinese renminbi? World Economy, 28, 465
489. Giannellis, N., & Papadopoulos, A. P. (2010). Nonlinear exchange rate adjustment in the enlarged Euro Zone: Evidence and implications for candidate countries. Review of International Economics, 18(4), 741
757. Goldstein, M., & Lardy, N. (2006). China’s exchange rate policy dilemma. American Economic Review, 96(2), 422
426. Gregory, R. P., & Shelley, G. (2011). Purchasing power parity and the Chinese yuan. Economics Bulletin, 31(2), 1247
1255. Guo, Q. (2010). The Balassa-Samuelson model of purchasing power parity and the Chinese exchange rates. China Economic Review, 21, 334
345. Hansen, B. (1996). Inference when a nuisance parameter is not identified under the null hypothesis. Econometrica, 64(2), 413
430. Hansen, B. (1997). Inference in TAR models. Studies in Nonlinear Dynamics & Econometrics, 2, 1
14. Ho, T.-W. (2005). Investigating the threshold effects of inflation on PPP. Economic Modelling, 22(5), 926
948. Kwiatkowski, D., Phillips, P. C. B., Schmidt, P., & Shin, Y. (1992). Testing the null hypothesis of stationarity against the alternative of a unit root: How sure are we that economic time series have a unit root? Journal of Econometrics, 54, 159
178. MacDonald, R. (2000). Concepts to calculate equilibrium exchange rates: An overview. Discussion Paper No. 3/00. Economic Research Group of the Deutsche Bundesbank. McKinnon, R. (2006). China’s exchange rate trap: Japan redux? American Economic Review, 96(2), 427
431. McKinnon, R., Lee, B., & Wang, Y. D. (2009). The global crisis and China’s exchange rate. Working Paper No. 391, Stanford Center for International Development. McKinnon, R., & Schnabl, G. (2009). The case for stabilizing China’s exchange rate: Setting the stage for fiscal expansion. China & World Economy, 17(1), 1
17. McKinnon, R., & Schnabl, G. (2012). China and its dollar exchange rate: A worldwide stabilizing influence? The World Economy, 35, 667
693. Newey, W. K., & West, K. D. (1994). Automatic lag selection in covariance matrix estimation. Review of Economic Studies, 61(4), 631
654. Ng, S., & Perron, P. (2001). Lag length selection and the construction of unit root tests with good size and power. Econometrica, 69, 1519
1554. Obstfeld, M. (2012). Financial flows, financial crises, and global imbalances. Journal of International, Money and Finance, 31, 469
480. Oliver, C. (2012). China’s yuan nearing fair value: PBOC’s Zhou. Market Watch: The Wall Street Journal. Retrieved from http://www.marketwatch.com/story/ chinas-yuan-nearing-fair-value-pbocs-zhou-2012-03-12

China’s International Competitiveness

183

Phillips, P. C. B., & Perron, P. (1988). Testing for a unit root in time series regression. Biometrica, 75, 335
346. Sarno, L., Taylor, M., & Chowdhury, I. (2004). Nonlinear dynamics in deviations from the law of one price: A broad-based empirical study. Journal of International Money and Finance, 23, 1
25. Subramanian, A. (2010). New PPP-based estimates of renminbi undervaluation and policy implications. Policy Brief No. PB10-8, Peterson Institute for International Economics. Taylor, M., Peel., D., & Sarno, L. (2001). Nonlinear mean-reversion in real exchange rates: Toward a solution to the purchasing power parity puzzles. International Economic Review, 42, 1015
1042. The People’s Bank of China Announcement [2012, No. 4]. Wang, Y., Hui, X., & Soofi., A. (2007). Estimating renminbi (RMB) equilibrium exchange rate. Journal of Policy Modeling, 29, 417
429. Young, A. (2000). The razor’s edge: Distortions and incremental reform in the people’s republic of China. The Quarterly Journal of Economics, 115(4), 1091
1135.

Linkages between the Eurozone and the South-Eastern European Countries: A VECMX* Analysis Minoas Koukouritakis, Athanasios P. Papadopoulos and Andreas Yannopoulos Department of Economics, University of Crete, Greece, e-mail: [email protected]

Abstract Purpose
In the present paper we assess the impact of the Eurozone’s economic policies on specific South-Eastern European countries, namely Bulgaria, Croatia, Cyprus, Greece, Romania, Slovenia, and Turkey. Design/methodology/approach
Since the countries under investigation are connected to the European Union (EU) or the Eurozone and the economic interdependence among them is evolving, we carried out our analysis using the VECMX* framework. Findings
Our results indicate that the transition economies in our sample react in a similar manner to changes in international macroeconomic policies. Cyprus and Greece react also in a similar way, but these responses are very small in magnitude. Finally, Turkey behaves in a different way, probably due to the inflationary pressures in its economy. In general, there is evidence of linkages and interdependence among the EU or Eurozone members of the region. Research limitations/implications
We did not construct a full structural model proposed by economic theory, but instead we estimated a reducedform model. Data limitation is one reason. The other reason is that our

International Symposia in Economic Theory and Econometrics, Vol. 23 G. P. Kouretas and A. P. Papadopoulos (Editors) Copyright r 2014 by Emerald Group Publishing Limited. All rights reserved ISSN: 1571-0386/doi:10.1108/S1571-038620140000023007

186

M. Koukouritakis et al.

sample countries are extremely heterogeneous. Also, for most of the sample countries there is an acute problem of structural uncertainty of their economies yet. Practical implications
The way that the economies under investigation react to changes in international macroeconomic policies, may influence the Eurozone policy makers regarding the implemented monetary policy. Originality/value
To our knowledge, the above methodology is implemented for the first time in the sample countries and provides a detailed investigation regarding their economic policies and the effects of the Eurozone policies. Keywords: South-Eastern Europe, Monetary Transmission, VECMX* Model, Generalized Impulse Response JEL Classifications: E43, F15, F42

Introduction The integration procedure of the South-Eastern European economies to the European Union (EU) is continuously evolving during the last decades. Some of the South-Eastern European countries are either already members of the EU or the Eurozone, or associated with the EU; while others are set to become EU members. This integration procedure implies that the EU affects the above countries in a more systematic way. It has also led to the expansion of economic transactions in the whole region. Consequently, there is a need of systematic and detailed research about the economic policies of the countries in this region, especially in our days when the current financial and debt crisis in the Eurozone is at stake. For instance, Greece, which is a Eurozone member since 2001, is in deep recession with high sovereign debt, and having signed the Memoranda I and II with the European Central Bank (ECB)
EU
International Monetary Fund (IMF), is in fiscal contraction and faces high unemployment. Also, the emerging economies of the South-Eastern Europe have relatively high current account deficits and are more vulnerable to the deterioration of the international economy, since they have been negatively affected by the reduction of external demand and the increase in the cost of borrowing from abroad. In this study, we attempt to investigate the monetary transmission mechanism of the Eurozone’s monetary policy for seven countries of SouthEastern Europe, namely Bulgaria, Croatia, Cyprus, Greece, Romania,

A VECMX* Analysis for the South-Eastern European Countries

187

Slovenia, and Turkey. Also, we explore the way that foreign macroeconomic variables affect their domestic counterparts, for each of the above countries. Especially for the transition economies (Bulgaria, Croatia, Romania, and Slovenia) this analysis will allow us (a) to understand how fast, and to what extent, a change in the central bank’s instruments modifies domestic variables, such as inflation, and (b) to evaluate whether monetary transmission operates differently in the transition economies. In general, the monetary policy transmission mechanism in Central and Eastern Europe has been analyzed by Coricelli, E´gert, and MacDonald (2006). These authors studied this mechanism through four channels: (i) interest rate channel, (ii) exchange rate channel, (iii) asset price channel, and (iv) broad lending channel. The literature about monetary policy transmission mechanism is quite large and extending. In general, the interest rate pass-through is usually investigated using an error correction model (ECM) framework. Regarding the transition countries of the Central and Eastern Europe, several researchers have studied the asymmetry of the adjustment process, in relation to the Eurozone countries, with mixed results (Crespo-Cuaresma, E´gert, & Reininger, 2004; E´gert, Crespo-Cuaresma, & Reininger, 2006; Horva´th, Kreko´, & Naszo´di, 2004; Opiela, 1999; Sander & Kleimeier, 2004). Additionally, a number of researchers have studied the long-run pass through. Their results indicate that both the contemporaneous and long-run pass-through increase over time, while the mean adjustment lag to full passthrough decreases, as more recent data have been used (Crespo-Cuaresma, E´gert, & Reininger, 2004; Horva´th, Kreko´, & Naszo´di, 2004; Sander & Kleimeier, 2004). Also, there is a number of studies that investigate the exchange-rate pass-through in the transition economies, using mainly vector autoregressive (VAR) and vector error correction models (VECMs) (see, for instance, Coricelli, Jazbec, & Masten, 2003; Darvas, 2001; Gueorguiev, 2003; Kara et al., 2005; Korhonen & Wachtel, 2005; Mihaljek & Klau, 2001). Since the economies of South-Eastern Europe are quite interdependent and influenced, as well, by the EU and the Eurozone, models that have been used for studying the domestic economies are not well suited, since they do not take into account the way economies react to economic and financial interdependencies. In the last decades, the use of VARs and the subsequent cointegration analysis have resulted in long-run relations between various variables in the same economy, as suggested by economic theory. However, many long-run relations in one country may be influenced and affected by variables from other regions. One of the problems with the VAR methodology is that it works with a limited number of variables. But in order to incorporate a reasonable number of variables to account for global effects, large-dimension systems are required. A very important step in this direction is the development of global vector

188

M. Koukouritakis et al.

autoregressive (GVAR) model developed by Pesaran, Schuermann, and Weiner (2004, henceforth PSW), which facilitated the study of international linkages. This methodology has been used to examine the interdependencies of economies worldwide. More specifically, it was used to investigate the changing degree of the dominance of the US economy and its effect on other regions (De´es & Sain-Guilhem, 2009), the analysis of the Swiss economy (Assenmacher-Wesche & Pesaran, 2009), the role of China and its increased influence around the world (Feldkircher & Korhonen, 2012), the linkages in the euro area (De´es, di Mauro, Pesaran, & Smith, 2005), world trade flows (Bussie´re, Chudik, & Sestieri, 2012), and regional financial effects across Europe (Galesi & Sgherri, 2009). In the present paper we investigate the impact of the Eurozone’s economic policies on specific economies of South-Eastern Europe, namely Bulgaria, Croatia, Cyprus, Greece, Romania, Slovenia, and Turkey. Today, these economies are interdependent, since they are with one way or another connected to the EU and the Eurozone.1 Thus, there is a need of detailed investigation of the economic policies of the above countries, as well as the effects of the Eurozone policies. To carry out this task, we followed the methodology developed first by Pesaran, Shin, and Smith (2000) and further extended by PSW (2004) in the GVAR framework. This methodology allows us to estimate country-specific VECMX* models and to evaluate econometric long-run relationships, including nonstationary foreign variables in each of them. Note here that we did not construct a full structural model with many equations in order to capture relationships proposed by economic theory. Data limitation is one reason. The other reason is that our sample countries are extremely heterogeneous. More specifically, some of them have been transformed from centrally planned to free market economies and probably they have not yet settled to a long-run pattern. Also, for most of the sample countries there is an acute problem of structural uncertainty of their economies yet. Thus, our model is a reduced-form one. In brief, our dynamic analysis indicates similar and expected impulse responses for Bulgaria, Croatia, Romania, and Slovenia. The same conclusion can be drawn for Cyprus and Greece, but in these two cases the impulse response functions are very small in magnitude. Finally for Turkey, even though the effects from the impulse response functions are .

1 Bulgaria and Romania joined the EU in 2007; Croatia joined the EU in 2013; Cyprus is a Eurozone member since 2008; Greece is a Eurozone member since 2001; Slovenia is a Eurozone member since 2007; and Turkey has settled a customs union with the EU in 1996 and is under negotiations for EU membership in the future. The latter country also had a stand-by agreement with the IMF for a number of years.

A VECMX* Analysis for the South-Eastern European Countries

189

expected, most of them do not converge to a stable level in the time horizon that we have used. A possible explanation could be the strong inflationary tendencies in the Turkish economy. The structure of the paper is organized as follows. The second section illustrates the framework of the VECMX* modeling, while the third section reports the data and the model specification. The fourth section analyses the empirical results, the fifth section presents the dynamic analysis, while the sixth section draws some concluding remarks.

Country-Specific Models The model developed by PSW (2004) begins with country-specific models and assumes that there exist N þ 1 countries in the global economy. These countries are indexed by i = 0; 1; 2; :::; N; adopting country 0 as the reference country. For each country, the country-specific variables are related to the global variables. The latter are measured as country-specific weighted averages of foreign variables. The weights will be analyzed in the following section. In general, deterministic variables and global (weakly) exogenous variables (such as oil prices) are also included in each country-specific model. In brief, for a first-order dynamic specification that relates the ki × 1 vector of country-specific variables (denoted by xit ) to a ki × 1 vector of foreign variables specific to country i (denoted by xit ), the VARX*(1,1) model is the following: xit = αi0 þ αi1 t þ Φi xi;t− 1 þ Λi0 xit þ Λi1 xi;t− 1 þ εit ; t = 1; 2; …; T; N = 0; 1; 2; …; N;

ð1Þ

where Φi is a ki × ki matrix of lagged coefficients, Λi0 and Λi1 are ki × ki matrices of contemporaneous and lagged coefficients related to foreign variables, respectively, and εit is a ki × 1 vector of idiosyncratic countryspecific shocks. In the special case where Λi0 = Λi1 = 0, the model reduces to a standard VAR(1) process. We also assume that the idiosyncratic shocks are serially uncorrelated with zero mean and a non-singular covariance matrix, or εit ∼ iidð0; Σii Þ.2

2

PSW (2004) indicate that for the idiosyncratic shocks there is allowance of limited correlation across countries, while the assumption regarding time invariance of the country-specific covariance matrices can be relaxed.

190

M. Koukouritakis et al.

The error correction representation of Equation (1) is given by   Δxit = αi0 þ αi1 t− Ιki − Φi xi;t−1 þ ðΛi0 þ Λi1 Þxi;t−1 þ Λi0 Δxit þ εit :   x Using zit = it , Equation (2) can be transformed to xit

ð2Þ

Δxit = αi0 þ αi1 t− ðAi − Bi Þzi;t− 1 þ Λi0 Δxit þ εit : ð3Þ   For country i we set the ki × ki þ ki matrix Πi = Ai − Bi , where its rank ðri Þ specifies the number of “long-run” (cointegrating) relationships among the domestic and the country-specific foreign variables (xit and xit , respectively). Thus, we have Ai − Bi = ai β0i ;

ð4Þ

where  ai is the ki × ri matrix of adjustment coefficients and βi is the ki þ ki × ri matrix of cointegrating vectors. Note that both matrices are of full column rank.

Data and Model Specification Our sample consists of monthly data for the period 2000:01
2011:12. We included Bulgaria, Croatia, Cyprus, Greece, Romania, Slovenia, and Turkey, along with EMU12 as the base country. We obtained data for real effective exchange rates based on consumer price index (RER), harmonized consumer price index (HCPI), index of industrial production (IP) and interest rates (IR). We used money market rates for all countries, except for Greece and Cyprus, for which these data are not available. For that reason, we used Treasury bill rates (TB) for Greece and government bond (GB) yields for Cyprus. We also obtained data for the nominal exchange rate of the euro against the US dollar (number of euros per US dollar
NER) and for the oil price (OIL). Almost all data were obtained from the International Financial Statistics of the IMF. The real effective exchange rate for Slovenia and Turkey, the HCPI for all countries, and the index of IP for the EMU (which excludes construction) were derived from the Eurostat. All data, except IR, were transformed into natural logarithms. In our model, we used seven countries, namely Bulgaria, Croatia, Cyprus, Greece, Romania, Slovenia, and Turkey. For each of these countries we set the vector of domestic variable xit = ðRERit ; HCPIit ; IRit ; IPit Þ0 , with ki = 4. EMU12 has been used as the reference country, in which we have also used NER and OIL as (weakly) exogenous variables. The vector xit of the foreign (“starred”) variables has been constructed from

A VECMX* Analysis for the South-Eastern European Countries

191

the domestic variables, using the following relations that are based on PSW (2004), Equation (4): xit = ðRERit ; HCPIit ; IRit ; IPit Þ0 ; XN wRER RERjt ; RERit = j = 0 ij XN HCPIit = wHCPI HCPIjt ; j = 0 ij XN IRit = wIR IRjt ; j = 0 ij XN IPit = wIP IPjt ; ð5Þ j = 0 ij For weights we based on trade weights. Trade data were obtained from the Comtrade database of the United Nations. Note that if we allow trade weights to vary over time, this could introduce an undesirable degree of randomness in the analysis. For this reason and based on the PSW (2004) analysis, we used fixed trade weights. These fixed trade weights were computed as averages of trade flows for the 2001
2006 period, and are presented in Table 1. The trade shares for each country are presented in columns and show the degree to which one country depends on the remaining ones. In our analysis, we estimate vector error correction models (VECMX*s) for each sample country, where the domestic macroeconomic variables (RER, HCPI, IR, and IP) are related to corresponding foreign (“starred”) variables constructed to match the international trade pattern of the country under consideration. The latter variables are treated as weakly exogenous for all sample countries. For Turkey we excluded domestic and foreign IR from the analysis. The reason is that the Turkish IR shows anomalies and extreme values for a long period of time, after the economic crisis of 2001 and the involvement of the IMF. For the VECMX* of the Eurozone, we Table 1: Trade Weights Country

EMU12 (ex. Greece) Bulgaria Croatia Cyprus Greece Romania Slovenia Turkey

EMU12 Bulgaria Croatia Cyprus Greece Romania Slovenia Turkey (ex. Greece) 0.0000

0.6738

0.8272

0.6906 0.8402

0.8539

0.8667

0.8884

0.0575 0.0688 0.0178 0.2036 0.1759 0.1183 0.3580

0.0000 0.0073 0.0049 0.1256 0.0538 0.0091 0.1256

0.0080 0.0000 0.0040 0.0079 0.0124 0.1269 0.0137

0.0121 0.0027 0.0000 0.2766 0.0117 0.0037 0.0027

0.0229 0.0047 0.0018 0.0305 0.0000 0.0078 0.0785

0.0062 0.0946 0.0003 0.0063 0.0125 0.0000 0.0133

0.0280 0.0023 0.0082 0.0260 0.0420 0.0052 0.0000

0.0446 0.0030 0.0295 0.0000 0.0293 0.0037 0.0497

Trade weights are computed as shares of imports and exports, shown in columns by country, such that each column sums to unity.

192

M. Koukouritakis et al.

used only the nominal exchange rate of the euro against the US dollar and the oil price as (weakly) exogenous variables.

Country-Specific Cointegration Models Before estimating each country-specific VECMX*, we tested each domestic, foreign and global variable for a unit root using the Augmented Dickey
Fuller (ADF) and Kwiatkowski, Phillips, Schmidt, and Shinn (1992, henceforth KPSS) unit root tests. In order to select the lag length in each regression of the ADF test, we started from 12 lags and employed the Akaike Information Criterion (AIC). The results are presented in Tables 2 and 3 and indicate that all of the variables under consideration have a unit root.3 Given the fact that almost all of the variables have a unit root, we individually estimate each country-specific cointegration model (VECMX*). Since we are dealing with a small number of time series observations relative to the number of unknown parameters in each model, we started for a VECMX* (3,3) model for each country and chose the lag specification for endogenous and exogenous variables based on the AIC. The cointegration results are presented in Table 4, while the selected VECMX* for each country is presented in column 1 of this table.4 The cointegration results imply that for each of Bulgaria, Croatia, Cyprus, Greece, Romania, Slovenia and Turkey there exist one cointegrating vector at the 5% level of significance. These results also show evidence of two cointegrating vectors for EMU12 at the 5% level of significance. Tables 5 and 6 report the solved cointegrating vectors normalized on the real effective exchange rate, while Tables 7 and 8 present the adjustment coefficients for the ECMs.5,6

3

We also tested all variables for a second unit root. This hypothesis was rejected in all cases. For saving space, these results are not presented here but are available upon request. 4 All estimations of the present paper were performed using the econometric package Microfit 5. 5 Note that it is commonly acceptable that the coefficients of the (Johansen) cointegrating vector are not easily interpretable in many times, without imposing (overidentified) restrictions from economic theory. PSW (2004) use their estimates to generate forecasts without insisting on economic interpretations. 6 The variables of the countries included in the model have probably experienced a number of structural shifts in their intercept or trend during the sample period, due to specific events that have taken place (e.g., the long transition period from centrally planned to free markets economies for Bulgaria, Croatia, Romania, and Slovenia, the involvement of the IMF in the Turkish Economy, and, of course, the current financial and debt crisis that affected all countries). Due to small sample and technical difficulties regarding the estimation of our model, we did not account for these potential structural breaks in the current analysis.

A VECMX* Analysis for the South-Eastern European Countries

193

Table 2: ADF Unit Root Test Results Variable

EMU12

Intercept and trend RER −1.40 HCPI −3.35 IR −1.94 IP −3.03 RER* −2.68 HCPI* −2.33 IR* −2.98 IP* −2.29 NER −1.74 OIL −2.21 Intercept RER −1.73 HCPI −3.71* IR −1.46 IP −1.98 RER* −2.32 HCPI* −1.83 IR* −2.36 IP* −1.78 NER −1.37 OIL −0.38

Bulgaria

Croatia

Cyprus

Greece

Romania

Slovenia

Turkey

−2.35 −1.40 −1.47 −2.21 −1.16 −3.09 −2.37 −2.63

−0.20 −2.36 −3.12 −0.50 −1.28 −2.89 −1.77 −1.05

−1.98 −2.56 −0.86 −1.74 −1.36 −2.89 −2.30 −2.55

−1.33 −2.49 −2.22 −0.20 −0.94 −2.89 −2.69 −1.04

−1.58 −1.93 −1.65 −2.02 −1.01 −2.96 −2.99 −1.07

−2.11 −2.83 −2.84 −1.89 −1.18 −2.85 −1.75 −1.01

−2.55 −3.03 NA −3.08 −1.32 −2.83 NA −1.00

−1.00 −1.73 −1.21 −1.91 −2.20 −3.85* −2.09 −2.36

−2.15 −0.63 −2.42 −2.33 −1.77 −3.69* −0.78 −1.09

−1.62 −0.50 −2.07 −1.47 −1.67 −4.57* −1.77 −2.01

−1.32 −1.16 −2.06 −0.67 −1.99 −3.69* −1.37 −1.03

−1.43 −0.81 −2.00 −1.22 −2.46 −3.73* −2.06 −1.03

−2.03 −2.14 −1.30 −1.94 −1.77 −3.76* −0.97 −1.11

−2.50 −1.81 NA −1.20 −1.69 −3.68* NA −1.05

The value in each cell is the ADF unit root test statistic. The 95% critical value for this test is −3.44 for regressions with intercept and trend, and −2.88 for regressions with intercept. *denotes rejection of the unit root hypothesis at the 5% level of significance. NA stands for nonavailable.

Also, we also tested our model for serial correlation in the residuals of the error correction regressions. Based on the VECMX*s specification, we used lag order 3 for EMU12, Croatia, Cyprus, Greece, Romania, Slovenia, and Turkey, and lag order 2 for Bulgaria. The F-statistics for the serial correlations test are reported in Table 9. As indicated in this table, 23 of the 31 regressions pass the serial correlation test, since for these cases the null hypothesis of no serial correlation cannot be rejected at the 5% level of significance. Finally, before proceeding with the dynamic analysis and the estimation of generalized impulse response functions (GIRFs), we estimated the persistence profiles for each cointegrating vector. Persistence profiles refer to the time profiles of the effects of system or variable-specific shocks on the cointegrating relations (Pesaran & Shin, 1996). They have a value of unity on impact, while they should tend to zero as the horizon n → ∞, if the vector under consideration is a valid cointegrating vector. The persistence profiles also provide information on the speed with which the cointegrating

194

M. Koukouritakis et al.

Table 3: KPSS Unit Root Test Results Variable

EMU12

Intercept and trend RER 0.29* HCPI 0.28* IR 0.25* IP 0.34* RER* 0.32* HCPI* 0.34* IR* 0.19* IP* 0.27* NER 0.26* OIL 0.23* Intercept RER 0.79* HCPI 1.25* IR 0.65* IP 0.54* RER* 0.99* HCPI* 1.36* IR* 1.08* IP* 1.23* NER 1.15* OIL 1.26*

Bulgaria

Croatia

Cyprus

Greece

Romania

Slovenia

Turkey

0.18* 0.29* 0.30* 0.32* 0.31* 0.28* 0.22* 0.19*

0.18* 0.23* 0.49* 0.36* 0.30* 0.28* 0.21* 0.28*

0.20* 0.17* 0.25* 0.36* 0.29* 0.28* 0.26* 0.23*

0.25* 0.17* 0.28* 0.34* 0.32* 0.28* 0.21* 0.31*

0.20* 0.19* 0.33* 0.26* 0.30* 0.28* 0.20* 0.29*

0.20* 0.31* 0.25* 0.21* 0.30* 0.28* 0.25* 0.31*

0.31* 0.34* NA 0.21* 0.30* 0.28* NA 0.30*

1.37* 0.79* 0.50* 1.06* 0.96* 1.27* 1.07* 0.59*

1.26* 1.41* 0.61* 1.13* 0.84* 1.26* 0.91* 0.73*

1.09* 1.41* 0.75* 0.76* 0.96* 1.27* 0.71* 0.64*

1.28* 1.41* 0.49* 0.72* 0.93* 1.25* 1.05* 0.99*

1.06* 1.41* 1.25* 1.43* 0.91* 1.26* 1.00* 0.84*

0.59* 1.36* 1.19* 0.93* 0.88* 1.26* 0.81* 0.77*

0.89* 1.32* NA 1.37* 0.88* 1.25* NA 0.66*

The value in each cell is the KPSS unit root test statistic. The 95% critical value for this test is 0.146 for regressions with intercept and trend, and 0.463 for regressions with intercept. *denotes rejection of the stationarity hypothesis at the 5% level of significance. NA stands for nonavailable.

relationships return to their equilibrium states. The estimated persistence profiles for each cointegrating vector of our model are presented in Figure 1. As shown, they all converge very fast to zero (except for the second cointegrating vector of the EMU12 and the cointegrating vector of Turkey) implying that our cointegrating vectors are valid.

Generalized Impulse Response Functions In this section we proceed with the dynamic analysis of our model using GIRFs, as they proposed by Koop, Pesaran, and Potter (1996) for nonlinear models and further developed in Pesaran and Shin (1998) for vector ECMs. The methodology of GIRFs differs from that of orthogonalized impulse responses (OIRs) developed by Sims (1980) in the following two ways. First, it does not require any a priori economic-based restrictions and its outcome

195

A VECMX* Analysis for the South-Eastern European Countries

Table 4: Cointegration Tests Results Model VARX*(3,3) restricted trend, unrestricted intercept

EMU12 (excluding Greece) p
r Trace Maxλ 4 186.44** 113.51** 3 72.93** 47.84** 2 25.09 16.89 1 8.20 8.20

CVa Trace 95% 94.60 63.70 38.53 20.06

90% 88.33 59.75 35.57 17.31

CV maxλ 95% 43.37 35.05 28.38 20.06

90% 40.78 32.58 24.75 17.31

Model VARX*(2,2) restricted trend, no intercept

Bulgaria p
r Trace maxλ 4 137.83** 72.11** 3 65.72 35.26 2 30.46 22.13 1 8.33 8.33

CV trace 95% 90% 101.62 97.05 72.05 66.57 44.78 40.81 22.31 19.75

CV maxλ 95% 90% 47.74 43.68 38.99 35.98 30.59 28.08 22.31 19.75

Model VARX*(3,1) restricted intercept, no trend

Croatia p
r Trace maxλ 4 149.02** 89.78** 3 59.24 27.38 2 31.86 17.86 1 14.01 14.01

CV trace 95% 90% 99.48 95.86 69.39 65.08 43.03 39.58 22.14 19.71

CV maxλ 95% 90% 45.38 42.54 37.88 35.76 30.57 27.54 22.14 19.71

Model VARX*(3,3) restricted trend, no intercept

Cyprus p
r Trace maxλ 4 108.23* 53.55** 3 54.67 36.76 2 17.91 11.07 1 6.84 6.84

CV trace 95% 90% 109.93 104.70 76.34 71.30 46.74 42.73 23.56 20.62

CV maxλ 95% 90% 50.78 47.05 41.86 38.60 32.28 29.28 23.56 20.62

Model VARX*(3,1) restricted trend, unrestricted intercept

Greece Slovenia CV trace p
r Trace maxλ Trace maxλ 95% 90% 4 120.77** 50.86** 154.25** 84.25** 110.02 105.07 3 69.91 32.13 70.00 34.24 79.28 72.63 2 37.78 23.13 35.76 22.80 48.80 45.37 1 14.65 14.65 12.96 12.96 24.44 22.04

CV maxλ 95% 90% 48.10 45.29 41.56 38.52 33.72 30.40 24.44 22.04

Model VARX*(3,3) restricted intercept, no trend

Romania p
r Trace maxλ 4 110.32** 57.46** 3 52.86 26.73 2 26.13 18.34 1 7.79 7.79

CV trace 95% 90% 108.88 103.71 74.66 69.92 46.26 42.57 23.66 20.83

CV maxλ 95% 90% 50.23 45.99 40.73 38.08 32.73 29.65 23.66 20.83

Turkey Trace maxλ 92.83** 62.78** 30.05 25.03* 5.02 5.02

CV trace 95% 90% 59.27 54.92 36.39 33.15 19.08 16.71

CV maxλ 95% 90% 35.18 31.70 26.47 24.40 19.08 16.71

Model VARX*(3,2) unrestricted intercept, no trend a

p
r 3 2 1

CV is for critical values. The 95% and 90% critical values are computed by stochastic simulations using 1000 replications. ** and * denote rejection of the null hypothesis at the 5% and the 10% level of significance, respectively.

Table 5: Estimated Coefficients of the Normalized Cointegrating Vectors Parameter estimates Bulgaria

Croatia

βRER βHCPI βIR βIP βRER* βHCPI* βIR* βIP* Intercept Trend

1.0000 1.0000 1.0000 0.1612 0.8885 0.7270 −0.0020 −0.0019 −0.0049 0.7068 0.2170 −0.1890 −0.0616 0.4194 0.4309 −0.0853 0.0354 −0.0596 −0.0066 0.0093 −0.0013 −0.4826 −0.5429 0.1794 3.5353 NA NA NA 0.0090 −0.0010

1.0000 0.7766 0.0165 −0.6984 0.0796 −0.0119 0.0018 0.7777 NA 0.0045

Cyprus

Greece

Romania Slovenia

Turkey

1.0000 −3.7554 −0.0090 8.9426 0.6293 −1.4906 −0.0257 −3.3158 −0.1386 NA

1.0000 −3.6776 NA −0.8282 −0.3454 3.1525 NA 1.4591 NA NA

1.0000 2.7287 0.0006 −2.2333 0.1677 −0.2500 0.0227 2.1138 NA −0.0028

β’s are the parameters of the solved cointegrating vectors, normalized on the real effective exchange rate. *indicates foreign variables. NA stands for nonavailable.

Table 6:

Estimated Coefficients of the Normalized Cointegrating Vectors

Parameter estimates

EMU12 (excluding Greece)

βRER βHCPI βIR βIP βNER βOIL Trend

1.0000 0.2127 0.0338 −1.2642 −0.7214 −0.0083 −0.0021

1.0000 0.0282 −0.0015 −0.1262 −0.5110 0.0496 −0.0021

β’s are the parameters of the solved cointegrating vectors, normalized on the real effective exchange rate.

Table 7:

Adjustment Coefficients

Parameter estimates

Bulgaria

αRER

0.0009 −0.0116 (0.0076) (2.0257) [0.930] [0.155] −0.0044 0.0061 (0.4348) (0.7017) [0.510] [0.402] 0.3315 0.5981 (0.1581) (0.1105) [0.691] [0.740] −0.4504* −0.4660* (78.9491) (108.1310) [0.000] [0.000]

αHCPI αIR αIP

Croatia

Cyprus

Greece

Romania

Slovenia

Turkey

−0.0470* (13.1155) [0.000] 0.0074 (0.8864) [0.346] −0.2173 (0.8561) [0.355] 0.2775* (27.3710) [0.000]

−0.0123* (15.8466) [0.000] 0.0034 (0.7440) [0.388] −0.0555 (0.1555) [0.693] −0.2051* (26.9618) [0.000]

−0.0094 (0.3135) [0.576] −0.0186* (4.9046) [0.027] 1.9402 (2.0764) [0.150] −0.2980* (52.6919) [0.000]

−0.0165* (12.4010) [0.000] −0.0130* (7.1491) [0.008] −0.0166 (0.0018) [0.966] −0.3690* (99.0004) [0.000]

−0.0803 (3.1310) [0.077] −0.0330* (38.1675) [0.000] NA

−0.1948* (8.1949) [0.004]

α’s are the adjustment coefficients. Numbers in parentheses are Wald test statistics for H0 : αi = 0 and numbers in brackets are the respective p-values. * denotes rejection of the null hypothesis at the 5% level of significance. NA stands for nonavailable.

A VECMX* Analysis for the South-Eastern European Countries

197

Table 8: Adjustment Coefficients Parameter estimates

EMU12 (excluding Greece)

αRER

−0.0354* (19.6598) [0.000] −0.0516* (21.0767) [0.000] −0.1144 (1.0938) [0.296] −0.0508 (0.5148) [0.473]

0.0026 (0.1035) [0.748] −0.0322* (8.1951) [0.004] 0.2578* (5.5604) [0.018] −0.7821* (121.8808) [0.000]

αHCPI αIR αIP

α’s are the adjustment coefficients. Numbers in parentheses are Wald test statistics for H0 : αi = 0 and numbers in brackets are the respective p-values. *denotes rejection of the null hypothesis at the 5% level of significance.

Table 9: Serial Correlation Tests of the VECMX* Residuals Country

Δ(RER)

Δ(HCPI)

Δ(IR)

Δ(IP)

EMU12 (excluding Greece)

5.0291* (0.003) 0.4780 (0.621) 0.7408 (0.530) 1.0129 (0.390) 2.8459* (0.040) 0.3116 (0.817) 0.2213 (0.881) 1.9182 (0.130)

2.4896 (0.064) 1.9464 (0.147) 1.8655 (0.139) 2.6043 (0.055) 1.4902 (0.220) 6.3188* (0.001) 0.4782 (0.698) 2.2266 (0.088)

5.1776* (0.002) 1.1492 (0.320) 0.2777 (0.841) 2.5879 (0.056) 2.4953 (0.063) 5.0579* (0.002) 1.2518 (0.294) NA

4.4748* (0.005) 2.4375 (0.091) 0.9020 (0.442) 1.1303 (0.340) 4.4555* (0.005) 1.0469 (0.375) 2.8930* (0.038) 1.2499 (0.295)

Bulgaria Croatia Cyprus Greece Romania Slovenia Turkey

The value in each cell is F-statistic for the null hypothesis of no serial correlation. Numbers in parentheses are the respective p-values. *denotes rejection of no serial correlation at the 5% level of significance.

is invariant to the ordering of the variables in the model, since it does not orthogonalize the residuals of the system. This methodology takes into account the historical correlations among the variables, summarized by the estimated variance-covariance matrix. Second, it cannot provide

36

48 0

12

36

48

48 0

12

12

Point Estimate for CV1

0

Point Estimate for CV1

24

Top 97.5% Level

24

Slovenia

Top 97.5% Level

36

48

36

Lower 2.5% Level

48

Lower 2.5% Level

0.0

0.0

0.2

0.4

0.6

0.8

1.0

1.2

–0.2

0

12

12

12 Point Estimate for CV1

0

Point Estimate for CV1

0

Point Estimate for CV1

Top 97.5% Level

24

Turkey

Top 97.5% Level

24

Greece

Top 97.5% Level

24

Bulgaria

Figure 1: Persistence Profiles of the Cointegrating Relations to System-Wide Shocks.

Lower 2.5% Level

–0.2

–0.2

Top 97.5% Level

0.0

0.0

Point Estimate for CV1

0.2

0.2

48

0.4

0.4

36

0.6

0.6

24

0.8

0.8

12

1.0

1.0

0

1.2

1.2

Romania

Lower 2.5% Level

0.0 –0.2

0.0

–0.2

Top 97.5% Level

0.4 0.2

0.2

0.2

Point Estimate for CV1

0.6

0.4

0.4

1.0 0.8

1.2

–0.2

0.6

Cyprus

Lower 2.5% Level

0.8

36

24 Top 97.5% Level

0.6

24

Point Estimate for CV2

0.8

0.0

1.0

Croatia

Lower 2.5% Level

1.2

12

24

Top 97.5% Level

0.0

1.0

0

12

0.2

1.2

0.0

0.2

0.4

Point Estimate for CV1

0.4

0.6

0

0.6

0.8

0.6 0.2

0.8

1.0

0.4

1.0

1.2

1.0

0.8

1.2

EMU12 1.4

EMU12

1.2

36

36

48

48

Lower 2.5% Level

Lower 2.5% Level

48

Lower 2.5% Level

36

198 M. Koukouritakis et al.

A VECMX* Analysis for the South-Eastern European Countries

199

information about the causal relationships among the variables because the shocks are not identified. However, the GIRFs methodology is preferable in VAR analysis, since in most cases there is no reasonable way to order the variables in the model. It is important to note here that our dynamic analysis is carried out on the levels of the variables, implying that the effects of a given shock are typically permanent. In the present analysis, we estimated GIRFs of one standard error (s.e.) shock of each of the foreign (“starred”) variables to each domestic variable. More specifically, for the EMU12 we investigated the propagation of a s.e. shock to the nominal exchange rate of the euro against the US dollar and to the oil price on the domestic variables. For each of Bulgaria, Croatia, Cyprus, Greece, Romania, Slovenia, and Turkey we explored the effects from one s.e. shock to each of the four “starred” variables on each of the domestic variables. The GIRFs, along with their ±2.5% bootstrap confidence intervals, are presented in Figures 2
9.7 As shown in these figures, we obtain almost similar impulse responses for Bulgaria, Croatia, Romania, and Slovenia. “Starred” RER seems to have expected responses on the domestic variables of Bulgaria, Croatia, and Slovenia, while for Romania there are some peculiarities. Also, “starred” HCPI and IP have expected results on the domestic variables of all these four countries. “Starred” IR has a more complicated picture, reflecting probably the differences on economic policy implementation in these countries, due to different stages of integration with the EU. For Cyprus and Greece, we obtain similar, but very small in magnitude, impulse response functions. Again, “starred” RER, HCPI and IP seem to have expected responses on the domestic variables of the model. “Starred” IR behaves differently, because for these two countries the corresponding IRs are GB yields (for Cyprus) and TB (for Greece). Note also that all GIRFs for the above six countries are moving quickly to equilibrium (less than twelve months for most of them) and thus, our model seems stable. Finally for Turkey, foreign RER, HCPI and IP have positive effects on domestic RER, and HCPI, and negative effects, as expected, on the domestic IP. For the domestic RER and HCPI, most of the impulse response functions do not converge to a stable level in the time horizon that we have used. A possible explanation could be the strong inflationary tendencies in

7 The GIRFs for the EMU12 model that are presented in Figure 2 were estimated with the assumption of a single cointegrating vector. The reason is that for the cases of RER and HCPI, the estimated GIRFs from the model with two cointegrating vectors do not stabilize on a certain level. This could be explained by the fact that the second cointegrating relation does not converge very fast to equilibrium, as shown by its persistence profile in Figure 1, and this might lead to model instability.

200

M. Koukouritakis et al. Generalized Impulse Response(s) to one S.E. shock in the equation for EURODOL

Generalized Impulse Response(s) to one S.E. shock in the equation for EURODOL

–0.010

0.010

–0.012

0.005

–0.014 0.000 –0.016 –0.005

–0.018

–0.010

–0.020 –0.022

0

6

12

18

–0.015

24

0

Top 97.5% Level Point Estimate for REREMU Lower 2.5% Level

6

12

18

24

Top 97.5% Level Point Estimate for HCPIEMU Lower 2.5% Level

Generalized Impulse Response(s) to one S.E. shock in the equation for EURODOL

Generalized Impulse Response(s) to one S.E. shock in the equation for EURODOL

0.05 0.00

0.01

–0.05

0.00

–0.10 –0.01 –0.15 –0.02

–0.20

–0.03

–0.25 0

6

12

18

24

0

Top 97.5% Level Point Estimate for MMREMU Lower 2.5% Level

6

12

18

24

Top 97.5% Level Point Estimate for IPEMU Lower 2.5% Level

Generalized Impulse Response(s) to one S.E. shock in the equation for OIL

Generalized Impulse Response(s) to one S.E. shock in the equation for OIL

0.004 0.002

0.005

0.000

0.000

–0.002 –0.005 –0.004 –0.010

–0.006

–0.015

–0.008 –0.010

–0.020 0

6

12

18

24

0

Top 97.5% Level Point Estimate for REREMU Lower 2.5% Level

6

12

18

24

Top 97.5% Level Point Estimate for HCPIEMU Lower 2.5% Level

Generalized Impulse Response(s) to one S.E. shock in the equation for OIL

Generalized Impulse Response(s) to one S.E. shock in the equation for OIL 0.03

0.4

0.02

0.3

0.01 0.2 0.00 0.1

–0.01

0.0

–0.02 0

6

12

18

24

Top 97.5% Level Point Estimate for MMREMU Lower 2.5% Level

Figure 2:

0

6

12

18

24

Top 97.5% Level Point Estimate for IPEMU Lower 2.5% Level

EMU12.

the Turkish economy. Provided that our model, as we think, is quite stable and accounts for the most significant economic facts, we argue that the Turkish economy sooner rather than later will face monetary and financial problems.

201

A VECMX* Analysis for the South-Eastern European Countries Generalized Impulse Response(s) to one S.E. shock in the equation for SBULRER

Generalized Impulse Response(s) to one S.E. shock in the equation for SBULRER

0.012 0.024

0.010

0.022 0.008

0.020 0.018

0.006

0.016

0.004

0.014 0.002

0.012 0.010

0.000 0

6

12

18

24

0

Top 97.5% Level Point Estimate for RERBUL Lower 2.5% Level

6

12

18

24

Top 97.5% Level Point Estimate for HCPIBUL Lower 2.5% Level

Generalized Impulse Response(s) to one S.E. shock in the equation for SBULRER

Generalized Impulse Response(s) to one S.E. shock in the equation for SBULRER

0.3

0.04

0.2

0.03

0.1

0.02

0.0

0.01

–0.1

0.00

–0.2

–0.01

–0.3

–0.02 0

6

12

18

24

0

Top 97.5% Level Point Estimate for MMRBUL Lower 2.5% Level

6

12

18

24

Top 97.5% Level Point Estimate for IPBUL Lower 2.5% Level

Generalized Impulse Response(s) to one S.E. shock in the equation for SBULHCPI

Generalized Impulse Response(s) to one S.E. shock in the equation for SBULHCPI

0.008

0.008 0.006

0.006

0.004

0.004

0.002

0.002

–0.002

0.000

–0.004

0.000

–0.006 –0.002

–0.008 0

6

12

18

0

24

Top 97.5% Level Point Estimate for RERBUL Lower 2.5% Level

6

12

18

24

Top 97.5% Level Point Estimate for HCPIBUL Lower 2.5% Level

Generalized Impulse Response(s) to one S.E. shock in the equation for SBULHCPI

Generalized Impulse Response(s) to one S.E. shock in the equation for SBULHCPI 0.04

0.1

0.03

0.0

0.02

–0.1

0.01

–0.2

0.00

–0.3

–0.01

0

6

12

18

Top 97.5% Level Point Estimate for MMRBUL Lower 2.5% Level

Figure 3:

24

0

6

12

18

24

Top 97.5% Level Point Estimate for IPBUL Lower 2.5% Level

Bulgaria (SBUL in Each Variable Denotes “Starred” Variable).

202

M. Koukouritakis et al. Generalized Impulse Response(s) to one S.E. shock in the equation for SBULMMR

Generalized Impulse Response(s) to one S.E. shock in the equation for SBULMMR

0.005

0.008

0.004

0.006

0.003

0.004

0.002 0.002 0.001 0.000

0.000

–0.002

–0.001 –0.002

–0.004 0

6

12

18

0

24

Top 97.5% Level Point Estimate for RERBUL Lower 2.5% Level

6

12

18

24

Top 97.5% Level Point Estimate for HCPIBUL Lower 2.5% Level

Generalized Impulse Response(s) to one S.E. shock in the equation for SBULMMR

Generalized Impulse Response(s) to one S.E. shock in the equation for SBULMMR 0.04

0.20

0.03 0.10 0.02 0.01

0.00

0.00 –0.10 –0.01 –0.02

–0.20 0

6

12

18

0

24

Top 97.5% Level Point Estimate for MMRBUL Lower 2.5% Level

6

12

18

24

Top 97.5% Level Point Estimate for IPBUL Lower 2.5% Level

Generalized Impulse Response(s) to one S.E. shock in the equation for SBULIP

Generalized Impulse Response(s) to one S.E. shock in the equation for SBULIP 0.008

0.006

0.006

0.004

0.004 0.002 0.002 0.000

0.000

–0.002

–0.002 –0.004

–0.004 0

6 12 24 18 Top 97.5% Level Point Estimate for RERBUL Lower 2.5% Level

0

6

12

18

24

Top 97.5% Level Point Estimate for HCPIBUL Lower 2.5% Level

Generalized Impulse Response(s) to one S.E. shock in the equation for SBULIP

Generalized Impulse Response(s) to one S.E. shock in the equation for SBULIP

0.3

0.08

0.2

0.07

0.1

0.06

0.0

0.05

–0.1

0.04

–0.2

0.03 0

6

12

18

Top 97.5% Level Point Estimate for MMRBUL Lower 2.5% Level

Figure 3: Continued.

24

0

6

12

18

Top 97.5% Level Point Estimate for IPBUL Lower 2.5% Level

24

203

A VECMX* Analysis for the South-Eastern European Countries Generalized Impulse Response(s) to one S.E. shock in the equation for SCRORER

Generalized Impulse Response(s) to one S.E. shock in the equation for SCRORER 0.006

0.006 0.005

0.004

0.004 0.002

0.003 0.002

0.000

0.001 –0.002

0.000

–0.004

–0.001 0

6

12

18

24

0

6

12

18

24

Top 97.5% Level Point Estimate for HCPICRO Lower 2.5% Level

Top 97.5% Level Point Estimate for RERCRO Lower 2.5% Level Generalized Impulse Response(s) to one S.E. shock in the equation for SCRORER

Generalized Impulse Response(s) to one S.E. shock in the equation for SCRORER 0.02

0.1 0.0 –0.1

0.01

–0.2 –0.3

0.00

–0.4 –0.5

–0.01

–0.6 –0.7

–0.02

–0.8 0

6

12

18

24

0

Top 97.5% Level Point Estimate for MMRCRO Lower 2.5% Level

6

12

18

24

Top 97.5% Level Point Estimate for IPCRO Lower 2.5% Level

Generalized Impulse Response(s) to one S.E. shock in the equation for SCROHCPI

Generalized Impulse Response(s) to one S.E. shock in the equation for SCROHCPI 0.012

0.007 0.006

0.010

0.005 0.008

0.004

0.006

0.003 0.002

0.004

0.001 0.002

0.000 –0.001

0.000 0

6

12

18

0

24

Top 97.5% Level Point Estimate for RERCRO Lower 2.5% Level Generalized Impulse Response(s) to one S.E. shock in the equation for SCROHCPI 0.05

0.2

0.04

0.0

0.03

–0.2

0.02

–0.4

0.01

–0.6

0.00

–0.8

–0.01 6

12

18

Top 97.5% Level Point Estimate for MMRCRO Lower 2.5% Level

Figure 4:

12

18

24

Generalized Impulse Response(s) to one S.E. shock in the equation for SCROHCPI

0.4

0

6

Top 97.5% Level Point Estimate for HCPICRO Lower 2.5% Level

24

0

6

12

18

24

Top 97.5% Level Point Estimate for IPCRO Lower 2.5% Level

Croatia (SCRO in Each Variable Denotes “Starred” Variable).

204

M. Koukouritakis et al. Generalized Impulse Response(s) to one S.E. shock in the equation for SCROMMR

Generalized Impulse Response(s) to one S.E. shock in the equation for SCROMMR

0.002

0.001 0.000

0.001 –0.001 0.000

–0.002

–0.001

–0.003 –0.004

–0.002 –0.005 –0.006

–0.003 0

6

12

18

24

0

Top 97.5% Level Point Estimate for RERCRO Lower 2.5% Level

6

12

18

24

Top 97.5% Level Point Estimate for HCPICRO Lower 2.5% Level

Generalized Impulse Response(s) to one S.E. shock in the equation for SCROMMR

Generalized Impulse Response(s) to one S.E. shock in the equation for SCROMMR

0.3

0.03

0.2

0.02

0.1 0.01 0.0 0.00 –0.1 –0.01

–0.2

–0.02

–0.3 0

6

12

18

0

24

Top 97.5% Level Point Estimate for MMRCRO Lower 2.5% Level

6

12

18

24

Top 97.5% Level Point Estimate for IPCRO Lower 2.5% Level

Generalized Impulse Response(s) to one S.E. shock in the equation for SCROIP

Generalized Impulse Response(s) to one S.E. shock in the equation for SCROIP 0.002

0.007 0.006

0.000

0.005

–0.002

0.004 –0.004 0.003 –0.006

0.002

–0.008

0.001

–0.010

0.000 0

6

12

18

0

24

Top 97.5% Level Point Estimate for RERCRO Lower 2.5% Level

6

12

18

24

Top 97.5% Level Point Estimate for HCPICRO Lower 2.5% Level Generalized Impulse Response(s) to one S.E. shock in the equation for SCROIP

Generalized Impulse Response(s) to one S.E. shock in the equation for SCROIP 0.2

0.08

0.0

0.07

–0.2 0.06 –0.4 0.05 –0.6 0.04

–0.8 –1.0

0.03 0

6

12

18

Top 97.5% Level Point Estimate for MMRCRO Lower 2.5% Level

Figure 4: Continued.

24

0

6

12

18

Top 97.5% Level Point Estimate for IPCRO Lower 2.5% Level

24

205

A VECMX* Analysis for the South-Eastern European Countries Generalized Impulse Response(s) to one S.E. shock in the equation for SCYPRER

Generalized Impulse Response(s) to one S.E. shock in the equation for SCYPRER

0.016

0.003

0.014

0.002 0.001

0.012

0.000 0.010 –0.001 0.008

–0.002

0.006

–0.003 –0.004

0.004 0

6

12

18

0

24

6

12

18

24

Top 97.5% Level Point Estimate for RERCYP Lower 2.5% Level

Top 97.5% Level Point Estimate for HCPICYP Lower 2.5% Level

Generalized Impulse Response(s) to one S.E. shock in the equation for SCYPRER

Generalized Impulse Response(s) to one S.E. shock in the equation for SCYPRER

0.20 0.02 0.10 0.01 0.00

0.00

–0.01 –0.10 –0.02 –0.03

–0.20 0

6

12

18

24

0

6

12

18

Top 97.5% Level Point Estimate for GBCYP Lower 2.5% Level

Top 97.5% Level Point Estimate for IPCYP Lower 2.5% Level

Generalized Impulse Response(s) to one S.E. shock in the equation for SCYPHCPI

Generalized Impulse Response(s) to one S.E. shock in the equation for SCYPHCPI

0.006

24

0.008 0.007

0.004

0.006 0.002

0.005

0.000

0.004 0.003

–0.002

0.002 –0.004

0.001

–0.006

0.000 0

6

12

18

24

0

6

12

24

18

Top 97.5% Level Point Estimate for RERCYP Lower 2.5% Level

Top 97.5% Level Point Estimate for HCPICYP Lower 2.5% Level

Generalized Impulse Response(s) to one S.E. shock in the equation for SCYPHCPI

Generalized Impulse Response(s) to one S.E. shock in the equation for SCYPHCPI

0.10 0.05 0.04 0.00

0.03 0.02 0.01

–0.10

0.00 –0.01 –0.02

–0.20 0

6

12

18

Top 97.5% Level Point Estimate for GBCYP Lower 2.5% Level

Figure 5:

24

0

6

12

18

24

Top 97.5% Level Point Estimate for IPCYP Lower 2.5% Level

Cyprus (SCYP in Each Variable Denotes “Starred” Variable).

206

M. Koukouritakis et al. Generalized Impulse Response(s) to one S.E. shock in the equation for SCYPMMR

Generalized Impulse Response(s) to one S.E. shock in the equation for SCYPMMR 0.006

0.004

0.005

0.002

0.004 0.000

0.003

–0.002

0.002 0.001

–0.004

0.000 –0.006

–0.001 –0.002

–0.008 0

6

12

18

0

24

6

12

18

Point Estimate for RERCYP Top 97.5% Level Lower 2.5% Level

Top 97.5% Level Point Estimate for HCPICYP Lower 2.5% Level

Generalized Impulse Response(s) to one S.E. shock in the equation for SCYPMMR

Generalized Impulse Response(s) to one S.E. shock in the equation for SCYPMMR

24

0.15 0.04 0.10 0.03 0.02

0.05

0.01 0.00

0.00 –0.01

–0.05

–0.02 –0.10

–0.03 0

6

12

18

0

24

Top 97.5% Level Point Estimate for GBCYP Lower 2.5% Level

6

12

18

24

Top 97.5% Level Point Estimate for IPCYP Lower 2.5% Level Generalized Impulse Response(s) to one S.E. shock in the equation for SCYPIP

Generalized Impulse Response(s) to one S.E. shock in the equation for SCYPIP 0.002

0.003

0.000

0.002

–0.002

0.001

–0.004

0.000

–0.006

–0.001

–0.008

–0.002

–0.010

–0.003

–0.012

–0.004 0

6

12

18

0

24

6

12

18

Top 97.5% Level Point Estimate for RERCYP Lower 2.5% Level

Top 97.5% Level Point Estimate for HCPICYP Lower 2.5% Level

Generalized Impulse Response(s) to one S.E. shock in the equation for SCYPIP

Generalized Impulse Response(s) to one S.E. shock in the equation for SCYPIP

0.05

24

0.10 0.09

0.00

0.08 0.07

–0.05

0.06 –0.10

0.05 0.04

–0.15

0.03 0.02

–0.20 0

6

12

18

Top 97.5% Level Point Estimate for GBCYP Lower 2.5% Level

Figure 5: Continued.

24

0

6

12

18

Top 97.5% Level Point Estimate for IPCYP Lower 2.5% Level

24

207

A VECMX* Analysis for the South-Eastern European Countries Generalized Impulse Response(s) to one S.E. shock in the equation for SGRERER

Generalized Impulse Response(s) to one S.E. shock in the equation for SGRERER 0.0030

0.009 0.0020 0.008 0.007

0.0010

0.006 0.0000 0.005 0.004

–0.0010 0

6

12

18

24

0

6

12

18

24

Top 97.5% Level Point Estimate for RERGRE Lower 2.5% Level

Top 97.5% Level Point Estimate for HCPIGRE Lower 2.5% Level

Generalized Impulse Response(s) to one S.E. shock in the equation for SGRERER

Generalized Impulse Response(s) to one S.E. shock in the equation for SGRERER 0.03

0.10

0.02 0.00 0.01 0.00 –0.10 –0.01 –0.20

–0.02 0

6

12

18

24

0

Top 97.5% Level Point Estimate for TBGRE Lower 2.5% Level

6

12

18

24

Top 97.5% Level Point Estimate for IPGRE Lower 2.5% Level

Generalized Impulse Response(s) to one S.E. shock in the equation for SGREHCPI

Generalized Impulse Response(s) to one S.E. shock in the equation for SGREHCPI 0.004

0.005 0.004

0.003

0.003 0.002

0.002

0.001

0.001

0.000 0.000

–0.001 –0.002

0

6

12

18

24

–0.001

0

6

12

18

24

Top 97.5% Level Point Estimate for RERGRE Lower 2.5% Level

Top 97.5% Level Point Estimate for HCPIGRE Lower 2.5% Level

Generalized Impulse Response(s) to one S.E. shock in the equation for SGREHCPI

Generalized Impulse Response(s) to one S.E. shock in the equation for SGREHCPI 0.04

0.2

0.03 0.1 0.02 0.01

0.0

0.00 –0.1 –0.01 –0.02

–0.2 0

6

12

18

Top 97.5% Level Point Estimate for TBGRE Lower 2.5% Level

Figure 6:

24

0

6

12

18

24

Top 97.5% Level Point Estimate for IPGRE Lower 2.5% Level

Greece (SGRE in Each Variable Denotes “Starred” Variable).

208

M. Koukouritakis et al. Generalized Impulse Response(s) to one S.E. shock in the equation for SGREMMR

Generalized Impulse Response(s) to one S.E. shock in the equation for SGREMMR

0.002

0.0010 0.0005

0.001

0.0000 0.000 –0.0005 –0.001

–0.0010

–0.002

–0.0015 0

6

12

18

0

24

6

12

18

24

Top 97.5% Level Point Estimate for HCPIGRE Lower 2.5% Level

Top 97.5% Level Point Estimate for RERGRE Lower 2.5% Level

Generalized Impulse Response(s) to one S.E. shock in the equation for SGREMMR

Generalized Impulse Response(s) to one S.E. shock in the equation for SGREMMR 0.02 0.10 0.01

0.08 0.06

0.00

0.04 0.02 0.00

–0.01

–0.02 –0.04

–0.02 0

6

12

18

24

0

6

12

18

Top 97.5% Level Point Estimate for TBGRE Lower 2.5% Level

Top 97.5% Level Point Estimate for IPGRE Lower 2.5% Level

Generalized Impulse Response(s) to one S.E. shock in the equation for SGREIP

Generalized Impulse Response(s) to one S.E. shock in the equation for SGREIP

0.005

0.003

0.004

0.002

0.003

0.001

0.002

0.000

0.001

–0.001

0.000

24

–0.002 0

6

12

18

24

0

6

12

18

Top 97.5% Level Point Estimate for RERGRE Lower 2.5% Level

Top 97.5% Level Point Estimate for HCPIGRE Lower 2.5% Level

Generalized Impulse Response(s) to one S.E. shock in the equation for SGREIP

Generalized Impulse Response(s) to one S.E. shock in the equation for SGREIP

24

0.30 0.08 0.07

0.20

0.06 0.05

0.10

0.04 0.03

0.00

0.02 –0.10

0.01 0

6

12

18

Top 97.5% Level Point Estimate for TBGRE Lower 2.5% Level

Figure 6: Continued.

24

0

6

12

18

Top 97.5% Level Point Estimate for IPGRE Lower 2.5% Level

24

209

A VECMX* Analysis for the South-Eastern European Countries Generalized Impulse Response(s) to one S.E. shock in the equation for SROMRER

Generalized Impulse Response(s) to one S.E. shock in the equation for SROMRER 0.003

0.010

0.002 0.005

0.001 0.000

0.000 –0.001 –0.002

–0.005

–0.003 –0.004

–0.010 0

6

12

18

0

24

Top 97.5% Level Point Estimate for RERROM Lower 2.5% Level

6

12

18

24

Top 97.5% Level Point Estimate for HCPIROM Lower 2.5% Level

Generalized Impulse Response(s) to one S.E. shock in the equation for SROMRER

Generalized Impulse Response(s) to one S.E. shock in the equation for SROMRER

0.2

0.02

0.0 –0.2

0.01

–0.4 –0.6

0.00

–0.8 –1.0

–0.01

–1.2 –1.4

–0.02 0

6

12

18

0

24

6

12

18

24

Top 97.5% Level Point Estimate for IPROM Lower 2.5% Level

Top 97.5% Level Point Estimate for MMRROM Lower 2.5% Level

Generalized Impulse Response(s) to one S.E. shock in the equation for SROMHCPI

Generalized Impulse Response(s) to one S.E. shock in the equation for SROMHCPI 0.015

0.010 0.008

0.010 0.006 0.005

0.004 0.002

0.000 0.000 –0.002

–0.005 0

6

12

18

24

0

Top 97.5% Level Point Estimate for RERROM Lower 2.5% Level

6

12

18

24

Top 97.5% Level Point Estimate for HCPIROM Lower 2.5% Level Generalized Impulse Response(s) to one S.E. shock in the equation for SROMHCPI

Generalized Impulse Response(s) to one S.E. shock in the equation for SROMHCPI 0.5

0.03

0.0

0.02

–0.5

0.01

–1.0

0.00

–1.5

–0.01

–2.0

–0.02 0

6

12

18

Top 97.5% Level Point Estimate for MMRROM Lower 2.5% Level

Figure 7:

24

0

6

12

18

24

Top 97.5% Level Point Estimate for IPROM Lower 2.5% Level

Romania (SROM in Each Variable Denotes “Starred” Variable).

210

M. Koukouritakis et al. Generalized Impulse Response(s) to one S.E. shock in the equation for SROMMMR

Generalized Impulse Response(s) to one S.E. shock in the equation for SROMMMR

0.008

0.006

0.006 0.004

0.004 0.002

0.002

0.000 0.000

–0.002 –0.004

–0.002

–0.006 –0.008

–0.004 0

6

12

18

24

0

6

12

18

24

Top 97.5% Level Point Estimate for RERROM Lower 2.5% Level

Top 97.5% Level Point Estimate for HCPIROM Lower 2.5% Level

Generalized Impulse Response(s) to one S.E. shock in the equation for SROMMMR

Generalized Impulse Response(s) to one S.E. shock in the equation for SROMMMR

1.4

0.03

1.2 0.02

1.0 0.8

0.01

0.6 0.00

0.4 0.2

–0.01

0.0 –0.2

–0.02 0

6

12

18

24

0

Top 97.5% Level Point Estimate for MMRROM Lower 2.5% Level

6

12

18

24

Top 97.5% Level Point Estimate for IPROM Lower 2.5% Level

Generalized Impulse Response(s) to one S.E. shock in the equation for SROMIP

Generalized Impulse Response(s) to one S.E. shock in the equation for SROMIP

0.010

0.010

0.008 0.008

0.006 0.004

0.006

0.002 0.004

0.000 –0.002

0.002

–0.004 –0.006

0.000 0

6

12

18

24

0

6

12

18

Top 97.5% Level Point Estimate for RERROM Lower 2.5% Level

Top 97.5% Level Point Estimate for HCPIROM Lower 2.5% Level

Generalized Impulse Response(s) to one S.E. shock in the equation for SROMIP

Generalized Impulse Response(s) to one S.E. shock in the equation for SROMIP

24

0.4 0.2

0.05

0.0 0.04 –0.2 0.03 –0.4 0.02

–0.6 –0.8

0.01 0

6

12

18

24

Top 97.5% Level Point Estimate for MMRROM Lower 2.5% Level

Figure 7: Continued.

0

6

12

18

Top 97.5% Level Point Estimate for IPROM Lower 2.5% Level

24

211

A VECMX* Analysis for the South-Eastern European Countries Generalized Impulse Response(s) to one S.E. shock in the equation for SSLORER

Generalized Impulse Response(s) to one S.E. shock in the equation for SSLORER 0.003

0.009 0.008

0.002

0.007 0.001

0.006 0.005

0.000

0.004 –0.001 0.003 –0.002

0.002 0

6

12

18

24

0

6

12

18

24

Top 97.5% Level Point Estimate for RERSLO Lower 2.5% Level

Top 97.5% Level Point Estimate for HCPISLO Lower 2.5% Level

Generalized Impulse Response(s) to one S.E. shock in the equation for SSLORER

Generalized Impulse Response(s) to one S.E. shock in the equation for SSLORER

0.10 0.01 0.00 0.00 –0.01 –0.10 –0.02 –0.20

–0.03 0

6

12

18

24

0

Top 97.5% Level Point Estimate for MMRSLO Lower 2.5% Level

6

12

18

24

Top 97.5% Level Point Estimate for IPSLO Lower 2.5% Level

Generalized Impulse Response(s) to one S.E. shock in the equation for SSLOHCPI

Generalized Impulse Response(s) to one S.E. shock in the equation for SSLOHCPI

0.003

0.007

0.002

0.006

0.001

0.005

0.000

0.004

–0.001

0.003

–0.002

0.002 0.001

–0.003 –0.004

0.000 0

6

12

18

24

0

Top 97.5% Level Point Estimate for RERSLO Lower 2.5% Level

6

12

18

24

Top 97.5% Level Point Estimate for HCPISLO Lower 2.5% Level

Generalized Impulse Response(s) to one S.E. shock in the equation for SSLOHCPI

Generalized Impulse Response(s) to one S.E. shock in the equation for SSLOHCPI 0.05

0.3

0.04

0.2

0.03 0.1 0.02 0.0 0.01 –0.1

0.00

–0.2

–0.01 0

6

12

18

Top 97.5% Level Point Estimate for MMRSLO Lower 2.5% Level

24

0

6

12

18

24

Top 97.5% Level Point Estimate for IPSLO Lower 2.5% Level

Figure 8: Slovenia (SSLO in Each Variable Denotes “Starred” Variable).

212

M. Koukouritakis et al. Generalized Impulse Response(s) to one S.E. shock in the equation for SSLOMMR

Generalized Impulse Response(s) to one S.E. shock in the equation for SSLOMMR 0.002

0.003 0.002

0.001

0.001 0.000 0.000 –0.001

–0.001

–0.002

–0.002 0

6

12

18

24

0

Top 97.5% Level Point Estimate for RERSLO Lower 2.5% Level

6

12

18

24

Top 97.5% Level Point Estimate for HCPISLO Lower 2.5% Level

Generalized Impulse Response(s) to one S.E. shock in the equation for SSLOMMR

Generalized Impulse Response(s) to one S.E. shock in the equation for SSLOMMR 0.04

0.08

0.03

0.06 0.02

0.04

0.01

0.02 0.00

0.00

–0.02 –0.01

–0.04

–0.02

–0.06 0

6

12

18

0

24

6

12

18

24

Top 97.5% Level Point Estimate for IPSLO Lower 2.5% Level

Top 97.5% Level Point Estimate for MMRSLO Lower 2.5% Level

Generalized Impulse Response(s) to one S.E. shock in the equation for SSLOIP

Generalized Impulse Response(s) to one S.E. shock in the equation for SSLOIP 0.004

0.006

0.003

0.005 0.004

0.002 0.003 0.001 0.002 0.000

0.001

–0.001

0.000 0

6

12

18

24

0

Top 97.5% Level Point Estimate for RERSLO Lower 2.5% Level

6

12

18

24

Top 97.5% Level Point Estimate for HCPISLO Lower 2.5% Level

Generalized Impulse Response(s) to one S.E. shock in the equation for SSLOIP

Generalized Impulse Response(s) to one S.E. shock in the equation for SSLOIP 0.11

0.20

0.10 0.09 0.10

0.08 0.07 0.06

0.00

0.05 0.04 0.03

–0.10 0

6

12

18

Top 97.5% Level Point Estimate for MMRSLO Lower 2.5% Level

Figure 8: Continued.

24

0

6

12

18

Top 97.5% Level Point Estimate for IPSLO Lower 2.5% Level

24

213

A VECMX* Analysis for the South-Eastern European Countries Generalized Impulse Response(s) to one S.E. shock in the equation for STURRER

Generalized Impulse Response(s) to one S.E. shock in the equation for STURRER 0.008

0.05

0.006

0.04

0.004 0.03 0.002 0.02 0.000 0.01

–0.002

0.00

–0.004 –0.006

–0.01 0

6

12

18

0

24

Top 97.5% Level Point Estimate for RERTUR Lower 2.5% Level

6

12

18

24

Top 97.5% Level Point Estimate for HCPITUR Lower 2.5% Level

Generalized Impulse Response(s) to one S.E. shock in the equation for STURRER

Generalized Impulse Response(s) to one S.E. shock in the equation for STURHCPI 0.10 0.08

0.01

0.06

0.00

0.04 –0.01 0.02 –0.02

0.00

–0.03

–0.02 0

6

12

18

24

0

Top 97.5% Level Point Estimate for IPTUR Lower 2.5% Level

6

12

18

24

Top 97.5% Level Point Estimate for RERTUR Lower 2.5% Level

Generalized Impulse Response(s) to one S.E. shock in the equation for STURHCPI

Generalized Impulse Response(s) to one S.E. shock in the equation for STURHCPI

0.025 0.02

0.020

0.01 0.015 0.00 0.010 –0.01 0.005

–0.02

0.000

–0.03 0

6

12

18

24

0

Top 97.5% Level Point Estimate for HCPITUR Lower 2.5% Level

6

12

18

24

Top 97.5% Level Point Estimate for IPTUR Lower 2.5% Level

Generalized Impulse Response(s) to one S.E. shock in the equation for STURIP

Generalized Impulse Response(s) to one S.E. shock in the equation for STURIP

0.08

0.025

0.07 0.020

0.06 0.05

0.015

0.04 0.03

0.010

0.02 0.01

0.005

0.00 –0.01

0.000 0

6

12

18

Top 97.5% Level Point Estimate for RERTUR Lower 2.5% Level

24

0

6

12

18

24

Top 97.5% Level Point Estimate for HCPITUR Lower 2.5% Level

Figure 9: Turkey (STUR in Each Variable Denotes “Starred” Variable).

214

M. Koukouritakis et al. Generalized Impulse Response(s) to one S.E. shock in the equation for STURIP 0.04 0.03 0.02 0.01 0.00 –0.01 0

6

12

18

24

Top 97.5% Level Point Estimate for IPTUR Lower 2.5% Level

Figure 9: Continued.

Concluding Remarks In this paper we assessed the impact of the Eurozone’s economic policies on specific South-Eastern European countries, namely Bulgaria, Croatia, Cyprus, Greece, Romania, Slovenia, and Turkey. To carry out our analysis, we used VECMX*s models, which allow the inclusion of nonstationary foreign variables that are treated as (weakly) exogenous. This approach seems quite appropriate, since it allows for the interdependencies that exist between national and international factors in a consistent manner. In general, our results indicate that for the transition economies of South-Eastern Europe, namely Bulgaria, Croatia, Romania, and Slovenia, changes in international (i.e., Eurozone’s) macroeconomic policies have expected effects on their domestic variables. Similar and expected results are obtained for Cyprus and Greece, but for these two countries these results are very small in magnitude. Also for Turkey, changes in Eurozone’s macroeconomic policies have expected results on the country’s IP, but in the cases of domestic RER and HCPI the GRIFs do not converge to a stable level in the time horizon that we have used. This anomaly could possibly be attributed to the strong inflationary tendencies in the Turkish economy. Overall, the above results indicate that there are linkages (a) among the economies of the South-Eastern Europe, and (b) between each of these economies and the Eurozone. Our evidence also implies that the international (i.e., Eurozone’s) economic policies affect the EU or Eurozone members of this region in the same way. On the other hand, the Turkish economy seems to behave relatively differently to Eurozone’s macroeconomic policies.

References Assenmacher-Wesche, K., & Pesaran, M. H. (2009). A VECX* model of the Swiss economy. Swiss National Bank Economic Studies, No. 6.

A VECMX* Analysis for the South-Eastern European Countries

215

Bussie´re, M., Chudik, A., & Sestieri, G. (2012). Modelling global trade flows: Results from a GVAR model. Federal Reserve Bank of Dallas Working Paper Series, No. 119. Retrieved from http://www.dallasfed.org/assets/documents/institute/wpapers/2012/0119.pdf Coricelli, F., E´gert, B., & MacDonald, R. (2006). Monetary transmission mechanism in central and Eastern Europe: Gliding on a wind of change. Discussion Paper 8. Bank of Finland, Institute for Economies in Transition. Retrieved from http://www. suomenpankki.fi/bofit_en/tutkimus/tutkimusjulkaisut/dp/Documents/dp0806.pdf Coricelli, F., Jazbec, B., & Masten, I. (2003). Exchange rate pass-through in acceding countries: The role of exchange rate regimes. Journal of Banking and Finance, 30, 1375
1391. Crespo-Cuaresma, J., E´gert, B., & Reininger, T. (2004). Interest rate pass-through in new EU member states: The case of the Czech Republic, Hungary and Poland. William Davidson Institute Working Paper, No. 671. Retrieved from http://wdi. umich.edu/files/publications/workingpapers/wp671.pdf Darvas, Z. (2001). Exchange rate pass-through and real exchange rate in EU candidate countries. Deutsche Bundesbank Discussion Paper, No. 10. Retrieved from http://www.bundesbank.de/Redaktion/EN/Downloads/Publications/Discussion_ Paper_1/2001/2001_07_19_dkp_10.pdf?__blob=publicationFile De´es, S., di Mauro, F. Pesaran, M. H., & Smith, L. V. (2005). Exploring the international linkages of the Euro Area: A global VAR analysis. European Central Bank Working Paper Series, No. 568. Retrieved from http://www.ecb.europa.eu/pub/ pdf/scpwps/ecbwp568.pdf De´es, S., & Sain-Guilhem, A. (2009). The role of the United States in the global economy and its evolution over time. European Central Bank Working Paper Series, No. 1034. Retrieved from http://www.ecb.europa.eu/pub/pdf/scpwps/ ecbwp1034.pdf E´gert, B., Crespo-Cuaresma, J., & Reininger, T. (2006). Interest rate pass-through in central and Eastern Europe: Reborn from ashes to pass away? Focus on European Economic Integration, No. 1. Feldkircher, M., & Korhonen, I. (2012). The rise of China and its implications for emerging markets
evidence from a GVAR model. Bank of Finland Discussion Papers, No. 20. Retrieved from http://www.suomenpankki.fi/pdf/170909.pdf Galesi, A., & Sgherri, S. (2009). Regional financial spillovers across Europe: A global VAR analysis. IMF Working Paper Series, No. 23. Retrieved from http://www. imf.org/external/pubs/ft/wp/2009/wp0923.pdf Gueorguiev, N. (2003). Exchange rate pass-through in Romania. IMF Working Paper, No. 130. Retrieved from http://www.imf.org/external/pubs/ft/wp/2003/ wp03130.pdf Horva´th, C., Kreko´, J., & Naszo´di, A. (2004). Interest rate pass-through: The case of Hungary. National Bank of Hungary Working Paper, No. 8. Retrieved from http://english.mnb.hu/Root/Dokumentumtar/ENMNB/Kiadvanyok/mnben_ mnbfuzetek/mnben_mf_2004_8/wp2004_8.pdf Kara, H., Ku¨c¸u¨k Tuger, H., O¨zlale, U¨., Tuger, B., Yavuz, D., & Yu¨cel, E. M. ˘ ˘ (2005). Exchange rate pass-through in Turkey: Has it changed and to what extent? Central Bank of the Republic of Turkey Working Paper, No. 4. Retrieved from http://www.tcmb.gov.tr/research/discus/WP0504ENG.pdf

216

M. Koukouritakis et al.

Korhonen, I., & Wachtel, P. (2005). A note on exchange rate pass-through in CIS countries. Bank of Finland Discussion Paper, No. 2. Retrieved from http://www. suomenpankki.fi/bofit_en/tutkimus/tutkimusjulkaisut/dp/Documents/dp0205.pdf Koop, G., Pesaran, M. H., & Potter, S. M. (1996). Impulse response analysis in nonlinear multivariate models. Journal of Econometrics, 74, 119
147. Kwiatkowski, D., Phillips, P. C. B., Schmidt, P., & Shin, Y. (1992). Testing the null hypothesis of stationarity against the alternative of a unit root. Journal of Econometrics, 54, 159
178. Mihaljek, D., & Klau, M. (2001). A note on the pass-through from exchange rate and foreign price changes to inflation in selected emerging market economies. Bank of International Settlements Paper, No. 8. Retrieved from http://www.bis.org/publ/ bppdf/bispap08c.pdf Opiela, T. P. (1999). The responsiveness of loan rates to monetary policy in Poland: The effects of bank structure. National Bank of Poland Materials and Studies, No. 12. Pesaran, M. H., & Shin, Y. (1996). Cointegration and the speed of convergence to equilibrium. Journal of Econometrics, 71, 117
143. Pesaran, M. H., & Shin, Y. (1998). Generalized impulse response analysis in linear multivariate models. Economics Letters, 58, 17
29. Pesaran, M. H., Schuermann, T., & Weiner, S. M. (2004). Modeling regional interdependencies using a global error-correcting macroeconomic model. Journal of Business and Economic Statistics, 22, 129
162. Pesaran, M. H., Shin, Y., & Smith, R. J. (2000). Structural analysis of vector error correction models with exogenous I(1) variables. Journal of Econometrics, 97, 293
343. Sander, H., & Kleimeier, S. (2004). Interest rate pass-through in an enlarged Europe: The role of banking market structure for monetary policy transmission in transition economies. University of Maastricht, METEOR Research Memoranda, No. 45, Maastricht. Sims, C. A. (1980). Macroeconomics and reality. Econometrica, 48, 1
48.

Predicting Economic Activity with Financial Market Data in a Small Open Economy: Revisiting Stylized Facts During Economic Turbulence Petri Kuosmanen and Juuso Vataja Department of Economics, University of Vaasa, FIN-65101 Vaasa, Finland, e-mail: [email protected]

Abstract Purpose
This paper examines the predictive content of financial variables above and beyond past GDP growth in a small open economy in the Eurozone. We aim to clarify potential differences in forecasting economic activity during periods of steady growth and economic turbulence. Design/methodology/approach
The out-of-sample forecasting analysis is conducted recursively for two different time periods: the steady growth period from 2004:1 to 2007:4 and the financial crisis period from 2008:1 to 2011:2. Findings
Our results from Finland suggest that the proper choice of forecasting variables relates to general economic conditions. During steady economic growth, the preferable financial indicator is the short-term interest rate combined with past growth. However, during economic turbulence, the traditional term spread and stock returns are more important in forecasting GDP growth. Research limitations/implications
The results highlight the importance of long-term interest rates in determining the level of the term spread when the central bank implements a zero interest rate policy. Moreover, during

International Symposia in Economic Theory and Econometrics, Vol. 23 G. P. Kouretas and A. P. Papadopoulos (Editors) Copyright r 2014 by Emerald Group Publishing Limited. All rights reserved ISSN: 1571-0386/doi:10.1108/S1571-038620140000023008

218

P. Kuosmanen and J. Vataja

economic turbulence, stock markets are able to signal the expected effects of unconventional monetary policy on GDP growth. Keywords: Term spread, short-term interest rate, stock market, forecasting, macroeconomy JEL Classifications: E37, E44, E47

Introduction What will be the weather in the next six hours, tomorrow, or next week? In absence of information on the current state of the atmosphere, barometric pressure, or the pressure tendency, probably the best starting point is likely to be that the weather will be the same as it is now. When the weather is in a steady state, it is a reasonable to predict that the current weather condition will persist. What will be the rate of GDP growth in your country in the next quarter or year? The chaotic natures of the atmosphere and the economy imply that substantial human input is necessary to select the most relevant predictors and forecast models. An economist would certainly prefer to have other predictors of economic growth than past performance. Financial market data are forward-looking aggregators of information that are easy to interpret and can be observed in real time without measurement errors. Therefore, the potential to use financial market information to forecast future economic activity has been actively explored, especially in the U.S. context. Financial variables such as interest rates, term spreads, and stock returns are examples of readily available and precise indicators; however, whether these variables can provide useful forecasts of future economic activity during both steady growth and more turbulent conditions remains an open question. Since the late 1980s, many studies have documented the usefulness of the yield curve or even the simple term spread for predicting economic activity (e.g., Estrella, 2005; Estrella & Hardouvelis, 1991; Harvey, 1988; Laurent, 1989; Stock & Watson, 2003). It has become standard in the United States to use the term spread between the 10-year Treasury note and the 3-month Treasury bill to predict recessions and future economic activity (e.g., Estrella & Mishkin, 1996; Haubrich & Dombrosky, 1996). The inversion of the term spread has been demonstrated to be a reliable “advance warning” of a subsequent recession; however, the ability of the term spread to forecast GDP growth rates is less clear. Many studies have found that since 1985, the term spread has been a less accurate predictor of U.S. output growth (e.g., Chinn & Kucko, 2010; Stock & Watson, 2003). This phenomenon

Revisiting Stylized Facts During Economic Turbulence

219

may reflect either the increased stability of output growth (the Great Moderation) and other macroeconomic variables since the mid-1980s or changes in the responsiveness of monetary policy to output growth and inflation (Wheelock & Wohar, 2009). If the central bank concentrates exclusively on controlling inflation, the term spread will most likely be a less accurate predictor of GDP growth (Stock & Watson, 2003). Thus, as the European Central Bank (ECB) focuses on controlling inflation, the term spread may not necessarily be the best single predictor of economic growth in the euro area. However, despite evidence that parameter instability may weaken the performance of the term spread as a predictor of growth, the term spread has gained wide acceptance as the single best indicator of economic activity (e.g., Estrella, 2005). Notwithstanding the predominance of this indicator, Ang, Piazzesi, and Wei (2006) found that the short-term interest rate had more predictive power for U.S. GDP growth during the period 1952
2001 than did any term spread. Stock prices are forward looking and thus represent another obvious financial indicator of future economic activity. Economists and investors have a well-known rule of thumb that stock market prices predict economic growth approximately half a year in advance. However, compared with the predictive content of the term spread, less empirical evidence exists regarding the ability of stock prices to predict economic performance (e.g., Stock & Watson, 2003). Chionis, Gogas, and Pragidis (2010) found that augmenting the yield curve with a stock index significantly improved the ability to predict GDP fluctuations in the euro area. Nyberg’s (2010) results supported this conclusion with respect to predicting recessions in Germany and the United States. Junttila and Korhonen (2011) discovered that both stock market dividend yields and short-term interest rates were relevant variables for future economic activity in the United Kingdom, the euro area, and Japan, particularly during turbulent times. Furthermore, Henry, Olekalns, and Thong (2004) emphasized that stock returns predict economic growth when the economy is contracting but that the predictive power of stock returns in non-recession periods is less clear. This mixed evidence is expressed in Samuelson’s (1966) famous note, “The stock market has predicted nine out of the last five recessions.” In any event, economic turbulence tends to strengthen the link between the stock market and economic activity. The case of Finland is interesting in many ways. The vast majority of the previous literature has examined larger (particularly G7) countries; however, the predictive content of financial variables is less known in smaller European countries. As a member of the Economic and Monetary Union (EMU), the Finnish economy is subject to the monetary policy of the ECB, which aggressively targets inflation. It has been argued that the predictive content of the term spread with respect to economic growth may

220

P. Kuosmanen and J. Vataja

weaken if inflation control is the primary concern of the central bank. Moreover, the monetary policy of the ECB is conducted on the basis of the entire euro area; therefore, interest rates in the euro area may be far from optimal for smaller euro countries that face asymmetric shocks. Indeed, evidence suggests that output shocks have been more country-specific in Finland than in other EU countries (e.g., Haaparanta & Peisa, 1997; Kinnunen, 1998), and the question of asymmetric shocks was among the central concerns when Finland considered EMU membership in the late 1990s. Therefore, the case of Finland is of particular importance for small EU countries that are considering joining the Eurozone. Moreover, of the Nordic countries Finland is the only economy that is subject to common European monetary policy. In general, if common monetary policy is found to be unsuitable for small Eurozone countries, expansion of the Eurozone may be less likely; in fact, the disintegration of the euro area is the more likely option. After recovering from an economic depression during the 1990s, Finland experienced continuous and sound growth until the global financial crisis plunged the Finnish economy into a deep recession at the end of 2008. A distinctive feature of this slump was its severity; in a single year, Finland’s GDP collapsed by an astonishing 10%, one of the largest decreases of economic activity among developed countries. Undoubtedly, the ups and downs of the Finnish economy pose a true challenge for forecasting economic activity. This paper contributes to the existing literature by explicitly addressing the predictive content of the classical term spread versus the short-term interest rate and stock returns in the context of a small open economy (SOE). Ang et al. (2006) found that compared with the term spread, short-term interest rates were a better predictor of economic activity in the U.S. context. Our aim is to test whether this result is specific to the United States or whether it also holds for smaller countries. Furthermore, we seek to clarify potential differences in forecasting economic activity between eras of steady growth and periods of economic turbulence, such as the financial and debt crises in Europe. Much of the previous literature has concentrated on the predictive content of a single financial indicator (e.g., Stock & Watson, 2003); however, we assess the predictive content of combinations of indicators. More broadly, this paper provides further information on the predictive ability of financial market indicators in smaller economies, a context that has rarely been examined in the previous literature. The remainder of this paper is organized as follows. In the second section we present the model setup and the data. The third section contains the empirical analysis of the study, and the fourth section concludes the paper.

Revisiting Stylized Facts During Economic Turbulence

221

The Model Setup and Data Forecasting Models In accordance with the previous literature, our financial market dataset consists of the following financial market variables: the term spread, stock returns, and the short-term interest rate. We use the following modeling strategy when constructing empirical forecasting models. Because we are interested in the forecasting ability of financial market indicators above and beyond past values of GDP growth, the informational content of financial indicators is combined with past economic growth (Equation (1)). However, during exceptional and turbulent periods, past economic activity may lose its predictive content. Therefore, the forecasting performance of Equation (1) is also compared to the predictive content of financial market information alone (Equation (2)). Finally, the forecasting ability of these models is compared to that of a simple autoregressive model, which serves as our benchmark (Equation (3)). lnyt þ h − lnyt Þ = α1 þ β01 Xt þ

Xm

γ 1 Δlnyt − i þ 1 i=1 i

þ u1t þ h ;

ðlnyt þ h − lnyt Þ = α2 þ β02 Xt þ u2t þ h ; ðlnyt þ h − lnyt Þ = α3 þ

Xm

γ 3 Δlnyt − i þ 1 i=1 i

ð1Þ ð2Þ

þ u3t þ h ;

ð3Þ

where y is the level of real GDP; X is the vector of financial market indicators consisting of the term spread (TS), stock returns (R), and the shortterm interest rate (i), that is, X = (TS, R, i)0 ; α is a constant term; β0 and γi are parameter estimates; and ut þ h;t is the error term. The subscript h refers to the forecast horizon. We focus on forecasts at the one-, two-, and four-quarter horizons (h = 1, 2, and 4). Much of the previous literature has only considered a single financial market indicator as a forecasting variable; however, in our view, a single indicator may be overly limited, and we therefore explore the forecasting ability of various combinations of financial indicators. Moreover, we consider the term spread and the short-term interest rate as alternative forecasting variables and thus do not include them in the same forecasting model. As is standard practice, we assume that all of the relevant predictive information for the financial market variables is included in the most recent observations of the financial indicators; therefore, lagged values of financial indicators are not included in the forecasting models. Regarding the specification of the dependence on past GDP growth, we assume that the relevant length of history dependence is directly related to

222

P. Kuosmanen and J. Vataja

the forecasting horizon.1 That is, when forecasting GDP growth one quarter ahead, we use the past quarter’s growth as the most relevant measure of history dependence. Thus, when we forecast GDP growth two quarters ahead, Δlnyt and Δlnyt − 1 are included, and for the four-quarter forecasting horizon, Δlnyt ; Δlnyt − 1 ; Δlnyt − 2 ; and Δlnyt − 3 are included. We begin the forecasting analysis with only one financial indicator and gradually proceed to richer model specifications until all of the relevant financial market forecasting variables are included in the forecasting model. This approach produces a total of 11 model specifications, including the autoregressive benchmark model.

Data The data are quarterly and span the 1988:1
2011:2 time period. Economic activity is measured using the logarithmic changes in the Finnish real GDP index. Nominal quarterly stock market returns are calculated as logarithmic changes in the Finnish general stock market index (OMX Helsinki PI). The short-term interest rate is the 3-month market rate. The term spread is constructed by calculating the difference between the 10-year government bond yields and the 3-month interest rates. The details of the data and data transformations are presented in Table 1. The time series properties of the data were explored using the two most efficient unit root tests, the DF-GLS test developed by Elliot, Rothenberg,

Table 1: The Data ðlnyt þ h − lnyt Þ × 100 Δlny = ðlnyt − lnyt − 1 Þ × 100 Rt = ðlnpt − lnpt − 1 Þ × 100 TSt=i10t
i3t

y = the Finnish gross domestic product index (volume, market prices). Source: OECD Economic Outlook database. p = the Finnish general stock market index (OMX Helsinki PI). Source: OECD Main Economic Indicators database. i10 = the Finnish 10-year government bond yield. i3 = the Finnish 3-month interest rate (1988:1
1998:4 Helibor 3; 1999:1
2011:2 Euribor 3). Source: OECD Main Economic Indicators database.

1 Alternatively, the number of AR terms could have been selected based on information criteria. We also conducted the analysis using the AIC criteria, which consistently suggested the AR(4) specification for Equations (1) and (3) irrespective of the forecasting horizon. Because the forecasting performance was worse, we preferred fixing the AR terms to the number of forecasting quarters. The results are available upon request.

Revisiting Stylized Facts During Economic Turbulence

223

and Stock (1996) and the Ng and Perron (2001) tests. The test results consistently suggested that all of the variables except for the short-term interest rate were stationary. The short-term interest rates were found to be nonstationary for the entire sample period; however, during the period of Finnish membership in the EMU (1999:1
), the test results suggested that the short-term interest rate was stationary.2 The nonstationary nature of the short-term interest rate for the entire sample period is likely to be reflective of the exceptionally high interest rates in the late 1980s and the beginning of the 1990s, which were caused by inflationary pressures and the defense of the national currency during the ERM crisis. Because the forecasting analysis consider EMU period, we estimated the forecasting models with short-term interest rates specified in levels.3 Annual GDP growth in Finland is presented in Figure 1. The dependence of economic activity on past growth is evident. The financial market indicators are illustrated in Figure 2. During the sample period, the Finnish economy experienced two major recessions, which are indicated by the shaded areas in Figure 2. It is noteworthy that the negative term spread (an inverted yield curve) provided an early warning of both recessions. Table 2 presents the descriptive statistics for the entire sample period (1988:1
2011:2) and the forecasting periods of the study (2004:1
2007:4 and 2008:1
2011:4). The former forecasting period is intended to represent 8

Forecast period I Steady growth

Forecast period II Turbulence

4 0 –4 –8 –12 88

90

92

94

96

98

00

02

04

06

08

10

Figure 1: Annual GDP Growth in Finland and the Forecast Periods.

2

The unit root test results are available upon request. We also estimated the models using the first differences of the short-term interest rate; in general, the level-based specifications exhibited better forecasting performance. 3

224 16

P. Kuosmanen and J. Vataja

(A)

TS

R

(B)

i

OMX Helsinki

12 8

6 5

4

2 1 0 –1 –2

4 0

3 2

–4 88 90 92 94 96 98 00 02 04 06 08 10

Figure 2: Finland.

88 90 92 94 96 98 00 02 04 06 08 10

The Financial Variable Values and Recessions (Shaded) for

Table 2: Descriptive Statistics

Mean

Std. Dev.

Max.

Min.

1988:1
2011:2 2004:1
2007:4 2008:1
2011:4 1988:1
2011:2 2004:1
2007:4 2008:1
2011:4 1988:1
2011:2 2004:1
2007:4 2008:1
2011:4 1988:1
2011:2 2004:1
2007:4 2008:1
2011:4

Δlny

i3

TS

R

0.50 1.04 −0.19 1.28 0.50 2.13 2.67 1.89 2.67 −5.63 −0.07 −5.03

5.26 2.91 2.09 4.08 0.94 1.72 15.81 4.72 4.98 0.66 2.06 0.66

1.23 0.97 1.55 1.55 0.79 1.29 4.67 2.19 2.80 −2.89 −0.41 −0.42

1.53 5.05 −3.63 13.65 7.08 13.08 41.73 11.94 12.57 −34.76 −15.56 −34.76

Notes: Δlny = quarterly GDP growth, i3 = 3-month interest rate, TS = term spread, R = quarterly stock returns. For details of the data, see Table 1.

a period of normal and steady economic growth, whereas the latter was a period of economic turbulence that was caused by the global financial crisis and its aftermath. The figures indicate that the relatively strong GDP growth collapsed due to the financial crisis. One interesting observation is that the sample period includes the exceptionally deep economic depression in Finland at the beginning of the 1990s; the greatest annual decline in the Finnish GDP (−10.7%) occurred as a result of the financial crisis. Moreover, the volatility of economic activity increased substantially as a result of the financial crisis. Large swings in performance are typical of Finnish stock markets (see Figure 2). Stock prices collapsed by 60
70% on three separate occasions (1989
1991, 2000
2002, and 2008) during the sample period. However,

Revisiting Stylized Facts During Economic Turbulence

225

stock market upswings (1993
1994, 1996
1999, and 2003
2007) were also exceptionally vigorous by international standards. Despite high volatility, the compound annual stock return during the sample period was a relatively normal nominal rate of 6.3%.

Empirical Analysis The forecasting analysis is conducted for two different time periods: the steady growth period from 2004:1 to 2007:4 and the financial crisis period from 2008:1 to 2011:2 (Figure 1). Separating the forecast periods in this way makes it possible to scrutinize the predictive content of financial market variables under different economic conditions. We estimate one-, two-, and four-quarter forecast models. To ensure that the forecasting procedure is realistic and practical, the forecasting analysis is conducted recursively. That is, for the first forecasting period (2004:1
2007:4), we first conduct regressions through 2003:4 and then use these estimates to compute forecasts for 2004:1, 2004:2, and 2004:4. The models are subsequently reestimated through 2004:1, and the new forecasts for 2004:2, 2004:3, and 2005:1 are computed. This process is continued throughout the forecasting period. The recursive forecasting scheme has the intuitive advantage that all of the available information is used to calculate each forecast.

In-Sample Analysis The initial parameter estimates are based on 1988:1
2003:4 sample for the first forecasting period and the 1988:1
2007:4 sample for the second forecasting period. The estimation method was OLS with heteroscedasticity
and autocorrelation
robust Newey
West standard errors. The estimation and the stability test results are presented in Table 3. The in-sample estimation results indicate that in the models, the term spread and the stock returns are positively correlated and the short-term interest rate is negatively correlated with economic activity. This result is consistent with the theoretical expectations. It is also noteworthy that all of the parameter estimates for the financial market indicator variables are consistently significant at least at the 10% significance level. With respect to the in-sample explanatory power of the various model specifications, the following notable results are observed. First, the model specifications with past values of GDP growth exhibit higher explanatory power than the model specifications without history dependence.

(10 )

(2)

(20 )

(3)

0.23 0.04 0.28

0.27

0.57

0.28 0.32

0.15 0.34

0.07 92:3

0.10 0.13 0.05 92:3

0.03

0.03

0.11

0.58

0.48

0.08 90:3

0.29

−0.14

1.54

Const. TSt Rt it Δyt Δyt − 1 R2 P(Chow) P(A
Q) Break

0.36 0.01 0.03 93:2

0.42

0.21

0.49 0.68

0.26 0.72

0.01 94:1

0.17 0.06 0.01 94:1

0.05

0.06

0.18

1.14

0.95

0.00 90:1

0.45

−0.30

3.20

(ii) Dependent (forecasted) variable: (yt þ 2 − yt )

Const. TSt Rt it Δyt R2 P(Chow) P(A
Q) Break

0.45 0.26 0.00 91:4

−0.28

3.02

0.30 0.55 0.13 91:1

−0.14

1.48

(30 )

0.36

0.48

0.27 0.63 0.03

0.77

0.30

0.16 0.30 0.02

(4)

In-Sample Regression Results

(i) Dependent (forecasted) variable: ðyt þ 1 − yt Þ

(1)

Table 3:

0.43 0.01 0.11

0.48 0.60 0.03

0.26 0.09 0.29

0.28 0.28 0.02

(40 )

0.03 91:2

0.52

0.03 −0.27

2.88

0.43

0.33

0.02 −0.13

1.39

(5)

0.52 0.26 0.02 91:3

0.03 −0.26

2.76

0.33 0.54 0.54

0.01 −0.12

1.36

(50 )

0.05 92:4

0.68 0.78 0.52

0.27

0.04 92:3

0.47 0.21

0.29

(6)

0.67 0.79 0.53 0.59 0.03 92:4

0.33

0.46 0.20 0.14 0.01 92:3

0.35

(60 )

0.28

0.41 0.62 0.61

0.02 0.40

0.11

0.26 0.31

0.11 0.25

(7)

0.46 0.66 0.60 0.26 0.33

0.13 0.35

0.28 0.28 0.14 0.19

0.20 0.23

(70 )

0.12

0.50 0.80 0.57

0.03

0.29

0.04 92:3

0.39 0.24

0.02

0.29

(8)

0.52 0.81 0.57 0.50 0.07 92:4

0.03

0.34

0.38 0.22 0.22 0.01 92:3

0.01

0.36

(80 )

0.08 91:1

−0.18 0.39 0.57 0.63

1.80

0.07 91:3

−0.11 0.17 0.33

1.17

(9)

−0.17 0.38 0.58 0.64 0.71 0.08 92:4

1.68

−0.11 0.21 0.33 0.43 0.06 92:3

1.17

(90 )

0.34

0.29 0.66 0.64

0.06 0.36 0.03

0.14

0.21 0.32

0.13 0.24 0.01

(10)

0.35 0.69 0.65 0.25 0.27

0.16 0.32 0.02

0.24 0.29 0.20 0.09 92:3

0.21 0.22 0.01

(100 )

0.29

0.03 −0.17 0.27 0.61 0.66

1.69

0.07 92:3

0.01 −0.11 0.18 0.34

1.14

(11)

0.02 −0.16 0.27 0.61 0.66 0.61 0.22

1.61

0.01 −0.11 0.16 0.34 0.51 0.06 92:3

1.15

(110 )

226 P. Kuosmanen and J. Vataja

0.45 0.00 0.01 93:2

0.49

0.04

0.85 1.39

0.47 1.43

0.00 93:2

0.19 0.03 0.00 93:2

0.11

0.11

0.22

2.22

1.77

0.00 93:2

0.54

−0.62

6.55

0.54 0.15 0.00 91:1

−0.58

6.07

0.13 93:2

0.57

0.44 1.26 0.07

0.52 0.01 0.03

0.84 1.23 0.07

0.00 93:2

0.62

0.07 −0.56

5.87

0.65 0.27 0.00 93:2

0.06 −0.53

5.57

0.02 93:2

1.52 1.36 0.20 −0.84 0.43

0.74

1.49 1.39 0.42 −0.75 0.43 0.36 0.00 93:2

1.01

0.07 93:2

0.68 0.87 0.20 −0.69 0.62

0.05 1.09

0.79 0.95 0.16 −0.62 0.58 0.04 0.02 93:2

0.40 0.98

0.09 93:2

0.96 1.29 0.43 −0.69 0.50

0.07

0.70

1.00 1.33 0.35 −0.61 0.49 0.31 0.04 93:2

0.06

0.99

0.11 93:2

−0.57 0.56 0.70 0.06 −0.73 0.70

5.51

−0.53 0.55 0.74 0.05 −0.68 0.70 0.55 0.03

5.12

0.09 93:2

0.32 0.85 0.38 −0.59 0.66

0.08 1.00 0.06 0.45 0.93 0.31 −0.51 0.62 0.04 0.03 93:2

0.43 0.91 0.05

0.25 93:2

0.05 −0.53 0.21 0.70 0.24 −0.63 0.73

5.18 0.05 −0.51 0.21 0.72 0.20 −0.58 0.74 0.61 0.20

4.93

Notes: Columns 1
11 present the regression results for the 1988:1
2003:4 sample and, columns 10
110 for the 1988:1
2007:4 sample. The bolded figures are statistically significant at least at the 10% significance level. P(Chow) refers to the p-value of the Chow-test statistic for the null hypothesis of constant parameter estimates between the 1988:1
2003:4 and 2004:1
2007:4 samples. P(A-Q) refers to the p-value of the Andrews
Quandt test for a single structural break at an unknown point within the sample (Andrews & Ploberger, 1994). The null hypothesis is that all of the parameter estimates are stable. Break refers to the break date suggested by the Andrews
Quandt test.

Const. TSt Rt it Δyt Δyt − 1 Δyt − 2 Δyt − 3 R2 P(Chow) P(A
Q) Break

(iii) Dependent (forecasted) variable: (yt þ 4 − yt )

Revisiting Stylized Facts During Economic Turbulence 227

228

P. Kuosmanen and J. Vataja

This phenomenon occurs consistently irrespective of the forecast window. Second, the highest explanatory power is obtained in the model specification with stock returns, the short-term interest rate, and past GDP growth (AR terms) as the explanatory variables. Third, the model specification that includes stock returns as the only predictor has the lowest explanatory power. Clearly, one should avoid utilizing stock returns as the sole predictor of output growth, as the short-term interest rate is a much better choice. Fourth, past growth alone is capable of explaining approximately 20
50% of the observed economic activity, and the parameter estimates confirm a fair amount of history dependence in economic activity. Parameter constancy is vital to the success of an empirical forecasting model. Because the forecasting is conducted out-of-sample, it appears reasonable to perform stability tests for both in-sample regressions (1988:1
2003:4 and 1988:1
2007:4). The Chow test results for parameter stability between the periods 1988:1
2003:4 and 2004:1
2007:4 suggest that the models are mostly stable; the null hypothesis of parameter stability is rejected in approximately one-fourth of the cases.4 The Chow test addresses the known potential break points. The introduction of the euro as the Finnish currency and the exceptionally deep depression in the Finnish economy in the 1990s, for example, appear to be other reasonable examples of known potential break points. Kuosmanen and Vataja (2011) tested possible instability due to the launch of the euro and found no evidence that supported this hypothesis. However, breaks may also exist at unknown points in time. This possibility was examined using the Andrews
Quandt test (Andrews & Ploberger, 1994; Hansen, 1997). The test calculates all possible break dates for a single structural break at an unknown date. The test results detect instability in approximately half of the estimated models.5 Interestingly, all of the suggested break dates relate to the Finnish economic depression of the 1990s. However, the number of unstable regressions detected is somewhat smaller than that found in Stock and Watson (2003). Stock and Watson regarded the instability of financial forecasting models as the norm. From that perspective, the instability detected in our models is hardly surprising.

4 Stability is rejected at least at the 10% significance level in 10/33 cases and at the 5% level in 8/33 cases by the Chow test. 5 The Andrews and Ploberger test rejects parameter constancy at least at the 10% level in 43/66 cases and 32/66 cases at the 5% level. The test was implemented using heteroscedasticity-consistent test statistics for the central 70% of the in-sample dates.

Revisiting Stylized Facts During Economic Turbulence

229

Out-of-Sample Forecasting Results The forecasting results are presented in Table 4. Forecast accuracy is measured in terms of the root mean squared forecast errors (RMSE). The forecasting period was divided into two approximately equal subperiods to examine the influence of the recent financial crisis on forecasting performance. The first forecasting subperiod (2004:1
2007:4) represents a steady growth period, whereas the second forecasting subperiod (2008:1
2011:2) incorporates exceptional economic turbulence. The first column in Table 4 displays the various forecasting model specifications employed in this study. The second and the fourth columns present the root mean square error (RMSE) of the forecasting model specification. The third and the fifth columns provide the Clark and McCracken (2001) MSE-F test statistics for the null hypothesis of equal forecast mean square errors between forecasting Equations (2) and (1).6 The rejection of the null hypothesis suggests that the forecasting performance of Equation (1) is significantly better than that of Equation (2). That is, the model specification with both financial market information and past growth (Equation (1)) is capable of yielding better forecasts than the model specification that only considers financial market information (Equation (2)). Certain general outcomes are evident from the forecasting results. As expected, forecast errors increase consistently with the forecast horizon. The accuracy of the forecasts collapses during the financial crisis, and the forecast errors are more than three times larger during the financial crisis than during the steady growth period. Under normal economic conditions, the differences in RMSEs between the best and the worst model specifications are rather limited in a short forecast horizon; however, the differences become more notable as the forecast window is extended to longer horizons. Thus, the selection of a proper model specification is far from inconsequential. The results also suggest that during the period of steady growth, past growth is clearly useful for forecasting purposes. The formal Clark and McCracken (2001) MSE-F test confirms this to a large extent (column 3). However, during the period of economic turbulence, the predictive power of lagged GDP growth effectively disappears for longer forecast horizons. This finding is also unambiguously supported by the MSE-F test results.

6

Clark and McCracken (2001)’s MSE-F test evaluates the null hypothesis that the forecasts generated by a pair of nested models generate equal mean square errors, in which the first model is a restricted version of the second. In the present case, model (2) is a restricted version of model (1). The test statistic is nonstandard and is tabulated in McCracken (2007).

230

P. Kuosmanen and J. Vataja

Table 4: Out-of-Sample Forecasting Results Forecast period: 2004:1
2007:4 (1)

(2) RMSE

Dep. var. → Expl. vars. ↓ Δyt Rt ; Δyt Rt TSt ; Δyt TSt it ; Δyt it Rt ,TSt ; Δyt Rt ,TSt Rt ,it ; Δyt Rt ,it

ðyt þ 1 − yt Þ

Dep. var. → Expl. vars. ↓ Δyt ; Δyt − 1 Rt ; Δyt ; Δyt − 1 Rt TSt ; Δyt ; Δyt − 1 TSt it ; Δyt ; Δyt − 1 it Rt ,TSt ; Δyt ; Δyt − 1 Rt ,TSt Rt ,it ; Δyt ; Δyt − 1 Rt ,it

ðyt þ 2 − yt Þ

Dep. var. → Expl. vars. ↓ Δyt ; Δyt − 1 ; Δyt − 2 ; Δyt − 4 Rt ; Δyt ; Δyt − 1 ; Δyt − 2 ; Δyt − 4 Rt TSt ; Δyt ; Δyt − 1 ; Δyt − 2 ; Δyt − 4 TSt it ; Δyt ; Δyt − 1 ; Δyt − 2 ; Δyt − 4 it Rt ; TSt ; Δyt ; Δyt − 1 ; Δyt − 2 ; Δyt − 4 Rt ,TSt Rt ; it ; Δyt ; Δyt − 1 ; Δyt − 2 ; Δyt − 4 Rt ,it

ðyt þ 4 − yt Þ

0.648 0.662 0.682 0.679 0.736 0.560 0.533 0.682 0.723 0.572 0.556

0.691 0.755 1.097 0.815 1.204 0.631 0.743 0.828 1.168 0.674 0.781

1.717 1.827 2.145 2.009 2.112 1.166 1.310 2.021 2.095 1.149 1.217

(3) MSE-F

Forecast period: 2008:1
2011:2 (4) RMSE

(5) MSE-F

ðyt þ 1 − yt Þ

0.981 2.763 −1.501 1.980 −0.863

1.923 1.751 1.850 1.847 2.043 2.057 2.290 1.730 1.834 1.907 2.021

1.630 3.134 3.354 1.726 1.720

ðyt þ 2 − yt Þ

17.756 18.973 6.205 15.844 5.480

4.092 3.735 3.279 3.509 3.412 3.950 4.042 3.290 3.078 3.645 3.564

−3.209 −0.760 0.656 −1.748 −0.619

ðyt þ 4 − yt Þ

6.051 1.685 4.226 1.179 1.947

7.309 6.577 5.189 5.262 4.999 6.610 6.681 4.791 4.387 5.967 5.782

−5.285 −1.367 0.302 −2.260 −0.856

Notes: MSE-F refers to the Clark and McCracken (2001) test statistics that compare the mean square errors (MSE) of forecasting Equations (1) and (2). The null hypothesis is that the MSEs are equal. If the null is rejected, the MSE of forecasting Equation (1) is significantly lower than the MSE of Equation (2). The bolded numbers indicate statistical significance at least at the 10% significance level.

Revisiting Stylized Facts During Economic Turbulence

231

What if one wishes to select a single financial market indicator to predict GDP growth? Our results demonstrate that the short-term interest rate is a better choice than the more traditional term spread or stock returns. It is also interesting to note that although stock returns and the term spread perform rather poorly as individual predictors of GDP growth, the combination of these variables proves useful for forecasting purposes. Although GDP growth appears to incorporate a degree of history dependence that is useful for forecasting purposes under normal economic circumstances, the usefulness of AR terms decreases considerably during economic turbulence at longer forecast horizons. The short-term interest rate is found to be the single most important financial market indicator for predicting economic activity during periods of steady growth; however, this finding does not hold during more turbulent times. According to our results, stock returns and the term spread are the appropriate financial market indicators for forecasting future growth under turbulent economic conditions.

Analysis of the Forecasting Results The previous literature suggests that financial market variables are useful for predicting economic activity but that their predictive content is not robust with respect to different countries and time periods (Stock & Watson, 2003). However, of the different financial market variables, the term spread has gained the status as the best single financial market indicator of future economic activity (e.g., Estrella, 2005; Wheelock & Wohar, 2009). The results by Kuosmanen and Vataja (2011) support this conclusion in the Finnish context. Stock and Watson (2003) emphasized that the marginal predictive content of financial variables above and beyond past economic growth has not been sufficiently accounted for in many previous studies. According to our results, this is a valid point, especially for steady growth periods. However, during turbulent periods, GDP growth is less serially correlated, which diminishes the predictive content of past growth. Therefore, forecasting models that contain purely financial variables have better forecasting performance during economic turbulence. The recent literature has suggested that the predictive content of the term spread has decreased since the mid-1980s. This decrease may be due to either the increased stability of economic activity (Wheelock & Wohar, 2009) or fundamental changes in the relationship between the term spread and economic activity across countries. These changes may have arisen as a result of a variety of factors, such as the birth of the European monetary union, the “Great Moderation,” the global savings glut, and the zero lower bound on nominal interest rates (Chinn & Kucko, 2010). Our results

232

P. Kuosmanen and J. Vataja

suggest, however, that the variable predictive content of the term spread is related to differences in economic circumstances rather than to irreversible, fundamental changes in the world economy. Although the term spread and stock returns represent the traditional financial market variables that are used to predict future economic activity, the results for the U.S. context that were obtained by Ang et al. (2006) provided evidence that short-term interest rates have greater predictive power than the term spread for GDP growth. According to our results this finding also holds for Finland. The result is novel for a small open economy and indicates that short-term interest rates can indeed help to predict future economic activity during steady economic growth. The importance of the short-term interest rate seems to be even more remarkable from the perspective of the euro area, as the monetary policy of the ECB targets the entire euro area and Finland represents only a tiny fraction of the Eurozone. Although the ECB concentrates exclusively on controlling inflation, short-term interest rates play a major role in forecasting economic activity in Finland under normal economic conditions. The proper choice of indicator variables changes notably during periods of exceptional growth. The forecasting ability of the short-term interest rate decreases during economic turbulence. Moreover, under turbulent conditions, the predictive content of past growth vanishes at longer forecast horizons. Instead, the traditional term spread and stock returns are found to be more appropriate indicator variables for future economic activity during turbulent times. The evidence that stock returns have predictive power with respect to GDP growth when the economy is contracting is consistent with Henry et al. (2004).

Conclusions The purpose of this study was to focus on the predictive content of readily available financial market variables that are observable in real time as tools for forecasting future GDP growth under both normal and exceptional economic circumstances. This study considers Finland, a small open economy in the Eurozone with a high propensity for external shocks and GDP fluctuations. Our results confirm the usefulness of financial market information for forecasting future economic activity. We find that the proper selection of financial market indicator variables is related to the general health of the economy and the conducted monetary policy. During steady growth periods, short-term interest rates augmented with past GDP growth play a dominant role in forecasting economic activity. In contrast, during

Revisiting Stylized Facts During Economic Turbulence

233

economic turbulence, the importance of the traditional term spread and stock returns increases notably. These results are not necessarily specific to small open economies or small member states in the Eurozone. More specifically, it appears that the predictive ability of the financial variables is not related to the size of the economy, that is, the results are relatively consistent with those of earlier studies in the U.S. context. Moreover, Finland’s membership in the Eurozone may not affect the predictive content of financial variables or the stability of forecast models, despite the ECB’s exclusive focus on controlling inflation. In general, the results of this study imply that instability in forecast models did not result from the establishment of the Eurozone; instead, it is a function of unstable economic circumstances. The results of this study are important because central banks all over the world, in essence, conducted a zero interest rate policy since the financial crisis. The results demonstrate that during economic turbulence, the term spread is a valid instrument for forecasting economic growth, although the term spread is nearly entirely determined by long-term interest rates. Moreover, even when short-term interest rates are close to zero and unconventional monetary policy tools are in use, stock markets can provide useful signals that forecast GDP growth. In contrast, as growth rates return to normal levels and central banks are able to conduct conventional monetary policy, short-term interest rates will likely regain their prior predictive ability and their important role in forecasting GDP growth.

References Andrews, D. W. K., & Ploberger, W. (1994). Optimal tests when a nuisance parameter is present only under the alternative. Econometrica, 62(6), 1383
1414. Ang, A., Piazzesi, M., & Wei, M. (2006). What does the yield curve tell us about GDP growth? Journal of Econometrics, 131, 359
403. Chinn, M. D., & Kucko, K. J. (2010). The predictive power of the yield curve across countries and time. NBER Working Paper No. 16398. National Bureau of Economic Research, Cambridge, MA. Chionis, D., Gogas, P., & Pragidis, I. (2010). Predicting European union recessions in the Euro era: The yield curve as a forecasting tool of economic activity. International Advances in Economic Research, 16(1), 1
10. Clark, T. E., & McCracken, M. W. (2001). Tests of equal forecast accuracy and encompassing for nested models. Journal of Econometrics, 105, 85
110. Elliott, G., Rothenberg, T. J., & Stock, J. H. (1996). Efficient tests for an autoregressive unit root. Econometrica, 64, 813
836. Estrella, A. (2005). The yield curve as a leading indicator: Frequently asked questions. Retrieved from http://www.ny.frb.org/research/capital_markets/ycfaq.pdf

234

P. Kuosmanen and J. Vataja

Estrella, A., & Hardouvelis, G. A. (1991). The term structure as a predictor of real economic activity. Journal of Finance, 46(2), 555
576. Estrella, A., & Mishkin, F. S. (1996). The yield curve as a predictor of U.S. recessions. Current Issues in Economics and Finance, 2(7). Federal Reserve Bank of New York. Haaparanta, P., & Peisa, P. (1997). Talouden rakenne ja ha¨irio¨t. Valtioneuvoston kanslian julkaisusarja, 21/97. Hansen, B. E. (1997). Approximate asymptotic P-values for structural change tests. Journal of Business and Economic Statistics, 15(1), 60
67. Harvey, C. R. (1988). The real term structure and consumption growth. Journal of Financial Economics, 22(December), 305
333. Haubrich, J. G., & Dombrosky, A. M. (1996). Predicting real growth using the yield curve. Federal Reserve Bank of Cleveland, Economic Review, 32(1), 26
35. Henry, O´. T., Olekalns, N., & Thong, J. (2004). Do stock market returns predict changes to output? Evidence from a nonlinear panel data model. Empirical Economics, 29, 527
540. Junttila, J., & Korhonen, M. (2011). Utilizing financial market information in forecasting real growth, inflation and real exchange rate. International Review of Economics and Finance, 20, 281
301. Kinnunen, H. (1998). The sources of output shocks in Finland and other EU countries. Bank of Finland Discussion Papers 3/98. Kuosmanen, P., & Vataja, J. (2011). The role of stock markets vs. the term spread in forecasting macrovariables in Finland. Quarterly Review of Economics and Finance, 51, 124
132. Laurent, R. (1989). Testing the spread. Federal Reserve Bank of Chicago Economic Perspectives, 13, 22
34. McCracken, M. (2007). Asymptotics for out-of-sample tests of Granger causality. Journal of Econometrics, 140, 719
752. Ng, S., & Perron, P. (2001). Lag length selection and the construction of unit root tests with good size and power. Econometrica, 69, 1519
1554. Nyberg, H. (2010). Dynamic profit models and financial variables in recession forecasting. Journal of Forecasting, 29, 215
230. Samuelson, P. (1966). Science and stocks. Newsweek, 19(September), 92. Stock, J. H., & Watson, M. W. (2003). Forecasting output and inflation: The role of asset prices. Journal of Economic Literature, 41(September), 788
829. Wheelock, D. C., & Wohar, M. E. (2009). Can the term spread predict output growth and recessions? A survey of the literature. Federal Reserve Bank of St. Louis Review, 91(September/October), 419
444.

What Drives the Bank
Firm Relationship? A Case Study of the Polish Credit Market Małgorzata Pawłowskaab, Krzysztof Gajewskia and Wojciech Rogowskiab a

The National Bank of Poland, Economic Institute, Warsaw, Poland, e-mail: [email protected] b Warsaw School of Economics, Warsaw, Poland

Abstract Purpose
The aim of this study is to understand the determinants of relationship between banks and nonfinancial corporations within Poland (which are considered relationship banking from this point onward). Design/methodology/approach
The main sources of data used in the study are the large credit database (credit register of the National Bank of Poland (NBP)) and other aggregated data, including data from the Warsaw Stock Exchange and the NBP. Econometric panel logit methods have been used to test how different factors affect bank
firm relationships. Three main groups of factors have been investigated: the characteristics of the firm (i.e., size, ownership type, and R&D activity); the characteristics of the financial sector (i.e., competition in the banking sector); and macroeconomic conditions. Findings
The findings demonstrate that Polish firms readily establish single-bank relationships, and firms with the highest quality of credit portfolios borrow often from multiple creditors. All conducted estimations demonstrated that the relationship between financing from a single bank and from foreign capital had a positive sign. Also, a decrease in concentration in the banking sector, which may be identified with an increase in competition, supports the establishment of relationship banking. Research limitations/implications
The study was performed using the data from large exposure database collected for supervisory purposes. International Symposia in Economic Theory and Econometrics, Vol. 23 G. P. Kouretas and A. P. Papadopoulos (Editors) Copyright r 2014 by Emerald Group Publishing Limited. All rights reserved ISSN: 1571-0386/doi:10.1108/S1571-038620140000023009

236

M. Pawłowska et al.

Exposures (credits, derivatives, etc.) larger than 500 thousand PLN (approx. 120 thousand EUR) were only considered. Future research on bank
firm relationships should focus on the influence of financing costs, maintaining relationships when the borrower is in a difficult financial position, and other unique features of banks using the strategy of relationship financing. Practical implications
The understanding of the characteristics of bank
firm relationships can help to improve banking practice and supervisory policy in Poland. Originality/value
This study makes a noticeable contribution to the understanding of the banking sector and its relationships with nonfinancial corporations in Poland. It is the first empirical study on such a large sample of panel data from Polish banking sector and industries, too. Keywords: Number of bank relationships, relationship banking JEL Classifications: G21, G30, G32, E21, C41

Introduction The financial crisis of 2007
2008 led to a reflection on banking operations and on the usefulness of regulatory system and supervision authorities. The reasons provided for the crisis included structural changes that had taken place in the banking sector over the past decades. Banks were becoming, to a large extent, universal banks or even financial conglomerates that operated in all segments of the financial markets. The asset structure of banks was changing, as the share of off-balance-sheet assets was growing. Banks focused on improving their profitability and efficiency using increasingly complex financial instruments; noninterest income was at par with interest income while the base of deposits decreased (Schildbach, 2011; Wallace & Herrick, 2009). The crisis called these trends into question and reversed them by drawing the attention of both bank managers and regulators toward sustainable banking that is based on stable customer relationships. Therefore, it appears today that banks want to develop relationship banking, thus providing a reason to take up the subject and examine the characteristics of the bank
firm relationship.1

1 However, one should bear in mind that relationship banking is a strategy of smaller banks; as banks grow they are increasingly inclined toward arm’s-length financing (Koch & MacDonald, 2010).

A Case Study of the Polish Credit Market

237

One might assume that the stability of the Polish banking sector during the crisis (banks in Poland recorded relatively high financial results throughout 2008
2009 and bankruptcy of banks was not observed) was to a certain extent due to the predominant cooperation between banks and enterprises in the form of relationship banking. Furthermore, the results of empirical research concerning the financial crises in Asian and Latin American countries as well as those concerning the latest financial crisis indicate that relationship financing may facilitate better access to funding even when a firm is experiencing financial issues (Abildgren, Buchholst, & Staghøj, 2011; Giovanni, Kang, & Kim, 2001; Hoshi, Kashyap, & Scharfstein, 1990). The literature seeks answers to questions about the benefits and costs of relationship banking, both for banks and enterprises, as well as its effect on the development of banking systems and the economy. The most important advantage of relationship banking is reduced agency dilemma on the principal
agent line (that occurs as a result of a contract between the lending bank and the firm), thanks to the bank obtaining additional information and reducing the costs associated with the issue of negative selection (Heffernan, 2007, p. 9). Theoretical models suggest that maintaining multiple relationships is expensive for firms, primarily due to transactional costs (Diamond, 1984). The benefits of relationship banking for firms are also extensively discussed in the literature and are concerned primarily with the reduction of negative effects and information asymmetry (cf. Diamond, 1991; Petersen & Rajan, 1994). However, empirical research provides no explicit answers whereas costs and benefits depend on many factors. The main purpose of this article is to identify factors that influence bank
firm relationships in Poland from the perspective of the firms, the financial sector and the macroeconomic environment. Relationship banking is identified in this study with a bank
firm relationship that involves the firm having commitments toward one bank (the so-called “single relationship”) during a specific time period. This is why the study places particular emphasis on analyzing the number of bank
firm relationships, defined as the number of banks providing credit facilities to a given enterprise. This is the first comprehensive paper of its kind focusing on relationship banking in Poland based on credit register data. The empirical analysis of relationship banking involves two stages. During the first stage the population of firms was divided according to the number of relationships they had with banks. The second stage involved selecting, from among enterprises maintaining single-bank relationships, those that maintain relationship banking.2 Information about the number

2

For the purpose of this study, the definition of relationship banking is provided in the third section.

238

M. Pawłowska et al.

of bank
firm relationships was obtained from bank reports submitted to the NBP. The study also utilized data received from other institutions (e.g., Warsaw Stock Exchange (WSE), Polish Financial Supervision Authority (PFSA)). The analysis covered only commercial banks and nonfinancial enterprises. Factors determining the number of relationships included both variables characterizing enterprises (such as the size of the firm) and variables characterizing the national financial sector (such as competition in the banking sector, credit risk and stock exchange development). The study also considered the effects of the macroeconomic environment on bank
firm relationships. This complete study includes four sections. The first section contains an extensive presentation of the notion of relationship banking and utilizes studies from an international body of literature. The second section provides a descriptive analysis of bank
firm relationships in Poland (the first stage of the analysis). The third section presents an econometric model used for empirical verification of the stated hypotheses concerning the determinants of relationship banking within Poland (the second stage of the analysis). The fourth section describes the results of estimations obtained on the basis of this model. It also provides the results of our hypotheses verification. The study ends with a conclusion, which presents our outcomes and guidelines for further research.

Literature Review Bank
firm relationships depend on numerous factors. The benefits and costs associated with such relationships often change depending on the business cycle and individual institutions, which provide an incentive to the interested parties to establish and support such relationships (Degryse, Kim, & Ongena, 2009). From the banks’ perspective, relationship banking reduces the negative selection associated with information asymmetry, which is particularly intensified during periods of financial crises or when there is a restrictive monetary policy. On the part of firms, relationship financing improves their access to external financing. The benefits of information about customers, which is a byproduct of relationship financing, may be among the main factors of banks’ profitability (Boot, 2000). Information obtained, thanks to relationship banking, allows banks to provide long-term, renewable, and flexible credit facilities to firms (who are their customers), which decreases the probability of banks’ bankruptcy as they conduct less risky (conservative) activity (Keeley, 1990). The scale of benefits for the bank depends on the quality of information about the customer; the diversified quality of this information determines

A Case Study of the Polish Credit Market

239

the level of a bank’s specialization when establishing relational contracts (learning by lending). However, a single-bank relationship may result when there is a monopoly rent in the form of the so-called “hold-up problem” (Rajan, 1992; Sharpe, 1990; Von Thadden, 1992, 1995). A monopoly of information about the customer reduces competition in the market of bank loans, which causes an increase in loan prices in the future (ex post) (Boot, 2000).

The Lender
Borrower Relationship: Definitions The lender
borrower relationship, which involves primarily a bilateral loan agreement between two interested parties, remains the subject of research with regards to its nature, propagation, and effects
for banks, enterprises, and the development of financial systems (Cull, Haber, & Imai, 2011; Elyasiani & Goldberg, 2004; Koch & Macdonald, 2010; Van Hoose, 2010). The most crucial research problems undertaken in the field of relationship banking include determining the scale of its propagation, its characteristics and evolution under the influence of changes in markets for banking services, changes in the structure of financial systems (including technological changes), its impact on the terms and conditions of financing firms, the goodwill of firms and banks, and other factors. Lender
borrower relationships are associated with the main tasks (functions) of banks: keeping the customer’s accounts and implementation of loan agreements or other banking services (deposits, guarantees, derivatives). These functions involve the exchange of benefits and information over time between the bank and the enterprise, thus leading to the development of certain information resources, both on the bank’s part (e.g., credit history, balance of funds on the account) and on the firm’s part (e.g., experience with consideration of credit applications). There is an assumption of a relational contract between a bank and its customer (firm) if they have an understanding that allows certain contract terms and conditions to be further specified over time. Over a long period of time the customer relies on the bank, which provides it with financial services, and the bank depends on repayment of its loans by long-term borrowers and the borrowers purchasing loan-related services (Heffernan, 2007). The theoretical background for relationship banking and relationship lending was established by the work of Stiglitz and Weiss (1981), which demonstrated the existence of this occurrence in the context of loan rationing and the risk of moral hazard. The benefit of this form of financing, involving diminished information asymmetry, has been shown in the theoretical works of Hodgman (1963), Wood (1975), Diamond (1991), and Boot, Greenbaum, and Thakor (1993).

240

M. Pawłowska et al.

Some of the first comprehensive definitions of relationship banking were formulated in the works of Ongena and Smith (2000a), Boot (2000), and Berger and Udell (2002). The authors of the above-mentioned definitions attempted to synthesize terms used, among others, in the works of Rajan (1992, 1996), Diamond (1984, 1991), and Petersen and Rajan (1994). In order to conduct a flexible firm financing policy (especially during difficult periods of development), the bank should be in a close relationship with the firm. This relationship is referred to as relationship banking. In the above-mentioned context, Ongena and Smith (2000a) define relationship banking as the relationship between a bank and a firm that comprises something more than a simple, anonymous financial transaction. Banks benefit from maintaining such relationships through better access to information about the firm, whereas the firm expects the bank to provide access to financing even when it is experiencing financial difficulty. Such flexibility is not possible, for example, in an anonymous securities market. According to Ongena and Smith (2000a), relationship banking can be described with more than two-dimensional detail. The first dimension is duration, since a measure of relationship banking is the depth of bank
firm cooperation (Rajan, 1998; Wood, 1975). The second dimension is the product scope for cooperation (Hodgman, 1963). Boot (2000) defines relationship banking as the provision of financial services by a bank that invests in access to specific information about a customer (which is frequently publicly unavailable) and assesses the profitability of this investment while taking into account the duration of cooperation with the customer and banking products provided. Berger and Udell (2002) define relationship banking (relationship lending) as bank’s ability to obtain information about the firm during cooperation. The information is then used to develop future business conditions for bilateral cooperation (loan availability, interest rates, and collaterals). In summary, relationship banking requires the fulfillment of four conditions: (1) banks must be in contractual relationship with enterprises; (2) the relationship must possess an enduring quality over time; (3) the collection of information must be more extensive than that available from public (open) resources; and (4) the collected information must be confidential (i.e., unpublished) (Berger, 1999 quoted from Dong & Li, 2010). However, it should be noted that the nature of cooperation between an enterprise and a bank is, in practice, more complex than it might appear. Therefore, it is difficult to provide strict definitions of relationship banking and transaction-oriented banking. To simplify, one can assume that if an enterprise uses the services of only one bank at a given time (the so-called “single relationship”), this cooperation takes the form of relationship banking (relationship financing). Otherwise, if there are more banks providing credit facilities, we are dealing with multiple relationships or

A Case Study of the Polish Credit Market

241

transaction-oriented cooperation. It should be noted, however, that multiple relationships might also assume the nature of relationship banking. As a result, banks may reduce their monopoly margin and offer better financing terms to current customers (Berger & Udell, 1995; Bonfim, Dai, & Franco, 2009; Degryse et al., 2009; Harhoff & Ko¨rting, 1998; Petersen & Rajan, 1994, 1995). This cooperation usually involves a main bank. Relationship banking is more frequently encountered in countries such as Japan and Germany where there are strong capital relationships between banks and enterprises from the nonfinancial sector (Ongena & Smith, 2000a). Close lender
borrower relationships were considered to be one of the key success factors for the economies of Japan and Germany in the previous century. However, in the 1990s the popularity of relationship banking dropped due to increased opportunities for enterprises to obtain funds and an increased number of players within the financial market (Heffernan, 2007).

Sources of Information about Relationship Banking In scientific research practice, information about bank
firm relationships is obtained from multiple sources. Researchers use data from lending registers of the banks themselves or systems for exchanging information about financial liabilities (Ongena & Smith, 2000b), survey data obtained from entrepreneurs, supplemented with data about enterprises and banks (Berger & Udell, 1995; Nam, 2004; Tymoczko & Pawłowska, 2007). Another possible source of knowledge about bank relationships may be information about material events (e.g., conclusion of a loan agreement), data submitted by stock-listed issuers of securities (Berg & Schrader, 2009; Chang, Liao, Yu, & Ni, 2010), as well as results of statistical research (individual and aggregated
Ogura & Yamori, 2008). Many research problems are analyzed through integration of individual data from different databases, including survey data (Castelli, Dwyer, & Hasan, 2010). Registers of major exposures (the so-called “central loan registers”) maintained by supervisory authorities provide a valuable source of information about bank
firm relationships (see e.g., Antao, Ferreirra, & Lacerda, 2011; Albertazzi & Marchetti, 2010; Gersl & Jakubı´ k, 2010; Jimenez & Saurina, 2004; Memmel, Schmieder, & Stein, 2007; Schmieder, 2006). The reason for existence of registers of major exposures is supervisory provisions restricting concentration of the bank’s receivables in a single entity. These provisions have been set forth in order to avoid a situation in which a bank would be exposed to a risk critical to its existence as a result of insolvency of the main creditor. The first regulations of this sort were established in

242

M. Pawłowska et al.

Germany in the 1930s and are currently common in international and national banking laws (Schmieder, 2006).3 Data included in supervisory registers may, apart from being used for prudence purposes (i.e., to determine the structure of receivables of individual banks and compliance with supervisory provisions), provide information about the behavior (relationships) of transaction entities, that is, banks and their major customers.

Determinants of the Number of Relationships with Banks The results of the empirical research to date have demonstrated numerous factors influencing an enterprise’s and/or bank’s decision to select a model of cooperation based on a single bank
firm relationship or on multiple banking relationships. Current research on the number of bank
firm relationships includes the features of enterprises, the characteristics of banks that provide credit facilities, the features of the financial sector, and the nature of the entire economy. Further on in this section the results of empirical research obtained from models where the variable describing relationship banking is the explained variable (endogenous variable) are described. Explanatory variables in these models are variables that determine competition in the banking sector, the size of firms, R&D expenditures, and macroeconomic factors. Competition Competition in the banking sector (measured by the concentration of lenders) is a crucial determinant of the number of bank
firm relationships. However, the influence of competitive conditions on the number of lending relationships and the distribution of relationship banking is somewhat ambiguous. On the one hand, growing competition reduces the consumer information available to the market due to the bank’s diminishing share of the market. On the other hand, as competition grows with the entrance of new banks into the sector, banks can invest in establishing relationships with customers in a given market segment in order to generate profit (Boot, 2000).

3 The first credit registers of the so-called “major loans” were established after World War II in France (1946), Belgium (1954), Spain (1962), and Germany (1962), see also Schmieder (2006).

A Case Study of the Polish Credit Market

243

Empirical research of the influence of competition in the banking sector on the functioning of relationship banking has been described extensively in literature dedicated to the microeconomy of the banking enterprise (cf. Berger & Udell, 1995, 2002; Berger, Goldberg, & White, 2001; Boot & Thakor, 2000; Degryse et al., 2009; Detragiache, Garella, & Guiso, 2000; Machauer & Weber, 2000; Petersen & Rajan, 1995; Presbitero & Zazzaro, 2010). On the one hand some studies have shown a positive relationship between competition in the banking sector and relationship financing (Boot & Thakor, 2000; Petersen & Rajan, 1995); on the other hand some studies have shown a negative relationship between the two (Memmel et al., 2007). However, the latest works suggest a U-shaped relationship between competition (concentration in the banking sector) and the number of bank relationships or the concentration of the enterprise’s debt (Presbitero & Zazzaro, 2010). When studying the influence of growing competition in the banking sector, Boot and Thakor (2000) suggested that this growth of competition results in an increased interest to provide credit facilities to enterprises that require a relationship (individual) approach. This supports the establishment of the bank
firm relationship and thus creates the opportunity to obtain benefits as a result of having unique data about the enterprise and its activity. Elsas (2005) and Degryse et al. (2009) have demonstrated a nonmonotonic dependency for the number of bank
firm relationships on the status of competition in the banking market. Single-bank relationships become more likely as competition increases; nevertheless, in markets with a high degree of concentration, lower competition works to the advantage of relationship banking. Due to the structure of the financial sector of different countries, research on relationship banking may not simply analyze competition within the banking sector itself but might also discuss more extensively the influence of competition on the part of other institutions in the financial market (e.g., the capital market). However, even in this case both banking and financing coexist on the balance sheets of enterprises (Rajan & Zingales, 1998). Size of Firms From the point of view of firms, an important determinant of the number of bank
firm relationships is the size of the borrower (Memmel et al., 2007; Neuberger & Rathke, 2006; Ongena & Smith, 2000b). Memmel et al. (2007) have demonstrated that the number of enterprises using the services of a single bank quickly diminishes as the size of the enterprise in the sample increases. A study conducted by Ongena and Smith (2000b) on a panel of several European countries has demonstrated a U-shaped dependency between the average number of relationships and the size of the enterprise.

244

M. Pawłowska et al.

Very small and very large enterprises are characterized by a high average number of single-bank relationships, although, in the case of small entities, firms using the credit facilities of a single bank are more frequent. Innovations Relationship financing recognizes the value of durable cooperation with customers that provides better two-way information flows. In such a way mutual expectations are established and atypical financial requirements can be handled with innovative activity (Neuberger & Rathke, 2006; Ogawa, Sterken, & Tokutsu, 2005). Investments in R&D, as well as financing new technologies, are associated with the strategy of relationship financing (cf. Bhattacharya & Chiesa, 1995; Brighi & Torluccio, 2009; Von Rheinbaben & Ruckes, 2004). The influence of innovative activity on relationship financing is ambiguous. On the one hand, if a firm is certain of its project’s success, it negotiates contractual conditions with a single bank, but a firm in a weaker position will seek financing of multiple borrowers (Yosha, 1995). The positive influence of innovativeness on multiple-borrower financing may be caused by a bank’s attempt to distribute the risk connected with financing innovative projects (Cosci & Meliciani, 2002) or a firm’s attempt to minimize risk due to liquidity issues of a bank providing the credit facility and an investment project prematurely terminated (Detragiache et al., 2000). On the other hand, Memmel et al. (2007) has confirmed the positive influence of innovative activity on relationship financing with a group of small, medium, and large enterprises in Germany. The empirical analysis has demonstrated that relationship banking has a positive correlation with the intensity of R&D expenditure in individual industries. Macroeconomic Factors The number of relationships may change over time in the business cycle; at different stages firms may change their manner of external financing from banks to something that is based on the capital market (Kashyap, Stein, & Wilcox, 1993). The few studies available in this field (Dietsch, 2003; Hommel & Schneider, 2003) provide no unequivocal results.

The Number of Bank
Firm Relationships in Poland: A Descriptive Analysis The following sections present the results of a descriptive analysis of the number of bank
firm relationships between 1997 and 2010. The basic

A Case Study of the Polish Credit Market

245

source of information about bank
firm relationships in Poland was provided by the data submitted to banking supervision bodies on NBP forms (i.e., the credit register). This data contains a list of customers in whose case the exposure of a given bank is considered major (i.e., it exceeds PLN 500 thousand in the case of commercial banks). The number of bank
firm relationships was defined as the number of banks specified on the NBP form of large exposure.4

Bank
Firm Relationships in Poland: Basic Information This analysis was carried out on a sample of 32,241 separate nonfinancial enterprises and covered the years 1997
2010 (see Table A.1). During this time period the average number of relationships with banks in a given population remained at a similar level of ∼1.6, which was partially due to domination of enterprises maintaining single-bank relationships; in individual years such enterprises constituted more than 60% of all analyzed (see Table A.2). Beginning in 2008 one could observe an increasing share of such enterprises, which translated to a slight drop in the average number of relationships from 2008 to 2010. The decreasing popularity of the strategy involving financing from multiple banks is confirmed by the fact that even those enterprises in the sample that used the services of more than one bank did not maintain a large number of relationships. Between 20% and 25% of examined enterprises in this time period maintained relationships with two banks, whereas relationships with three banks were maintained by far less than 10% of the enterprises. Contact with more than five banks was rare in the examined sample and occurred in ∼1% of enterprises. There was more volatility in the maximum number of banks whose services were used by a given enterprise
depending on the period it varied between 16 and 32. These, however, were individual cases. Enterprises maintaining single-bank relationships were the dominant portion of the analyzed sample (more than 60%); however, their share of total debt was considerably smaller and remained between 29.4% and 40.4%. This may suggest that the group of enterprises maintaining a relationship with multiple banks is composed of the largest enterprises with the

4 The bank’s exposure (loan) is considered “major” for supervisory purposes if it exceeds the value of 10% of the minimum admissible value of the bank’s own funds. In European and Polish provisions of the banking law, this amount is EUR 5 million (i.e., the threshold of 10% is EUR 500 thousand, approximately PLN 2 million).

246

M. Pawłowska et al.

biggest lending requirements.5 This is also suggested by data on the most active entities, which have relationships with more than five banks. In 1997
2010 such enterprises constituted ∼1% of enterprises; however, their share in debt was ∼20% (see Table A.2).

Panel Data Analysis On the basis of the literature review presented in the first section, this section formulates a number of hypotheses concerning relationship banking that will be subject to subsequent empirical verification. It should be noted that the literature review was quite broad, whereas the empirical part of this study, concerning the Polish market only, addresses only those issues that are permitted by virtue of access to data. Nevertheless, the study conducted in this second stage allowed for a relatively broad analysis of the occurrence of bank
firm relationships and its determinants in Poland. This study used an approach that involves combining data from different available sources that is commonly used in literature on the subject (e.g., Memmel et al., 2007; Ogawa et al., 2005). This is one of the first comprehensive studies of relationship financing in Poland. Previous empirical studies concerning the Polish banking sector have covered only survey data for one year
2005 (Tymoczko & Pawłowska, 2007). The results of the descriptive analysis of the number of bank relationships, presented in the previous section, have demonstrated that the dominant strategy of Polish enterprises is to maintain single-bank relationships. For this reason the study assumed that enterprises that maintain singlebank relationships also use what has been defined as relationship banking.6 However, there may be situations where an enterprise has exposure to a single bank but this bank changes every year; this situation does not testify to relationship lending. Nevertheless, according to the data presented in Table A.2), single-bank relationships are relatively durable; in the majority of studied periods the share of enterprises that maintained single-bank relationships with one and the same bank both in a given year and three years prior varied between 63.9% and 82.6%. Taking into account the above assumptions, the analysis of literature, and the database structure, the following simplified definition of relationship banking was adopted for the purposes of this study: a firm is considered

5 Due to the debt threshold of PLN 2 million, this research sample does not include small enterprises. 6 See definitions of relationship banking presented in the first section.

A Case Study of the Polish Credit Market

247

to use relationship financing if it has liabilities toward only one bank (balance sheet and off-balance sheet liabilities, including loans) over a period of three years. A similar definition, taking into account both the number of banks providing credit facilities and the duration of cooperation, was applied, among others, in the works of Elsas (2005) and Memmel et al. (2007).

Research Method and Data Set According to the adopted definition, relationship financing depends on the number of banks with which a firm maintains lending relationships in subsequent periods. In this situation, given the same number of relationships with banks in a given year, a firm may engage in relationship banking or not depending on its behavior in periods preceding the analyzed period. In the case of the adopted definition of relationship banking, only two options are possible: the presence of relationship banking (when the firm remains in a lending relationship with only one and the same bank over three subsequent periods) or the absence of relationship banking (all other cases). According to the adopted assumptions, relationship banking is binary: there are no intermediate states and there is no scale to assess the strength of the relationship. For this reason, further analyses employ the logit model where the dependent variable may assume one of the following values: 1 (success, the case a situation where relationship banking is present) or 0 (failure, the lack of relationship banking in contacts between a given firm and banks). Therefore, the results of this model may be interpreted in the category of probabilities
a positive coefficient for a given variable means that it increases the likelihood of success. Such an approach is commonly applied in literature and the majority of authors studying bank
firm relationships apply different versions of models with discrete dependent variables including in particular logit models (see Table 1). A majority of studies employ models based on cross-section data (taking into account a single period of analysis) whereas this study uses longitudinal (panel) data. The advantage of panel data is that the analysis takes into account both diversity between firms and the changes that take place over time. The temporal dimension may be of particular significance in developing economies (such as Poland) where very fast and deep economic transformations are frequent and may be reflected in bank
firm relationships. Sufficiently long time series also allow for the inclusion of different phases of the economic cycle. Using the panel model for this analysis necessitates a specification of the type of individual effect used

248

M. Pawłowska et al.

Table 1: Tools Used to Study Bank
Firm Relationships Authors Detragiache, Garella, and Guiso (2000)

Ongena and Smith (2000b)

Machauer and Weber (2000) Cosci and Meliciani (2002) Yu and Hsieh (2003)

Guiso and Minetti (2004)

Berger et al. (2005)

Ogawa et al. (2005)

Memmel et al. (2007)

Type of data

Method

Cross-sectional Two-step estimation: Probit (choice between single-bank relationships and multiple-bank relationships) and LSM (the optimum number of relationships for firms maintaining contacts with multiple banks). For simplicity the authors treat the number of relationships as a continuous variable, which enables them to apply LSM. Cross-sectional The Tobit model assesses the influence of certain features of firms, industries, and characteristics of individual countries on the number of relationships for comparative data between countries. LSM will assess the influence of different factors at a national level. Panel The Poisson model with random effects to explain the number of relationships. Cross-sectional The negative binomial model to explain the number of bank
firm relationships. Cross-sectional The logit model to assess the influence of a firm’s features on the choice between maintaining a single-bank relationship or multiple-bank relationships. Cross-sectional The Heckman two-step procedure. In the first step, the Probit model for assessment of the probability that the firm chooses financing from multiple banks is used. This is followed by an assessment of diversity of the number of relationships for firms maintaining relationship with more than one bank. Cross-sectional The Heckman two-step procedure. In the first step there is an assessment of the probability that a firm maintains multiple-bank relationships and subsequently an analysis of the number of relationships. Moreover, the Poisson model is used to explain the number of relationships maintained with banks, and the Probit model assesses the probability that the firm will diversify its form of ownership in terms of banks financing it. Cross-sectional The logit model assesses the decision to maintain a single-bank relationship. The multinominal logit model is used for a detailed analysis of the decision to maintain contact with multiple banks, and the Tobit model analyzes the concentration of loans. Panel The logit model with random effects specifies the factors determining the maintenance of a single-bank relationship.

Source: Author’s own study.

A Case Study of the Polish Credit Market

249

and the choice between fixed and random effects (see also Ciecielag & Tomaszewski, 2003). For the purposes of this study the specific nature of the sample had a major impact on the choice of the effects. Its dominant portion comprises enterprises that continuously maintain relationship banking
for such entities, the explained variable will be constant over time (its value will be 1 in every period). An attempt to estimate the model with fixed effects would cause their removal from the study (as a result of the within transformation7), which would contradict the purposes of the study. In an attempt to assess the factors determining relationship banking we would remove firms most willing to apply such a banking philosophy from the sample. Due to differences in the size of the samples and different assumptions regarding the type of individual effect, it is impossible to apply the popular Hausman test8 and thus formally decide on the choice between fixed and random effects. Eventually, to specify the factors determining relationship banking, a panel logit model with random effects was used (a similar approach was applied in work by Memmel et al., 2007).9 The panel data covers the years 1999
201010 and when using this data we made an attempt to verify hypotheses concerning the determinants of relationship financing on the part of the banking sector and the macroeconomic environment. Additionally, tests covered hypotheses concerning the sector of firms for which variables were available, among others, in the lending register. Panel construction also used annual data concerning the financial sector that was obtained from the NBP (BIS), PFSA, CSO, and WSE databases.

Main Hypothesis The information in the panel data was used to test the hypotheses concerning the sector of firms, information about what was included in the Large

7

The within transformation involves the mean value over time of a given variable being subtracted from that variable (cf. Wooldridge, 2001). For other fixed variables the mean value is equal to the variable itself and the difference amounts to zero. 8 The Hausman test compares estimators obtained from models with fixed and random effects (cf. Baltagi, 2001). If models have been estimated based on samples of different sizes, the estimators cannot be compared. 9 The detailed form of the model has been presented later in this study. 10 The data panel had to be shortened with respect to the set used for descriptive analysis as a result of the definition of relationship banking.

250

M. Pawłowska et al.

Exposure database (the facility size and credit risk), as well as structures of ownership and innovative activity that were obtained from other sources. Variables in panel data concerning the financial sector and the macroeconomic environment served to verify subsequent hypotheses. With respect to the firms sector, the following hypotheses were tested: Hypothesis 1. Firms with larger lending requirements are more inclined toward arm’s-length financing (in multiple banks). This fact may be connected with conducting larger investment activity, which requires multiple financial partners (the anticipated sign for this variable is negative). Hypothesis 2. The growth of credit risk supports maintaining a single-bank relationship. Typically, firms with the highest quality of credit portfolios borrow from multiple creditors (the anticipated sign for this variable is positive). Hypothesis 3. The more innovative the firm, the greater the importance of relationship banking (the anticipated sign for this variable is positive). It appears that conducting innovative activity requires a relationship approach, which allows for financing of atypical requirements such as R&D expenditures. Hypothesis 4. Foreign capital supports maintaining a single-bank relationship (the anticipated sign for this variable is positive). This fact may be connected with the daughter company using the same bank as the parent company. The choice of the manner of financing depends on the factors characterizing the situation in the financial sector and the macroeconomic environment. The financial system in Poland is based primarily on banks whose share in assets of the entire financial sector in 2010 amounted to 69.6% (see also NBP, 2010, 2011). The Polish banking sector is clearly dominated by commercial banks,11 although the role of other financial institutions is systematically growing. Among the features of the banking sector, the factors that influence relationship banking include the financial standing of banks and concentration (competition). Empirical results concerning the influence of competition (measured using concentration measures) on relationship banking are ambiguous. On the one hand, a negative dependency has been demonstrated (cf. Memmel et al., 2007), while on the other hand a positive one dependency has been demonstrated (cf. Boot & Thakor, 2000; Petersen & Rajan, 1995). Some studies also indicate that the average level

11

The share of cooperative banks in the banking sector is approximately 6% (cf. NBP, 2011).

A Case Study of the Polish Credit Market

251

of competition is most advantageous for relationship financing (Dinc, 2000; Yafeh & Yosha, 2001).12 With respect to factors on the part of the financial system, the following hypotheses have been formulated: Hypothesis 5. Increasing competition/decreasing concentration in the banking sector supported maintaining a single-bank relationship (the anticipated sign for this variable is negative). The Polish banking sector is moderately concentrated, which demonstrates a moderate competition level. Hence, in this case the growth of competition supports relationship financing and reduces the so-called “hold-up problem.” The Warsaw Stock Exchange was developing dynamically during the analyzed period (see also NBP, 2010, 2011). Boot and Thakor (2000) demonstrated that a more competitive capital market reduces relationship banking. Hence, the following hypothesis concerning the influence of competition in the capital market has been formulated:13 Hypothesis 6. The growth of competition in the capital market reduces interest in relationship banking (the anticipated sign for this variable is negative). Research on crisis points suggests that during periods of economic slowdown there is increased interest in relationship financing. However, this is due to the financial condition of banks. It should be noted that the use of banking services during the period analyzed (measured by the ratio of sector assets to GDP) increased systematically; throughout the years 1997
2010 banks were developing more quickly than economic growth. This is why the following hypotheses have been formulated with respect to the Polish economy: Hypothesis 7. Increasing the role of the banking sector in the economy supports the establishment of a single-bank relationship (the anticipated sign for this variable is positive). Hypothesis 8. Economic slowdown supports the establishment of relationship banking (the anticipated sign for the pkb variable is negative
the probability that a firm has a loan with a single bank increases as GDP decreases).

12

Markets for which HHI is below 0.1 are considered unconcentrated. When the value of the index exceeds 0.18 the market is considered concentrated (cf. ECB, 2005). It is assumed that the market is moderately concentrated when HHI ranges are between 0.1 and 0.18. 13 Providing credit facilities to a firm by a bank improves its credibility in the capital market, which has positive impact on its goodwill. This also applies to a guaranteed issue of its shares. See also Petersen and Rajan (1995).

252

M. Pawłowska et al.

Table 2: Explanatory Variables Group of variables

Variables characterizing the enterprise sector

Name of variable lk foreign

hightech

Credit risk

nplp

Concentration/ competition in the banking sector

CR5, HHI

lb

Competition in the financial/capital market Economic development

lgpw

Development of the banking sector

aktb_GDP

GDP

Description of variable

Logarithm of the credit exposure of a given firm i during period t Binary variable characterizing the type of property of a given firm i during period t, depending on the type of property: 1
foreign, 0
other cases Binary variable of a given firm i during period t, determining whether the firm operated in the high-tech sector according to OECD: 1
the firm operates in the high-tech sector, 0
other cases The share of nonperforming loans in total debt of firm i during period t
1 Indices of concentration in the banking market (market share of the five largest banks, Herfindahl
Hirschman concentration index) Number of banks and branches of lending institutions operating in Poland Number of firms listed on the stock exchange (main market) and NewConnect market GDP growth

Relationship of banking sector assets to GDP

Data source

NBP NBP

OECDa

NBP PFSA and own calculations

NBP, PFSA

WSE

Central Statistical Office NBP, PFSA

Source: Author’s own study. OECD (2003). This classification also includes knowledge-intensive services.

a

On the basis of the above hypotheses, explanatory variables were chosen in the model. These are presented in Table 2 (the basic statistics concerning this data is presented in the appendix). The estimations also used binary variables for individual industries (as control variables).14

14

Analysis examined the enterprise sector in general; hence, control variables have been grouped as processing, services, transport, and construction. In panel A control variables were taken into account as divisions into individual sections.

A Case Study of the Polish Credit Market

253

The following econometric model has been developed to verify Hypotheses 1
8:   n X pðREit = 1Þ βj ðFIRMit Þ þ βn þ 1 ðR Kit − 1 Þ ln = αo þ 1 − pðREit = 1Þ j=1 þ βn þ 2 ðK Bt Þ þ βn þ 3 ðΔPKBt Þ þ νi þ εit ;

ð1Þ

where REit assumes two values: 1 or 0. Hence, the explained variable assumes a value of 1 when an event constituting a so-called success has occurred meaning that the enterprise has a so-called single-bank relationship (i.e., its total liabilities are toward a single bank over a period of three years) with probability pi. The explained variable assumes a value of 0 when an event constituting a so-called failure has occurred meaning that the opposite event has occurred with a probability of 1
pi. The following explanatory variables have been defined in the model: FIRMit: a matrix of variables characterizing the sector of firms described in Table A.4; R_Kit−1: credit risk, measured by share of nonperforming debt to total debt for each firm i during the period t
1 (see also Degryse, Masschelein, & Mitchell, 2005); K_Bt: competition in the banking sector during the period;15 GDPt: GDP growth over the period t;16 νi : individual random effect, εit : pure random effect.

Estimation Results Based on Panel Data This section discusses the empirical results obtained from econometric models verifying hypotheses that were formulated in the third section. Detailed results of the estimations are presented in the appendix. The results of five estimations, allowing for verification of Hypotheses 1
8 on the basis of Equation (1), are presented in Table A.8. The hypotheses presented in the section “Main Hypothesis” were verified through an assessment of materiality and coefficients accompanying individual variables (i.e., using Student’s t-test). The selection of variables for individual estimations was based on Equation (1) and the results of correlations between individual explanatory variables (see Table A.5).

15

An alternative estimation was also carried out for competition in the financial/ capital market. 16 An alternative estimation was also carried out with a share of assets of the banking sector to GDP instead of GDP growth.

254

M. Pawłowska et al.

Negative coefficients for the variable specifying the loan amount (lk) indicate that as the debt grows, the inclination to maintain a single-bank relationship decreases. Results for five estimations (see Table A.8, estimations (1)
(5)) allowed for positive verification of Hypothesis 1 concerning the influence of the size of credit exposure on the type of financing. A positive sign for the variable characterizing credit risk (nplp) may indicate that firms experiencing financial difficulty are inclined to establish relationships with a bank (see also Table A.8, estimations (1)
(5)) and allowed for positive verification of Hypothesis 2. The results are in line with the results of Bolton and Scharfstein (1996) who demonstrated that firms with the highest quality of credit portfolios borrow from multiple creditors. On the other hand, von Thadden (2004) has demonstrated the opposite dependency. Estimations based on panel data have also demonstrated the interaction between relationship financing and firms conducting innovative activity. The coefficient for the variable (hightech), characterizing innovative industries, was negative (see also Table A.8, estimation (3)). The above results negated Hypothesis 3 concerning the firms sector. The results may confirm the fact of sharing the risk associated with financing innovative projects (Cosci & Meliciani, 2002; Detragiache et al., 2000). All estimations conducted on panel data demonstrated the relationship between financing from a single bank and foreign capital (the coefficient for the variable characterizing foreign capital (foreign) had positive sign in all alternative specifications, as shown in Table A.8, estimations (1)
(5)). The above results allowed for a positive verification of Hypothesis 4 concerning the firms sector. Moreover, on the basis of coefficients for variables characterizing competition in the banking sector, an attempt was made to verify Hypothesis 5 concerning the impact of the situation in the banking sector on relationship banking. The results have shown that a decrease in concentration in the banking sector, which may be identified with the growth of competition, supports the establishment of relationship banking. It is worth noting that for each measure of concentration (HHI, CR5) this coefficient proved significant and negative (see also Table A.8, estimations (1) and (2)). The results are in line with the studies of Petersen and Rajan (1995) and Boot and Thakor (2000) who each established that the growth of competition in the banking sector results in increased interests of banks in providing credit facilities to enterprises that require a relationship (individual) approach and thus supports the establishment of bank
firm relationships. This also creates the opportunity to obtain benefits as a result of having unique data about the enterprise. Confirmation of this thesis is also supported by the value of the coefficient obtained in additional estimations (i.e., the

A Case Study of the Polish Credit Market

255

number of banks (lb)). The results demonstrate a positive coefficient for the variable lb, specifying the number of banks and branches of lending institutions in Poland (see also Table A.8, estimations (3)). Hence, Hypothesis 5 has been verified positively with panel data, confirming the influence of competition on relationship banking and with a number of measures specifying the level of competition in the banking sector. The results that proved contrary to the existing literature (see also Boot & Thakor, 2000; Ongena, 2000b) were those concerning competition on the part of the capital market and Hypothesis 6 concerning the influence of competition in the capital market. This was negated; the sign of the estimated coefficient for the variable (lgpw) for relationship banking proved positive. See also Table A.8, estimation (5). It appears that the results involved the specific nature of the Polish financial sector, which is developing dynamically. At the same time, competition within the banking sector and in the capital market is growing. In particular, during the period of financial crisis (2008
2009), the stock exchange in Poland was developing quite dynamically, generating competition for banks. The results of estimations concerning the influence of banks in the economy on relationship financing (see Table A.8, estimation (5)) proved to be in line with the expectations, as confirmed by the positive sign on the estimated coefficient for the variable (aktb_GDP). The result allowed for a positive verification of Hypothesis 7 and demonstrated that the growth of a bank’s role in the economy makes firms more inclined to establish singlebank relationships. The sign that proved contrary to the literature was the one for the variable describing the business cycle
GDP growth (GDP) (see also Table A.8, estimations (1), (3), and (5)). The results regarding the influence of the business cycle on relationship banking have shown that the probability of a firm having a loan with a single bank increases as GDP increases. By contrast, in estimation (2) the coefficient accompanying the variable (GDP) proved insignificant. A positive sign for the variable determining GDP growth (see Table A.8, estimations (1) and (3)) means that the occurrence of GDP is procyclical
positive economic trends supported the establishment of in-depth bank
firm relationships while the economic slowdown provided a stimulus for financing in multiple banks. Hypothesis 8, concerning the influence of the business cycle on relationship banking, was negated, as demonstrated by the positive coefficient for variable (GDP). In order to verify the correctness of this result additional estimations were made to determine the influence of individual years of analysis on the situation of relationship banking. The results have demonstrated a negative sign for years 2001
2002 and a positive sign for years 2003
2010 (cf. Table A.9). The years 2001
2002 are ones of economic

256

M. Pawłowska et al.

slowdown. This means that in those years the probability that firms used relationship financing was less than in years 2003
2010. The result negates Hypothesis 8.

Conclusion Bank
firm relationships depend on many factors, both microeconomic and macroeconomic, which provide incentives to the interested parties to establish and support relationships. The results presented in this study are predominantly compliant with the results of this type of research for countries where the financial sector is based on banks (cf. Bonfim et al., 2009; Memmel et al., 2007). The empirical results of the first stage of analysis presented in this article have demonstrated that Polish firms readily establish relationships with banks. The descriptive analyses have shown that in 1997
2010 more than 60% of enterprises maintained single-bank relationships. Additionally, the results of the second stage of analysis using an econometric model allowed us to determine that relationship financing depends on factors associated with the firms themselves, banks providing credit facilities to firms, and the macroeconomic environment. In general, smaller enterprises that have smaller lending requirements are less profitable, are characterized by a higher credit risk, and are more inclined toward relationship financing. Moreover, relationship financing is influenced by increased competition in the financial sector and the business cycle. The results of estimations using panel data confirmed the hypotheses regarding the influence of the size of credit exposure. These results demonstrated that as the debt grows, the inclination to maintain a single-bank relationship decreases. This may be associated with larger investment activity. A positive sign for the variable characterizing credit risk may indicate that firms with the highest quality of credit portfolios borrow from multiple creditors. The fact of sharing risk with other banks was confirmed by the results concerning innovative activity (the coefficient for the variable characterizing innovative industries is negative in all alternative specifications). All conducted estimations demonstrated the relationship between financing from a single bank and from foreign capital. The results of analysis using this data have also confirmed that a decrease in concentration in the banking sector, which may be identified with an increase in competition, supports the establishment of relationship banking. The results that proved to be contrary to the literature (cf. Boot & Thakor, 2000; Ongena, 2000b) were those concerning competition on the part of capital market. Empirical analysis has demonstrated the positive influence of competition in the capital market on relationship banking.

A Case Study of the Polish Credit Market

257

Moreover, the results have shown that the growth of the role of banks in the economy has caused firms to become more involved in relationship financing. A positive sign for the variable determining GDP growth (business cycle) means that positive economic trends support the establishment of indepth bank
firm relationships. Conversely, economic slowdown provided a stimulus for financing in multiple banks (i.e., arm’s-length financing). However, it should be noted that the results of estimations are always determined by the definition of the explanatory variable; hence, the results of the above study should be treated as preliminary. The occurrence of relationship banking requires further in-depth studies using different definitions. Future research on bank
firm relationships should focus on the influence of financing costs, maintaining relationships when the borrower is in a difficult financial position, and other unique features of banks using the strategy of relationship financing.

Acknowledgment This article includes personal views of the authors and does not necessarily represent the position of the NBP. The authors are responsible for any and all errors.

References Abildgren, K., Buchholst, B. V., & Staghøj, J. (2011). Bank
firm relationships and the performance of non-financial firms during the financial crisis 2008
2009: Microeconometric evidence from large-scale firm-level data. Denmark National Bank Working Paper No. 73. Antao P., Ferreirra, M. A., & Lacerda A. (2011). Banks’ corporate control and relationship lending: Evidence from retail loans. Banco de Portugal Working Paper No. 17. Baltagi, B. (2001). Econometric analysis of panel data (2nd ed.). Chichester: Wiley. Berger, A. (1999). The ‘Big Picture’ of relationship finance. In J. L. Blanton, A. Williams, & S. L. Rhine (Eds.), Business access to capital and credit (pp. 390
400). A Federal Reserve System Research Conference. Berger, A. N., Goldberg, L. G., & White, L. J. (2001). The effects of dynamic changes in bank competition on the supply of small business credit. European Finance Review, 5, 115
139. Berger, A. N., & Udell, G. F. (1995). Relationship lending and lines of credit in small firm finance. Journal of Business, 68, 351
381. Berger, A. N., & Udell, G. F. (2002). Small business credit availability and relationship lending: The importance of bank organizational structure. Economic Journal, 112, F32
F53.

258

M. Pawłowska et al.

Bhattacharya, S., & Chiesa, G. (1995). Proprietary information, financial intermediation and research incentives. Journal of Financial Intermediation, 4, 328
357. Bolton, P., & Scharfstein, D. S. (1996). Optimal debt structure and the number of creditors. The Journal of Political Economy, 104(1), 1
25. Bonfim, D., Dai, Q., & Franco, F. (2009). The number of bank relationships, borrowing costs and bank competition. Banco de Portugal Working Paper No. 12. Boot, A. W. A. (2000). Relationship banking: What do we know. The Journal of Financial Intermediation, 9, 7
25. Boot, A. W. A., Greenbaum, S. I., & Thakor, A. V. (1993). Reputation and discretion in financial contracting. The American Economic Review, 83, 1165
1183. Boot, A. W. A., & Thakor, A. V. (2000). Can relationship banking survive competition? The Journal of Finance, 55(2), 679
713. Brighi, P., & Torluccio, G. (2009). The financing menu of R&D and traditional investments. RCEA Working Paper No. WP-09. Castelli, A., Dwyer, G. P., & Hasan, I. (2010). Bank relationships and firms’ financial performance: The Italian experience. European Financial Management, 18(1), 28
67. Chang, C., Liao, G., Yu, X., & Ni, Z. (2010). Information from relationship lending: Evidence from loan defaults in China. European Banking Center Discussion Paper No. 2009-10S. Cosci, S., & Meliciani, V. (2002). Multiple banking relationships: Evidence from the Italian experience. Manchester School Supplement, 70, 37
54. Cull, R., Haber, S., & Imai, M. (2011). Related lending and banking development. The World Bank Development Research Group, Finance and Private Sector Development Team Policy Research Working Paper No. 5570. Degryse, H., Kim, M., & Ongena, S. (2009). Microeconomics of banking: Methods, applications and results. Oxford: Oxford University Press. Degryse, H., Masschelein, N., & Mitchell, J. (2005). Staying, dropping, or switching: The impacts of mergers on small firms. The Review of Financial Studies, 24(4), 1102
1140. Detragiache, E., Garella, P., & Guiso, L. (2000). Multiple versus single banking relationships: Theory and evidence. The Journal of Finance, 55(3), 1133
1161. Diamond, D. (1984). Financial intermediation and delegated monitoring. Review of Economic Studies, 51, 393
414. Diamond, D. (1991). Monitoring and reputation: The choice between bank loans and directly placed debt. Journal of Political Economy, 99, 689
721. Dietsch, M. (2003). Financing small businesses in France. European Investment Bank Working Paper No. 93-119. Dinc, I. S. (2000). Bank reputation, bank commitment, and the effects of competition in credit markets. Review of Financial Studies, 13(5), 781
812. Dong, B., & Li, Y. (2010). The relationship banking paradox: No Pain No Gain versus raison d’etre. School of International Trade and Economics Working Paper. EBC. (2005). Consolidation and diversification in the euro area banking sector. Monthly Bulletin, May.

A Case Study of the Polish Credit Market

259

Elsas, R. (2005). Empirical determinants of relationship lending. Journal of Financial Intermediation, 14, 32
57. Elyasiani, E., & Goldberg, L. G. (2004). Relationship lending. A survey of the literature. Journal of Economics and Business, 56, 315
330. Gersl, A., & Jakubı´ k, P. (2010). Relationship lending, firms’ behaviour and credit risk: Evidence from the Czech Republic. IES FSV Charles University Working Paper No. 22/2010. Giovanni, F., Kang, T., & Kim, I. (2001). The value of relationship banking during financial crises: Evidence from the Republic of Korea. The World Bank Working Paper No. 2553. Harhoff, D., & Ko¨rting, T. (1998). Lending relationships in Germany: Empirical results from survey data. Journal of Banking and Finance, 22, 1317
1353. ´ Warsaw: PWN. Heffernan, S. (2007). Nowoczesna bankowosc´. Hodgman, D. R. (1963). Commercial bank loan and investment policy. Bureau of Economic and Business Research, University of Illinois, Urbana-Champaign, IL. Hommel, U., & Schneider, H. (2003). Financing the German mittelstand. Investment Bank Papers, 8, 53
80. Hoshi, T., Kashyap, A., & Scharfstein, D. (1990). The role of banks in reducing the costs of financial distress in Japan. Journal of Financial Economics, 27(1), 67
88. Jimenez, G., & Saurina, J. (2004). Collateral, type of lender and relationship banking as determinants of credit risk. Journal of Banking and Finance, 28, 2191
2212. Kashyap, A. K., Stein, J. C., & Wilcox, D. W. (1993). Monetary policy and credit conditions: Evidence from the composition of external finance. American Economic Review, 83(1), 78
98. Keeley, M. C. (1990). Deposit insurance risk and market power in banking. American Economic Review, 80(5) 1183
1200. Koch, T. W., & MacDonald, S. S. (2010). Bank management. Mason, OH: SouthWestern Cengage Learning. Machauer, A., & Weber, M. (2000). Number of bank relationships: An indicator of competition, borrower quality, or just size? CFS Working Paper No. 2000/06. Memmel, C., Schmieder, C., & Stein, I. (2007). Relationship lending
Empirical evidence for Germany. Deutsche Bundesbank Discussion Paper No. 14/2007. Nam, S.-W. (2004). Relationship banking and its role in corporate governance. ADB Institute Research Paper No. 56. Neuberger, D., & Rathke, S. (2006). Microenterprises and multiple bank relationships: Evidence from a survey among professionals. University of Rostock Institute of Economics Thuenen-Series of Applied Economic Theory, No. 61. OECD. (2003). OECD science, technology and industry scoreboard 2003. Retrieved from http://www.oecd-ilibrary.org/content/book/sti_scoreboard-2003-en. doi: 10.1787/sti_scoreboard-2003-en Ogawa, K., Sterken, E., & Tokutsu, I. (2005). Bank control and the number of bank relations of Japanese firms. CESifo Working Paper No. 1589. Ogura, Y., & Yamori, N. (2008). Lending competition and relationship banking: Evidence from Japanese prefectural level data. Retrieved from http://papers.ssrn. com/sol3/papers.cfm?abstract_id = 1488667

260

M. Pawłowska et al.

Ongena, S., & Smith, D. C. (2000a). Bank relationships: A survey. In P. Haker & S. A. Zenios (Eds.), The performance of financial institutions. London: Cambridge University Press. Ongena, S., & Smith, D. C. (2000b). What determines the number of bank relationships. Journal of Financial Intermediation, 9(January), 26
56. Petersen, M. A., & Rajan, R. G. (1994). The benefits of lending relationships: Evidence from small business data. The Journal of Finance, 49(1), 3
37. Petersen, M. A., & Rajan, R. G. (1995). The effect of credit competition on lending relationships. The Quarterly Journal of Economics, 110(2), 406
443. Rajan, R. G. (1992). Insiders and outsiders: The choice between informed and arm’s-length debt. Journal of Finance, 47, 1367
1400. Rajan, R. G. (1996). Why banks have a future: Towards a new theory of the commercial bank. Journal of Applied Corporate Finance, 9(2), 114
128. Rajan, R. G. (1998). The past and future of commercial banking viewed through an incomplete contract lens. Journal of Money, Credit and Banking, 30, 524
550. Rajan, R. G., & Zingales, L. (1998). Financial dependence and growth. The American Economic Review, 88(3), 559
589. Schildbach, J. (2011). Home, sweet home? International banking after the crisis. Deutsche Bank Research EU Monitor, No. 80. Sharpe, S. A. (1990). Asymmetric information, bank lending, and implicit contracts: A stylised model of customer relationships. The Journal of Finance, 45(4), 1069
1087. Stiglitz, J. E., & Weiss, I. (1981). Credit rationing in markets with imperfect information. American Economic Review, 71(3), 393
410. ´sci kredytu Tymoczko, I., & Pawłowska, M. (2007). Uwarunkowania dostepno ˛ bankowego
analiza. Bank i Kredyt, 38(6), 47
68. Van Hoose, D. (2010). The industrial organization of banking, bank behavior, market structure, and regulation. Berlin Heidelberg: Springer-Verlag. Von Rheinbaben, J., & Ruckes, M. (2004). The number and the closeness of bank relationships. Journal of Banking and Finance, 28, 1597
1615. Von Thadden, E. L. (2004). Asymmetric information, bank lending and implicit contracts: the winner’s curse. Finance Research Letters, 1(1), 11
23. Wallace, V., & Herrick, A. (2009). The future of banking: reappraising core capabilities after the crisis. Booz & Company.Retrieved from http://www.booz.com/ media/uploads/Future_of_Banking.pdf Wood, J. (1975). Commercial bank loan and investment behavior. London: Wiley. Wooldridge, J. (2002). Econometric analysis of cross-section and panel data. MIT Press. Yafeh, Y., & Yosha, O. (2001). Industrial organization of financial systems and strategic use of relationship banking. European Finance Review, 5(1-2), 63
78.

261

A Case Study of the Polish Credit Market

Appendix: Breakdown of Relationships with Banks and Estimation Results Table A.1: Year

The Number of Bank-Firm Relationships
Basic Information Number of enterprises in the sample

Number of relationships Mean/Median

Maximum

1997 1998 1999 2000 2001 2002 2003 2004 2005 2006 2007 2008 2009 2010

5,012 6,254 7,562 8,832 9,247 9,277 9,877 9,625 10,169 11,421 13,586 15,603 15,498 16,356

1.6 1.6 1.6 1.7 1.6 1.6 1.6 1.6 1.6 1.6 1.6 1.5 1.5 1.5

29 32 18 20 17 19 19 18 16 17 19 22 21 21

1997
2010

32,241

1.6/1

32

Source: Own study. Note: The data include nonfinancial enterprises, toward which the exposure of banks providing credit facilities exceeds PLN 2 million. The adopted threshold of PLN 2 million concerns total debt of the firm in all banks (indicated on NBP form credit register). Failure to take the threshold into account would cause overestimation of the number of firms deciding to maintain single-bank relationships since this group would include both firms for which this strategy is a matter of choice and those which must be specified on forms only in one bank, due to the structure of the form. For instance, a firm which has a loan with a bank in the amount of PLN 2.5 million could theoretically be specified on forms by five banks (by taking out loans of PLN 500 thousand in each bank)
being restricted to one institution is therefore the effect of its choice. On the other hand, a firm with a loan with a single bank, amounting for example to PLN 700 thousand, may be specified on form in only one bank (since the reporting threshold is PLN 500 thousand).

66.1 63.1 62.1 60.4 61.3 61.1 61.9 61.9 62.7 64.2 64.7 66.0 67.0 67.5

39.9 34.7 33.4 29.4 31.5 35.8 32.9 32.0 35.1 35.5 36.0 38.3 40.4 40.3

22.1 24.6 25.5 26.5 26.5 26.6 26.3 26.0 25.4 24.9 24.1 23.5 22.8 22.4

17.4 19.5 19.8 22.4 22.6 20.4 24.0 21.9 19.0 18.6 18.4 18.1 17.9 18.3

Share in debt

Share in the number of enterprises

Share in the number of enterprises

Share in debt

Relationships with two banks

Relationships with one bank

7.1 7.4 7.2 7.7 7.5 7.7 7.5 7.8 7.6 6.6 6.8 6.4 6.4 6.4

Share in the number of enterprises 11.1 10.9 11.5 12.1 12.0 11.8 10.6 12.6 12.3 10.3 10.6 9.5 10.2 9.7

Share in debt

Relationships with three banks

2.5 2.4 2.6 2.7 2.5 2.5 2.3 2.2 2.1 2.2 2.5 2.2 2.1 2.0

Share in the number of enterprises 6.1 7.7 7.9 7.8 6.3 7.1 6.3 5.8 6.4 6.6 7.1 7.1 6.4 6.6

Share in debt

Relationships with four banks

1.1 1.2 1.1 1.1 1.0 1.1 0.9 1.0 1.1 0.9 0.9 1.0 0.9 0.9

Share in the number of enterprises 5.3 5.1 4.5 4.6 5.9 5.8 6.0 6.7 5.2 5.8 5.7 5.6 4.7 5.2

Share in debt

Relationships with five banks

Breakdown of the Number of Relationships with Banks in 1997
2010 (in %)

1.2 1.2 1.5 1.6 1.2 1.0 1.0 1.2 1.1 1.0 1.0 0.9 0.8 0.8

Share in the number of enterprises

20.3 22.1 22.8 23.7 21.6 19.0 20.3 21.1 22.0 23.2 22.2 21.4 20.5 19.8

Share in debt

Relationships with more than five banks

Source: Own study. Note: The data include nonfinancial enterprises, toward which the exposure of banks providing credit facilities exceeds PLN 2 million.

1997 1998 1999 2000 2001 2002 2003 2004 2005 2006 2007 2008 2009 2010

Year

Table A.2:

262 M. Pawłowska et al.

263

A Case Study of the Polish Credit Market

Table A.3:

Durability of Single-Bank Relationships

Year

t
1

t
2

t
3

t
4

t
5

1998 1999 2000 2001 2002 2003 2004 2005 2006 2007 2008 2009 2010

95.9 87.7 94.8 81.9 93.4 94.1 93.7 92.8 92.8 86.1 94.6 95.5 95.1

82.6 80.0 77.5 77.1 86.6 86.5 84.5 84.0 78.2 81.4 88.5 89.5

75.7 63.9 72.4 70.2 78.1 78.2 75.9 70.4 74.0 76.7 82.6

60.6 60.7 65.8 63.4 71.6 70.4 64.5 66.3 68.8 70.8

58.6 56.4 59.5 58.2 64.8 60.1 60.1 61.7 63.6

Source: Own study. Notes: The table shows the percentage share of enterprises which had a single-bank relationship in a given year and also had a relationship with only one and the same bank the year before (column t
1), 2 years before (column t
2), etc. For instance among firms which had a single-bank relationship in 1999, 87.7% had a relationship with only a single bank also the year before, 82.6% also had a relationship only with this one bank 2 years before. Only enterprises which were in the database in both analyzed periods are taken into account.

Table A.4:

Panel Data: Macrovariables

Year

GDP change (in %)

CR5

HHI

lb

aktb_GDP

lgpw

1999 2000 2001 2002 2003 2004 2005 2006 2007 2008 2009 2010

6.6 2.7 0.5 2.2 4.7 4 4.4 6.6 6.6 3.2 3.3 3.8

47.7 46.5 54.7 53.4 52.3 50.2 48.7 46.5 46.6 44.6 44.5 43.9

0.079076 0.076136 0.089419 0.087696 0.083016 0.076490 0.073029 0.071468 0.072136 0.062060 0.065841 0.064997

77 73 69 59 58 57 61 63 64 70 67 70

59.1 59.2 61.8 57.7 58.0 58.2 59.7 64.4 68.0 85.0 82.0 82.0

221 225 230 216 203 230 255 284 375 458 486 585

Source: NBP, PFSA, and own calculations.

264

M. Pawłowska et al.

Table A.5:

Table of Correlations of Variables lk

nplp

HHI

CR5

lb

lgpw

hightech foreign

lk 1 nplp −0.0092 1 HHI −0.0305 0.1789 1 CR5 −0.0197 0.1814 0.9146 1 lb 0.0132 −0.1832 −0.5655 −0.6274 1 lgpw 0.0202 −0.1924 −0.8753 −0.8831 0.5553 1 hightech 0.0064 0.068 0.0627 0.0565 −0.0196 −0.0638 foreign 0.2252 −0.019 −0.0038 0.0033 −0.0082 −0.0004 GDP 0.0204 0.0439 −0.1401 −0.0897 −0.4566 0.0861 aktb_GDP 0.0278 −0.2062 −0.9033 −0.8056 0.6799 0.9085

1 0.0555 −0.0228 −0.0596

1 0.017 0.002

GDP

aktb_GDP

1 −0.0346

1

Source: Own calculations.

Table A.6:

Diversification of the Explained Variable

Total variability Between group variability Within group variability

0.4799648 0.2959134 0.3723836

Source: Own calculations.

Table A.7:

Panel Data: Basic Statistics of Variables

Year

No. of observations

lk

nplp

nplba

1999

Mean Standard Deviation Min Max

18,833

5,045.917 48,071.69 0 5,305,561

14.646 0.346415 0 38.24053

15.9076 15.86631 0 100

2000

Mean Standard Deviation Min Max

21,874

5,067.838 44,027.15 0 4,906,411

17.6251 0.423835 0 211.7132

19.1532 16.26596 0 100

2001

Mean Standard Deviation Min Max

22,680

5,186.148 43,232.21 0 4,518,261

23.7727 0.419852 0 100.0006

24.98892 17.10383 0 100

2002

Mean Standard Deviation Min Max

22,357

4,995.149 31,689.5 0 1,605,729

30.0239 0.452532 0 100.1232

32.16102 18.77899 0 100

265

A Case Study of the Polish Credit Market

Table A.7:

(Continued )

Year

No. of observations

lk

nplp

nplba

2003

Mean Standard Deviation Min Max

23,405

4,902.986 33,036.66 0 1,735,430

32.3382 0.463188 0 100

33.8905 20.08616 0 100

2004

Mean Standard Deviation Min Max

22,459

4,867.611 29,733.97 0 1,636,510

24.695 0.427233 0 100

22.26393 18.13393 0 100

2005

Mean Standard Deviation Min Max

23,117

4,886.803 30,613.9 0 1,672,580

18.8775 0.388068 0 100

14.69442 15.08236 0 100

2006

Mean Standard Deviation Min Max

25,703

5,137.899 33,806.06 0 2,376,191

13.2169 0.335416 0 100

9.257541 11.7928 0 100

2007

Mean Standard Deviation Min Max

30,793

5,334.951 33,613.15 0 2,786,907

9.1219 0.284095 0 100

6.253481 9.71584 0 100

2008

Mean Standard Deviation Min Max

35,779

5,873.391 42,749.05 0 4,395,883

7.2911 0.255945 0 100

6.160959 9.10171 0 100

2009

Mean Standard Deviation Min Max

35,524

5,744.941 40,547.73 0 4,167,259

9.4257 0.286759 0 100

10.04738 11.03561 0 100

2010

Mean Standard Deviation Min Max

37,969

5,251.597 34,121.49 0 3,517,422

9.6251 0.29049 0 100

9.910863 9.867466 0 100

Source: NBP and own calculations. a Share of nonperforming loans for individual industries.

HHI

Concentration/competition in the banking sector

foreign

Form of ownership

Control variables (characteristics of industries)

_cons

Bud

Tr

Us

hightech .62131*** (.0461012) .52541*** (.0384873) .0888 (.0770593) .3431*** (.0594074) −5.5158*** (.1455485)







aktb_GDP

Development of the banking sector Innovativeness .623542*** (.0274439) .5243 (.0191982) .0911 (.0345509) .3394*** (.030823) −3.67457 (.1435193)

−.00469 (.005909)


.012642** (.0057716)


GDP







Competition on the part of capital market Economic development

4.0671*** (.1325039)







−.19661*** (.0457999) .62978*** (.0485353)


0.0920*** (.0058232)


2.397*** (0.0920)***





−0.0983*** (0.00308)





−.63049*** (.0152683) 1.01912*** (.0343286)


−.06111*** (.0145433) 1.0125*** (.01954)


−.65825*** (.006353) 1.1874*** (.0939739) −2.9489*** (.054192)


lgpw

lb

CR5

nplp

Credit risk

lk

(3)a

(2)

(1)

Panel Data: Estimation Results and Significance Tests

Size of exposure

Table A.8:

.65302*** (.046852) .51411*** (.039082) .12458 (.078273) .29856*** (.0603357) −2.241*** (.1934697)




1.101*** (.0262215) .046684*** (.0056709)








−.68665*** (.0148565) 1.2325*** (.033634)


(4)

.0461185*** (.0461185) .00919*** (.0384849) .52212 (.077036) .0777*** (.0594287) 2.0097*** (.1336531)

.03085*** (.0009639)














−.66445*** (0146118) 1.2251*** (.0332193)


(5)

266 M. Pawłowska et al.

−59,250.462 (0.0000) χ2(24)=3,911.2 (0.0000) 118,520.9/ 118,617.4

−59,252.707 (0.0000) χ2(24)=3,910.8 (0.0000) 118,525.4/ 118,621.9

−53,680.292 (0.0000) χ2(24)=2,723.9 (0.0000) 105,270.4/ 105,423.2

−59,680.292 (0.0000) χ2(24)=2,782.9 (0.0000) 105,270.4/ 105,423.2

−58,997.11 (0.0000) χ2(24)=4,224.7 (0.0000) 118,013.6/ 118,110.1

Source: Own calculations. *Significant at the level of 10%, **5%, ***1%; standard errors provided in brackets. Us, services; Tr, transport; Bud, construction. The variable specifying industrial processing has been removed due to colinearity (it constitutes the base level). a The data panel did not include firms from the construction industry (Bud).

AIC/BIC information criteria

Wald test

Value of credibility function

A Case Study of the Polish Credit Market 267

268

Table A.9:

M. Pawłowska et al.

Panel Data: Control of Correctness of the Results

Size of exposure Credit risk Innovativeness Forms of ownership Control variables

lk nplp hightech foreign 2001 2002 2003 2004 2005 2006 2007 2008 2009 2010 _cons

(1)

(2)a

−.69833*** .0151394 1.115197*** (.0344596)


−.731843*** (.016342) 1.08479*** (.0194619) −.111129** (.0483615) .618368*** (.026984) −.120395** (.0511692) −.137071* (.0516118) .4455662*** (.0511469) .3180451*** (.0521232) .4785411*** (.0518318) .5903543*** (.0512696) .4287337*** (.0512696) .5995026*** (.0496514) 1.081105*** (.0494979) 1.554292*** (.0500409) 4.441323*** (.1417306)

.6313174*** (.0479941) −.144938** (.0490875) −.115478* (.0495022) .4398138*** (.0490665) .317012*** (.0499922) .4815298*** (.0496405) .5935254*** (.0490274) .3875255*** (.0490274) .5423408*** (.0481308) 1.059071*** (.0472145) 1.558778*** (.0469707) 4.145872*** (.1313846)

Source: Own calculations. *Significant at the level of 10%, **5%, ***1%; standard errors provided in brackets. a

The data panel did not include firms from the construction industry.