Contagious Currency Crises

Transcription

1 Western Michigan University ScholarWorks at WMU Dissertations Graduate College Contagious Currency Crises Fasika Damte Haile Western Michigan University Follow this and additional works at: Part of the Economics Commons Recommended Citation Haile, Fasika Damte, "Contagious Currency Crises" (2003). Dissertations This Dissertation-Open Access is brought to you for free and open access by the Graduate College at ScholarWorks at WMU. It has been accepted for inclusion in Dissertations by an authorized administrator of ScholarWorks at WMU. For more information, please contact

2 CONTAGIOUS CURRENCY CRISES by Fasika Damte Haile A Dissertation Submitted to the Faculty of The Graduate College in partial fulfillment of the requirements for the Degree of Doctor of Philosophy Department of Economics Western Michigan University Kalamazoo, Michigan April 2003

3 CONTAGIOUS CURRENCY CRISES Fasika Damte Haile, Ph.D. Western Michigan University, 2003 Currency crises, prior to the 1990s, were thought to be the result of inconsistencies between domestic macroeconomic policies and the exchange rate commitment. But the collapse of the European Exchange Rate Mechanism in 1992, the 1997 Asian crisis and the most recent crisis in Latin America have shifted the focus to models based on self-fulfilling expectations and on contagion. This has resulted in the development of different theoretical models suggesting different mechanisms by which contagion works. But there is still relatively little empirical consensus on how crises spread across countries. My dissertation is intended to fill in the void by testing for contagion and identifying the transmission channels for crises. With this objective. I have estimated a panel probit model using quarterly data ( ) from 37 advanced and emerging market economies. Two main points make my work different from other studies on contagion. First, crises are identified using a relatively more objective method based on extreme value theory. The main argument for this approach is that exchange rate changes during crisis periods (like the 1997 Asian crisis) are outliers, making the distribution of exchange rate changes

4 fat-tailed. Extreme value theory allows us to determine the tail mass or observations measured via the tail index, and hence we take all the extreme outliers as indicators of crisis. Since we do not know the specific parametric distribution to which the crisis index belongs, the threshold to determine the tail observations is determined using Monte Carlo simulation. Second, I have allowed for crises to spread on a broader basis among (a) major trade partners/competitors, (b) countries with strong financial linkages, (c) countries with similar macroeconomic fundamentals, and (d) neighbors. Results from my estimations reveal that countries face currency crises because of unsustainable macroeconomic fundamentals and contagion. In all cases considered, contagion works via the trade linkages channel. The results also show that the probability of a crisis in a given country increases as the number of its neighboring countries in crisis increases implying the presence of the neighborhood effects in the contagious spread of crisis. Because countries with sound macroeconomic fundamentals are still vulnerable to currency crisis, to prevent contagion countries, mainly major trade partners, need to consider alternative policies such as fixing their exchange rates collectively in a more firm and credible way. At the extreme, major trade partners may adopt a regional currency, the scheme followed by some of the European countries in creating the euro, to prevent contagion among members.

5 INFORM ATION TO USERS This manuscript has been reproduced from the microfilm master. UMI films the text directly from the original or copy submitted. Thus, some thesis and dissertation copies are in typewriter face, while others may be from any type of computer printer. The quality of this reproduction is dependent upon the quality of th e copy submitted. Broken or indistinct print, colored or poor quality illustrations and photographs, print bleedthrough, substandard margins, and improper alignment can adversely affect reproduction. In the unlikely event that the author did not send UMI a complete manuscript and there are missing pages, these will be noted. Also, if unauthorized copyright material had to be removed, a note will indicate the deletion. Oversize materials (e.g., maps, drawings, charts) are reproduced by sectioning the original, beginning at the upper left-hand comer and continuing from left to right in equal sections with small overlaps. ProQuest Information and Learning 300 North Zeeb Road, Ann Arbor, Ml USA

6

7 UMI Number: UMI* UMI Microform Copyright 2003 by ProQuest Information and Learning Company. All rights reserved. This microform edition is protected against unauthorized copying under Title 17, United States Code. ProQuest Information and Learning Company 300 North Zeeb Road P.O. Box 1346 Ann Arbor, Ml

8 Copyright by Fasika Damte Haile 2003

9 ACKNOWLEDGMENTS I would like to thank the many people who have helped me in the development of my dissertation from the beginning to its final stage. First, I would like to take this opportunity to express my heartfelt gratitude to Professor Susan Pozo, Chair of my Dissertation Committee, who has been a constant source of encouragement and excellent guidance. I have benefited from her constructive and useful comments and suggestions. She has also been very helpful by reviewing and proofreading draft after draft in an amazing speed and punctuality. I am also very grateful to Professor Matthew Higgins, member of my Dissertation Committee, for his insightful and constructive comments and suggestions on the technical and other aspects of my dissertation. I am also indebted to Professors Sisay Asefa and Ahmed Hussen for their constant encouragement and the constructive advice they have given to me while writing my dissertation. All the remaining errors are, however, mine. I should also thank the Department of Economics and Western Michigan University for having provided me with all the resources that I need as a graduate student. Last but not least. I would like to thank my wife, Nigist, who has been by side in both bad and good days. Fasika Damte Haile ii

10 TABLE OF CONTENTS ACKNLOWLEGEMENTS... LIST OF TABLES... LIST OF FIGURES... ii v viii CHAPTER 1. INTRODUCTION LITERATURE REVIEW CURRENCY CRISES AND CONTAGION: TESTS USING A PANEL OF 20 OECD COUNTRIES Introduction Methodology Model Data Empirical Results Conclusions INDENTIFICATOIN OF CURRENCY CRISES USING THE EXTREME VALUE THEORY Introduction Theory iii

11 Table of Contents-Continued 4.3. Estimation of the Tail Index Identification of Currency Crises Conclusions HOW DO CURRENCY CRISES SPREAD? NEW EVIDENCE USING THE EXTREME VALUE THEORY Introduction Methodology The Model Data Estimation Results Conclusions CONCLUSIONS AND IMPLICATIONS APPENDIX A. Data Description REFERENCES iv

12 LIST OF TABLES 1. Number of Currency Crises: 1960:1-1998: Sample of Trade Linkage Weights Among OECD Countries (in % ) Sample of Macroeconomic Similarity Weights Among OECD Countries: 1995Q1 (in %) Sample of Financial Linkage Weights Among OECD Countries (in % ) Common Threshold for Crisis Identification: Full Sample (1960:1-1998:4), Panel Probit Model Country Specific Threshold for Crisis Identification: Full Sample (1960:1-1998:4), Panel Probit Model Common Threshold for Crisis Identification: Eichengreen et al. (1996) Sample (1960:1-1993:4), Panel Probit Model Country Specific Threshold for Crisis Identification: Eichengreen et al. (1996) Sample (1960:1-1993:4), Panel Probit Model Common Threshold for Crisis Identification: Eichengreen et al. (1996) Sample (1960:1-1993:4) & No Financial Linkages, Panel Probit Model Country Specific Threshold for Crisis Identification: Eichengreen et al. (1996) Sample (1960:1-1993:4) & No Financial Linkages, Panel Probit Model v

13 List of Tables-Continued 11. Sample of G ARCH Estimation Output for the Three Components of EM P Properties of the EMP Distribution fora Sample of Countries (60:1-98:12) Optimal Number of the Highest Order Statistics (m*) for Different Sample Size and Student-t Distribution: Result from Simulated Random Samples Hill-Tail Index Estimate of the U.S. EMP: 1960:1-1998:12 (Sample Size 467) Number of Tail Observations of the Actual EMP Distribution of the Sample Countries Number of Currency Crises in Quarter Summary of the Crises Identified Based on the Extreme Value Approach Periods of High Exchange Market Pressure on the U.S. dollar Identified by the Extreme Value Approach Number of Quarters in Crisis Identified by the Extreme Value and the Standard Approaches Sample of Trade Linkage Weights Among OECD & Emerging Countries (in %) Sample of Macroeconomic Similarity Weights Among OECD & Emerging Countries: 1992Q1 (in %) vi

14 List of Tables-Continued 21. Sample of Financial Linkage Weights Among OECD & Emerging Countries (in % ) Test of Contagion: 20 OECD Countries (1960:1-1998:4), Panel Probit Model Test of Contagion with the Neighborhood Effects Captured by a Dummy Variable: Extended Data from 37 Countries (1960:1-1998:4), Panel Probit Model... I ll 24. Test of Contagion with the Neighborhood Effects Captured by the Number of Neighbors in Crisis: Extended Data (1960:1-1998:4), Panel Probit Model Test of Contagion with Heteroskedastic Residuals: 20 OECD Countries (1960:1-1998:4), Panel Probit Model Test of Contagion with Heteroskedastic Residuals: Extended Data from 37 Countries (1960:1-1998:4), Panel Probit Model vii

15 LIST OF FIGURES 1. Percentage of Countries in Currency Crisis: 20 OECD Countries (Common Threshold) Percentage of Countries in Currency Crisis: 20 OECD Countries (Country Specific Threshold) Percentage of Countries in Currency Crisis: 37 Countries (Extreme Value Approach) Percentage of Countries under Currency Crisis: 20 OECD Countries (Extreme Value Approach) viii

16 CHAPTER 1 INTRODUCTION The main objective of this dissertation is to test whether currency crises are contagious and to identify the channels through which crises are transmitted across countries. Many economists have now realized that a role was played by contagion in propagating the financial and currency crises of advanced and emerging market economies in the 1990s. Different theoretical models have been recently developed suggesting different mechanisms by which crises have been transmitted across countries. But there is relatively little empirical consensus on how crises spillover across countries while the frequency and intensity of crises points to the urgency of additional empirical work to come up with solutions for crisis prevention, crisis management and crisis resolution1. Currency crises prior to 1990s did not appear to spread across countries with the virulence and speed observed recently. The earlier literature tried to explain the crises as the result of inconsistencies between fiscal and monetary policies and the existing exchange rate commitment (so called first generation models of currency crises such as Krugman, 1979). The collapse of the European Exchange Rate 1See White (2000) for the different solutions suggested in the literature with respect to these three objectives: prevention, management and resolution. 1

17 Mechanism (ERM) in 1992/93, the Mexican Peso crisis in 1994 and the Asian crisis in 1997 have, however, shifted the focus to models based on self-fulfilling expectations (Obstfeld, 1995) and on contagion (see Dombusch et al., 2000, Wolf, 1999 and Pericoli and Sbracia, 2001 for a comprehensive survey of models of contagion). The Asian currency crisis, for example, began in July 1997 with the Thai baht devaluation. It then spread to Malaysia, the Philippines, and Indonesia in the third quarter of Prior to the 1997 crisis, all these countries had a few common characteristics: an appreciating real exchange rate, large current account deficits and financial sector squeezes linked to overexposure to a property market whose prices had fallen sharply (see Masson, 1998 and Krugman, 1998). But the currency pressures also quickly spread to Hong Kong, Singapore and Korea, economies with strong current account and fiscal positions. The crisis even jumped surprisingly to several emerging markets outside the region, notably to Brazil and Russia (see IMF, 1998). This experience coupled with the earlier crises in 1992/1993 and 1994/1995 led economists to suspect that crises in the 1990s were contagious. Surprisingly, there is not yet one generally accepted definition for contagion. Forbes and Rigobon (1999a) define contagion as a significant increase in crossmarket linkage after a crisis in one or more countries while Eichengreen et al. (1996) define it as an increase in the probability of a crisis in a given country conditional on the occurrence of a crisis somewhere else. Contrary to the above two definitions, Masson (1998) defines contagion to mean only those transmissions of crises that 2

18 cannot be identified with observed changes in macroeconomic fundamentals. In this dissertation, contagion is defined as an increase in the probability of a crisis in one country given a crisis elsewhere as this allows to test for the existence of contagion and to empirically identify the transmission channels of crisis. Like the multiplicity of the definitions of contagion, there are different theoretical models that show how crises end up spreading across countries. Some of the major models of contagion are based on trade linkages (Gerlach and Smets, 199S) and macroeconomic similarities (Goldstein, 1998 and Eichengreen et al., 1996). Other models are based on financial linkages, neighborhood effects, or exogenous shifts in investors beliefs (herd behavior) (Masson, 1998, Calvo, 1999, Kaminsky and Reinhart, 2000 and Calvo and Mendoza, 2000). What is lacking is a general consensus on empirical findings on the relevant contagion channels. Existence of contagion has important implications. Because no open economy can insulate itself from what is happening in the rest of the world, to prevent contagious currency crises countries may need to adopt regionally or globally coordinated measures. But the specific measures that should be taken to prevent the spread of currency crisis presuppose knowledge of the relevant contagion channels. If the trade contagion channel is relevant, countries may need to fix their exchange rates collectively in a more firm and credible way in order to avoid speculative attacks following loss of international competitiveness caused by currency depreciation by one or more of their major trading partners. At the extreme, international cooperation of the countries may lead to the creation of a common 3

19 currency. If, on the other hand, the financial contagion channel is relevant, countries may need to impose capital controls. Some suggest that a lender of last resort, like the IMF or the World Bank, would need to be instituted to neutralize the financial contagion channel by providing liquidity support. As the foregoing discussion points out, the rise in the frequency, intensity and time clustering of the crises has now forced both policy makers and academics to focus on contagion as an explanation for the observed sequence of events. Why do currency and financial crises hit selected countries within a very close time period? Are those countries simultaneously under crisis hit by common shocks? Or do they have unsustainable fiscal and monetary policies or unsustainable current account positions to the extent that both countries face crises simultaneously? If each of these is not the case, why and how does a crisis in one country transmit to other selected countries that have sound macroeconomic fundamentals? This dissertation seeks to address some of the above questions. More specifically, it has attempted to provide answers to two related questions- does a currency crisis in one country spread to other selected countries? What are the channels through which crises spread across countries? To address these questions, my dissertation is structured into a series of inter-related chapters in which I sequentially develop the statistical and economic methodology to analyze currency crisis periods and how they spread across countries. Following the second chapter on literature review, the third chapter of this dissertation re-estimates the probit model of Eichengreen et al. (1996) using data 4

20 from 20 OECD countries. The probit model in Eichengreen et al. (1996) tests for contagion through trade linkages and macroeconomic similarity channels. But my study extends Eichengreen et al. (1996) in two ways. First, it allows for a third contagion channel, the financial linkages channel. That is, in addition to the trade and the macroeconomic similarity channels considered in Eichengreen et al. (1996), the financial linkages channel is considered. My second extension of Eichengreen, et al. paper is to employ two different but commonly used procedures to identify periods of currency crises. This allows for checking the sensitivity of the results to the identification of periods of currency crises. The first method that I use to identify periods of currency crisis is the Eichengreen et al. (1996) procedure in which one divides the sample into crisis and tranquil periods using a threshold that is common to the entire countries in the sample. The second method uses the Kaminsky and Reinhart (2000) procedure that applies a country specific threshold to identify periods of crisis in each country. In both procedures, a country experiences a currency crisis at time t if its index of currency crisis given by the exchange market pressure (EMP) at time t is above the specified level of threshold. Overall, the estimation results of various probit models in chapter three indicate that currency crisis is contagious at least among members of the OECD countries. However, the channel by which contagion operates appears sensitive to the identification of currency crises. When currency crises are identified by the Kaminsky and Reinhart (2000) procedure, countries macroeconomic similarity appears to be the relevant contagion 5

21 channel. But there is no single contagion channel when the Eichengreen et al. (1996) procedure of crisis identification is applied. In the latter case, the macroeconomic similarity contagion channel appears to explain transmission of crises when a relatively liberal/moderate threshold value is used to identify periods of crisis. The contagion channel switches to trade linkages when currency crises are identified by a relatively higher threshold value. This inconclusive result related to the relevant contagion channel points to the need to carefully identify periods of currency crisis. In the fourth chapter, the study applies an alternative and relatively more objective method of crisis identification based on the extreme value theory. The main argument for this approach is that exchange rate changes during crisis periods (like the ERM attacks of 1992 and the Asian financial crisis of 1997) are outliers, making the distribution of exchange rate changes fat-tailed". The standard approach to identify periods of currency crisis is to set a threshold based on the mean and the standard deviations of the crisis index. But there is no theoretical justification or consensus in setting the level of threshold. Given this problem, the fourth chapter proposes to identify the tail observations of the EMP as indicators of currency crisis using an alternative approach. One way of determining the tail observations of the crisis index is to employ one of the fat-tailed distributions (example, student-t, non-normal sum-stable and the ARCH processes) (Jansen and De Vries, 1991, Boothe and Glassman, 1987, and Koedijk et al., 1990). Another way of identifying the frequency of outliers is to estimate the tail mass of the distribution of the crisis index using the extreme value 6

22 approach as first proposed in Pozo and Amuedo-Dorantes (2002). The advantage of the extreme value approach over the first alternative is that no specific parametric distribution is assumed in estimating the tail index. The extreme value theory (EVT), which has wide applications in measuring various forms of risks (both natural and in the financial markets), allows us to determine the tail mass measured via the tail index. Since the specific parametric distribution of the crisis index (i.e., the exchange market pressure) is not known, the threshold to determine the tail observations is determined using Monte Carlo simulation. According to the application of the EVT to a sample of 20 OECD countries, a large number of countries were in a crisis in the early 1970s, 1978/79, 1987, and 1992/932. The early 1970s and 1978/79 correspond to the collapse of the Bretton Woods system and the snake, respectively. The years 1987 and 1992/93 correspond to the October 1987 U.S. stock market crash and the European Exchange Rate Mechanism (ERM) crisis, respectively. Based on the standard approach of currency crisis identification (using the mean plus l.s standard deviations as one possible cutoff point as applied in the third chapter), a large number of countries were found in crisis in the early 1970s and 1992/93. Identification of crisis using the extreme value approach allows us to identify additional major crises in 1978/79 and 1987 during which many countries were affected. 2 The analysis in the fourth chapter is actually based on an expanded data set from 37 countries. The 20 OECD member countries are part of the larger data set. 7

23 In the fifth chapter of my dissertation, I re-estimate the panel probit model specified in chapter three. But there are two major changes. First, the test for contagion is undertaken using crises identified by the extreme value approach. Second, other countries from Asia and Latin America are also added to the OECD sample to form an expanded data set representing many different regions of the world. This allows for testing contagion on a broader basis while also allowing for contagion to operate through a fourth channel to capture the neighborhood effects of a crisis. Using EVT and the objectively identified crises, the fifth chapter, therefore, tests whether there is contagion among i) major trade partners/competitors, ii) countries with strong financial linkages such as having common creditors, iii) countries with similar macroeconomic fundamentals and/or iv) neighbors. According to the estimation results from different model specifications, currency crises are contagious. In all cases considered, contagion works through the trade channel. Moreover, the estimation results point out that the probability of a currency crisis in a given country increases as the number of its neighboring countries in crisis increases. The macroeconomic similarity and financial linkage channels turn out significant in none of the estimations. The last chapter of my dissertation provides overall conclusions and implications of the results. The main conclusion of my dissertation is that currency crises are contagious. Contagion is regional and more specifically it operates through the trade channel. The main implication of this result is that countries could prevent contagion by fixing their exchange rates collectively in a more firm and credible way 8

24 in order to avoid speculative attacks following loss of international competitiveness. At the extreme, major trade partners may adopt a regional currency, which is the track followed by some of the European countries in creating the Euro, to prevent contagion among members. 9

25 CHAPTER 2 LITERATURE REVIEW A large number of studies have concluded that the Mexican crisis of 1994/95, the Asian crisis of 1997, the Russian crisis of 1998 and even the earlier ERM crisis of 1992/93 were contagious. Despite a general consensus that contagious currency crises are important phenomena, there is not yet a uniform definition of what constitutes contagion. In what follows a discussion of contagion definitions is presented. Forbes and Rigobon (1999a) define contagion as a significant increase in cross-market linkages after a crisis in an individual country (or a group of countries) without taking a stance on how this shift occurred. They name this shift-contagion. Eichengreen et al. (1996) argue that there is contagion if the probability of a crisis in a given country increases conditional on the occurrence of a crisis elsewhere, after controlling for the standard set of macroeconomic fundamentals. In contrast to these two definitions, Masson (1998) defines contagion to mean only those transmissions of crises that cannot be identified with observed changes in macroeconomic fundamentals. Contagion according to Masson (1998) involves changes in expectations that are not related to changes in a country s macroeconomic fundamentals. Coinciding with the various definitions of contagion, there exist a variety of 10

26 economic models that explain how crises are propagated internationally. Following Masson (1998), these models are divided in the literature into two major categories3. In the first category, crises spread resulting from economic interdependence among different countries. Accordingly, a crisis in one country spreads by changing the macroeconomic fundamentals of other countries. Some of the factors considered in this category for the simultaneous occurrence of currency crises are: common shocks, trade and direct financial linkages between countries. These are generally termed as fundamentals-based contagion models (see Calvo and Reinhart, 1996 and Kaminsky and Reinhart, 2000). In the second category, contemporaneous crises are modeled through shifts in the behavior of investors or other financial agents. Here, a crisis in one country triggers a crisis elsewhere without having any impact on their macroeconomic fundamentals. The crisis spreads because of changes in market sentiment or interpretation of existing information about the economy held by investors. In what follows, a review of some of the individual models of contagion in each category is presented. Common shocks: a common shock, be it regional or global, may serve as the cause for the simultaneous occurrence of crisis across countries. Calvo and Reinhart (1996), for example, cited the sharp increase in the U.S. interest rates in the early 1980s and 1994 as one major cause for the two Mexican crises in 1982 and Another recent 3 A comprehensive survey of the literature is provided in Dombusch, et al. (2000), Pericoli and Sbracia (2001) and Wolf (1999). 11

27 example is the large appreciation of the dollar between 1995 and 1997 and the long lasting slowdown in Japanese growth that might together have contributed to the Asian crisis by weakening the external sector of Asian countries simultaneously (see Baig and Goldfajn, 1998). Trade linkage: trade linkages involving both bilateral and third party market competition could explain contagion through the possibility of loss of international competitiveness (price effects) and income effects (see Gerlach and Smets, 1995 and Glick and Rose, 1999). When a country experiences a crisis marked by a significant currency depreciation, its major trade partners are negatively affected both through loss of competitiveness and through the fall in demand in the crisis country if the latter is experiencing economic downturn, too. The two effects, price and income, work in both the bilateral and third party export markets of the major trade partners. The impact of the spillover through the trade link could be even larger if we consider the possibility of cascading effects4. Macroeconomic similarity: due to incomplete information, investors treat all countries that look alike in their macroeconomic fundamentals as equal. Therefore, once one country is hit by a crisis, investors take this as a wake-up call and view 4 Consider three countries: A, B and C. Assume A and B are major trade partners. For some reason, A's currency depreciates due to a crisis. Due to the price and income effects, B will be affected automatically. Country C is then affected through its trade linkages with both A and B. 12

28 this as new information on what will happen in other countries with some similarities. Investors, then, attack these other countries that have macroeconomic fundamentals similar to those in the crisis country (Goldstein, 1998 and Eichengreen et al., 1996). In this model, crisis spreads to the second country without necessarily having experienced deterioration in its macroeconomic fundamentals. Investors perception is what links crisis from one to another country. Financial linkage: there are many mechanisms by which cross-border spillovers work through financial linkages. Some of the major ones cited in the literature include direct financial linkages, liquidity and incentive problems and herd behavior (see Dombusch et al., 2000 for an elaborate classification). In some of these cases, crisis spreads by changing the fundamentals of other countries when there is, for example, direct financial linkage and in others without any impact on the fundamentals of the non-crisis countries. Calvo (1999), for example, has built a model for contagion based on margin calls for liquidity5 and asymmetric information. In this model, the market is populated with informed and uninformed investors. Given this, a large depreciation of the currency and decline in equity prices in one country may lead to a large capital loss to some informed investors. These losses may induce these investors to sell off good securities in other emerging markets in order to raise cash in anticipation of a higher frequency of redemption. But the uninformed may misread this action of informed 5 A model of contagion based on liquidity is also given in Valdes (199S). 13

29 investors as a signal for low returns in this market. The action of the uninformed on account of a change in their perceptions then depresses equity and other asset prices in the country with healthy fundamentals. Liquidity problems may also be faced by commercial banks that have their lending concentrated in particular regions. If these banks experience a marked deterioration of the quality of loans to one country, they may attempt to reduce the overall risk of their loan portfolio by reducing exposures in other higher risk investments elsewhere, including other countries. Kaminsky and Reinhart (2000) and Van Rijckeghem and Weder (1999) term this the role of common lenders for the contagious spread of crisis. Calvo and Mendoza (2000) present a model of utility maximizing investors where the presence of fixed costs to gather and process country-specific information leads to herd behavior thereby creating the room for the contagious spread of crisis. Due to the fixed costs, most small investors may find it more advantageous to follow the investment patterns of large informed investors. According to their model, globalization increases contagion through herding by weakening incentives for gathering costly information while at the same time strengthening the incentives for imitating arbitrary market portfolios. Empirical Evidence A great deal of the empirical literature on the test for contagion focuses on 14

30 whether there is a fundamental change in the propagation of the transmission mechanism and on the identification of the contagion channels. The studies have looked at the co-movement of asset returns, volatility, and capital flows across countries using cross-market correlation coefficients, ARCH, Logit/Probit and VAR models6. Tests based on cross-market correlation coefficients are the most common and widely used approach to test for contagion. Under this approach, a significant increase in the correlation coefficient of asset returns between two markets after a crisis in one of them is considered as evidence of contagion. This is applied, among others, by Calvo and Reinhart (1996) and Baig and Goldfajn (1998). Using this approach, Calvo and Reinhart (1996) have shown an increase in the co-movement of weekly returns on equities and Brady bonds for Asian and Latin American emerging markets after the 1994 Mexican crisis. Baig and Goldfajn (1998), on the other hand, have provided evidence for a significant rise in the cross-country correlation among currencies and sovereign spreads of Indonesia, Korea, Malaysia, the Philippines and Thailand during the East Asian crisis period. Forbes and Rigobon (1999a) and Rigobon (2001), however, argue that a marked increase in correlation among different countries markets may not be a sufficient proof of contagion. The rise in the post-shock correlation may be due to an increase in volatility following a crisis in one market. Thus, an increase in unadjusted correlation could simply be a continuation of strong transmission mechanisms that 6 A good review of the various methods applied in contagion is provided in Forbes and 15

31 exist in more stable periods. When volatility adjusted correlation of stock indices of 28 countries is used, Forbes and Rigobon (1999b) claim that there is no evidence of contagion during the 1987 U.S. stock market crash, the 1994 Mexican peso crisis, and the 1997 East Asian crisis. The second most commonly used methodology to test for contagion, introduced in Eichengreen et al. (1996), is to examine whether the likelihood of crisis is higher in a given country when there is a crisis elsewhere. One advantage of this approach is that it readily allows statistical tests for the existence of contagion. The method also helps to investigate the channels through which contagion may occur, distinguishing, among others, trade and financial linkages. Eichengreen et al. (1996) use a probit model and a panel of quarterly macroeconomic and political data covering 20 OECD countries from 1959 through 1993 to test for contagious currency crises. The results of their estimation show that the probability of a domestic currency crisis increases with a speculative attack elsewhere and that contagion is more likely to spread through trade linkages than through macroeconomic similarities. Trade as the relevant contagion channel is also supported by Glick and Rose (1999). But the trade weights in Eichengreen et al. (1996) reflect only bilateral trade linkages while Glick and Rose (1999) allow crises to spread only from the Ground Zero (the first crisis) country without allowing for the possibility of cascading effects. Rigobon (2001) and Rigobon (2001). 16

32 The same approach as that of Eichengreen, et al. (1996) is also applied by De Gregorio and Valdes (1999) who have examined whether the crisis indicator of a country is explained by the initial macroeconomic conditions of the country and the weighted average of crisis indicators of other countries during the 1982 debt, the 1994 Mexican and the 1997 Asian crises. Their results indicate that there is a strong neighborhood effect. Trade links and similarity in pre-crisis growth also explain, to a lesser extent, which countries suffer more contagion. The evidence shows that the 1982 debt crisis was as contagious as the Asian crisis, while the Mexican crisis was considerably less so. Finally, both debt composition and exchange rate flexibility limit to some extent contagion, whereas capital controls do not appear to curb it. Some of these results are, however, refuted by Caramazza, Ricci and Salgado (2000) who investigate the ERM, the 1994 Mexican, the Asian, and the Russian crises using the same approach. According to their results, the contagious nature of the Mexican, Asian and Russian crises does not differ much. Fundamentals, including trade and financial (common creditor) linkages and financial fragility, are highly significant in explaining crises, while exchange rate regimes and capital controls do not seem to matter. Using a slightly different approach (by comparing the conditional and unconditional probability of a crisis), Kaminsky and Reinhart (2000) examine how trade and financial sector links influence the pattern of fundamentals-based contagion using monthly data from 20 industrial and developing countries from 1970 through According to their results, contagion is more regional and highly nonlinear. 17

33 Furthermore, when the number of crises in a given cluster is high, the financial sector link via common bank lenders is a more powerful channel of fundamental-based contagion than are trade linkages. The importance of common bank lender is also supported in Van Rijckeghem and Weder (1999) and Hemadez and Valdes (2001). The results in the first study are obtained from a probit model while the second study estimates a pooled OLS model using data from 17 emerging market economies. 18

34 CHAPTER 3 CURRENCY CRISES AND CONTAGION: TESTS USING A PANEL OF 20 OECD COUNTRIES 3.1. Introduction The main objective of this paper is to test whether currency crises are contagious and to identify the channels through which crises are transmitted across countries. Many economists have now realized that a role was played by contagion in propagating the currency crises of many countries in the 1980s and 1990s. Different theoretical models have been developed suggesting different mechanisms by which crises have been transmitted across countries. Transmission channels that have been proposed include trade linkages (Gerlach and Smets, 1993), macroeconomic similarities (Goldstein, 1998 and Eichengreen et al., 1996), financial linkages, neighborhood effects, and exogenous shifts in investors beliefs (herd behavior) (Masson, 1998, Calvo, 1999 and Calvo and Mendoza, 2000). What is lacking is a general consensus on empirical findings on the relevant contagion channels. Testing for contagion and identifying the relevant channels have important implications on how to control and prevent currency crises. Most of all, contagion signifies the need for open economies to adopt regionally or globally coordinated measures to prevent contagious currency crises. But the specific measures that should 19

35 be taken to prevent the spread of currency crisis presuppose knowledge of the relevant contagion channels. If contagion works via the trade channel, countries may need to fix their exchange rates collectively in order to avoid exchange rate/currency pressures following loss of international competitiveness as a result of the currency depreciation of major trading partners. At the extreme, international cooperation of the countries may lead to the creation of a common currency. If, on the other hand, the financial contagion channel is relevant, countries may need to impose capital control. Some suggest that a lender of last resort, like the IMF or the World Bank, would need to be instituted to neutralize the financial contagion channel by providing liquidity support. The 1990s have witnessed on the one hand an increase in economic integration among world economies and on the other hand recurrent currency crises. The major crises during this period include the European Exchange Rate Mechanism (ERM) attacks of 1992/93, the Mexican Peso collapse of 1994/95, the East Asian crisis of 1997, the Russian collapse of 1998, and the Brazilian devaluation of Each of these crises began as country-specific event but appeared to have been rapidly transmitted to other countries that had even sound macroeconomic fundamentals. The Asian currency crisis, for example, began in July 1997 with the Thai baht devaluation. It then spread to Malaysia, the Philippines, and Indonesia in the third quarter of Prior to the 1997 crisis, all these countries had few common features: 20

36 an appreciating real exchange rate, large current account deficits and financial sector problems linked to overexposure to a property market whose prices had fallen sharply (see Masson, 1998 and Krugman, 1998). But the pressures also spread to Hong Kong, Singapore and Korea that had strong current account and fiscal positions. The crisis even jumped to several emerging markets outside the region, notably to Brazil and Russia (see IMF, 1998). When the currencies in Asia reached their low points in January 1998, the Indonesian rupiah had fallen (relative to its July 1997 level) by 81%, the Malaysian ringgit by 46%, and the Thai baht by 55% (IMF, 1998, p. 2). The intensity and time clustering of the crises causes both policy makers and academics to focus on contagion as a principal culprit in the ensuing discussion. Contagion refers to the spread of crisis from one country to the other. This is reflected in the co-movements of exchange rates, stock prices, sovereign spreads and capital inflows. A number of questions have been raised in the literature. Why do currency and financial crises hit selected countries within a very close time period? Are those countries simultaneously under crisis hit by common shocks? Or do they have unsustainable fiscal and monetary policies or unsustainable current account positions to the extent that both countries face crises simultaneously? If each of these is not the case, why and how does a crisis in one country transmit to other selected countries that have sound macroeconomic fundamentals? The literature identifies different transmission channels of shocks from the crisis to the non-crisis countries. Trade and financial linkages, macroeconomic similarities (resulting from internal and external imbalances), and herd behavior are major contagion channels discussed in the 21

37 literature. This paper is motivated by two inter-related questions. Does a currency crisis in one country spread to other selected countries? What are the channels through which crisis spreads across countries? To address these questions, this chapter reestimates the probit model of Eichengreen et al. (1996) using data from 20 OECD countries. My study has, however, extended Eichengreen et al. (1996) in two ways. First, it includes a third contagion channel through financial linkages in addition to the trade and macroeconomic similarity channels considered in Eichengreen et al. (1996). Second, it employs the two commonly used procedures to identify periods of currency crises. This allows for checking the sensitivity of the results to the identification of periods of currency crises. The first method uses the Eichengreen et al. (1996) procedure that divides the sample into crisis and tranquil periods using a threshold common to the entire countries in the sample. The second method uses the Kaminsky and Reinhart (2000) procedure that applies a country specific threshold to identify periods of crisis in each country. Overall, the estimation results of various probit models indicate that currency crisis is contagious. But the relevant contagion channel through which crisis spreads across countries turns out to be sensitive to the identification of periods of crises. This suggests the need for additional research to clarify the definition and determination of periods of currency crisis. In the next chapter, I have applied an alternative and relatively more objective method of crisis identification based on the extreme value theory. 22

38 The rest of this chapter is divided into three sections. Section two discusses the method of study and data sources. Section three provides analysis of the empirical findings and outlines extensions of the study. The last section is devoted to conclusions Methodology The main objective of my dissertation is to provide answers to the two questions raised earlier: is the probability that a country faces a currency crisis affected by a crisis elsewhere? What are the most relevant contagion channels? To address these questions, this chapter re-estimates the probit model of Eichengreen et al. (1996) using data from the same 20 OECD countries. The model is re-estimated first, to see whether the results in Eichengreen et al. (1996) still hold when the estimation period is extended from to Second, most recent studies using different methodologies point out the importance of financial linkages in the contagious spread of currency crises. In response to that, a third contagion channel through financial linkages is added to the trade linkages and macroeconomic similarities contagion channels considered in Eichengreen et al. (1996). Last, there are two commonly used methods to identify currency crisis. This chapter applies the two methods to identify crisis. This allows us to check the sensitivity of the 7 The actual estimation period in Eichengreen et al. (1996) is from 1959:1 to 1993:4. But there are no macroeconomic data for most countries starting 1959:1. Thus, the estimation period in my paper starts from 1960:1. 23

39 estimation results to the identification of crisis before we resort to a new method of crisis identification. As a first step, the paper constructs a currency crisis index to capture the strength of a speculative attack. The term speculative attack is loosely used and it simply means pressure on the existing exchange rate or currency emanating from a process by which investors change the composition of their portfolios, reducing the proportion of domestic currency and raising the proportion of foreign currency (Krugman 1979, p. 312). While investors8 trade their assets so as to balance their portfolios and trading-off risk and return, they provoke a currency crisis when no one in the market is willing to acquire domestic currency at the prevailing price (given by the pegged exchange rate)" (Eichengreen et al., 1996, p.7). In the process of high pressure against the currency of a country, the central bank of the country may allow the currency to depreciate. In some other cases, the bank may defend the currency by running down its foreign exchange reserves or by raising interest rate. Thus an index of currency crisis should capture these different manifestations of speculative attacks, be they successful (resulting substantial currency depreciation) or otherwise. Following Girton and Roper (1977), the exchange market pressure, a weighted average of exchange rate changes, reserve changes, and interest rate changes, is used 8 Investors, broadly, include professional currency traders, foreign investors (who normally withdraw their deposits/investments by changing to foreign exchange when they anticipate currency depreciation), and even domestic firms and households who, in anticipation of currency depreciation, may attempt to maintain their real wealth by converting their domestic money holdings into foreign currency. 24

40 to measure speculative pressure on a given currency (see Eichengreen et al., 1996 and Kaminsky and Reinhart, 2000)9. The exchange market pressure for country i at time t is computed as: EMPu = [X(%Aeu) + KA(i«-igt) - 8(%Arit - %Argt)] (3.1) Where e is the price of the German DM in terms of country i s currency, i;t is the nominal interest rate of country i, igt is the nominal interest rate of Germany and r is the ratio of reserve to M. Following Eichengreen et al. (1996), Germany is taken as the center of reference. It is believed that Germany is a key player in Europe where most of the sample countries are located. In addition, Germany has a relatively stable economy so that its idiosyncratic shocks do not have much influence on the EMP index. X, k and 5 are weights selected to equalize the volatilities of the three components of EMPit so that one component does not dominate the index. 5 in equation (3.1) is one, X is the ratio of the standard deviations of the third to the first components of EMP while k is the ratio of the standard deviations of the third to the second components of EMP. Since the conditional variances of the three components of EMP may not be constant, the weights need to be time varying. In that case, they could be estimated using GARCH models. In this chapter the weights are computed from the unconditional standard deviations of the three components of EMP so that results of my study compare with those from previous studies. However, 9 Since most of the sampled countries did not have market determined interest rates in the 1970s and early 1980s, the crisis index in Kaminsky and Reinhart (2000) incorporates only reserve losses and depreciation. 25

41 the main essence of the conclusions of this chapter does not change when I use time varying weights instead of constant weights (results are not reported here). In the next chapters, time varying weights derived from GARCH models are used to equalize the volatilities of the three components of the EMP. Given the crisis index in equation (3.1), currency crises are defined or associated with unusually large exchange market pressures. The main problem with this methodology is in defining the threshold that determines the largeness of the index. The approach used varies from study to study. For Eichengreen et al. (1996), a crisis occurs if the value of the exchange market pressure is 1.5 standard deviations above the mean of the full panel crisis index. By contrast, Kaminsky and Reinhart (2000) set the cutoff point at 3 standard deviations above the mean value of the own country s crisis index. In Eichengreen et al. (1996), mean and standard deviations are computed for the entire panel of countries in the sample while Kaminsky and Reinhart (2000) compute mean and standard deviations for each individual country. Thus, the cutoff value is common to all countries in Eichengreen et al. (1996) but it is unique for each country in Kaminsky and Reinhart (2000). Each method of crisis identification has its own advantage and disadvantage. The inclusion and exclusion of some countries may change the panel mean and standard deviations. In that case, the Eichengreen et al. s (1996) procedure may or may not indicate a crisis to one country at one point in time depending on the composition of the panel countries. Kaminsky and Reinhart (2000), on the other hand, force each country to have crises for a certain percentage of the 26

42 time even if it is generally agreed that crises rarely if ever take place in that country. In order to check the robustness of the results, crises in this chapter are identified using the two approaches. To see how the two approaches differ in practice, the number of crises in each country using the two crises identification procedures (common mean and standard deviations vs. country specific mean and standard deviations) is reported in Table 1. For example using the common threshold value of 1.5 standard deviations greater than the entire panel s mean value, Australia experiences 7 periods of crises. However, when the threshold is defined with respect to Australia s mean and standard deviations, it experiences only 6 periods of crises. Similarly, Japan experiences 8 and 0 periods of crises if the thresholds are country specific mean plus 1.5 standard deviations and common mean plus 1.5 standard deviations, respectively. Time plots of the frequency of crises are displayed in Figures 1 and 2. The percentage of countries experiencing crisis in each quarter is plotted in Figure 1 where crises are defined using the common mean and standard deviation while Figure 2 plots the percentage of countries experiencing crises using the definition of crises derived from country specific means and standard deviations10. Note that most of the crises are concentrated during 1970, 1974/75 and 1992/93. These periods seem to correspond to the break down of the Bretton Woods system (1970 and 1974/75), and the European Exchange Rate Mechanism (ERM) crisis (1992/93). 10 The panel is unbalanced, i.e., the sample period for some countries does not start in 1960Q1. So, the percentage of crisis at each point in time is computed as number of countries under crisis out of the total countries that have EMP data at that point in time. 27

43 Table 1 Number of Currency Crises: 1960:1-1998:4 Country Threshold = Mean+l.5s.d. Threshold = Mean+2s.d. Threshold = Mean+3s.d. Common to Country all countries specific Common to Country all countries specific Common to Country all countries specific Australia Austria Belgium Canada Denmark Finland France Greece Ireland Italy Japan Netherlands Norway Portugal Spain Sweden Switzerland UK USA Total A crisis may last well over one quarter. Following Eichengreen et al (1996), this paper assumes that once a crisis occurs, its effect may extend to the next quarter.

44 o co o i^o CO o m T O CO o CM luaojad * I * *861. * I Z2.6L * I 10*2.61. = L *08961 V * *o w n c (0 <D e Figure I. Percentage of Countries in Currency Crisis: 20 OECD Countries (Common Threshold) 29

45 o 03 O 00 o o o co 10 luaojsd o o o o CO CM» * * I * e * E *OSZ6l. IOPL61 I 2D2Z61. 6 O 0 Z 6 I. * I.O Z 9 6 I * I mean+1.5sd Figure 2. Percentage of Countries in Currency Crisis: 20 OECD Countries (Country Specific Threshold) 30

46 To avoid counting the same crisis more than once, we exclude observations corresponding to the adjoining quarter The Model Once crises are identified, the next step is to look at whether the probability of a crisis in an individual country is affected by a weighted crises elsewhere variable while controlling for the initial macroeconomic conditions in the country under question. This paper empirically tests for the relevance of four contagion channels through which crises may spread across countries. According to a number of theoretical models reviewed in chapter two, currency crises may be contagious among major trade partners/competitors (trade), countries that have common lenders (finance), countries that have similar macroeconomic fundamentals (macrosim), and/or neighbors (neigh). To empirically identify the relevant contagion channel, each of these channels is captured by a weighted crises elsewhere variable where the weight is constructed to reflect the strength of that contagion channel. A non-structural model is used to estimate the probability of a crisis in country i and period t. This is specified to be: P(CU= 1) = prob[f}o + (5 X a + yytradeu + y 2financea + Y3macroSimu + Y4neighu + u > 0] (3.2) 11 Eichengreen et al. (1996) apply the one-quarter-exclusion window to the tranquil or noncrisis periods, too. But the results of the estimation, not reported in this chapter, do not change when the same exclusion window is applied to the tranquil period, too. 31

47 where i=l,..,n, t=l,..,t ande (equals Ui+Vit) is the sum of the group or heterogeneity effect (Ui) and an idiosyncratic error (V;t). Vj, is assumed to be standard normal and uncorrelated across countries and over time. Unless Ui (the group effect) is zero, estimation of a pooled probit model of (equation 3.2) ignoring Ui will result in inconsistent estimates. If Ui and one or more of the regressors are correlated, we have a fixed effect probit model. In the latter case, the Uj s are treated as parameters to be estimated along with the (3Sand ys- But estimation of the Uj s along with the ps and ys introduces an incidental parameters problem in addition to being computationally difficult in a situation where there are a large number of groups (see Wooldridge, 2002). When N is large in a linear model, the group effects can be removed using the results of partitioned regression. But the strategy to remove the fixed effects cannot be extended to nonlinear probit models. Due to the computational difficulty in estimating the fixed effect probit model in the presence of a large number of groups, the literature resorts to a random effect probit specification. Here Ui is treated as an unobserved random variable drawn along with the other variables. Further, it is assumed that U,/Z,~N(P, or;) (3.3) where Zj is a vector of the right-hand side variables in equation 3.2. In this 32

48 specification, the relative importance of the unobserved group effect is given by: p = cru/(\+ o t) (3-4) The random effect probit specification is justified if p is statistically significant. If not, equation (3.2) is estimated by pooled probit. Xjt is a vector of macroeconomic variables. The variables in Xlt are growth rate of money (m2), inflation from the CPI, growth rate of domestic credit, growth rate of real GDP, percentage of government budget balance relative to GDP, percentage of current account relative to GDP, and the unemployment rate. Each variable is entered as deviation from the corresponding variable of the center country, Germany. The variables in vector Xu are included in line with the arguments of the first generation models of speculative attacks. These models predict co-movements between speculative attacks and adverse developments in the fundamental determinants of exchange rates. In these models, diverging fundamentals are viewed as being inconsistent with a given parity and are interpreted by market participants as a signal that realignment will occur sooner or later. This expectation leads to an immediate speculative attack/pressure against the currency and possibly resulting a crisis in country i independently of the contagious spread of crises from other countries. Next, on the right hand side of equation 3.2, are variables that capture the various channels by which contagion may take place. Tradeu is the trade contagion 33

49 channel. It is measured by the weighted average of crises elsewhere, n -1 y=i, j*i, and with n representing the number of countries in the sample. The weight ( m,'adt) is constructed to reflect the extent of trade linkages or competition between country i and country j. When crises happen in some countries (say j, j+1, and j+3) at time t, all may not have an equal impact on the probability of a currency crisis in country i. Given this, different weights are assigned to crises in other countries based on the extent of trade linkages or competition between i and each of the other countries. Thus, the coefficient on the trade linkage variable, y,, measures the trade-weighted effect of crises elsewhere on the probability of a currency crisis in the representative sample country. If y, is statistically different from zero, it will be taken as evidence for the existence of contagious spread of currency crisis through trade linkages. Finance* is a measure of the financial linkage weighted crises elsewhere and is n-1 given by, j*i. If external lending banks are closely/equally important to countries i and j as source of credit, the financial linkage weight {m ^umce) is assigned a larger value. Based on this, the statistical significance of y2is taken as evidence for the existence of contagious spread of currency crisis through common lenders or financial linkages. Similarly, macrosim«is a measure of the macroeconomic similarity weighted 34

50 n -1 crises elsewhere and is given by ^ m jaerocjt, j*i. The macro weight ( m" cro) is >=i assigned a larger value if countries i and j have similar macroeconomic fundamentals. Based on this, the statistical significance of y3 is taken as evidence for the existence of contagious spread of currency crisis through macroeconomic similarity. Lastly, neighu is the neighborhood effect dummy, and it takes a value of 1 at time t for country i if one of i s neighbors is in a crisis at time t or zero otherwise. Since most of the OECD countries are in Europe, the neighborhood effect dummy is not included in the estimation of this chapter. Weights for each of the remaining three channels are constructed using the methodologies applied in Eichengreen et al. (1996) for macroeconomic similarity, Glick and Rose (1999) for trade linkages, and Van Rijckeghem and Weder (1999) for financial linkages. Trade linkages: following Glick and Rose (1999), the trade weight is computed by taking the importance of bilateral and third party markets for the exports of countries i and j. Exports from the U.S. and Canada, for example, compete in other countries (third party markets like the U.K. and France) and the two countries do also compete in their respective markets. The weight assigned to country j on the extent of bilateral and third party markets trade competition with i is given by m"adt. The trade weight may be computed using the absolute bilateral trade flows and hence an absolute trade weight (m ^ s~,rade) or using bilateral trade flow shares and hence a relative trade weight (m r'l ~,radt). In both cases, the final trade weights are re-scaled so 35

51 that they add-up to one for each country i. The absolute trade weight is computed as: mabs. trade _ L ( )/( + ^ _ L _ } L + x ) 1 y 1 y j1 1 j I y /* I V J1 J + + ^ i)l } (3.5) k where or,* denotes exports from i to k (k*ij) and xt is total exports from country i. The total trade weight as shown in (3.5) is the sum of bilateral trade weight (the first component) and third market trade weight (the second component). The third market trade weight is the weighted average of the importance of exports to country k (all the other export markets of i and j) for countries i and j. This weight is greatest if country k is of equal importance to both i and j (i.e.,jca and xlk are nearly equal) and market k accounts a significant proportion of the exports of the two countries. The same argument holds for the bilateral trade weight. Due to lack of data, all previous studies assume only one time invariant trade weight. This paper also constructs one trade weight for each pair of countries i and j. But the trade weights are constructed based on the average annual trade flow data from 1980 to This allows us first, to get relatively more representative trade linkage weights and second, to avoid the problem of selecting one specific year instead of the others. 36

52 To control for the impact of varying size of countries, relative trade weights based on trade shares, instead of absolute bilateral trade flows, are also computed. These weights are computed as: m r i-,radr + X j, ) /(X, + X, )] [1 - ((X ;, IX,) ~ (* / X, )) /( (X / ) + (X / X, ))] + ^ l(x it+xjk)/(xl+xj)][l-\((xjt/xj)-(xik/xi))\/((xjk/xj) + (xlk/xl))] } (3.6) k where k* i,j. A sample of absolute and relative total trade weights that reflect trade linkages or competition of the U.S. with other OECD member countries are reported in the first four columns of Table 2. Based on the absolute trade weights, the major trade partner/competitor of the U.S. is Japan while its least trade competitor is Greece. Accordingly, a crisis in Japan is given a weight of 11.62% while a crisis in Greece is weighted by 0.92% to construct the absolute trade contagion variable, tradej,. When the sizes of the countries are taken into account, Australia becomes the major trade partner of the U.S. But Japan is still one of the top trade competitors of the U.S. as shown by the relative trade weights. Table 2 also reports weights assigned for the trade partners of France. Macroeconomic Similarities: if a crisis occurs in one country, it is hypothesized that investors take this as a wake-up call and attack other countries that have similar macroeconomic fundamentals with the crisis country. This paper 37

53 Reproduced with permission of the copyright owner. Further reproduction prohibited without perm ission. Table 2 Sample of Trade Linkage Weights Among OFCD Countries (in %) W eights to USA s Trade Partners W eights to France's Trade Partners A bsolute Trade W eights Relative Trade W eights Absolute Trade W eights Relative 'Trade W eights Japan Australia 7.08 Italy Spain 6.95 UK Japan 7.02 Netherlands Italy 6.67 France Switzerland 6.77 UK 9.90 Switzerland 6.64 Italy 9.78 UK 6.15 Belgium 8.61 Sweden 6.48 Netherlands 8.99 Italy 6.11 Switzerland 6.67 Denmark 6.38 Belgium 8.76 Sweden 6.05 USA 6.59 Netherlands 6.37 Sw itzerland 5.43 Spain 5.83 Japan 6.17 Greece 6.37 Spain 4.65 Denmark 5.77 Sweden 5.69 Portugal 6.21 Sweden 4.56 Finland 5.64 Spain 5.13 Belgium 6.17 Australia 3.78 France 5.56 Austria 4.65 Austria 5.96 Austria 3.63 Ireland 5.01 Canada 4.11 Ireland 5.79 Norway 3.44 Austria 5.00 Norway Finland 5.77 Denmark 3.37 Norway 4.91 Denmark 3.96 UK 5.77 Canada 3.26 Belgium 4.91 Australia 3.45 Norway 5.41 Finland 2.67 Netherlands 4.88 Finland 3.19 Japan 3.60 Ireland 2.19 Greece 4.66 Ireland 2.87 USA 3.58 Portugal 1.36 Canada 4.55 Portugal 1.67 Australia 3.44 Greece 0.92 Portugal 4.08 Greece 1.24 Canada 2.43 USA 0.00 USA 0.00 France 0.00 France 0.00

54 considers seven macroeconomic fundamentals to construct the similarity weights: domestic credit growth, money growth, inflation, current account balance relative to GDP, government budget balance relative to GDP, real GDP growth and the unemployment rate. Following Eichengreen et al. (1996), the weight for macroeconomic similarity between countries i and j is computed by:... m acrosim «m J - 1 ~ (3.7) where is a cumulative distribution function of the standardized normal function, xjt is the average of the growth rate of M2, growth rate of domestic credit, growth rate of real GDP, inflation, unemployment rate, and percentages of current account and government budget balance relative to GDP for country i at time t. The growth rates of real GDP, and percentages of current account and government budget balances relative to GDP are multiplied by negative one so that higher values for each of the seven-macro variables reflect greater risk in terms of currency crisis, fi, is the average value of x at time t for all countries in the sample, and a, is the sample standard deviation of the variable x at time t. According to this construction, if country j has similar standardized growth rates of the relevant macroeconomic variables with country i, it is assigned a higher weight for this contagion channel. Again, the macro weights are re-scaled. Unlike the trade and the financial linkages weights, the macroeconomic 39

55 similarity weights vary with time. As an example, the 1995Q1 weights to reflect macroeconomic similarity of the U.S. with each of the other OECD countries are reported in the first two columns of Table 3. According to the results in Table 3, Sweden has the highest macroeconomic similarity with the U.S. and hence a crisis in Sweden if there is any in 1995Q1 is weighted by 7.83% to construct the macroeconomic similarity contagion channel for the U.S. Table 3 also reports the macroeconomic similarity weights assigned to partners of France and Australia for the first quarter of Table 3 Sample of Macroeconomic Similarity Weights Among OECD Countries: 1995Q1 (in %) Weights to reflect Macro Similarity with USA Weights to reflect Macro Similarity with France Weights to reflect Macro Similarity with Australia Sweden 7.83 UK 6.51 Canada 7.10 Italy 7.58 Belgium 6.41 Netherlands 7.03 Netherlands 7.39 Denmark 6.41 Italy Australia 7.19 Greece 6.41 USA 6.41 Canada 7.07 Ireland 6.41 Sweden 6.18 Belgium 5.56 Japan 6.41 Belgium 5.75 Denmark 5.56 Norway 6.41 Denmark 5.75 Greece 5.56 Portugal 6.41 Greece 5.75 Ireland 5.56 Switzerland 6.41 Ireland 5.75 Japan 5.56 Spain 5.77 Japan 5.75 Norway 5.56 Canada 5.14 Norway 5.75 Portugal 5.56 Australia 5.04 Portugal 5.75 Switzerland 5.56 Netherlands 4.88 Switzerland 5.75 UK 5.44 Italy 4.72 UK 5.64 France 5.12 Finland 4.47 France 5.36 Spain 3.92 USA 4.29 Spain 4.29 Finland 2.36 Sweden 4.08 Finland 2.90 Austria 1.62 Austria 3.84 Austria 2.24 USA 0.00 France 0.00 Australia

56 Financial linkages (common creditors): if a bank is confronted with a marked rise in non-performing loans in one country under crisis, it will more likely try to reduce overall risk by pulling out from other high-risk projects located in other countries, too. This may lead to cross-border spillover of crisis. To capture contagion through this channel, a weight to reflect the commonality of bank creditors for country i and country j is constructed using the methodology of Van Rijckeghem and Weder(1999). V r bj+b, ( \b -b h 1 _ Jc b +b JC II (3.8) where bjcis the stock of debt of country i from banks in country c and bt = ^ b ir,i.e. f country i s total borrowings. The weights are computed based on the absolute magnitude of financial flows from common creditors to i and j and hence the name absolute financial linkage weights. The first part of the weight measures the importance of banks in country c for countries i and j together whereas the second component measures the difference in the importance of this center for the two countries. The weight is larger if first, banks in country c are equally important to i and j as source of credit and second, the banks are major source of funds for countries i and j. To control for the size of countries, relative financial linkage weights are also computed based on shares from total borrowings. These weights, which are also re- 41

57 scaled so that they add-up to one for each country i, are computed as: m rel-finance _ V b )c + K j _ [ ( K /fc, ) bj +b, Ib^bJ + ibjb,) (3.9) A sample of absolute and relative financial linkage weights assigned to partners of the U.S. and France is reported in Table 4. According to the absolute financial linkage weights, as reported in the first two columns of Table 4, the U.K. is ranked number 1 in terms of having common bank lender with the U.S. Accordingly, a crisis in the U.K. is weighted by 23.81% in constructing the absolute financial contagion channel for the U.S. When size is taken into account, Australia and Canada take the lead in terms of having common creditors with the U.S Data This chapter uses quarterly data from 1960 to 1998 for a panel of 20 OECD Countries. Countries in the sample are Australia, Austria, Belgium, Canada, Denmark, Finland, France, Germany, Greece, Ireland, Italy, Japan, Netherlands, Norway, Portugal, Spain, Sweden, Switzerland, UK, and the U.S. The data are obtained from three sources. Data to construct the exchange market pressure and all the macroeconomic fundamentals are taken from International Monetary Fund, International Financial Statistics CD-ROM. 42

58 Table 4 Sample of Financial Linkage Weights Among OECD Countries (in %) Weights to USA s partners Weights to France s partners Absolute Financial Linkage Weights Relative Financial Linkage Weights Absolute Financial Linkage Weights Relative Financial Linkage Weights UK Australia 7.37 Italy Ireland 6.73 Italy Canada 6.59 Netherlands 9.76 Greece 6.11 France UK 6.59 Japan 8.61 Spain 6.11 Japan 9.58 Sweden 6.38 Australia 6.92 Netherlands 5.91 Netherlands 7.50 Japan 6.26 Belgium Italy 5.88 Belgium 4.96 Ireland 5.98 Spain 6.80 UK 5.78 Spain 4.71 France 5.84 Ireland 6.26 Sweden 5.78 Ireland 3.75 Netherlands 5.65 Canada 6.14 Denmark 5.72 Australia 3.60 Belgium 5.49 Austria 5.79 Belgium 5.57 Switzerland 3.59 Norway 5.38 UK 5.49 Switzerland 5.54 Austria 3.36 Denmark 5.25 Switzerland 5.40 Finland 5.48 Canada 3.13 Finland 5.19 USA 3.88 Norway 5.35 Sweden 2.32 Spain 5.11 Sweden 3.81 Australia 5.32 Portugal Italy 4.84 Portugal 3.39 Portugal 5.13 Denmark 2.08 Greece 4.81 Denmark 3.23 Austria 5.04 Greece 1.59 Austria 4.69 Greece 2.90 Canada 5.02 Norway 1.52 Portugal 4.35 Norway 2.36 Japan 4.84 Finland 1.35 Switzerland 4.22 Finland 1.95 USA 4.71 USA 0 USA 0 France 0 France 0 Trade weights are constructed using the average annual aggregate bilateral trade flows to 160 countries in the world. The data are obtained from the Center for International Data at University of California, Davis. The data are annual and cover the period 1980 to 1992.

59 To construct weights for financial linkages, data on consolidated bank lending from 19 major industrial countries12 to the sample countries are collected from the Bank for International Settlements (BIS). The lending institutions in each creditor country include commercial banks, saving banks, saving and loan associations, credit unions or cooperatives, building societies, and post office savings banks or other government-controlled savings banks. The principal forms of resources these institutions lend to other countries are deposits and balances placed with banks, loans and advances to banks and non-banks, holdings of securities and participation. Lending is to the public, banks and non-bank private sector. The data on consolidated bank loan statistics to almost all of the sample countries are, however, available only starting So, the financial linkage weights are constructed based on the average claims of lending institutions in each reporting country to each of the sample countries over the period 1999:2, 1999:4 and 2000:1 to 2001: Empirical Results The probit model, specified in equation (3.2), is estimated using quarterly data from a panel of 20 OECD countries over the period 1960 to The estimation results are reported for two cases. In the Erst case, we use a common threshold value for all countries in the sample to identify periods of crises. This is the procedure 12 The 19 countries that report bank lending by nationality of lending institutions are: Austria, Belgium, Canada, Denmark, Finland, France, Germany, Ireland, Italy, Japan, Netherlands, Norway, Portugal, Spain, Sweden, Switzerland, Taiwan, UK and USA. 44

60 applied in Eichengreen et al. (1996). The second case applies country specific cutoff values to identify crises in line with Kaminsky and Reinhart (2000). Estimation results for the two cases stated earlier are reported in Tables 5 to 10. In all cases, the estimated coefficients are marginal effects evaluated at the mean values of the regressors. The tables also report the associated z-statistics (in parentheses) and some diagnostic tests. The diagnostic tests include joint tests for the significance of all the coefficients, joint test for the significance of all the contagion channels and test for the significance of the group effect (or p). In all cases, the null of no group (or random) effect cannot be rejected. Based on this test result, the model is estimated by pooling the data from all the sample countries. The results also show that the null hypotheses that all the contagion channels are jointly insignificant are rejected in most of the cases. Results of the probit estimation based on the first approach of crisis identification are reported in Table S. Columns A and B, where the trade and financial linkage weights are absolute and relative respectively, provide results when the common threshold value is given by mean plus 1.5 standard deviations of the panel EMP index. In this case, we find contagion when absolute trade and financial linkage weights are used. Crisis spreads among countries that have similar macroeconomic fundamentals. Trade and financial linkages are statistically insignificant. Among all the control variables, the growth rate of domestic credit and inflation rate are statistically significant. As suggested in the first generation models of currency crisis, 45

61 Reproduced with permission of the copyright owner. Further reproduction prohibited without perm ission. Table 5 Com m on Threshold for Crisis Identification: Full Sample ( I MO: I -1998:4)', Panel Probit Model. Common Cutoff Value is Mean +1,5s.d. Mean + 2s. d. Mean + 3s. d. A B C D p: F Constant E-0I (-14.18) (-14.22) (-9.14) (-9.49) (-3.40) (-3.49) Trade Contagion 1.0E E ** 0.44** (0.07) (-0.01) (-0.91) (0.2 7 ) (1 9 8 ) (2.19) Macro Contagion 0.24** 0.16 q 4 ^*** 0.30** ** (2.27) (0.93) (4.18) (2. 11) (-1.57) (-2.26) Financial Contagion -0.45E E-0I E-0I -0.90E-02 -O.83E-0I (-0.03) (0.39) (-1.49) (-0.23) (0.1 5 ) (-0.52) Growth o f M2 0.I0 E E E E E E-03 (1.09) ( 1. 10) (0.53) (0.50) (-1.23) (-0.96) G rowth o f D. Credit 0.22E-02** 0.22E-02** 0.1 IE E E E-03 (2.16) (2.16) (1-59) (1.58) (0.66) (0.71) Growth o f Real G D P -0.84E E E E E E-03 HOI) (-0.99) ( -1.M ) (-0.96) (1.19) (1.31) Inflation 0.86E-02*** 0.87E-02*** 0.38E-02** 0.39E-02** 0.15E-02** 0.14E-02** (3.82) (3.85) (2.30) (2.30) (2.20) (2.18) % of C un t. A ccount/gd P -0.79E E E E E-04 0.I2E-03 (-0.78) (-0.80) (0.13) (0.13) (0. 11) (0.50) % or Govt. D eficit/g D P E-02 -O.I7E E E-03-0.I2E E-05 (-1.54) (-1.52) (0.29) (0.31) (-0.03) (0.01) Unem ploym ent rate 0.I5 E I5E IE E E E-03 (1.27) (1.31) (0.85) (0.87) (0.51) (0.52)

62 Reproduced with permission of the copyright owner. Further reproduction prohibited without perm ission. -j Table 5-Conlinucd Number of Observ. Joint tests for slopes 2 ( 10) Joint tests for no contagion x 'O ) Test for the significance of Common Cutoff Value is Mean +.5s.d. M ean + 2s.d. Mean + 3s.d. A B c: D Li F P a. Due to missing observations for some countries, the data are unbalanced. Figures in parenthesis are z-statistics. Critical values from the standard normal table: *** 1% (=2.575), ** 5% (=1.96), and * 10% (=1.645).

63 relative increase in domestic credit and inflation both increase the probability of a currency crisis. The result of macroeconomic similarity as a significant contagion channel is not consistent with what is found in Eichengreen et al. (1996). Columns C and D of Table 5, absolute and relative trade and financial linkage weights respectively, provide the probit estimation results when the common threshold value changes from mean plus l.s standard deviations to mean plus 2 standard deviations. The results suggest that contagion takes place. The relevant contagion channel is still via macroeconomic similarity. Columns E and F of Table S use mean plus 3 standard deviations as a common threshold value. Again the hypothesis that contagion is a factor is supported by the results. But this time crises spread among major trade partners/competitors as the marginal effect of the trade contagion channel is statistically significant. The marginal effect of the macro contagion channel is also statistically significant when relative trade and financial linkage weights are used but it has unexpected sign. Table 6 provides the estimation results when country specific cutoff values, computed using each country's mean and standard deviations, are used to identify periods of crises. When crises are identified by country specific mean plus l.s and 2 standard deviations, contagion works via macroeconomic similarity channel. The results reported in Table 6 from columns A to D are similar with those in Table 5 at least in terms of the test results for the relevant contagion channels. The two tables differ in their last columns where no contagion is detected when the cutoff point is country specific mean plus 3 standard deviations. 48

64 Reproduced with permission of the copyright owner. Further reproduction prohibited without perm ission. Table 6 Country Specific 'Threshold lor Crisis Identification: Tull Sample ( I : 1- I998:4), Panel Piobil Model. Country Specific Cutoff Value is Mean + 1.5s.d. Mean + 2s.d. Mean + 3s. d. A B C D E I* Constant E-0I -0.99E-0I -0.36E-0I -0.36E-0I (-14.60) (-14.64) (-8.59) (-8.53) (-3.94) (-3.92) 'Trade Contagion 0.88E-0I E-0I E-0I 0.21 (0.61) (-0.40) (0.70) (0.49) (-0.20) (0.92) M acro Contagion 0 ^9*** 0.45*** 0.19* 0.23* E-02 (3.45) (2.88) (1.85) (164) (0.96) (-0.06) Financial Contagion E-0I I2E (-1.57) (-0.38) (-0.95) (-0.83) (-O.(M)I) (-0.55) G row th o f M2-0.58E E E E E E-03 (-0.63) (-0.64) (0.71) (0.72) (-134) (-1.29) G row th o f D. C redit 0.33E-02*** 0.33E-02*** 0.10E E E E-03 (3.27) (3.31) (1.42) (1.40) (0.48) (0.54) G row th o f Real GDP -0.51E IE E E E E-04 (-0.59) (-0.59) (-0.93) (-0.99) (-0.16) (-0.13) Inflation 0.52E-02** 0.52E-02** 0.32E-02** 0.32 E-02** 0.13 E-02 0.I3E-02 (2.09) (2. 11) (2.02) (2.03) (1.58) (1.60) % o f Currt. A ccount/gd P 0.56E-04 0.I5E E E-03-0.I2E IE-03 (0.05) (0.15) (-0.59) (-0.54) (-0.29) (-0.29) % of Govt. Deficit/GDP 0.83E E IE * * E-03 (.07) (0.09) (-0.86) (-0.87) (0.28) (0.32) Unemployment rale 0.47H H i.ot:-o * 03 0.I9E-03 (039) (0.44) (132) (1.27) (0.54) (0.49)

65 Reproduced with permission of the copyright owner. Further reproduction prohibited without perm ission. L/l o Table 6-Continucd Number of Obscrv. Joint tests for slopes x'( 10) Joint tests for no Mean + 1.5s.il. A B Country Specific C utoff Value is Mean + 2s. il. C D contagion 2(3) Test for the significance of () a. Due to m issing observations for some countries, the i ata arc unbalanced. Figures in parenthesis are ^-statistics. Critical values from the standard normal table: *** 1% (=2.575), ** 5% (=1.96), and * 10% (=1.645) Li Mean + 3s.d

66 The story does not change when the sample is from the first quarter of 1960 to the fourth quarter of 1993, a sample similar to that in Eichengreen et al. (1996). The results are reported in Tables 7 and 8. Overall, the results of the different probit estimations consistently indicate that currency crises are contagious in most of the cases. But the results with regard to the relevant contagion channels are sensitive to the identifications of periods of crisis. It appears that the macroeconomic similarity contagion channel works when currency crises are identified using the Kaminsky and Reinhart (2000) procedure. But there is no single contagion channel when the Eichengreen et al. (1996) procedure of crisis identification is applied. In the latter case, the macroeconomic contagion channel appears to work when a relatively liberal/moderate threshold value, allowing even weaker crises, is used to identify periods of crisis. The contagion channel switches to trade linkages when currency crises are the severe ones; identified by a relatively higher threshold value. The shift in the contagion channels may indicate that moderate and weaker currency crises in some countries may not have significant real sector impact in other countries and hence the trade contagion channel is dominated. Instead, investors who cannot perfectly isolate the characteristics of countries in moderate or weaker currency crises from those of other countries may simply put pressures on currencies of other countries that have similar selected macroeconomic fundamentals with those of the crisis countries and hence the working of the macroeconomic similarity contagion channel. 51

67 Reproduced wild permission of the copyright owner. Further reproduction prohibited without perm ission. U3 IO Table 7 Com m on Tine 2 ' lor Crisis Identification: Eichengreen ct al. (1996) Sc 25, l- (1960:1-1993:4) \ Panel Probit Model. Common Cutoff Value is Mean + l.5s.d. Mean + 2s.d. Mean + 3s.d. A B C D E F Constant E-0I -0.35E-01 (-14.87) (-14.91) (-9.49) (-9.85) (-3.68) (-3.56) Trade Contagion 0.3 IE E-0I E-0I 0.20 ** 0.48** (0.03) (-0.09) (-0.89) (-0.29) (1.99) (2.21) M acro Contagion 0.26 ** *** 0.33** (2.23) (0.94) (4.16) (2.07) (-1.59) (-2.28) Financial Contagion 0.98E E E E-01 (0.01) (0.47) (-147) (-0.20) (-0.15) (-0.52) Growth o f M2 0.I2E-02 0.I2 E E E E E-03 ( 112) (1 1 4 ) (0.54) (0.52) (-1.23) (-0.97) Growth o f D. Credit 0.23E-02** 0.23E-02** 0.12E-02 0.I2E E IE-03 (2.05) (2.04) (1.50) (1.50) (0.63) (0.67) Growth o f Real G D P -0.95E E IE E E IE-03 (-1.03) ( ) (-109) (-0.95) ( 120) (132) Inflation 0.89E-02*** 0.90E-02*** 0.38E-02** 0.40E-02** 0.16E-02** 0.15E-02** (3.55) (3.58) (2. 10) (2. 10) (2. 11) (2. 10) % of Currt. Account/GDP -0.86E E-03 0.I0E-03 0.I1E E-04 0.I3E-03 (-0.77) (-0.80) (0.13) (0.13) (0. 10) (0.49) % of Govt. Deficit/GDP -0.20E E E E IE E-05 (-1.57) (-1.55) (0.28) (0.31) (-0.02) (0.02) Unemployment rate 0.I4E-02 0.I4E IE E-03 0.I7E E-03 (0.98) (1.03) (0.62) (0.63) (0.38) (0.39)

68 Reproduced with permission of the copyright owner. Further reproduction prohibited without perm ission. Table 7-Contiiuieil Common Cutoff Value is Mean + 1.5s.d. Mean + 2s.d. Mean + 3s.d. A B C I) E F N um ber o f Obscrv Joint tests for slopes 3(10) Joint tests for no contagion %1(3) Test for the significance o f p Table 8 Country Specific Threshold for C risis Identification: Eichengreen et al. ( 1996) Sample ( :1-1993:4)a, Panel Probil Model. Country Specific Cutoff Value is Mean +l.5s.d. Mean + 2s.d. Mean + 3s.d. A B C D E F Constant E-0I -0.39E-0I (-14.87) (-14.93) (-8.90) (-8.85) (-4.03) (-4.01) Trade Contagion 0.89E-0I E IE (0.58) (-0.49) (0.70) (0.46) (-0.20) (0.93) M acro Contagion 0.42*** 0.48*** 0.20* IE-01 (3.43) (2.83) (1.83) (161) (0.95) (-0.07) Financial Contagion E E (-1.53) (-0.27) (-0.94) (-0.79) (0.002) (-0.55)

69 Reproduced with permission of the copyright owner. Further reproduction prohibited without perm ission. Table 8-Continucd Country Specific C utoff Value is Mean + 1.5s.d. Mean + 2s.d. Mean + 3s.d. A li C D E F G rowth of M2-0.62E E E E E E-03 (-0.62) (-0.63) (0.74) (0.75) (-1.33) ( ) Growth of D. Credit 0..35E-02*** 0.35E-02*** 0.1 IE I0E E E-03 (3.19) (3.22) (1.32) (1.30) (0.44) (0.51) Growth of Real GDP -0.52E E E E E E-04 (-0.55) (-0.55) (-0.93) (-0.99) (-014) (-0. 11) Inflation E-02** 0.56E-02** 0.32E-02* 0.32E-02* 0.13 E E-02 (2.04) (2.07) (1.82) (1.83) (1.44) (1.46) % o f Currt. A ccount/g D P 0.8 IE-04 0.I8E E E E-03-0.I2E-03 (0.07) (0.16) (-0.58) (-0.54) (-0.29) (-0.28) % o f Govt. D eficit/g D P 0.83E E E E-03 0.I6E-03 0.I9E-03 (0.06) (0.09) (-0.88) (-0.89) (0.28) (0.33) Unemployment rate 0.I2E IE-03 0.I0E E E E-03 (0.08) (0.15) ( 110) (1.06) (0.41) (0.36) N um ber o f Observ Joint tests for slopes (10) Joint tests for no contagion x20) Test for the significance o f p a. Due to m issing observations for som e countries, the data arc unbalanced. Figures in parenthesis arc /.-statistics. Critical values from the standard normal table: *** 1% (=2.57.*)), ** 5% (=1.96), and * 10% (=1.645).

70 Tables 9 and 10 report the results of the probit estimation when we include only the trade and the macroeconomic similarity contagion channels, the only two channels considered in Eichengreen et al. (1996). Hence, columns A and B of Table 9 are replications of data and methodology as originally carried out by Eichengreen et al. (1996). As before, the macroeconomic contagion channel is statistically significant when a relatively liberal/moderate threshold v 'ue is applied to define periods of crisis. Trade becomes the relevant contagion channel only when a relatively higher threshold value defines periods of crisis as reported in columns E and F of Table 9. The story changes when country specific threshold values are used to identify crises. In the latter case, there is contagion only when a relatively liberal threshold value defines currency crises. As reported in columns A and B of Table 10, crises spread through the macro similarity channel. Extension o f the study As stated earlier, most of the results of the different panel probit estimations consistently indicate that currency crises are contagious. But the results with regard to the relevant contagion channels are sensitive to the identifications of periods of crisis. This result regarding the changes in the way crises propagate across countries points out that at times of moderate and weak crises, investors, with incomplete information, spread crises by putting pressures on other countries that have similar macroeconomic fundamentals with those in countries in crises. Conversely, a severe crisis in one 55

71 Reproduced with permission of the copyright owner. Further reproduction prohibited without perm ission. I'ablc 9 Common 'Threshold for Crisis Identification: Sample: (1960:1-1993:4)" and No Financial Linkages; Panel Probit M odel Common Cutoff Value is Mean + 1.3s.d. Mean + 2s.d. Mean + 3s.d. A B C D E F Constant E-0I -0.35E-0I (-14.87) (-14.87) (-9.64) (-9.87) (-3.68) (-3.38) 'Trade Contagion 0.38E IE ** ** 0.41*** (0.03) (0.41) (-2.14) (-0.7.3) (2.43) (2.67) Macro Contagion 0.26** *** 0.32** ** (2.28) (113) (3.92) (2.06) (-1.60) (-2.33) Growth o f M2 0.12E-02 0.I2 E E E E E-03 (LI2) (112) (0.34) (0.32) ( - 122) (-0.91) G rowth o f D. Credit 0.23E-02** 0.23 E-02** 0.I3E E E E-03 (2.03) (2.03) (1.33) (1.30) (0.62) (0.64) Growth o f Real GDP -0.93E E E E E E-03 (-1.03) ( 1-02) (109) (-0.94) (119) (1.32) Inflation 0.89E-02*** 0.89E-02*** 0.40E-02** 0.4 IE-02** 0.I6E-02** 0.I6E-02** (3.36) (3 36) (2.19) (2.13) (2. 12) (2.21) % of C unl. Account/GDP -0.86E E-03 0.I6E IE E E-03 (-0.77) (-0.77) (0.19) (0.13) (0. 10) (0.49) % o f Govt. D eficit/g D P -0.20E E E E-03-0.I2E E-03 (-1.37) (-1.33) (0.27) (0.30) (-0.02) (0.02) Unemployment rate 0.14 E E E E E E-03 (0.98) ( 1.00) (0.37) (0.63) (0.37) (0.40)

72 Reproduced with permission of the copyright owner. Further reproduction prohibited without perm ission. -J Table 9-Continucd Common Cutoff Value is Mean + l.5s.d. Mean + 2s.d. Mean + 3s.d. A B C D E F N um ber o f Obscrv Joint tests for slopes 3 (9) Joint tests for no contagion J 3 (2) Test for the significance of p Table 10 Country Specific Threshold for Crisis Identification: S 98,,c:(I960:I-I993:4) 1 and N o Financial Linkages; Panel Probit Model Country Specific Cutoff Value is Mean +l.5s.d. Mean + 2s.d. Mean + 3s.d. A B C D H F Constant E E-01 (-15.07) (-14.98) (-8.99) (-8.96) (-4.03) (-4.04) Trade C ontagion -0.73E-0I E E IE (-0.64) (-116) (0.10) (-0.22) (-0.29) (0.89) M acro C ontagion 0.35*** 0.47*** E-0I (3.09) (2.85) (1.60) (1.43) (1.03) (-0.32) Growth o f M2-0.70E E IE IE E E-03 (-0.70) (-0.63) (0.76) (0.76) (-1.33) (-1.28)

73 Reproduced with permission of the copyright owner. Further reproduction prohibited without perm ission. Table 10-Continucd Coi nlry Specific Cutoff Value is M ean +1.5s.d. Mean + 2s.d. Mean + 3s.d. A B C I) I* 1* G rowth o f 1). Credit 0.361*02*** 0.35K-02*** 0.1 OH * K *03 (3.30) (3.22) (1.29) (1.29) (0.44) (0.49) G rowth of Real G D P * H * * *04 O.43K-04 (0.5 0 ) (-0.52) (-0.96) (0.9 6 ) (0.1 4 ) (0. 12) Inflation 0.55K-02** 0.56E-02** 0.321*02* 0.32K-02* 0.131* *02 (2.04) (2.07) (1.83) (1.85) (1.44) (147) % of Currt. Accounl/GDP 0.26K 03 0.I7K * *03-0.I3K *03 (0.23) (0.15) (-0.54) (0.5 5 ) (-0.29) (-0 30) % o f Govt. D eficit/g D P 0.151* * * IK KO *03 (0. 10) (0.09) (-0.86) (-0.87) (0.28) (0.33) Unemployment rale * l* * * * E-03 (0.19) (0.17) (III) (1 10) (0.41) (0.42) N um ber o f Obscrv Joint tests for slopes 2 (9) Joint tests for no contagion X2(2) % Test for the significance o f p a. Due to m issing observations for som e countries, the data arc unbalanced. Figures in parenthesis arc /.-statistics. Critical values from the standard normal table. *** 1% (=2.575), ** 5% (=1%), and * 10% (=1.645).

74 country spreads contagiously by affecting the real sectors of its major trade partners through their trade linkages. Similar results about differences in the way different regimes of return (large negative vs. small negative stock returns) spread contagiously are also reported in Bae et al. (2000). The different regimes of returns lead, according to them, to different levels of panic among investors, irrational outcomes, and excess volatility. Given the above explanation for the existence of different possible contagion channels depending on the severity of crises, one plausible extension of the study may be to apply a nonlinear Markov Chains switching regression model which does not even require encoding the crisis index as zero one values. This modeling strategy may enable us to test for contagion and to identify the relevant contagion channels by fully capturing the different regimes of crisis propagation13. To get a more robust and consistent results, this chapter is, however, extended in the next two chapters in two other major ways. First, more countries in particular emerging market economies from Asia and Latin America are added. This allows us to add the other contagion channel to capture the neighborhood effect of a crisis. The second extension of the study is to identify crises using a relatively more objective method based on the extreme value theory. This approach is also applied in Pozo and Amuedo-Dorantes (2002). Given my findings that the relevant contagion channel is sensitive to the identification of crisis, it becomes more important to use a relatively mote objective method to identify periods of currency crisis. 59

75 The main argument for using the extreme value theory is that exchange rate changes during events like the October 1987 stock market crash, the Asian currency crisis of 1997, and the hedge fund (or Russian) crisis of 1998 are outliers making the distributions of exchange rate changes and other asset returns fat-tailed. The next chapter intends to determine the frequency of large exchange market pressures from the tail shape of the distribution. The tail index, which measures the tail shape of the distribution, is a good indicator for the mass in the tails. The extreme value theory allows us to determine the tail mass through the tail index and hence we can take the frequency of outliers as indicators of currency crisis Conclusions This chapter of my dissertation estimates a panel probit model for 20 OECD countries using quarterly data from 1960 to 1998 to test whether currency crises are contagious and to identify the channels through which crises spread across countries. It extends Eichengreen et al. (1996) in two ways. First, it includes a third contagion channel through financial linkages in addition to the trade and macro-similarities channels considered by Eichengreen et al. (1996). Second, it employs the two commonly used procedures of currency crisis identification. This allows for checking the sensitivity of the results to the identification of periods of currency crises. A number of different model specifications are estimated. All results consistently indicate that currency crises are contagious at least among members of 131 am very grateful to Professor Higgins for pointing out this as a possible modeling strategy to fully account for the different regimes of crisis transmission. 60

76 the OECD countries. However, the channel by which contagion operates appears sensitive to the identification of currency crises. When currency crises are identified by the Kaminsky and Reinhart (2000) procedure, countries macroeconomic similarity appears to be the relevant contagion channel. But there is no single contagion channel when the Eichengreen et al. (1996) procedure of crisis identification is applied. In the latter case, the macroeconomic similarity contagion channel appears to work when a relatively liberal/moderate threshold value, allowing even weaker crises, are used to identify periods of crisis. The contagion channel changes to trade linkages when currency crises are the severe ones, identified by a relatively higher threshold value. These results may indicate that a severe crisis in a country spreads to its major trade partners by significantly affecting their real sectors. Moderate and weaker crises, on the other hand, may not have tangible impact on the real sectors of other non-crisis countries. Instead investors, with incomplete information, may contagiously spread these crises by putting pressures on non-crisis countries that have similar selected macroeconomic fundamentals with those in the crisis countries. This inconclusive result related to the relevant contagion channel points to the need to either carefully define and identify periods of currency crisis or to employ a technique so as to model by fully capturing the different regimes of crisis propagation without censoring the crisis index as zero one values. 61

77 CHAPTER 4 IDENTIFICATION OF CURRENCY CRISES USING THE EXTREME VALUE THEORY 4.1. Introduction Results from a number of different model estimations in chapter three consistently indicate that currency crises are contagious at least among members of the OECD countries. But the result with regard to the relevant contagion channel appears sensitive, to the level of threshold used to identify periods of currency crisis. Due to this inconclusive result regarding the relevant contagion channel, this chapter uses an alternative and relatively more objective method to carefully define and identify periods of currency crisis using the extreme value theory. The main argument for this approach is that exchange rate changes during crisis periods (like the ERM attacks of 1992 and the Asian financial crisis of 1997) are outliers, making the distribution of exchange rate changes fat-tailed. Extreme value theory allows us to determine these tail observations via the tail index and hence we take or identify the frequency of outliers as indicators of crisis. The standard approach in contrast, followed by Eichengreen et al. (1996) and Kaminisky and Reinhart (2000) is to identify periods of currency crises associated with values of the exchange market pressure (EMP) larger than a given threshold. But no theoretical justification is provided in setting the threshold, which uses the 62

78 mean and the standard deviation of the EMP. In addition, both the mean and the standard deviations are sensitive to very large outliers. Thus, some large exchange market pressures, which could be indicators of crisis, may be excluded or some middle values that are not the results of intense currency pressures may be included due to the influence of outliers on both sides of the distribution of the exchange market pressure. Furthermore, the use of variance as the appropriate measure of dispersion is not warranted as stated below. A number of studies have shown that the distributions of exchange rate changes and other asset returns have fat-tails reflecting the prevalence of extreme returns or outliers14 (see Pownall and Koedijk, 1999). These outliers are the results of events like the October 1987 stock market crash, the Asian financial crisis of 1997, and the hedge fund (or Russian) crisis of Given the possibility of tail fatness, the distributions of asset returns may not even have finite variance further complicating the use of mean and standard deviation in setting the level of the threshold (see Boothe and Glassman, 1987). One way of determining the frequency with which extreme price changes are expected to occur is to employ one of the distributions that have been advanced in the literature for modeling stock price returns (Jansen and De Vries, 1991, p. 18). The 14 Since the two components of EMP are strictly asset returns, this paper assumes that EMP has a fat-tailed distribution. This is tested in section 3 of this chapter. 15 In statistics, extremes of a random process refer to the lowest observation (the minimum) and the highest observation (the maximum) over a given period. In financial markets, extreme price movements correspond to stock market crashes, bond market collapses or currency crises during extraordinary periods. These observations are frequently named as extreme values (see Longin, 2000 and Kellezi and Gilli, 2000). 63

79 literature has offered a number of alternative fat-tailed distributions for exchange rate changes and other asset returns such as student-t, non-normal sum-stable and the ARCH processes (see Boothe and Glassman, 1987 and Koedijk et al., 1990). Alternatively, we can determine the frequency of extreme price changes by estimating the tail mass via the tail index of the return distribution. The extreme value theory shows that the extreme values of some distributions have limiting distributions. The form of the asymptotic distribution of the extreme returns is independent of the process generating the returns. In line with this, the tail behavior of the alternatives can be parameterized by the so-called tail index y from the limit law of the distribution of the maxima (Koedijk et al., 1990). The tail index does then provide information about the underlying distribution and thereby allows for discriminating between the alternative hypotheses (Hols and De Vries, 1991). This paper intends to determine the frequency of large exchange market pressures from the tail shape of the distribution of EMP. The tail index is a good indicator for the mass in the tails. The extreme value theory develops a method of estimating the tail index of the distribution. The idea is that the tail of the distribution of EMP includes the outliers, which are the results of successful and unsuccessful pressures against the currency of the country16. But the tail index as shown in section two is estimated based on these extreme values/outliers. If we choose few 16 The extreme value approach is widely used to determine the extreme returns (or outliers) in a given sample in order to compute the Value at Risk of a given position (see Longin, 2000 and Danielsson and De Vries, 1997). Longin and Solnik (2001) also use this approach to identify the largest returns in order to compare the correlations of assets during periods of low and high returns. 64

80 outliers/extreme values, the tail index estimates will be inefficient. Alternatively, choosing too many EMP values induces biased parameter estimates as the result of including observations not in the tail in the estimation process. To optimize the tradeoff between bias and inefficiency, we use Monte Carlo simulation to determine the optimal number of outliers. The procedures of estimating y (the tail index) along with the optimal number of extreme values or tail observations are discussed in section 4.2. The rest of this chapter is divided into three sections. The next section provides a brief discussion of the extreme value theory and the estimation method for the tail index. The third section reports the estimation results of the tail index for a sample of thirty-seven developed and emerging market economies. Identification of periods of crises for the sample countries is provided in section four. The last section is devoted to a brief conclusion Theory Consider a stationary sequence Xi. X:, X3,... of independent and identically distributed random variables17 with a common distribution function F (d.f. F). Suppose one is interested in the probability that the maximum Mn = max (X,. X2 Xn) (4.1) 17 Danielsson and De Vries (1997), Quintos et al. (2000), and Leadbetter et al. (1983) point out that the results shown here do still hold true under the weak assumption of stationarity only; and even with serially dependent data. 65

81 of the first n variables is below a certain level x. This probability is given by P{M < x} = P{X,<x, X2<x Xn<x}= F (x) (4.2) As n tends to infinity, it is clear that for any fixed value of x, 1, i f F(x) = 1 lim p{m n < Jt}= n >00 0, if F(x) < 1 (4-3) which is a degenerate distribution. Extreme value theory studies the limiting distribution of the appropriately scaled order statistic M. To find a limiting nondegenerate distribution of interest if there is any, the maximum order statistic, M. should be reduced with a location parameter b and a scale parameter a > 0 such that i.e. P {a (M - b ) < x } G(x) (4.4) F(x/an+ bn) ^ G(x), where w stands for weak convergence and G(x) is one of the three possible asymptotic distributions for the extreme values- Gumbel, Frechet, and Weibull distributions (see Kotz and Nadarajah, 2000, Jansen and De Vries, 1991, Longin, 66

82 1996, and Embrechts et al., 1997 for detailed discussions). The three Extreme Value distributions that G(x) may take (up to the location and scale changes) are: Type 1, (Gumbel-type distribution): G( x) = exp(-e *), - o o < J C < o o ; (4.5) Type 2, (Frechet-type distribution): 0, X < 0 G ( jc) = exp(-x a ), for some a >0, X >0 ^4 6^ Type 3, (Weibull-type distribution): G(x) = exp(-(-x )a ), for some a > 0, X < 0 1, X > 0 (4.7) with a representing the shape parameter. The shape parameter reflects the weight of the tail of the distribution of the parent variable X. For type 1, the tail of the distribution F(x) declines exponentially. For type 2, the tail of the distribution F(x) declines by a power. Type 3 distribution has a right tail limit. For the first and third cases, all moments of the distribution of X are well defined. For the second case the 67

83 shape parameter a corresponds to the maximal order moment: the moments of order r greater than a are infinite, the moments of order r equal to or less than a are finite; the distribution of X in this case is fat-tailed (see Longin, 1996). Since the distributions of exchange rate changes and stock returns are fattailed and unbounded in principle, the limit law given by the Frechet-type is the only relevant one for our case. The literature focuses on directly estimating a without any prior hypothesis and identifies the relevant F(x) s, which are nested within the Frechet limit law, by the values of a. The leptokurtic stable hypothesis requires a < 2 while the Student -t class and the ARCH process allow a > 2 (see Hols and De Vries, 1991). There are two procedures to estimate the tail index y (= 1/a). The first A method is to estimate it by maximum likelihood. The ML estimator, y ^, is consistent and is asymptotically normally distributed: ( r l - y ) V ^ ~ m a + y ) 2) (4.8) where m is the number of highest observations from a sample of size n>m (see Koedijk et al., 1992). This estimator is based on the hypothesis that the limit law given by the Frechet distribution is the correct model, whereas it holds only in an approximate sense for finite sample sizes (Koedijk et al., 1992, p. 465). Thus, the estimator may not be efficient. 68

84 The second procedure is a non-parametric estimation method initiated by Hill (1975). The Hill index is estimated non-parametrically based on a certain number of the largest order statistics. A number of studies have shown that the Hill index is a more efficient estimator than the maximum likelihood estimator as the former has a smaller asymptotic variance ( y 1 Im ) compared to the variance of the ML estimator ((1 + y): / m ) (see Jansen and De Vries, 1991, Koedijk et al., 1992 among others). To compute the Hill estimator, define X<i) s X<2) s -s X<n) as the ascending order statistics from a sample Xi, X2,..., X of n consecutive exchange market pressures, X,. The proposed Hill estimator is given by: / a m (4.9) where m is the largest order statistics used to compute y H and n is the sample size. A y H is a consistent estimator of y and is asymptotically normally distributed (see Koedijk, et al., 1992): - w (0,y2) (4.10) Generally, m in both the MLE and the Hill-estimator should increase with the sample size n but it should be small relative to the overall sample size n (see Jansen and De Vries, 1991 and Wagner and Marsh, 2000). This is given by the following asymptotic condition: 69

85 , X m(n) A m(n)» oo, » 0, as n >«>. n (4.11) The Hill estimator has, however, the problem of selecting the m largest order statistics that goes into the estimation of the tail index. A number of studies exploit A the properties of y H t0 select m. One way is to compute yh for different m and to A select an m value in the region over which the estimated y is more or less constant (see Hols and De Vries, 1991). Jansen and De Vries (1991), Hols and De Vries (1991), Koedijk et al. (1990), and Longin and Solnik (2001), among others, use simulation to select the optimal m. They conduct Monte Carlo experiment to find the A level of m, conditional upon a sample size n and d.f. F(x), for which the MSE for y is the minimum18. A good estimate of the tail index is obtained if only the correct tail observations given by m are known. If we choose only few tail observations, we will have inefficient parameter estimates. On the other hand, if we choose many observations, the parameter estimates will be biased as a result of including observations not in the tail in the estimation process. The Monte Carlo simulation method is, therefore, helpful in optimizing the trade-off between the bias and the inefficiency mentioned above. The main focus of this part of my dissertation is to determine the largest exchange market pressures as indicators of currency crises. The number of these 18 Danielsson, et al. (2001) employ a two-step sub-sample bootstrap method to select the sample fraction that minimizes the asymptotic mean-squared error. Unlike previous methods, their study claims that prior knowledge of the second-order parameter is not required. 70

86 extreme values is given by m, which is used to locate the specific times at which the tail observations occur and thereby identify the periods of crisis. The paper follows the simulation method suggested by Jansen and De Vries (1991), Koedijk et al. (1990), and Longin and Solnik (2001) to determine m and the corresponding periods of currency crisis. Jansen and De Vries (1991) use this approach in order to judge events like the October 1987 stock market crash against the expected frequency/probability of extreme price changes. Koedijk et al. (1990) employ this approach to determine which class of the fat-tailed distribution is most appropriate to model the foreign exchange rate. Longin and Solnik (2001), on the other hand, apply this approach to derive the asymptotic distribution of conditional tail correlation in order to test whether international equity market correlation increases in volatile periods or times. My paper, on the other hand, applies this approach to identify periods of currency crises and test finally whether currency crises are contagious. The test whether currency crises are contagious is done in the next chapter. The spirit of the simulation method by Jansen and De Vries (1991), Koedijk et al. (1990), and Longin and Solnik (2001), among others, is to conduct a Monte Carlo experiment to generate pseudo random numbers (or simulated time-series) drawn from different known theoretical d.f.s F(x). For each simulated series of size n, the tail index is calculated for different values of m. Each experiment is replicated S times so that the MSE of the S tail indexes obtained for a given m and given d.f. F(x) of sample size n is computed. The MSE criterion is used to pick one optimal m-level for each theoretical d.f. F(x). y is known for each theoretical d.f. F(x) and is given by 71

87 . These optimal m s are then used to estimate y for the actual EMP data assuming a different d.f.s F(x) of sample size n. Finally, one optimal level of m is chosen corresponding to the point where the estimated y computed from the actual EMP data is statistically the closest to the known y corresponding to the assumed theoretical d.f. F(x). The paper adopts Longin and Solnik s (2001) simulation steps to determine m. The Outline of the steps are given below: 1. Simulate S time-series containing n observations of exchange market pressure from each known Student-t distributions with a degrees of freedom where a ranges from 1 to K. The theoretical distributions include a sum stable distribution for a equals 1 (see Jansen and De Vries, 1991). The lower the degree of freedom, the fatter the distribution as the tail index y is related to a by y - l/a (Longin and Solnik, 2001). 2. For different number m of extreme exchange market pressures, estimate a tail index y (m a >corresponding to the s* simulated time-series from the student-t distribution with a degree of freedom. Longin and Solnik (2001) allows m to vary from 1% to 2 0% of n. n is the sample size of the actual exchange market pressures data. A 3. Compute the mean square error mse ((V (m,a)) )Of the S tail index * S m «L h J estimates for a particular student-t distribution with a degree of freedom and a 72

88 particular value of m. Repeatedly compute the MSE for different values of m but for a particular student-t distribution of a degrees of freedom. Then, select the optimal m, denoted by m*(oc), that minimizes the MSE for that particular studentt distribution with a degrees of freedom. Then, repeatedly choose optimal values of m for different student-t distributions. A total of K optimal values of m, (m*(a))osi...k, will be selected for the K possible theoretical distributions. 4. Using each of the K optimal values of m obtained in step 3, compute the tail index estimate using the actual exchange market pressure data, y (m*(a)) using Hill A (1975) tail index estimator. K number o fy tail indexes are estimated from the actual exchange market pressure data, for a varying from 1 to K. 5. The main objective of the whole exercise is to get one optimal number of extreme exchange market pressures, m**. Select that single number of extreme exchange market pressures, m**, corresponding to which the tail index estimate from the actual data (from step 4) is statistically the closest to the corresponding tail index of the theoretical distribution. Once m** is obtained, take all the times corresponding to the m** largest observations as periods of crisis Estimation of the Tail Index As discussed in the previous chapter, the exchange market pressure is used as an index of currency crisis. Following Girton and Roper (1977) and Eichengreen et al. (1996), the index is computed as the weighted average of changes in the exchange 73

89 rate, reserve, and interest rate. The exchange market pressure for country i at time t is computed as: EMPit = [X(%Aeit) + KA(iir ig,) - 8(%Aru - %Arg,)] (4.12) where e is the price of the German DM in terms of country i s currency, i«(ip) is the nominal interest rate of country i (Germany) and r is the ratio of reserve to Mj. Germany is taken as the center of reference, following Eichengreen et al. (1996). X, k, and 5 are weights selected to equalize the volatilities of the three components of EMPit so that one component does not dominate the index19. All previous papers such as Eichengreen et al. (1996) and Kaminisky and Reinhart (2000) use time invariant weights constructed from the standard deviations of the three components of EMP. But the conditional variances of the three components of EMP, as shown below, are not constant, and hence the weights need to be time varying. To account for this possibility, the conditional standard deviations that are used as time varying weights are estimated using GARCH models. GARCH models are fitted for each of the three variables (changes in exchange rate, changes in reserves and changes in interest rate differentials) and for each of the thirty-six countries (excluding Germany) in my sample20. The conditional standard 19 8 is one, X is the ratio of the standard deviations of the third to the first components of EMP while K is the ratio of the standard deviations of the third to the second components of EMP. 20 The 36 countries, excluding the center Germany, in my sample are Argentina, Australia, Austria, Belgium, Brazil, Canada, Colombia, Denmark, Ecuador, Finland, France, Greece, Indonesia, Ireland, Israel, Italy, Japan, Korea, Malaysia, Mexico, Netherlands, New Zealand, Norway, Peru, Philippines, Portugal, South Africa, Singapore, Spain, Sri Lanka, Sweden, 74

90 deviations are derived from a GARCH (1,1) model because the latter is believed to represent most of the financial time series data (see Bera and Higgins, 1993). A sample of estimation results of the GARCH (1,1) model for the three variables are reported in Table 11. In most of the cases, the coefficients of the ARCH and the GARCH terms are statistically significant. In few other cases, the GARCH coefficients are insignificant. In the latter cases, an ARCH (1) model is fitted. The results derived from both the GARCH and the ARCH models confirm that the conditional variances of the three components of EMP are not constant. Except with very few cases, the LM tests for any remaining ARCH/GARCH effects are insignificant suggesting that GARCH (1,1)/ARCH (1) is a good fit for each of the three components of EMP. Before estimating the tail index of the EMP distribution, a number of tests are made to check the tail-fatness of its distribution. Table 12 reports some of the properties of the EMP distribution for a sample of countries. According to the results reported in Table 12, the EMP distribution has a fatter tail than the normal distribution as indicated by its excess kurtosis. The Jarque-Bera and Shapiro-Wilk tests also indicate that the EMP has a non-normal distribution. In few cases, the EMP is highly skewed to the right. This result, however, shows one limitation of this chapter in which my simulation assumes the EMP is from a symmetric but fat-tailed student-t distribution. Switzerland, Thailand, Turkey, UK, and USA. Choice of these countries is driven mainly by the availability of data. Monthly data from 1960:1 to 1998:12 (if available) are collected for exchange rate, short term interest rate given by money market interest rate whenever available or the discount rate otherwise, and ratio of international reserve to Ml. Monthly data are used to get sufficient observations for the application of the extreme value theory. 75

91 Reproduced with permission of the copyright owner. Further reproduction prohibited without perm ission. Country Sample Period Argentina 1979:3-1998:12 Australia 1969:7-1996:6 Brazil 1965:3-1998:12 Canada 1960:1-1998:12 Denmark 1972:1-1998:12 France 1974:1-1998:12 Japan 1963:3-1998:12 M alaysia 1968:1-1998:12 New 1967:9- Zealand 1998:12 South 1965:1- Africa 1998:12 Singapore 1973:8-1998:12 Tuhlc 11 Sample of GARCI I Hslimalion Output for the Three Components of BMP ARCH coefficient (2.72) 0.238" (3.33) (6.61) (5.05) (9.13) (5.34) (4.91) (4.56) (3.03) (9.90) (2.69) %Ac A(i,i-iKl) (%Ai j, - %Aiel) GARCH ARCH ARCH GARCH ARCH ARCH GARC ARCH coefficient LM Test Coeffi coeffi LM Test coeffi Hcocffi LM l est (order 6) cient cient (order 6 ) cient -cicnt (order 6) (12.84) (12.76) (79.55) (2. 11) (22.78) (10.28) (2.06) (16.49) (25.26) (21.05) (2.34) (7.64) (0. 10) (9.21) (9.92) (9.18) (8.37) (7.80) (9.28) (6.79) (4.32) (-1.98) (44.29) ( ) (33.36) (50.3) (237.31) (36.66) (25.00) (56.06) (31.78) (25.63) (6.59) (8.48) (3.44) (5.27) (2.90) 3.123* 0.512" (5.23) (3.56) (5.33) (4.53) (4.98) " (4.09) (79.29) (8. 10) (25.18) (23.90) (7.75) (3.27) (16.57) (66.47) (15.81) *

92 Reproduced with permission of the copyright owner. Further reproduction prohibited without perm ission. Table 11-Continued Country Sample Period Sweden 1966:1-1998:12 Thailand 1977:1-1998:12 UK 1972:1-1998:12 USA 1960:1-1998:12 ARCH coefficient 0.209" (5.39) (2.77) (4.73) (6.61) %Ac A(iiria) (%Ar - %Ar,.,) ARCH ARCH GARCH ARCH OARC LM Test Coefficienciencient coeffi coeffi Hcocffi (older 6) -cicnl GARCH coefficient (7.12) (2.59) (88.97) * (37.72) (4.49) (8.98) (10.80) ( 12.86) (33.14) (34.26) ARCH I,M 'lest (order 6) (5-2) " (2.65) " (6.63) 2.04 "** (5.25) Figures in bracket are /-statistics. "Significant at 1%; ** Significant at 5%; and *** Significant at 10%. * The GARCH coefficient is insignificant and hence an ARCH ( I ) model is fitted (5.77) (.07) ARCH LM Test (order 6) 6.658*

93 The results reported in Table 12 point-out that the tail of the EMP falls asymptotically within the Frechet distribution. The tail index, which is a good indicator for the tail mass or outliers, can then be estimated using the Hill-estimator. The latter is highly dependent on the number of highest order statistics used to compute the index. Following the steps in Longin and Solnik (2001), the optimal number of highest order statistics (m*) for each possible theoretical distribution is determined using Monte Carlo simulation. Thirty-one possible theoretical distributions (student-t from 1 to 5 degree of freedom in increment of 0.2 and studentt from 5 to 10 degree of freedom in increment of 0.5) are considered21. A random sample of size n is drawn 10,000 times. For each level of highest order statistics-m (ranging from 1% to 20% of n), the tail index is estimated for the 10, simulated/random samples. The mean squared error (MSE) of the tail index estimator is then computed for a given value of m. For each possible theoretical student-t distribution with a degree of freedom (d.f.), an optimal m* is chosen based on the smallest MSE value of the estimator. Table 13 reports the optimal m* for different sample size n (matching the sample period for each of the countries in my sample) and student-t with a d.f. 21 a in my simulation is actually allowed to vary from 1 to 15. But the results for student-t with degrees of freedom from 10 to 15 are the same (see also the pattern from Table 13). 78

94 Table 12 Properties of the EMP Distribution for a Sample of Countries (60:1-98:12) Sample Exchange Market Pressure Test for Normality of EMP Country Size (EMP) Mean Skewness Kurtosis Jarqu-Bera Shapiro-Wilk Argentina * 0.494* Australia * 0.950* Brazil * 0.884* Canada * 0.948* France * 0.839* Indonesia * 0.692* Japan * * Malaysia * 0.946* New * 0.955* Zealand South * 0.951* Africa Singapore * 0.927* Sweden * 0.565* Thailand * 0.982* U.K * 0.753* U.S.A * 0.938* According to the results in Table 13 for a sample size of 171 from a student-t with 2 degree of freedom, the right tail begins at the 91st percentile. So the right tail includes 17 observations. If the sample is from a student-t with 6 degree of freedom instead, the right tail starts at the 97th percentile. In the latter case, the right tail includes 7 observations. The optimal m* for each sample size and each possible student-t with a d.f., reported in Table 13, is used to calculate the tail index of the Thus, we report simulation results from student-t distributions with degrees of freedom from 1 to

95 actual EMP of each of the sample countries included in this study. As expected we can see from Table 13 that student-t s with lower degree of freedom have fatter-tails or larger number of tail observations. The remaining part of this section uses the optimal m* reported in Table 13 to estimate the Hill-tail index of the actual EMP of each of the 36 countries in the sample. Table 14, for example, reports the estimated tail index of the U.S. EMP. If the U.S. EMP is assumed to be from a student-t with 1 degree of freedom (where the corresponding true tail index is the inverse of the degree of freedom), the estimated Hill-tail index is 0.538, computed based on the optimal m* equals 70 determined by the earlier simulation. If the EMP is instead from a student-t with 5 degree of freedom, the estimated Hill-tail index is Of all the estimated tail indexes, one is statistically the closest to the true index given by the inverse of the degree of freedom of the assumed student-t distribution. The results of the test with the null of estimated tail index equals with the true index are given in the last four columns of Table 14. The closest estimated tail index to the true tail index is selected by the highest p-value of the t-test for their equality. Accordingly, the U.S. EMP is from a student-t with 2 degree of freedom with 42 tail observations. The procedures in Table 14 are repeatedly applied for each of the remaining sample countries. Table 15 reports the number of tail observations of the actual EMP distribution of all the sample countries in this study. According to the results in Table 15, the monthly EMP distribution for Argentina has twenty-one tail observations 80

96 Optimal Number of ihc Highest Older Statistics (m*) lor I)iIIcrcut Sample Size and Student-t Distribution: Result Irom Sim ulated Random Samples. Student-t Sample Si/.c With DF (0.82) 38(0.81) 42(0.81) 46 (0.82) 48 (0.82) 50 (0.82) 51 (0.84) 52 (0.85) " (0.84) 34 (0.83) 32 (0.86) 34 (0.87) 38 (0.86) 37 (0.87) 45 (0.86) 48 (0.86) (0.87) 27 (0.87) 25 (0.89) 34 (0.87) 33 (0.88) 37 (0.87) 39 (0.88) 36 (0.90) (0.87) 19(0.91) 23 (0.90) 34 (0.87) 28 (0.90) 29 (0.90) 33 (0.90) 36 (0.90) (0.91) 19(0.91) 23 (0.90) 22 (0.92) 28 (0.90) 29 (0.90) 27 (0.92) 36 (0.90) 2 17(0.91) 17(0.92) 23 (0.90) 22 (0.92) 23 (0.92) 21 (0.93) 27 (0.92) 26 (0.93) (0.91) 17(0.92) 23 (0.90) 22 (0.92) 18(0.94) 21 (0.93) 24 (0.93) 19(0.95) (0.94) 15(0.93) 13(0.95) 17(0.94) 18 (0.94) 16(0.95) 21 (0.94) 19(0.95) (0.94) 15(0.93) 13(0.95) 17(0.94) 18(0.94) 16(0.95) 18 (0.95) 19(0.95) (0.94) 15(0.93) 11 (0.96) 17(0.94) 18 (0.94) 16(0.95) 18(0.95) 19(0.95) 3 12(0.94) 13(0.94) 11 (0.96) 10(0.97) 13 (0.96) 16(0.95) 18 (0.95) 19(0.95) (0.94) 11 (0.95) 8 (0.97) 10(0.97) 13 (0.96) 16(0.95) 15(0.96) 19(0.95) (0.97) 11 (0.95) 8 (0.97) 10(0.97) 13 (0.96) 16(0.95) 15 (0.96) 19(0.95) (0.97) 10 (0.96) 8 (0.97) 10 (0.97) 13 (0.96) 16(0.95) 15(0.96) 19(0.95) (0.97) 10 (0.96) 8 (0.97) 10(0.97) 13 (0.96) 8 (0.98) 15(0.96) 10 (0.98) 4 7 (0.97) 10 (0.96) 8 (0.97) 10 (0.97) 13 (0.96) 8 (0.98) 12(0.97) 48 ::: v 8) (0.97) 10 (0.96) 6 (0.98) 10(0.97) 8 (0.98) 8 (0.98) 12(0.97) 10(0.98) (0.97) 8 (0.97) 6 (0.98) 10(0.97) 8 (0.98) 8 (0.98) 12(0.97) 10 (0.98) (0.97) 8 (0.97) 6 (0.98) 10 (0.97) 8 (0.98) 8 (0.98) 12(0.97) 10 (0.98) (0.97) 8 (0.97) 6 (0.98) 10(0.97) 8 (0.98) 8 (0.98) 12(0.97) 10 (0.98) 5 7 (0.97) 8 (0.97) 6 (0.98) 10(0.97) 8 (0.98) 8 (0.98) 12 (0.97) 10 (0.98) (0.97) 8 (0.97) 6 (0.98) 10 (0.97) 8 (0.98) 8 (0.98) 12 (0.97) 6 (0.99)

97 Os' O' S O' S O' S o o o' o S o o' o o S CN W' '.MS sc sc O o nc voo <0 sc O oc O' S O' S O' S o ' 00s o 0? o oo o S /-N o o o CN o 0, o o 0, 0, e O' O' O' o o o o sc sc 00 0? 00 0? 0? S' O ' O' O' o o o o' o o S o w 71.. CN ' ' '. - o oc ir, ir. ir, 00 oc CJ CN O ' O'. S N it. O'. ' 00 S o' o od o S o' o o. S o. rr CN d d d o ir, w. ir. i/-., 2 53 C/3 S S ST o' O' O ' O ' o S o o' o S o o' o o' o CN 0. 0, 0, o 52, IT; vr. I/-. V/. I/-. IT. sr. I/-. ir. S ' S ' ' CN O' O ' o o o' o' o o' o o' o o o' o' o CN 0, o 0, 52, 'T '3' T V -S' '3- u o p - 00 S 00 ' c ^ oc rn fn N r c^ T ir c^. ir. CN ir o.!^1 * a! A '"T 'W' 'W >C- w - CN CN CN! 00 oc ir*. CN CN oc CN 1 m TT CN /-V NC On r, oo o c ^ r QN? r CN r PT CN p^. C ' o 2 ^ 2 o 00 CN CN TT ^r r - NC p^. r^, p^i p^. p>~ CN On ' ' W W' w v~. r~ TT 00 ir. r r s oo 00 oc S ' _ o o o s. 2. r r p^ l/. "T T 3 c/ 2 ir. o c 00 o o s o c r - r". C' TT c^ W' W' <N CN r - r r s s sc 'T sc r r 00 p^. r r ^T c^ C' CN CN CN ^r CN CN ir O'. S o CN CN CN CN CN ' 3* O CN CN o CN ir, O 00 CN o c CN CN r r CN CN ir. CN CS ir. o ir. o o c r CN PT CN PT CN r~ O ' S oc S ' O' o o o o 00 o oo" o 00 o 0, 0, c 0, o 52, VCsC 'O >o sc sc so in r r Pri 00 (N s c NC o p^, CN pn CN p^*. C' ir. O it, O o CN ir. 00 f^. 0 0 CN 0 0 CN o c CN CN CN Tabic 13-Continucd f Studcnt-t With DF r~ 0 0 O ' O' o 00 o oc* o 00 o o oo S o ' 00 r* o o 0. 0, 52, r~- r- V. ir. ir. v. 1/-. o ir. sd r I/-. oc' o 1/ -. d o o 6 U l cn _o «Q 53 1 S -a uc ir*, v r, 0 0 o c o o CN CN c«~, c^ CN O O 50- o o v-. LL. Q _ CN "T r. r^. r - CN p^, CN p^. CN r r CN CN CN CN CN O CN 82

98 (0.95) 28(0.95) 23 (0.96) 14(0.98) 14 (0.98) 14 (0.98) (86 0) FI 14 (0.98) (86 0) FI (86 0) FI 14 (0.98) 14(0.98) (86 0) FI 14 (0.98) (66 0) 6 (66 0) 6 (6 6 0 )6 (66 0) 6 ( )6 (660) 6 (86 0) FI (86 0) FI Table 13- C ontinued (0.96) (96 0) (0.94) 17(0.97) Sample Size (0.94) 21 (0.95) 23 (0.95) 20 (0.96) l/t rr rr (96 0) 0Z Studcnl-l with DF 2.8 it. O d rr CN UT Os CN (960) L1 rr 22 (0.96) 22 (0.96) (86 0) e I (96 0) ZZ (96 0) ZZ 17(0.97) 17 (0.97) 13(0.98) 13(0.98) 16(0.97) 16(0.97) l/t O ' rr CN 19(0.96) o>. vc s vc O', O' (86 0) (0.98) 13 (0.98) 13 (0.98) 13(0.98) (660)6 (660)6 (660)6 (660)6 (660)6 (660)6 (86 0)11 (8 6 0 )1 1 13(0.98) c*_ O' oc oc 16(0.97) 12(0.98) 12(0.98) 12(0.98) 12(0.98) 12(0.98) 12(0.98) 12(0.98) 15(0.97) 15(0.97) 15(0.97) 15(0.97) 15(0.97) 14(0.97) 14 (0.97) 14(0.97) 14(0.97) 14(0.97) 14 (0.97) 14(0.97) 14(0.97) 14 (0.97) 14(0.97) 1 13(0.97) 13 (0.97) o rr rr r- o d rr oc rr 13 (0.97) 13 (0.97) 13(0.97) 13(0.97) 13(0.97) 13 (0.97) 7 (0.99) (0.99)(660)6 1 (660)6 (66 0)6 (660)6 (660)6 (66 0) 8 (66 0) 8 (66 0) 8 (66 0) 8 (66 0) 8 (660) 8 (66 0) 8 (66 0) 8 (66 0) 8 (66 0) 8 8 (0.99) (660) 8 (66 0) 8 (66 0) 8 (66 0) 8 (66 0) 8 (66 0) 8 (66 0) 8 (66 0) 8 (66 0) 8 (66 0) 8 (66 0) 8 (86 0) 11 (860) 11 (86 0) 11 (860) 11 (860) 11 (86 0) 11 (86 0) 11 (86 0) 11 (66 0) L (66 0) L (66 0) L (66 0) L (66 0) L (66 0) L (66 0) L \ (66 0) L (66 0) L (66 0) L 00 'T l/t ur SC UT c - (66 0) L (660) L (66 0) L (66 0) L (660) L (66 0) L (660) L (660) L (66 0) L vr OC VT OO O' i/r Os 01 Note: *' bcrs in parentheses are the percentile at which the tail of the distribution begins

99 c Table 14 Hill-Tail Index Estimate of the U.S. EMP: 1960:1-1998:12 (Sample Size 467) Student-t with d.f. (a) Optimal m* Hill-index A Y -< > II Y Ho Y = Y t-valuea p-value t-valueb p-value 1 70 (0.86) (0.86) (0.89) (0.92) (0.92) (0.92) * * (0.95) (0.95) (0.95) (0.95) (0.95) (0.96) (0.98) (0.98) (0.98) (0.98) (0.98) (0.98) (0.98) (0.98) (0.98) (0.98) (0.98) (0.98) (0.98) (0.99) (0.99) (0.99)

100 Table 14-Continued Student-t with d.f. (a) Optimal m* Hill-index A a SU "< > II Ho Y=Y t-value* p-value t-valueb p-value y 9 9 (0.99) (0.99) (0.99) Note: (y-y)m1/2 ~ N(0,y); *r-value = (y-y)/se(y) where sely) = -^(y2/ m) \and y=l/a. 't-value = (y~ Y )l se(y) where seiy)- Yz/m while there are sixteen tail observations for the Korean monthly EMP distribution. Note that all countries do not have the same sample period Identification of Currency Crises The tail observations of the actual EMP distribution of the 36 sample countries are determined in section 4.3 with the help of Monte Carlo simulations. These observations are the results of successful and unsuccessful speculative pressures against the currency of each country. The corresponding times at which these observations occur define the periods of currency crisis. 85

101 Reproduced with permission of the copyright owner. Further reproduction prohibited without perm ission. OO ON Country Table 15 Number of Tail Observations of the Actual LMP Distribution of the Si 23, 'c Countries. Sample Period D.F of Closest Studcnt-l Tail Obscrvs. Country Sample Period Hillindex Hillindex D.F of (. loses t Sludcnt-t Tail Observs. Y Y Argentina* 79:3-98: M alaysia* 68:1-98: Australia 69:7-96, M exico 81:4-98: Austria 67 :1-98: Netherlands* 60:1-97: Belgium 60:1-98: New Zealand* 67:9-98: Brazil 65:1-98: Norway 64:1-98: Canada 60:1-98: Peru 65:1-98: Colom bia 65:1-98: Philippines 65:1-98: Denmark 72:1-98: Portugal 60:1-98: Ecuador 70:1-98: South Africa 65:1-98: Finland 77:12-98: Singapore* 7 3:8-98: France 74:1-98: Spain 74:1-98: Greece* 60:1-98: Sri Lanka 78:1-98: Indonesia 83:1-98: Sweden* 66:1-98: Ireland 67:1-98: Sw itzerland 64:1-98: Israel 82:3-96: Thailand 77:1-98: Italy 71:1-98: Turkey* 65:1-96: Japan 63:1-98: UK 72:1-98: Korea* 76:8-98: U.S.A 60:1-98: * Optimal m* for these countries is set to that levc Table 14. of m* determined in the simulation for (he closest sample size shown in

102 The main objective of the whole exercise is to test whether currency crises are contagious and to determine the relevant contagion channels using the objectively identified crises. The final test is done using quarterly data. But the extreme value approach in this chapter is applied to monthly EMP data due to the large data requirement necessary to undertake the simulation. As proved in the literature, the tail index is invariant to the frequency of the data or time aggregation (see Longin, 2000, and Martien, Hols and De Vries, 1991) though, as expected, the lower frequency distribution has a smaller number of tail observations. Thus crises, corresponding to the extreme values/tail observations, are identified from the monthly EMP data and then aggregated to their quarterly counter parts. In the aggregation, there is currency crisis in a quarter if crisis, which corresponds to the tail observation determined in section 4.3 above, occurs in one or more of the corresponding three months within that quarter. The frequency of quarters in currency crisis for each country in our sample is reported in Table 16. As in Eichengreen et al. (1996) and Kaminsky and Reinhart (2000), crises are not allowed in consecutive quarters because the impact of one crisis may extend beyond one quarter. The figures in Table 16 indicate the total number of either full fledge currency crises or intense pressure against the currency of a country but defended by the central bank by running down its reserves and/or raising the domestic interest rate. Using the extreme value approach, six periods of crises are identified for Argentina while 87

103 Reproduced with permission of the copyright owner. Further reproduction prohibited without perm ission. Table lb Number of Currency Crises in Quarter: Summary of the Crises Identified Based on the Hxtrcme Value Approach Sam ple Period Number of quarters in Crisis Incidence (%) Country Sample Period Number of quarters in Crisis Incidence (%) Country Argentina 79:2-98: Malaysia 68:1-98: Australia 69:3-96: M exico 81:2-98: Austria 67:1-98: Netherlands 60:1-97: Belgium 60:1-98: New Zealand 67:4-98: Brazil 63:1-98: Norway 64:1-98: Canada 60:1-98: Peru 65:1-98: Colom bia 65:1-98:4 II 8.09 Philippines 65:1-98: Denmark 72:1-98: Portugal 60:1-98: Ecuador 70:1-98: S. Africa 65:1-98: Finland 78:1-98:4 II Singapore 73:3-98: France 74:1-98: Spain 74:1-98: Greece 60:1-98: Sri Lanka 78:1-98: Indonesia 83:1-98: Sweden 66:1-98: Ireland 67:1-98: Switzerland 64:1-98: Israel 82:2-96: Thailand 77:1-98: Italy 71:1-98: Turkey 65:1-96: Japan 63:1-98: u k 72:1-98: Korea 76:3-98: USA 60:1-98:

104 the U.S. experiences twenty-four periods of crises. Note that the sample periods over which crises are identified for the U.S. and Argentina are different22. It may seem odd that currency crises are more prevalent in the U.S. than in Argentina. However, such a result is not unwarranted. To see why consider Table 17 which presents the specific time periods during which the U.S. dollar was under intense pressure. Most of the U.S. currency crises occurred in the 1960s. Since 1944, the value of U.S. dollar was fixed in terms of gold ($35 an ounce). Most nations then fixed their currencies to the U.S. dollar and kept a major part of their international reserves in dollars. But the 1960s saw a series of political and economic events in the U.S. and abroad stressing the so-called dollar exchange or Bretton Woods system. Since the U.S. issued the key currency, it enjoyed the benefits of deficits without tears, running balance of payment deficits and consuming a multitude of consumer products from abroad. In addition, the U.S. foreign policy goal of containing communism (e.g., Vietnam War ( )), waging the Cold War, and striving for decolonization kept dollars out-flowing and leaving large dollar reserves circulating abroad. At the same time, the level of U.S. gold reserves backing those dollars circulating around the world steadily dwindled. Foreign central banks and governments held over 14 billion U.S. dollars by At the same time, the U.S. had $13.2 billion in gold reserves of which only $3.2 billion was available to cover foreign dollar holdings. The remainder was 22 To account for the variations of the sample period, crisis incidence measured by the 89

105 Table 17 Periods of High Exchange Market Pressure on the U.S. dollar Identified by Extreme Value Approach Periods of Currency Crisis 1961QI, 1962Q4, 1964Q2, 1964Q4, 1966Q1, 1966Q3, 1967Q1, 1968Q2, 1968Q4, 1969Q2 1970Q2, 1970Q4, 197IQ2, 1972Q3, 1973QI, 1974Q2 Political and Economic Events -Cuban Missile crisis -Assassination of J. F. Kennedy -American Civil Right Movements -The Vietnam War ( ) -The 1968 Dollar Devaluation (Gold parity changed) -The collapse of the Bretton Woods system 1977Q4, 1978Q4, 1979Q3 -The 1977/78 European currency crisis 1987Q1, 1988Q4 (Snake) -Dollar was depreciating to the point of creating conflict with Germany and Japan and finally leading to the 1987 Louvre Accord -The 1987 US stock market crash 1992Q3, 1993Q2 -The 1992/93 ERM crisis 1994Q1 -The U.S. interest rate hike, which may have led to the 1994/95 Mexican crisis needed to cover domestic dollar balances (see IMF, 2001). This had the effect of putting the dollar under a lot of pressure for most of the 1960s because many international financial lenders suspected that the U.S. would be forced either to devalue the dollar or stop redeeming dollars for gold. Many countries such as France percentage of crisis occurrence are given in the 4 and 8th columns. 90

106 were also openly requesting the U.S. to redeem their dollar reserves into gold. Eventually, the U.S. devalued the dollar by increasing the price of gold from its historical level of $35 to $70 per ounce in When this did not solve the pressure on U.S. reserves, the U.S. altogether abandoned the convertibility of the dollar into gold on August 15, These events accord well with 16 of the 24 crisis quarters found in the U.S. EMP data. The remaining 8 currency crises in the U.S. are divided into four groups as reported in Table 17. In some cases they appear to reflect contagion from elsewhere (as in the Snake (1977/78) and ERM (1992/93) crises). But in all cases, major events can be identified during each time period. For example, the crisis in 1994Q1 may reflect the U.S. interest rate hike, which according to Calvo and Reinhart (1996) might have led to the 1994 Mexican crisis, later to engulf the whole of Latin America. Figures 3 and 4 plot the percentage of countries in our sample of countries in crisis across time. Figure 3 plots all the thirty-seven countries together while Figure 4 covers a subset of twenty OECD countries23. In both Figures 3 and 4, a large number of countries were in crisis in the early 1970s, 1978/79, 1987 and 1992/93. The early1970s and 1978/79 correspond to the collapse of the Bretton Woods system and the Snake, respectively. The years 1987 and 1992/93 appear to correspond to the October 1987 U.S. 23 Since Germany is used as the center of reference, the actual number of countries in Figure 3 is 36 while the actual number in Figure 4 is

107 KMX) WMX) g (M X ) o & 4().(X) 20.(X) ().(X) mm iiiiiiinniiiiimiiiimmmiffiimfrmiiihnifniimmmmmnmmmihmimmmiimnimim GisisALL Pigure 3. Percentage of Countries in Cuircney Crisis: 37 countries (lixtrcmc Value Approach)

108 j 53 Cl. c. < u _3 > u X U s i U U au a o 3 tn s. s. 'Z u >* u 3 c 3 u s. u U u =0 3 U u 1U3DJ3Q

109 stock market crash and the European Exchange Rate Mechanism (ERM) crisis, respectively. Based on the standard approach of currency crisis identification (using the mean plus 1.5 standard deviations), a large number of countries were found in crisis in the early 1970s and 1992/93 (results are not reported here). Identification of crisis using the extreme value approach, as reported in Figures 3 and 4, captures additional major crisis years in 1978/79 and 1987 during which many countries were affected Conclusion In this chapter, periods of crisis are identified based on the extreme value theory. This is a relatively more objective method and is also a good alternative to the more commonly used approach in identifying periods of currency crises. The standard approach that is widely applied in the literature is more subjective in setting the threshold to identify periods of crisis. There are also variations in the standard approach as to whether the threshold should be common to all countries or specific to each country in the sample. Further, the threshold set using the mean and the standard deviations of the EMP is sensitive to outliers leaving the possibility of ignoring periods with a reasonably high exchange rate pressures or including some middle values that are not really the results of high pressures. For the purpose of comparison, Table 18 reports crises identified based on the extreme value and standard approaches. In both cases, crises are identified using 94

110 monthly EMP data and then aggregated to get the reported quarterly crises. Figures in parentheses are the crisis incidence rates. As can be seen in Table 18, the extreme value approach identifies more crises than the standard approach for the majority of the countries. But it also appears that crises identified using the extreme value approach are closer to those identified based on a cut-off point of mean plus l.s standard deviations. But identification of currency crisis using the mean and standard deviation has its own limitation as discussed earlier. Given these results, the approach followed in this paper lends itself as a relatively more objective alternative to the standard approach in identifying periods of currency crisis. 95

111 COUNTRY 'I ahlc IX Number of Quarters in Crises Identified by the Extreme Value and the Standard Approaches Period Crisis Determined according to the EVT Country Specific Threshold (the Kaminsky and Reinhart (2000) Procedure) Common Threshold (the Eichcngrccn et al. (1996) procedure) mean+1.5sd mean+2sd mean+3sd mcan+l.5sd mcan+2sd mcan+3sd Argentina 79:2-98:4 6 (7.6) 4 (5.1) 3 (3.8) 2 (2.5) 5 (6.3) 5 (6.3) 4(5.1) Australia 69:3-96:2 6 (5.6) 13 (12.0) 8 (7.4) 3 (2.8) 7 (6.5) 3 ( 2.8) 0 (0.0) Austria 67:1-98:4 21 (16.4) 8 (6.3) 4(3.1) 3 (2.3) 4(3.1) 3 ( 2.3) 1 (0.8) Belgium 60:1-98:4 26(16.8) 14 (9.0) 9 (5.8) 5 (3.2) 9(5.8) 7(4.5) 3(1.9) Brazil 65:1-98:4 7 (5.2) 13(9.6) 13(9.6) 8 (5.9) 19(14.0) 13(9.6) 9 (6.6) Canada 60:1-98:4 17(10.9) 18(11.5) 11(7.1) 5(3.2) 14 (9.0) 8(5.1) 2(1.3) Colombia 65:1-98:4 11 (8.1) 13(9.6) 9 (6.6) 4 (2.9) 12(8.8) 8 ( 5.9) 4 (2.9) Denmark 72:1-98:4 14 (13.0) 14(13.0) 6 (5.6) 3 (2.8) 8(7.4) 6 ( 5.6) 2(1.9) Ecuador 70:1-98:4 13(11.2) 18(15.5) 13(11.2) 4 (3.4) 17(14.7) 11(9.5) 3 (2.6) Finland 78:1-98:4 II (13.1) 10(11.9) 6(7.1) 4 (4.8) 10(11.9) 6(7.1) 4(4.8) France 74:1-98:4 20 (20.0) 11 (MO) 9 (9.0) 6 (6.0) 9 (9.0) 6 ( 6.0) 3 (3.0) Greece 60:1-98:1 18(11.8) 16(10.5) II (7.2) 5 (3.3) 15(9.8) 11 ( 7.2) 4 (2.6) Indonesia 83:1-98:4 9(14.1) 6 (9.4) 3 (4.7) M I.6) 6 (9.4) 3(4.7) 1 ( 1.6) Ireland 67:1-98:4 24(18.8) 18(14.1) 10(7.8) 4(3.1) 12(9.4) 8(6.3) 1 (0.8) Israel 82:2-96:2 9(15.8) 6(10.5) 5 (8.8) 3 (5.3) 6(10.5) 4 ( 7.0) 2(3.5) Italy 71:1-98:4 8(7.1) 12(10.7) 9 (8.0) 5 (4.5) 12(10.7) 9 (8.0) 5 (4.5) Japan 63:1-98:4 15(10.4) 11 (7.6) 5 (3.5) 3(2.1) 4 (2.8) 3(2.1) 0 (0.0) Korea 76:3-98:4 12 (13.3) 8 (9.0) 5 (5.6) 3(3.4) 3 (3.4) 3(3.4) 2 (2.2) Malaysia 68:1-98:4 16(12.9) 17(13.7) 7 (5.6) 3 (2.4) 3 (2.4) 2 ( 1.6) 0 (0.0) Mexico 81:2-98:4 12(16.9) 5 (7.0) 3 (4.2) 2 (2.8) 11(15.5) I0( 14.1) 4 (5.6) Netherlands 60:1-97:4 24 (15.8) 3 (2.0) 1 (0.7) 1 (0.7) 4 (2.6) 2 ( 1.3) 1 (0.7)