RESEARCH includes research articles that focus on the analysis and resolution of managerial and academic issues based on analytical and empirical or case research Executive Summary Measurement of International Currency Crises: A Panel Data Approach using Composite Indices K V Bhanu Murthy and Anjala Kalsie The literature on crisis has often developed indices for measurement of crisis. The indices that have been developed in the earlier studies suffer from three problems: • Conceptually, they include only exchange related variables and not other relevant variables that are crucial for international trade and international finance. • The extant studies do not use a causal framework as a methodology for the selection of variable. • Empirically, they do not use more evolved statistical tools such as Principal Component Analysis for constructing a composite index. This paper seeks to measure the international currency crisis of 1997 in Asia. It has taken the case of the A5 countries in 1997 and has developed a methodology meant to measure and explain currency crisis. The study has constructed composite indices for capturing the causes – macro-economic and financial – as well as an index of crisis. It uses Jha & Murthy’s (2006) approach for constructing composite indices. It combines continuous and discrete approaches for defining and measuring crises and uses India as a ‘control’ which enables international and inter-temporal comparisons during crisis. An attempt to explain a widespread and complex phenomenon in terms of a single dependent variable would be incomplete and partial, where the dependent variables which represents the crisis are themselves a conglomerate of many factors. Since it is a complex phenomenon, it cannot be represented by one single variable. Moreover, these variables tend to be correlated. With the help of two composite indices – one for financial variables and another for macro variables, a fixed effects panel regression model is developed for explaining the crisis. The paper measures the decomposition effects of causal financial and macro variables on the crises. KEY WORDS International Currency Crises Composite Index It has been found that Index of crises consists of exports, exchange rate, and interest rate. The financial index contains risk rating, domestic financing, and stock traded. The index of macro variables is based on GDP, capital formation, and budget balance. Macro index negatively influences crisis while financial index influences crisis positively. Asian Crises India as the base country clearly brings out the contrast between crises affected countries and neutral countries with the help of this methodology. The crisis window shows up very clearly, as it combines a discrete and continuous definition. Principal Component Analysis Finally, the differences amongst the five Asian countries during crises are also brought out by this methodology. Panel Regression VIKALPA • VOLUME 38 • NO 4 • OCTOBER - DECEMBER 2013 13 T he 1990s were marked by a number of rather unusual financial and economic crises, such as the Mexican Peso Crisis of 1994-95 and the Asian Crisis of 1997. While these crises ranged from the “garden variety” of currency crises to rather esoteric real estate bubbles, studies of such crises exhibit empirical and theoretical commonalties. The literature distinguishes three varieties of financial crises: currency crises, banking crises, and debt crises. The analysis in this study is primarily focused on the currency crises. The task at hand is to analyse and measure the currency crises in the A5 countries1 during 1997. A currency crisis is an episode of intense foreign exchange market pressure. It can be defined simply as an incident in which a country experiences a substantial nominal devaluation or depreciation. This criterion, however, would exclude instances where a currency came under severe pressure but the authorities successfully defended it, by intervening heavily in the foreign exchange market, by raising interest rates sharply, or by both. Extant approaches have sometimes constructed an index of speculative market pressures (Kaminsky, Lizondo, & Reinhart, 1998; Edison, 2000; Goldstein, Kaminsky, & Reinhart, 2000). The indices that have been developed in the earlier studies suffer from three problems: 1. Conceptually, they include only the exchange-related variables2 and not the other relevant variables that are crucial for international trade and international finance. 2. They do not use causal framework as a methodology for the selection of variable. 3. Empirically, they do not use more evolved statistical tools such as Principal Component Analysis for constructing a composite index. This paper is a part of a larger study that looks into a new approach to measure and analyse international currency crises. A crucial part of the study is to develop a set of new composite indices, for both causal variables and impacted/dependent variables, so as to correlate them in a regression framework. These indices are based on a large number of variables and involve a three-stage procedure, 1 Malaysia, Philippines, Korea, Thailand, and Indonesia. 2 Weighted average of ER changes, weighted average of RER changes, Reserves changes, and Interest rate changes. 14 which shall be discussed later. RATIONALE An attempt to explain a widespread and complex phenomenon in terms of a single dependent variable would be incomplete and partial, if the dependent variables which represent the crisis are themselves a conglomerate of many factors. Since a crisis is a complex phenomenon, it cannot be represented by one single variable. Moreover, the variables tend to be correlated. Thus, the ordinary regression framework results in the problem of multicollinearity. Therefore, it is necessary to measure the phenomenon of the crisis with the help of composite indices, which would adequately represent the complex phenomenon. This applies to both the causes and the effects of a crisis. Different studies have identified a variety of factors related to crisis. This study conducts a causality test to determine which of these factors are causes and which are effects. The final selection of variables is done on the basis of an elaborate procedure, which ensures that the variables which are entering in the construction of the index of crisis are the ones that are theoretically relevant, as they are drawn from extant studies and are empirically sound as they are tested for causality. In addition, they are appropriate because they have been checked for data redundancy. CONCEPTUAL ISSUES Prior Procedure Several steps were followed as a part of the larger study to ensure the above considerations: 1. A literature survey of empirical and conceptual studies (Moreno, 1995; Berg & Pattillo, 1999; Frankel & Rose, 1996) was undertaken on the basis of which a data set consisting of a large number of crises variables (30 variables including financial and macro variables) could be arrived at. 2. The set of available variables were checked for data redundancy. Many defined variables in the data set were different versions of the same variable. The authors used their judgment to drop a certain version and retain another version. For instance, if a variable was defined in terms of PPP $, US $ or local currency units, only one of them was chosen. MEASUREMENT OF INTERNATIONAL CURRENCY CRISES: A PANEL DATA APPROACH ... 3. A correlation exercise was carried out on these 30 variables. The purpose of this exercise was to establish that crisis variables were ordinarily correlated3. 4. A dummy variables exercise4 was also undertaken wherein the data series for six countries for each of the 30 variables were tested to see whether there was any structural break during 1997 and 1998, which was the crises window of the Asian currency crisis. However, the prior correlation analysis and dummy variable exercise did not indicate anything about the causality amongst the variables. The dummy variable exercise is a univariate analysis that does not capture the complexity of the phenomenon. Measurement of Crisis After having undertaken the above empirical steps, the authors proceeded with the measurement of crises. The first part of the measurement exercise consisted of construction and measurement of the indices while in the second part, these indices were used to model and predict the crises and predict. ses. On the other hand, the explanatory variables are so endowed that they are capable of inducing discrete change in the indicator of crises. The methodology used in this study captures and measures the discrete change through the independent variable and not through the dependent variable. The argument is that certain continuous changes can be captured through the causal variables, which lead up to the crises. This continuous change is manifest in the volatility of the crises variables. In effect, it implies that those countries which were crises-ridden had experienced a continuous trend of volatility to the extent that this built up continuously to discrete change resulting in the crises. The merit in using a continuous crisis definition lies in its ability to capture both the continuous influence on the crises variable (dependent variable) as well as its ability to explain discrete change which is occurring in the crises variables during crises, due to the causal variable(s). Crises Window Crisis Definition In certain studies, the crisis variable itself has been defined in discrete terms (Eliasson & Krevter, 2000) with the understanding that the dependent variable had a builtin discrete change or kink. It should not be forgotten that the dependent variable is an effect. The authors’ understanding is that whatever change comes in the dependent variable is on account of the causal variables, including changes in the intercept, because the intercept also contributes towards the explanatory power of the equation. In extant literature, the distinction between discrete and continuous crises definition has been captured through either discrete or continuous dependent variable. In a causal framework, the mechanism to capture discrete change in the phenomenon has to necessarily rest upon the causal variables and not the dependent variables. Accordingly, this study follows the methodology that, the impetus to discrete change, during crises, is caused by the indicator of the crises and would necessarily come from the causal variables. This methodology has been tailored in such a manner that it does not pre-suppose the nature of dependent variable that represents cri- 3 The result of correlation exercise is not reported. 4 The result of dummy variable exercise is not reported. VIKALPA • VOLUME 38 • NO 4 • OCTOBER - DECEMBER 2013 The years 1997 and 1998 were taken as the crises window. The crises developed in November 1997 and peaked in 1998 in many countries. Therefore, neither 1997 nor 1998 could be ignored. This was vindicated by the dummy variable exercise which showed a significant structural break across variables during this crises window (Malik, 2008). Some of the extant studies have used monthly data and defined the crises window in terms of particular months. As this study was using annual data, it was neither possible to have a crises window that pinpointed the precise period of crises, nor was there any interest in the process of the crises. Therefore, the crises window in this study is defined in annual terms. Relevance of Control An important issue of research design was the introduction of a control. In the case of extant studies, with a discrete crises definition, the control was established with reference to the pre-crises period, since a control is meant to represent a normal period or normal observation. In the present study, by ‘control’ is meant a ‘benchmark country’ that was not affected by the crises. For identifying the control, dummy variable exercise was used, wherein it was found that in the case of India, none of the relevant variables (those variables which were identified 15 through literature review) were affected by the crises; and therefore, India was chosen as the control. In other words, none of the variables showed any structural break, during the crisis period 1997 and 1998, in India. Such an approach has the advantage of allowing both inter-temporal as well as international comparisons. The extant studies permit only inter-temporal comparisons (Malik, 2008). Since India was established as a control, it was reasonable to expect that all the relevant variables would display a normal behaviour in terms of cause and effect. Hence the causality tests were applied to the relevant variables only in the case of India. LITERATURE REVIEW In view of the complex set of issues involved in the literature review, the discussion is structured in the form of a table (See Appendix). METHODOLOGY The data was taken from the World Development Report and World Development Indicators (World Bank) for the years 1987 to 2002. To account for such a conception of the phenomenon of crises and causes of the phenomenon, there was a need for evolving an appropriate methodology. The two most desirable features of the methodology are that, firstly, it should capture the volatility or variance in the relevant variables, because it is this volatility that leads up to and results in crisis; secondly, it should enable working with a large number of related variables because a crisis according to our understanding is a complex phenomenon resulting from a large number of intercorrelated variables. The third dimension of methodology is that given the constraints of the data point and degree of freedom, the methodology should allow working with a parsimonious set of variables. The methodology which possesses all these features is Principal Component Analysis (PCA). Unlike Ordinary Least Squares (OLS), wherein the procedure is to minimize the sum of the squares of deviations, in the case of PCA, the procedure is to maximize the variance. Another feature of PCA is that it segregates inter-correlated variables into the separate orthogonal factors or principal components. Thirdly, PCA can be used for developing a composite index which collapses a set of variables into a single variable that represents a complex phenomenon like currency crisis. 16 Procedure of Estimation The empirical procedure involved five distinct steps: 1. Granger causality test was applied to the data on India in respect of all the relevant variables to identify the causal variables and those that they impacted. This was done in case of India, since it was the control. 2. Correlation analysis was used to segregate the variables into impacted or dependent variable and causal or explanatory or independent variables. Once the variables were separated into dependent and independent variables, independent variables were found to be correlated. 3. To deal with this problem, Principal Component Analysis was applied. PCA helped in (a) data reduction and (b) making the dependent variables uncorrelated. 4. The next step was the formation of a composite index. It helped in representing the crises phenomenon which was manifest in a large number of variables. This is the unique feature of this study. 5. The next step was to specify the model which helped in merging the two methodological issues, namely measuring of the crises and comparing the crises-hit countries with a non-crises hit country. Merging was done by using panel regression and treating India as a control. Causality Test For developing a causal framework, it is essential to adopt a procedure by which causal variables can be distinguished from impacted/dependent variables. Once the set of crises variables are sorted, through this procedure, it would be possible to develop an index of crisis. In the true spirit, causality test indicates the precedence of one variable over the other; it is therefore sometimes cautioned that the result of such tests may not be interpreted as cause and effect relationship. Here the authors would like to point out that the present study is not dependent on the Granger Causality Test. After using the test and sorting the variables as causal and impacted variables, a structured causal framework was used with the appropriate regression technique for establishing cause and effect. Granger Causality Test: For carrying out the granger causality test, the following two equations were estimated for all the variables: MEASUREMENT OF INTERNATIONAL CURRENCY CRISES: A PANEL DATA APPROACH ... Ho: X does not cause Y. Ho: Y does not cause X. We test the null hypothesis against an F stat, namely, F= {(RSSr-RSSur)/m}/ (RSSur/ (n-k)) The degrees of freedom are m and (n-k). The restrictions are respectively: ∑ai = 0 and ∑di = 0 Yt = ∑ ai Xt—i + ∑ bi Yt—i + u1t (1) Xt= ∑ ci Xt—i + ∑ di Yt—i + u2t (2) i =1, 2 RSSr =Residual sum of square restricted RSSur = Residual sum of square unrestricted m = Number of linear restrictions n = Number of observation k = Number of parameters in the unrestricted regression hence, no information contained in the points in the event space is lost. The normality assumption is not essential in PCA and with such a dispersed set of outcomes, this feature is useful. PCA is ideally suited because it maximizes the variance rather than minimizing the least square distance and this study is interested in capturing variance because the authors believe that volatility is the basic cause of a crisis. Since the objective is to develop a composite ‘Index of Crisis’ and relate it to two other indices of financial variables and macro variables, there is a need to choose the essential variables and arrive at relative weights for the purpose of consolidating these variables into a single index. This is facilitated by PCA. PCA linearly transforms an original set of variables into a smaller set of uncorrelated variables representing most of the information in the following form: (3) The first principal component is defined such that the variance of y1 is maximized. Consider the p random variables x1, x2,.. xp subject to the constraint that the sum of It was also ensured that there was no two-way causality among the relevant variables. As a consequence of the causality test, three sets of variables could be identified: (i) pure causal variable; (ii) pure impacted variable; and (iii) common variables which alternatively appeared as causal and impacted variables, although not as two-way causality. squared weights is equal to 1, i.e. Correlation Matrix optimal weight vector (a11, a12, a1p) and the associated variance of y1 (which is denoted as λ1). As mentioned earlier, the correlation exercise was done with a view to sorting the causal and impacted variables. After identifying the causal and impacted variables, through the causality test, some variables were found to be common, that is, they were both impacted as well as causal variables thus making it difficult to decide which variables to be selected for constructing an index of crisis. To solve this problem, correlation matrix was used. There are two objectives of studying the correlation matrix: • To segregate the set of dependent and independent variables • To identify a set of crisis variables. Principal Component Analysis Principal Components Analysis (henceforth PCA) (Lewis-Beck, 1994) is based on a linear transformation of the variables so that they are orthogonal to each other; VIKALPA • VOLUME 38 • NO 4 • OCTOBER - DECEMBER 2013 . If the vari- ance of y1 is maximized, then the sum of the squared correlations, i.e. is also maximized. PCA finds the If the objective is a simple summary of the information contained in the raw data, the use of component scores is desirable. It is possible to represent the components exactly from the combination of raw variables. The scores are obtained by combining the raw variables with weights that are proportional to their component loadings. In this case, the component scores were used for determining the weight of each of the raw variables in constructing a composite index. As more and more components are extracted, the measure of the explanatory power increases but it is necessary to strike a balance between parsimony and explanatory power. The goal of PCA is to reveal how different variables change in relation to each other, or how they are associated. This is achieved by transforming correlated original variables into a new set of uncorrelated (orthogonal) underlying variables (termed principal components) using the covariance matrix, or its standardized form – the correla- 17 tion matrix. The lack of correlation in the principal components is a useful property because it means that the principal components are measuring different “statistical dimensions” in the data. The new variables are a linear combination of the original ones and are sorted into descending order according to the amount of variance that they account for in the original set of variables. Each new variable accounts for as much of the remaining total variance of the original data as possible. Cumulatively, all the new variables account for 100 percent of the variation. PCA involves calculating the eigen values and their corresponding eigen vectors of the covariance matrix or correlation matrix. Each eigen value represents the total remaining variance that the corresponding new variable accounts for. The expectation from conducting PCA is that correlations among original variables are large enough so that the first few new variables or principal components account for most of the variance. If this holds, no essential insight is lost by further analysis or decision-making, and parsimony and clarity in the structure of the relationships are achieved. Each factor is a combination of variables which are correlated with the principal component. This methodology has two purposes. As reported early, both the sets of macro and financial variables were correlated within each set. Under such circumstance, it is not possible to use the variables in a regression framework on account of multicollinearity. Secondly, when there are a large number of impacted variables, they need to be collapsed into a single dependent variable (Jha & Murthy, 2006). There is a relevance of using PCA analysis in the modeling. It allows for data reduction. The reduced data set would contain the maximum information in all the variables, which were being considered as dependent variable. As a result of PCA, the reduced data set consists of variables which were not correlated to each other, since the principal components are orthogonal (perpendicular) to each other. Finally, the PCA methodology enables us to construct a composite index. The crux of this methodology is to represent complex interrelated phenomena such as a crisis with the help of a single composite index, which could act as a unique dependent variable. It may be argued that there are many other factors that influence crises, but our methodology ensures that the variables which were cho- 18 sen to construct crises index, effectively represent the impact of all the crises variables. Since the ‘principal variables’ are highly correlated to the principal components, they contain the same information. One measure of the explanatory power of the index formed by this procedure is given by the variation explained by the retained principal component. Composite Index Method for Construction of the Index The main aim of this empirical work is to evolve a composite Index of Crisis (IOC). Hence, there is a need to choose the essential variables by a data reduction procedure and arrive at relative weights for the purpose of consolidating these variables into a single index. (4) Xj = Retained variables Wj = Component scores (weights) j= 1, 2, 3. For determining the weights, the Joliffe criterion was used. According to this procedure, those principal variables whose component scores were the highest with respect to the retained components were retained. The process involved a seriatim selection of variables along with their scores moving from the highest score in the first component to the next components and selecting the variable with its highest score, and so on. Scale and Code It must be ensured that the index does not suffer on account of problems relating to scale and code. The problem of scale arises out of the difference in scale of the variables that were components of the composite index. In this case, the problem of the scale was handled by normalizing the variables5. As a result, variables were expressed in proportionate terms. The code of the variable refers to the interpretation of the direction of change with respect to the value or the measure of the index. For instance, a high number in the index should represent an increase in the phenomena that the composite index stands for and higher number should also be generated by the 5 All relevant variables have been expressed as a proportion of GDP. MEASUREMENT OF INTERNATIONAL CURRENCY CRISES: A PANEL DATA APPROACH ... higher value of the component of the index, which implies a rise in the phenomena. For instance, if exchange rate is expressed as a local currency unit per dollar, then a sharp devaluation or depreciation of the currency is expected during the crises. Hence, the magnitude of the variable should rise. It is therefore necessary to have consistency between the magnitude of the code and the interpretation of direction of change of the phenomenon, on the one hand, and consistency within the code in relation to the individual variables that constitute the index, on the other. Therefore, a composite index would be representative only if the components of the index are representative and both the scale and the code are consistent with the value of the index that is generated. It should yield a unique and consistent interpretation of the index. Advantage of the Composite Index To ascertain whether composite indices function better than the individual variables, a regression equation was estimated by including the principal variable directly in the regression6. The results were not satisfactory, and to the contrary, a composite index performed better. Apparently, the complexity of the phenomenon was better represented by a composite index that represented the combined information content. At the same time it reduced the number of variables and permitted higher degrees of freedom. Panel Data Analysis A methodology which allowed a combination of a continuous and discrete crisis definition, was adopted. The intention of this paper is to capture a mix of the impact of continuous independent variables and fixed effects in terms of crisis window (discrete effects), as well as country effects in the form of country dummies. This is an essential part of the research design and objective. Therefore, the need for a random effects model was not justifiable. In general, random effects models are meant for capturing generalized least square (GLS) estimators which measure the generalized effects and not the fixed effects (country effects) as has been the purpose of this study. A fixed effects panel model was used where the slope coefficients were constant, but the intercept varied over country as well as time. This study constructs a model which includes n-1 county dummies but only one time 6 dummy which represents the crises period (CW). Although there are three periods (pre-crises, during crises, or post-crises ), only one time dummy was included in the final model instead of two because the focus of this study was on crises period. The idea was to have more degrees of freedom. The recovery period was not so much of interest to the authors as the crises period. The general form of the equation that was estimated is as follows: IOCit = A +b1* (CW) t + b2d2 + b3d3 + b4d4 + b5d5 + b6d6 + b7* X1it + b8* X2it+U1it (5) IOCt = Dependent variable = Log (Index of Crisis) A = intercept b1 = Coefficient of crisis window. b2 = Intercept dummy for Thailand b3 = Intercept dummy for Philippines b4 = Intercept dummy for Korea Republic b5 = Intercept dummy for Indonesia b6 = Intercept dummy for Malaysia (CW)t = Dummy for crisis window: ‘0’ for all years & ‘1’ for 1997, 1998 b7 = Elasticity coefficient for index of macro variables IMV = Index of macro variables b8 = Elasticity coefficient for index of financial variables IFV = Index of financial variables Since the methodology is based on a comparison of crises-hit countries with a non-crises hit country (that is, the control), this objective can be met only through panel regression. Secondly, panel regression leads to more efficient estimators and the result of the panel regression supports this objective. Thirdly, the fixed effect model allows measurement of the precise impact of crises on a differential basis among the crises-ridden countries, and between the crises-ridden country and control. Apart from the above-mentioned justification for choosing the fixed effects model in line with the ‘mixed approach’ to crises definition, an important basis of choosing the fixed effects model is that different dummies were taken with India as a base country. Therefore, the very basis of this study lies in making a comparison between India which was relatively unaffected by the crises and A5 countries which were severely affected. The research design essentially lends itself to a fixed effects model with a view to capture discrete country effect on a comparative basis. This is known as Principal Variable Regression VIKALPA • VOLUME 38 • NO 4 • OCTOBER - DECEMBER 2013 19 Fixed effects models are not without their drawbacks. They may frequently have too many cross-sectional units of observations requiring too many dummy variables for their specification. Too many dummy variables may sap the model of sufficient number of degrees of freedom for adequately powerful statistical tests. Moreover, a model with many such variables may be plagued with multicollinearity, which increases the standard errors and thereby drains the model of statistical power to test parameters. If these models contain variables that do not vary within the groups, parameter estimation may be precluded. Although the model residuals are assumed to be normally distributed and homogeneous, there could easily be country-specific (group wise) heteroscedasticity or autocorrelation over time that would further plague estimation. To avoid this kind of problem, only country-specific dummies and one time dummy were used in the final modeling, as mentioned earlier. Time effect was considered as fixed. To deal with the problem of multicollinearity, the index of the impacted/ dependent variables was constructed and it was ensured that there was no correlation among those indexes. The problem of auto-correlation was taken care of by measuring the Durbin-Watson statistic which happens to be in the normal range. RESULT AND ANALYSIS This section interprets the results of various empirical procedures which were applied to the data set in order to arrive at some relevant insights and conclusions. Causality Test The Granger causality test consists of testing pairs of equations expressed below: Yt = a1 Xt-1 + a2 Xt—2 + b1 Yt—1 + b2 Yt—2 + u1t (6) Xt = c1 Xt—1 + c2 Xt—2 + d1 Yt—1 + d2 Yt—2 + u2 (7) In Equation 6, Xt = Independent variable Yt = Dependent variable Xt-1 = First lag of independent variable Xt-2 = Second lag of independent variable Yt-1 = First lag of dependent variable Yt-2 = Second lag of dependent variable a1 = Co-efficient of first lag of independent variable a2 = Co-efficient of second lag of independent variable 20 b1 = Co-efficient first lag of dependent variable b2 = Co-efficient second lag of dependent variable u1t = Error term of first equation In Equation 7, Xt = Dependent variable Yt = Independent variable Xt-1 = First lag of dependent variable Xt-2 = Second lag of dependent variable Yt-1 = First lag of independent variable Yt-2 = Second lag of independent variable c1 = Co-efficient of first lag of dependent variable c2 = Co-efficient of second lag of dependent variable d1 = Co-efficient of first lag of independent variable d2 = Co-efficient of second lag of independent variable u2t = Error term of second equation The procedure consists of constructing sets of pairs of two variables amongst which the causality is being tested for. Alternatively, each variable acts as the dependent and independent variable. The lags of both the variables are included on the swap as either dependent lagged variables or independent lagged variables. If the lags of the independent variable significantly impact the dependent variable, then the independent variable is known to have caused the dependent variable. The procedure is repeated by swapping the dependent and independent variable for testing for reverse causality. If the causality is found in both directions, then it is known as a bi-way causal relationship ; else it is known as a one-way causal relationship. Causality test was conducted on a set of 30 variables. (See Tables 1 and 2 that describe the causal macro and financial variables). There were a total of 435 combinations. Accounting for our own covariance, which was 15 (in pair) in number, there were 420 combinations. Since the procedure of testing involved testing in pair, it implies that 210 causality tests were applied. On account of transitivity, the number of combination was halved7. The variables with two-way causality were dropped As in such cases, one cannot identify which variable is to be taken as dependent variable and which variable as independent variable. The result of the causality exercise shows 19 causal variables. On the other hand, there were 16 impacted/dependent variables. (See Tables 3 and 4 that describe the impacted macro and financial variables). 7 Tested at 10% level of significance. Result not reported MEASUREMENT OF INTERNATIONAL CURRENCY CRISES: A PANEL DATA APPROACH ... Table 1: Macro Causal Variables (Code and Name of the Variable) Code Abbreviation of the Variable Name of the Variable M1 BN.RES.INCL.CD Changes in net reserves (Bop, current US$) M2 BN.CAB.XOKA.GD.ZS Current account balance (% of GDP)- CAB M3 NE.EXP.GNFS.ZS Exports of goods and services (% of GDP)-EOGD M4 NY.GDP.MKTP.KD.ZG GDP growth (annual %)- GDPG M5 NY.GDP.PCAP.KD.ZG GDP per capita growth (annual %)- GDPPC M6 NE.GDI.TOTL.ZS Gross capital formation (% of GDP)- GCF M7 NE.IMP.GNFS.ZS Imports of goods and services (% of GDP)- IOGS M10 PA.NUS.FCRF Official exchange rate (LCU per US$, period average)- OER % change over previous year M11 GB.BAL.OVRL.GD.ZS Overall budget balance, including grants (% of GDP)- OBB M13 FR.INR.RINR Real interest rate (%)- RI Table 2: Financial Causal Variables (Code and Name of the Variable) Code Abbreviation of the Variable Name of the Variable F2 FS.AST.PRVT.GD.ZS Domestic credit to private sector (% of GDP)- DCTPS F3 GB.FIN.DOMS.GD.ZS Domestic financing, total (% of GDP)- DFT F4 BX.KLT.DINV.DT.GI.ZS Foreign direct investment, net inflows (% of gross capital formation)- FDI F6 IQ.ICR.RISK.XQ ICRG composite risk rating (0=highest risk to 100=lowest)- CRR F8 FR.INR.LEND Lending interest rate (%)- LR F10 CM.MKT.LCAP.GD.ZS Market capitalization of listed companies (% of GDP)- MC F12 DT.DOD.DSTC.ZS Short-term debt (% of total external debt)- STD F13 CM.MKT.TRAD.GD.ZS Stocks traded, total value (% of GDP)- ST F15 DT.TDS.DECT.GN.ZS Total debt service (% of GNI)- TDS GNI Table 3: Impacted Macro Variables (Code and Name of the Variable) Code Abbreviation of the Variable Name of the Variable M1C BN.RES.INCL.CD Changes in net reserves (BoP, current US$)- CINR M2C BN.CAB.XOKA.GD.ZS Current account balance (% of GDP)- CAB M3C NE.EXP.GNFS.ZS Exports of goods and services (% of GDP)- EOGD M5C NY.GDP.PCAP.KD.ZG GDP per capita growth (annual %)- GDPPC M7C NE.IMP.GNFS.ZS Imports of goods and services (% of GDP)- IOGS M10C PA.NUS.FCRF Official exchange rate (LCU per US$, period average)- OER % change over previous year M13C FR.INR.RINR Real interest rate (%)- RI M9 FP.CPI.TOTL.ZG Inflation, consumer prices (annual %) Table 4: Impacted Financial Variables (Code and Name of the Variable) Code Abbreviation of the Variable Name of the Variable F2C FS.AST.PRVT.GD.ZS Domestic credit to private sector (% of GDP)- DCTPS F4C BX.KLT.DINV.DT.GI.ZS Foreign direct investment, net inflows (% of gross capital formation)- FDI F5C BG.KAC.FNEI.GD.ZS Gross private capital flows (% of GDP)- GCF F6C IQ.ICR.RISK.XQ ICRG composite risk rating (0=highest risk to 100=lowest)- CRR F8C FR.INR.LEND Lending interest rate (%)- LR F12C DT.DOD.DSTC.ZS Short-term debt (% of total external debt)- STD F13C CM.MKT.TRAD.GD.ZS Stocks traded, total value (% of GDP)- ST F15C DT.TDS.DECT.GN.ZS Total debt service (% of GNI)- TDS GNI VIKALPA • VOLUME 38 • NO 4 • OCTOBER - DECEMBER 2013 21 The common variables which occurred both as a cause and impact variable are shown in Table 58. In all, there were 15 common variables. Thus out of 16 impacted variables, only one variable was left as pure impacted variable, that is, M9 (FP.CPI.TOTL.ZG–Inflation, consumer prices (annual %)). Correlation Analysis Out of the 15 common variables, which ones would be retained as impacted variables and form a part of the index of crisis was sorted out through correlation analysis. First, correlation among the pure impacted variable M 9, inflation, consumer prices (annual %) and the 15 common variables was calculated9. Out of the list of 15 common variables, only 14 were retained. Variable F15 - Total debt service (% of GNI) was dropped because of non-availability of data in case of some countries. The list of 15 common variables are given in Table 5. Common variables that were correlated with the single impacted variable M9 inflation, consumer prices (annual %) were retained as impacted/dependent variable. One can argue that correlation could be high in case of impacted variable, since a composite index was constructed out of it with the help of PCA that ensured that the correlation was removed. After applying correlation analysis, the variables which were found to be highly correlated with the pure impacted variables have been reported in Table 6. Formation of Index of Crises PCA was applied on impacted variables. The final procedure for the formation of the index involved the following steps: 1. Determination of number of principal components to be retained. For this, the Kaiser criteria was used and three principal components where eigen value was greater than one, were retained. Table 7(a) shows the total variance explained by the extracted principal component. It is evident that over 72 percent of the information is captured by the retained component. 2. Rotation of components: With the help of Varimax rotation with Kaiser normalization, the components were rotated. This was done with a view to obtain the clear interpretation of the components. This resulted in a set of component scores for each of the nine variables with respect to the three retained components. Table 7(b) reports the component scores coefficient matrix. 3. Selection of principal variables: The Joliffe procedure (explained earlier) was used to select the principal variables. Three variables were selected in the de- Table 5: List of Common Variables Code Abbreviation of the Variable Name of the Variable M1C BN.RES.INCL.CD Changes in net reserves (BoP, current US$)- CINR M2C BN.CAB.XOKA.GD.ZS Current account balance (% of GDP)- CAB M3C NE.EXP.GNFS.ZS Exports of goods and services (% of GDP)- EOGD M5C NY.GDP.PCAP.KD.ZG GDP per capita growth (annual %)- GDPPC M7C NE.IMP.GNFS.ZS Imports of goods and services (% of GDP)- IOGS M10C PA.NUS.FCRF Official exchange rate (LCU per US$, period average)- OER % change over previous year M13C FR.INR.RINR Real interest rate (%)- RI F2C FS.AST.PRVT.GD.ZS Domestic credit to private sector (% of GDP)- DCTPS F4C BX.KLT.DINV.DT.GI.ZS Foreign direct investment, net inflows (% of gross capital formation)- FDI F5C BG.KAC.FNEI.GD.ZS Gross private capital flows (% of GDP)- GCF F6C IQ.ICR.RISK.XQ ICRG composite risk rating (0=highest risk to 100=lowest)- CRR F8C FR.INR.LEND Lending interest rate (%)- LR F12C DT.DOD.DSTC.ZS Short-term debt (% of total external debt)- STD F13C CM.MKT.TRAD.GD.ZS Stocks traded, total value (% of GDP)- ST *F15C DT.TDS.DECT.GN.ZS Total debt service (% of GNI)- TDS GNI * Variable which was dropped due to non-availability of data in case of some countries. 8 Marked as C in Table 5. 9 By using SPSS 15 22 MEASUREMENT OF INTERNATIONAL CURRENCY CRISES: A PANEL DATA APPROACH ... Table 6: Variables Significantly Correlated with Variable M9 Code Abbreviation of the Variable Name of the Variable M13 FR.INR.RINR Real interest rate (%) M1 BN.RES.INCL.CD Changes in net reserves (Bop, current US$) M3 NE.EXP.GNFS.ZS Exports of goods and services (% of GDP) M2 BN.CAB.XOKA.GD.ZS Current account balance (% of GDP) M7 NE.IMP.GNFS.ZS Imports of goods and services (% of GDP) M10 PA.NUS.FCRF Official exchange rate (LCU per US$, period average) % change over previous year F2 FS.AST.PRVT.GD.ZS Domestic credit to private sector (% of GDP) F8 FR.INR.LEND Lending interest rate (%) *M9 FP.CPI.TOTL.ZG Inflation, consumer prices (annual %) Table 7(a): Total Variance Explained Extraction Sums of Squared Loadings Rotation Sums of Squared Loadings Component Total % of Variance Cumulative % Total % of Variance Cumulative % 1 3.143 34.921 34.921 3.013 33.482 33.482 2 2.352 26.137 61.058 2.300 25.553 59.035 3 1.052 11.687 72.745 1.234 13.709 72.745 Extraction Method: Principal Component Analysis Table 7(b): Component Score Coefficient Matrix Component 1 2 3 CINRM1 -0.055 -0.072 0.366 CABM2 0.236 -0.002 0.485 EOGDM3 0.315 0.022 -0.019 IOGSM7 0.297 0.021 -0.062 ICPM9 OERM10 RIM13 -0.098 0.361 0.048 0.009 0.402 -0.030 -0.053 -0.384 0.147 DCTPSF2 0.299 -0.051 0.104 LRF8 0.069 -0.044 0.694 Extraction Method: Principal Component Analysis. Rotation Method: Varimax with Kaiser Normalization Table 7(c): Correlations EOGDM3 EOGDM3 OERM10 LRF8 Pearson Correlation 1 0.048 -0.268* Sig. (2-tailed) . 0.654 0.011 N OERM10 90 90 90 Pearson Correlation 0.048 1 0.094 Sig. (2-tailed) 0.654 . 0.379 N LRF8 Pearson Correlation Sig. (2-tailed) N 90 90 90 -0.268* 0.094 1 0.011 0.379 . 90 90 90 * Correlation is significant at the 0.05 level (2-tailed). VIKALPA • VOLUME 38 • NO 4 • OCTOBER - DECEMBER 2013 23 scending order beginning with the largest component. Accordingly, the three principal variables selected were; M 3 Exports of Goods and Services (% of GDP), M 10 Official Exchange Rate (LCU per US$, period average, % change over the previous year), and F 8 Lending Interest Rate (%). 4. Construction of index of dependent variable. By using the weights from the component score coefficient matrix (given in step 3 of the PCA analysis), the index of dependent variables was constructed. The composite index of impacted variables (Y variable/the LHS variable) was calculated by multiplying the variables with their respective weights. (8) The code of the variable refers to the value and direction of each included variable in relation to the value and direction of the index. IOC measures the crises; therefore, a higher value of the index should represent a higher degree of crises. Percentage change in the official exchange rate over the previous year10 is expressed as LCU/$; therefore, a rise in its value would represent depreciation of domestic currency. In effect, a higher value implies an increase in the degree of crises. With depreciation, it may be expected that value of exports of goods and services as a percentage of GDP would increase which would also add to the value of the crises index. Similarly, a higher lending interest rate could also be expected during the period of crises. Thus, all the three variables conform to the desired code of IOC. That is, they rise in value term when the crises increase; so also does the index of crises. Thus, the code of the components of the index and the crises index share the same interpretation. During the crises, in general, there would be a tendency of inflation to rise. Secondly, there could be a speculative bubble; therefore, after the monetary authority resorts to tight money policy, the interest rate is likely to increase. There was a dual purpose for adopting this elaborate procedure. Firstly, it was aimed at developing a composite index. Secondly, it was important to ensure that a correlation amongst retained variables was minimized which is a 10 Since we had to take % change over previous year for normalization of the variable, we had to forego one data point, that is, 1987. The final period therefore is 1988 to 2002. 24 merit of PCA methodology. After having constructed an index, it was necessary to verify the degree of correlation. Table 7(c) shows that the correlations amongst the retained principal variables that have been used for constructing an index were low and not statistically significant. Index of Macro Variables In the second stage of creating of composite index, the index of causal macro variable was calculated. We had a combined list of 10 causal variables, for both financial and macro variables (Table 8). Table 8: Combined List of Causal Variables Code Abbreviation of the Variable Name of the Variable M4 NY.GDP.MKTP.KD.ZG GDP growth (annual %) M6 NE.GDI.TOTL.ZS Gross capital formation (% of GDP) M11 GB.BAL.OVRL.GD.ZS Overall budget balance, including grants (% of GDP) F3 GB.FIN.DOMS.GD.ZS Domestic financing, total (% of GDP) F10 CM.MKT.LCAP.GD.ZS Market capitalization of listed companies (% of GDP) M5 NY.GDP.PCAP.KD.ZG GDP per capita growth (annual %) F13 CM.MKT.TRAD.GD.ZS Stocks traded, total value (% of GDP) F4 BX.KLT.DINV.DT.GI.ZS Foreign direct investment, net inflows (% of gross capital formation) F6 IQ.ICR.RISK.XQ ICRG composite risk rating (0=highest risk to 100=lowest) F12 DT.DOD.DSTC.ZS Short-term debt (% of total external debt) In the first place, out of the four macro variables, GDP per capita growth was dropped on account of data redundancy. To make a composite index, PCA was applied on the list of macro causal variables. The results of PCA are shown in Tables 9(a) – (c). The Kaiser criterion was applied for choosing components whose eigen value was greater than one. Accordingly, three components were retained and the total variance explained by these components was 100. MEASUREMENT OF INTERNATIONAL CURRENCY CRISES: A PANEL DATA APPROACH ... Table 9(b) shows the rotated component score coefficient of the three macro independent variables that were being sought to collapse into an index. The coefficients show the factor loading of these three variables with respect to the three retained principal components. Since all the three principal variables were being retained, the component score matrix was used to identify the weights of each of these variables. Joliffe procedure (explained earlier) was used to select the principal variables. (9) The index of macro variables was calculated as the value of the variables multiplied by its weights which were reported in the component score coefficient matrix. The selected variables are, M4 NY.GDP.MKTP.KD.ZG—GDP growth (annual %), M6 NE.GDI.TOTL.ZS—Gross capital formation (% of GDP), and variable M11 GB.BAL. OVRL.GD.ZS — Overall budget balance, including grants (% of GDP). The final step in the procedure was to examine the correlation matrix because the study was dealing with inter- correlated variables that were sorted through a long process including causality tests. Presented in Table 9(c) is the correlation matrix of the components of the index of macro variables. M11 and M4 seem to have a high correlation, but as it is statistically not significant, it would not cause mulicollinearity of any serious order. The code of the variable refers to the interpretation of the value of the direction of change with respect to the value of the measure of the index of macro variables. As the growth rate of GDP and the Gross Capital Formation decline, the crises can be expected to increase. When the overall budget balance (OBB) increases under conditions of deficit budget, OBB would fall further and would lead to a crisis. Hence, a negative relationship among the variables similar to the macro variable index was expected. Index of Financial Variables In the third stage of creating of composite index, the index of causal financial variable was calculated. To make a composite index, PCA was applied on the list of financial causal variables. The results of PCA are given in Tables 10 (a)-(c). Table 9(a): Total Variance Explained Extraction Sums of Squared Loadings Rotation Sums of Squared Loadings Component Total % of Variance Cumulative % Total % of Variance Cumulative % 1 2.156 71.869 71.869 1.186 39.531 39.531 2 0.700 23.339 95.207 1.048 34.919 74.450 3 0.144 4.793 100.00 0.766 25.550 100.00 Extraction Method: Principal Component Analysis Table 9(b): Component Score Coefficient Matrix Component 1 GDPGM4 -0.684 GCFM6 OBBM11 2 3 -0.254 1.873 -0.036 1.108 -0.380 1.507 -0.024 -1.079 Extraction Method: Principal Component Analysis; Rotation Method: Varimax with Kaiser Normalization Table 9(c): Correlation Matrix EOGDM3 Correlation OERM10 LRF8 GDPGM4 1.000 0.504 0.838 GCFM6 0.504 1.000 0.351 OBBM11 0.838 0.351 1.000 VIKALPA • VOLUME 38 • NO 4 • OCTOBER - DECEMBER 2013 25 Table 10(a): Total Variance Explained Component Initial Eigen Values Extraction Sums of Squared Loadings Rotation Sums of Squared Loadings Total % of Variance Cumulative % Total % of Variance Cumulative % Total % of Variance Cumulative % 1 1.828 45.706 45.706 1.828 45.706 45.706 1.535 38.384 38.384 2 1.101 27.514 73.220 1.101 27.514 73.220 1.203 30.082 68.465 3 0.817 20.429 93.650 0.817 20.429 93.650 1.007 25.184 93.650 4 0.254 6.350 100.00 Extraction Method: Principal Component Analysis. Table 10(b): Component Score Coefficient Matrix Component 1 F3 -0.047 2 3 0.122 1.025 F6_MOD 0.678 0.274 -0.014 F13 0.144 0.864 0.092 F10 -0.460 0.294 0.056 Extraction Method: Principal Component Analysis; Rotation Method: Varimax with Kaiser Normalization Table 10(c): Correlation Matrix Correlation F3 F6_MOD F13 F3 1.000 0.070 -0.224 -0.198 F6_MOD 0.070 1.000 0.011 -0.596 F13 -0.224 0.011 1.000 0.440 F10 -0.198 -0.596 0.440 1.000 The criterion for choosing three components was applied. By the Kaiser criterion, the coverage was not adequate (68.465). Accordingly, three components were retained and the total variance explained by these components was 93.650. Table 10(b) shows the rotated component score coefficient of the four financial independent variables that were being sought to collapse into an index. The coefficients show the factor loading of these four variables with respect to the three retained principal component. The Joliffe criterion was used to retain the three principal variables. The component score matrix identifies the weights of each of these variables. (10) 26 F10 The index of financial variables was calculated as the value of the variables multiplied by its weights which were reported in the component score coefficient matrix. The three retained variables were: F3 - GB.FIN.DOMS.GD, ZS—Domestic financing (% of GDP), F6 - IQ.ICR.RISK. XQ—ICRG composite risk rating (0=higher risk to 100=lowest risk) and F13 - CM.MKT.TRAD.GD.ZS— Stocks traded, total value (% of GDP). These were multiplied by the respective scores. As in the case of macro variables, the final step in the procedure was to examine the correlation matrix because the study was dealing with inter-correlated variables that were sorted through a long process including causality test. Table 10(c) presents the correlation matrix of the component of the index of financial variables. There was some moderate correlation between F10 and F6_MOD; however, none of the correlation was statistically significant. Therefore, there was no problem of mulicollinearity. MEASUREMENT OF INTERNATIONAL CURRENCY CRISES: A PANEL DATA APPROACH ... The code of the variable refers to the interpretation of the value of the direction of change with respect to the value of the measure of the index of financial variable. The total value of stocks traded is an indicator of the capital market activity. The total value would be influenced by the turnover as well as the value of assets traded. When the markets are on a high, both these aspects are likely to be high. There may be a high degree of overvaluation of shares. During a crisis, the turnover may also be high. And, if the fundamentals of the economy are not strong, such an excessive bubble could lead to crises. Therefore, it may be expected that an increase in the value of stocks traded would promote crises. MODEL SPECIFICATION AND ESTIMATION Domestic financing to private sector is to be interpreted as the availability of excess credit. In the event that this credit is feeding the asset bubble, it would result in greater volatility and an inflation of asset prices that is unrelated to the fundamental of the economy. This would therefore be a major influence in precipitating crises. A model was constructed which included n-1 countries dummies and only one time dummy which represented the crises period. Difference dummy was used for the intercept dummy in the final model. Although there were three periods (pre-crises, crises, or post–crises), only one time dummy was used in the final model instead of two because the study was interested only in the crises period. More degree of freedom was desirable. Such a model is represented by equation 11: The third variable which was included in the index of financial variable was (F6) ICRG - composite risk rating (0=highest risk to 100=lowest risk). The code of this variable was running contrary to the direction of the index because a lower value represents a higher risk. The final form of this variable, which was a risk indicator, had to bear a positive relationship with the index of crises. Therefore, the variable was modified by defining it as the complement of the original variable F6. This was achieved by subtracting the value of each data point from 100. We tested the modified F6 variable by applying PCA and found that the result did not change either in terms of variable selection form principal component or as component scores. The advantage therefore is that the index of financial variable would now consist of all the three variables that move in the same direction. Thus, a higher value of the index clearly defines a greater level of the factors that belong to the financial index. In this sense, when the modified F6 is larger, the risk would be larger and hence the potential of crises is larger. It may also be recalled that higher risk implies greater volatility which is the single most important factor responsible for crises. On the basis of the above consideration, we expect that the index of financial variables would have a positive relationship with the index of crises. VIKALPA • VOLUME 38 • NO 4 • OCTOBER - DECEMBER 2013 Once it was clear which variable was to be taken as a dependent variable and which other variables were to be taken as independent variable, the choice was between the appropriate model, functional form, and estimation technique. As far as model specification is concerned, crisis was considered as a function of 16 macro variables, 14 financial variables, and a set of exogenous variables. Linear functional form was adopted for the model. The log linear form could not be taken as some data points were negative and some were in percentage terms. LIOCt = A +b1 * (CW) t + b2d2 + b3d3 + b4d4 + b5d5 + b6d6 +b7 * LIDXFVit + b8* LIDXMVit +U1it (11) IOCt = Dependent variable = Log (Index of Crisis) A = intercept b1= Coefficient of crisis window b2 = Intercept dummy for Thailand b3 = Intercept dummy for Philippines b4 = Intercept dummy for Korea Republic b5 = Intercept dummy for Indonesia b6 = Intercept dummy for Malaysia (CW)t = Dummy for crisis window: ‘0’ for all years & ‘1’ for 1997, 1998 b7 = Elasticity coefficient for index of macro variables IMV = Index of macro variables b8 = Elasticity coefficient for index of financial variables IFV = Index of financial variables U1it = Error term i = 1 to 6 by country t = 1988 to 2002 Determinants of Crises The main purpose of the foregoing analyses, discussion on methods, and model specification is to estimate a structured model which seeks to explain currency crises along 27 with individual differences among countries and other ‘fixed effects’ through a set of explanatory variables. These explanatory variables are the determinants of crises in the three policy periods provided that their parameters are found to be statistically significant. For this purpose, the following fixed effects panel regression was estimated. The results are reported in Table 11. The result of Equation 11, the final model is reported as follows: LIOCit = A + 0.36 (CW) t + 0.51 (d2) + 0.41 (d3) + 1.37 (d4) + 0.54 (d5) + 0.98 (d6) + -0.20 * LIDXFVit + (0.26) * LIDXMVit + U1it (12) (12) In the estimated double log model, financial variables index is significant at 10 percent which is an acceptable threshold. This is presumably so due to higher volatility of financial variables; therefore, standard error is high. So far, as a determinant, the impact of financial variables on the crisis variables is less definite in comparison to the macro index. The value of R square is fairly high at 0.72. However, in general, the index of financial variables has a smaller elasticity of 0.21 approximately, but it is posi- tive which indicates that according to our continuous definition of crises, financial variables are responsible for escalating the crisis index. The positive signs clearly show that financial variables are responsible for enhancing the crises on a continuous basis. The index of macroeconomic variables, on the other hand, has slightly larger magnitude of 0.255. However, macroeconomic variables index bears a negative sign. This implies that macro variables in general reduce the magnitude of crises. The macro variables therefore stabilize the economy and hence reduce the intensity of crisis. Yet, it needs to be pointed out that both the indices have elasticities that are less than one. This has a significant bearing on our measurement and analysis of crises. In terms of magnitude, the intercept is large and significant. This implies that there is a large domain of unknown variables which have been omitted whose influences on crises are substantial and significant. This means that our understanding and measurement of crises has a large element of uncertainty. It must be remembered that the intercept contributes to the determination of the predicted variable-index of crisis. Table 11: Final Model Regression Statistics Multiple R 0.85231 R Square 0.726433 Adjusted R Square 0.699414 Standard Error 0.229639 Observations 90 ANOVA df 28 SS MS F Significance F 26.88602 8.00433E-20 Regression 8 11.34244 1.417805 Residual 81 4.271447 0.052734 Total 89 15.61389 Coefficients Standard Error t Stat P-value Intercept 2.872204 0.568448 5.052712 2.64E-06 CW 0.362984 0.073842 4.915711 4.55E-06 1.437612841 D2 0.516381 0.096763 5.336526 8.43E-07 1.675950899 D3 0.406131 0.084158 4.825837 6.46E-06 1.500999912 D4 1.374522 0.177197 7.757031 2.26E-11 3.953185275 D5 0.549734 0.085405 6.436757 8.05E-09 1.732792008 D6 0.980138 0.103953 9.42863 1.13E-14 2.664823967 LIDXFV 0.209432 0.124327 1.684531 0.095929 1.232977862 LIDXMV -0.25518 0.07065 -3.61194 0.000526 0.774774645 17.67593008 MEASUREMENT OF INTERNATIONAL CURRENCY CRISES: A PANEL DATA APPROACH ... Secondly, the country differentials such as culture, policy, historical circumstances, etc. are all captured by the country dummy. Since all these are significant and positive, it implies that country differential plays an important role in determining crisis. It is difficult to further segregate these effects that are country-specific. However, the methodology adopted in the paper takes into account the information content in the large set of variables. It is thus most comprehensive in terms of its explanatory power and its potential for measurement. There were various models for the estimation of panel data. This study used fixed effects approach in which the slope coefficients were constant but the intercept varied across countries, that is, fixed effects or least square dummy variable (LSDV) regression model. The fixed effects arise due to the special features of different countries such as economic structure, political condition, conditions of macro and financial fundamentals, countries’ linkages with the rest of the world in terms of the trade, capital market, and money market linkages. The term fixed effect denotes that although the intercept may differ across countries (six in the present study), each individual country’s intercept does not vary over time; that is, it is time invariant. It may be noted that fixed effect model assumes that the slope coefficient of the independent variables such as index of macro and financial variables do not vary across countries or over time. D2 =1, if the observation belongs to Thailand, 0, otherwise and so on for other countries, seriatum. Since the study considered six countries, only five dummies were used to avoid falling into the dummy variable trap (that is, the situation of perfect co-linearity). There was no dummy for India. In other words, the overall intercept (A) represents the intercept of India and A+d2 represents the differential intercept coefficient of Thailand and so on. It tells by how much the intercept of the rest of the countries differ from the intercept of India, since India is the base country. Time effect was not included in the model. Hence, the crises function does not shift over time. There is only one dummy to represent the crisis window which takes the value ‘0’ for all years except 1997 and 1998 where it takes the value ‘1’. One of the main findings of the estimation confirms our conception of crises and vindicates our understanding that the crises definition needs to be built into the causal variables. Accordingly, we had introduced a dummy variVIKALPA • VOLUME 38 • NO 4 • OCTOBER - DECEMBER 2013 able to capture a discrete change in the index of crises. This variable was termed as CW. It basically implies a discrete jump in the index of crises during the period of crises. Therefore, while the continuous explanatory variable explains the buildup to the crises, it is the set of discrete intercept dummies and the CW variable that capture the discrete aspects of crises. The country dummies enable international comparison while CW captures intertemporal comparison. It is therefore clear that the crises phenomenon is a combination of continuous as well as discrete change. The continuous effect arises on account of a set of variables incorporated in the two indices. The speciality of our approach is that it captures the genesis of crises which is the volatility in the causal and dependent variables. This has been achieved by constructing the variables through PCA methodology which maximizes the variance. Therefore, the interpretation of the beta coefficients of the continuous variables is that it shows the relationship between the volatility of causal variables and the effects on the dependent variable in the form of volatility in the crises variables. The crises window was highly significant; hence the definition of the crises window is correct as it captures the crises as is clear from the p value. All the coefficients are significant. As far as the individual country’s shocks are concerned, that may be due to omitted variables which may be important in respect of that country. For example, in case of Korea, withdrawal of short-term debt was one of the major problems. Likewise in case of Thailand, depletion of foreign reserves was the major problem. Korea was most affected by the crises in this analysis as it is clear from the intercept of D4 which is 21.629. It turns out to be an economy that has undergone severe shock. The p value in case of Korea was very low; it means that the level of significance was high. It could be due to the size of the economy — among all the A5 countries, it was Korea which was giant in size. In fact, it was the world’s 11th largest economy in terms of GDP at that time. Another possible explanation of this might be that all the crises-hit countries, except Korea, had abounded their earlier monetary system; if they were on the peg system like Thailand, they would have announced a managed float, like the Philippines which allows its peso to move in a wider range against the dollar and Indonesia which allows its rupiah to float. In November of 1997, it was 29 only Korea which was on the existing exchange rate system. Moreover, by that time all the rest of the countries had approached the International Monetary Fund for technical as well as financial assistance or International Monetary Fund itself had offered a financial support package to the crises-hit countries. For instance, in July 1997, IMF had offered $1.1 billion as financial support to the Phillipines. IMF had also approved $16.7 billion credit for Thailand. Korea had not approached it until the first week of December 1997. It tried to handle the problem on its own, because the financial assistance from IMF was coming with a cost of rising interest rate. The Korean economy could not have sustained it since it was already affected by the banking crises in January of 1997. Hanbo Steel, a large Korean Chaebol, collapsed under a $6 billion debt load; it was the first Korean conglomerate in a decade. In November, the sharp rise of dollar against the Japanese’s yen in global trade boosted the currency against won in Korea. In South Korea, sentiments were negative; media report in the Western press also stated that the South Korean economic crisis was set to get worse. The Korean government defended itself against the foreign press report which suggested that the country’s financial turmoil could surpass the recent market meltdowns in Thailand, Indonesia, and the Philippines. The reports in major international publication also reported that South Korea’s foreign reserves were dwindling and that the country might need assistance from the IMF. South Korea’s woes began when its trade deficits ballooned in the year 1996 and the value of its currency slipped after the South East Asian market shock waves reached the country. That came at the time when the market was saddled with billions of dollar of bad loans in the wake of a series of major bankruptcies. The international rating agencies downgraded Korea’s foreign debt. Interest rates soared and the stock market plummeted 28 percent in the year 1997 to reach a five-year low at the end of October 1997. South Korea was the world’s 11th largest economy larger than Thailand, Indonesia, and Malaysia put together and the prospects of the financial crises had put everyone from the Japanese exporters to investors in Latin America on edge. Over 50 percent of the Korean exports compete directly with Japanese products. The won’s move against the dollar often mirrors the yen trading movements versus the US currency. As local exporters tried to make their products more attractively priced compared with the products of their Japanese rivals, any 30 positive effect of won’s depreciation was effectively cancelled out by the yen decline vis-à-vis the US dollar. Korea in its own way was right in not abandoning its current system; for doing this, they had to completely liberalize the market. It was too early to do so in view of their economic status and in view of the fact that won was not fully convertible. By mid-November 1997, the uncertainty surrounding Korea had pressurized the regional currencies, the hardest hit being the Thai baht, the Philippine peso, the Malaysian ringgit, and the Indonesian rupiah. According to one view, the Korean currency turmoil was unlikely to stop at its own borders. If the won depreciated, the yen was likely to depreciate as well. An IMF bailout was a humiliating necessity for South Korea, proud of its meteoric rise from a war-torn nation to the world’s 11th largest economic powerhouse. IMF bailout required stringent economic reforms and policy controls. We can say that South Korea was the latest wounded tiger to suffer heavy losses, its economic plight and delays in signing a loan agreement with the IMF being the reasons for the same. In the case of Indonesia, the intercept was higher (19.408) than the base country (India 17.67). The p value was also very low implying that the level of significance was high. Indonesia abolished its system of managing the exchange rate through the use of band and allowed it to float. IMF gave Indonesia $23 billion financial support package. The second most affected country by our analysis was Malaysia whose p value was quite low. It is also clear from the intercept value of D6. The Malaysian Central Bank restricted loans to property and stocks to head off a crisis. In case of Thailand, the intercept was higher than the base country. The Thai baht was hit by the attack by speculators who considered Thailand’s slowing economy and political instability as the right time to sell. The move to save Finance One, Thailand largest finance company, failed. The Thai Central Bank suspended operations of 16 cash-strapped finance companies and ordered them to submit merger or consolidation plan. On July 2 1997, the Bank of Thailand announced a managed float of the baht and called IMF for technical assistance. The announcement effectively devalued the baht by about 15-20 percent. This was triggered by the East Asian Crises. Our analysis clearly reveals that it was not most hit by the crises. In case of the Philippines, the intercept was higher than the base country indicating that although it was affected by the crises, the impact was not severe. The cen- MEASUREMENT OF INTERNATIONAL CURRENCY CRISES: A PANEL DATA APPROACH ... tral bank raised the overnight rate by 1.34 percent to 13 percent and again to 24 percent thus dumping the dollar. The Philippines Central Bank allowed the peso to move in a wider range against the dollar. The IMF backed the move and the board approved the Extended Fund Facility. Hence the effect was not distinct. It was low and uncertain. In summary, the main purpose of the regression was to identify a set of dependent variables that formed the index of crises and a set of independent variables which formed the index of financial variables as well as the index of macro variables. Table 12 summarizes the cause and effect relationship that explains the crises. The Table does not indicate the countries’ dummies and the time dummy.11 From the main model, the predicted value of the index of crises was estimated. The first observation is that during the crises window, all countries, including India, have been clearly affected. There is evidence of a discrete jump in the predicted index across countries. However, it can be seen that the impact on India was the minimal. The highest index was that of Indonesia which stood at 60.83157, while India, with 23, was the lowest. Most of the countries during the crises were in the range of 60’s. The maximum rise was in the case of Indonesia that was around 24 points. In Korea, on an average, the jump was just about one point. Similarly, in the case of Thailand and the Philippines, the appreciation was 11 and 7 points respectively. During the recovery phase, the patterns were more stable. In the case of India, there was a decline down to 40 percent and the recovery was almost complete except for a marginal overall rise in comparison to the pre-crises period. Both the Philippines and Thailand experienced a halving of the index after crises and a mild decline towards pre-crises levels in the next three years. In the case of Korea, while the dip in the index was down by one– third, there was a marginal rise and a stable trend which resembled that of the late 1980s. In both Malaysia and Indonesia, the decline was less than half and there was a mild tendency towards a falling index which approximated their state at the end of 1980s and beginning of 1990s (Table 12). This pattern is depicted in Figure 1. CONCLUDING REMARKS The first objective of this study was to identify the causal and impacted variables and the second was to sort them into LHS and RHS variables for which correlation analysis was used. Both the objectives were duly met. The third objective – i.e. to form indices – was met through PCA Table 12: Predicted Index of Crisis Year Thailand Philippines Korea Indonesia Malaysia India 1988 21.55736 25.96764 41.03978 27.37488 38.32451 16.36509 1989 20.6508 25.18948 40.64091 25.83559 36.86196 17.32047 1990 20.30497 26.81612 41.77969 25.84149 33.37887 17.91071 1991 20.58842 29.20521 42.27586 23.13533 32.34299 19.33262 1992 20.54555 25.74871 42.23782 23.5756 32.17903 16.69407 1993 20.20048 23.58768 42.29674 23.35578 30.22425 16.90031 1994 20.09617 21.84762 42.81112 23.01697 29.74991 15.10235 1995 19.43599 22.28141 42.6515 22.59393 29.65968 14.57693 1996 19.2316 21.15191 41.89849 23.00288 29.59737 15.15495 1997 37.09209 30.72588 60.01316 36.47265 46.58788 23.03583 1998 40.58957 37.11995 60.62232 60.83157 66.57738 23.11819 1999 29.18055 24.66319 41.00381 36.3185 36.59223 15.51987 2000 24.4165 23.74109 41.58963 30.83911 34.28066 16.73032 2001 25.44351 24.47137 41.43097 30.0384 39.19026 15.98588 2002 23.65691 24.39335 41.4666 30.89146 36.50731 15.9514 11 We also tested the model in which we first considered the interaction between time and crises window. We also tested the equation in which we considered individuals variables and not the indexes. The results were not significant. VIKALPA • VOLUME 38 • NO 4 • OCTOBER - DECEMBER 2013 31 Figure 1: Predicted Indices of Crisis Prediction of Crisis 70 60 50 Thailand Philippines IOC 40 Korea Indonesia 30 Malaysia 20 India 10 data reduction and weighting procedure. This was the crux of the methodology. The indices captured the volatility of the underlying variables, successfully summarizing the information of the macro and financial causal variables, while the model captured the effect of these causal variables on the crises phenomenon. The next objective was to specify a model which reflected various aspects of the problem, namely crises definition (continuous and discrete), crises window, individual country effect, inter-temporal and international comparison with respect to the benchmark country and better estimators. This was achieved with the help of a panel data fixed effects model in difference form. This confirms our notion of the crises phenomenon as follows: • The crises develops continuously on account of volatility of variables. • The crises period can be captured and explained with the help of certain discrete effects. • The crises can be understood with reference to a base country. • A set of variables, both financial and macro, acts as a composite index and determines crises in a continuous way. Various other models which were tried out only reaffirmed our conceptualization and methods of measurement of crises because those models could not perform as well as 32 2002 2001 2000 1999 1998 1997 1996 1995 1994 1993 1992 1991 1990 1989 1988 0 our chosen model. In the process, the problem of multicollinearity could be overcome, although the effect of a large number of interrelated variables was incorporated. It has been found that the index of crises consists of exports, exchange rate, and interest rate. The financial index contains risk rating, domestic financing, and stock traded. The index of macro variables is based on GDP, capital formation, and budget balance. Macro index negatively influences crisis and financial index influences crisis positively. The decomposition effects show that the major influence is that of financial variables amongst which stock traded and domestic financing are the most important causes of crisis. Amongst the macro variables, GDP growth is the major influence. The main contribution of this paper lies in developing an appropriate methodology that is capable of explaining such a complex phenomenon as crisis in terms of two indices of macro-economic and financial variables. In this methodological paper, India is treated as the base country. This shows that this methodology is capable of making a comparative analysis between those countries that are directly affected by crisis and, those like India, which are indirectly affected. In the present context, this means that a similar study can be done by taking India as a base and comparing it with some of the developed econo- MEASUREMENT OF INTERNATIONAL CURRENCY CRISES: A PANEL DATA APPROACH ... mies like US, the UK, and the EU. Secondly, the PCA along with Granger causality would help in identifying causal and impacted variables. Also, the methodology for composite index formation can be applied for summarizing a larger number of macro-economic and financial variables which were affected during the global financial crises. The concept of innovatively combining ‘discrete and continuous crises definitions’ in this paper can also be applied successfully to the current international crises. Hence, this paper can easily be extended to the present financial crises. Appendix: Summary of Literature Review Study Kaminsky, Lizondo, & Reinhart (KLR) (1998) Dataset Monthly data for 20 countries, 1970-1995 (Berg & Pattillo or BP, 1999) Crisis Definition Weighted average of ER changes and reserve changes; threshold is mean +3* standard deviation; separate treatment for high inflation countries. Exclusion Window None Indicators 15 indicators capturing external balance, monetary factors, output and equity movements Approach “Signals” set thresholds for indicators to minimize noise-to-signal ratio Results RER (deviation from deterministic trend), M2/reserves “excess” M1 growth, reserves, exports, domestic credit, M2 multiplier (all % changes) good predictors Asian Crisis Results None Notes BP (1999) suggest using level of M2/reserves Study BP (1999) Frankel & Rose (1996) Dataset KLR (1998); 5 European economies, + 8 emerging markets; 1970-April 1995 (24months before Thai crisis) Annual data from 1971-1992 for 105 countries Crisis Definition Same as KLR (1998) ER change > 25% and which are at least 10% higher than the last ER change; (to allow for high inflation countries) Exclusion Window None Yes; 3 years Indicators KLR (1998) + 2 more indicators (M2) (level) Reserves and CA/GDP) binary base and weighted sum a la Kaminsky (1998) 3 macroeconomic, 4 external balance, 2 foreign and 7 composition of capital flows Approach KLR (1998) signalling and probit model Probit Results Signalling approach produces similar results as KLR (1998). The 2 new indicators noted above are informative. Probit model shows that the probability of crises increases when the following variable exceeds the thresholds: real exchange rate deviation, the CA, reserve growth, export growth, level and growth rates of M2/reserves Output growth, DC growth, foreign interest rate, FDI/ Total debt significant; reserves, RER somewhat significant? Asian Crisis Results Both approaches perform significantly better than pure guessing in terms of out-of-sample prediction of the crisis in 1997 Not good, according to BP Notes Probit models seem to have better out-ofsample predictability than KLR Many variables were unavailable for 1996, even as of mid-1998 Study Eichengreen, Rose, & Wyplosz (1996) Caramazza, Ricci, & Salgado (2000) Dataset Quarterly data for 20 industrial countries from 1959-1993 41 emerging markets and 20 industrial economies (covers Mexican, Asian, and Russian Crises) (19901998) Crisis Definition Weighted average of ER changes, reserve changes and interest rate changes; threshold is mean +1.5* standard deviation An index of speculative market pressure = weighted average of detrended monthly exchange rate changes and reserve changes. The weights are chosen so that the conditional variance of the two components of the index is equal, and the trends are country-specific. Several thresholds are used. VIKALPA • VOLUME 38 • NO 4 • OCTOBER - DECEMBER 2013 33 Alternative indices that include other financial market variables are also used Exclusion Window Yes; one quarter None Indicators Monetary, fiscal and external balance variables, equity prices, labor market variables, and contagion variables Real variables (real GDP growth, average 3-year real GDP growth, etc.); Monetary variables (12-month inflation rate, real interest rate, etc.); Fiscal variables (3-year log change of 12 month average of foreign reserves, etc.); Financial variables (short-term share of debt to BIS banks, etc.); Dummy variables (Mexican, Asian, Russian crises dummy, etc.) Approach Probit Panel Probit Results Contagion is significant (this was the focus of the study) Indicators of vulnerability to international financial spillover (common creditor) and of financial fragility (reserve adequacy) are highly significant after controlling the role of domestic and external fundamentals and trade spillovers. Exchange rate regimes and capital control does not seem to matter Asian Crisis Results None Same as above Notes The recent pattern of crises in emerging market economies does not appear to be different across crises episodes (Mexican, Asian, and Russian). Financial linkages through a common creditor appear to explain the common pattern Study Edison (2000) Glick & Moreno (1999) Dataset 20 KLR countries and 8 emerging economies that have experienced any currency crises during 1970-95, and then expanded up to 1998 monthly data, 24-month crisis window East Asian and Latin American countries (January, 1972-October 1997) Crisis Definition Similar to KLR. Index = % change in USD or DM/unit currency weight times the % change in foreign reserves. Weight is the ratio of standard deviation of USD or DM /unit currency and foreign reserves. Crisis occurs if the index is > 2.5 times the (standard deviation of the index) + mean of the index Crisis = the percentage in the exchange rate exceeds the mean plus two standard deviations. Different definition for the economies with hyperinflation (inflation rate greater than 150% in the previous 6 months a la KR (1996) Exclusion Window None 12 months Indicators In addition to KLR seven new indicators like US output and G-7 output, US interest rates, oil prices, the level of M2/foreign exchange reserves, the change in short-term debt /foreign exchange reserves and the level of short-term debt/foreign exchange reserves. Focus on money, credit (nominal and real M2, M2/ reserve money multiplier, M2/ foreign reserves, and nominal and real domestic credit, all in growth rates), trade and competitiveness (deviations in trend in the real trade, Weighted ex-change rate, export revenue growth, and the trade balance) variables Approach KLR (1998) signalling Event-study graphs and multivariate probit Results It terms of noise-to-signal ratio, exchange rate, the ratio of short-term debt to reserves, M2 / reserve, loss in FX reserve, and sharp decline in stock price do well. Based on the share of crises called correctly, the 12 months percent change in the ratio of short-term debt to reserves and export growth do well. Reductions in real domestic credit and in foreign reserves and appreciation in the real exchanges rate imply increases in the probability of a crisis. Asian Crisis Results The results for out-of-sample analysis are mixed, but still can be used as a diagnostic tool. Interpretation should be complemented by other methods such as country surveillance There is a distinct rise in the predicted crisis probability in East Asian economies Notes 34 MEASUREMENT OF INTERNATIONAL CURRENCY CRISES: A PANEL DATA APPROACH ... Study Goldstein, Kaminsky, & Reinhart (2000) Milesi-Ferretti & Razin (1998) Dataset 87 currency crises and 29 banking crises occurred in a sample of 25 emerging economies and smaller industrial countries over 1970-1995 and then conduct out-of-sample analysis for 1996-1997 monthly and annual 48 African, 26 Asian economies, 26 from Latin America and Caribbean and 5 European 1970-1996 Crisis Definition Currency crises; KLR (1998) crisis index is the weighted sum of the rate of the change of the exchange rate (e) and of reserves (R). I= (de/e)-s. d. e)/ (s. d. R)* (dR/R). Use different critical value for the economies with hyperinflation 4 similar definitions based on Frankel and Rose (1996) Exclusion Window None 3 years Indicators 15 indicators in KR (1999) + 9 indicators including CA as a share of GDP, short-term capital inflows, FDI, and the overall budget deficit, the growth rates in general government consumption, central bank credit to the public sector, net credit to the public sector and the CA Macroeconomic variables (economic growth, real consumption growth, etc.); Financial variables (ratio of M2 to GDP, etc.); External variables (CA, real effective exchange rate, etc.); Debt variables (ratio of external debt to output, etc.); Foreign variables (real interest rate in the US, the rate of growth of OECD economies, etc.); Dummy variables (regional dummies, exchange rate regime dummy, etc.) Approach Signalling The extension of Frankel and Rose (1996) (longer sample and alternative definition of currency crises) multivariate probit Results Monthly: Appreciation of the real exchange rate relative to trend, a banking crisis, a decline in stock prices, a fall in exports, a high ratio of M2 to international reserves, and a recession. Annual: A large CA deficit relative to both GDP and investment Similar to Frankel and Rose (1996) highlighting the importance of degree of overvaluation, the level of reserves, growth and interest rate in industrial countries, and the terms of trade Asian Crisis Results There was a significant increase in the predicted probability for Thailand, South Korea, the Philippines, and Malaysia but Indonesia was completely missed NA Notes Examines the predictors of CA reversal and currency crises. Then compare their effects over the output afterwards – emphasis is put on the comparison. REFERENCES Berg, A., & Pattillo, C. (1999). Predicting currency crises: The indicators approach and an alternative. Journal of International Money and Finance, 18, 561-586. Carramazza, F., Ricci, L., & Salgado, R. (2000). Trade and financial contagion in currency crises. IMF WP 00/55, March. Accessed through www.imf.org Edison, H. (2000). Do indicators of financial crises work? An evaluation of an early warning system. Board of Governors of the FRS International Finance. Discussion Paper 675. Eichengreen, B., Rose, A. K., & Wyplosz, C. (1996). Contagious currency crises. NBER Working Paper 5681. Eliasson, A.-C., & Krevter, C. (2000). On currency crisis model: A continuous crisis definition. Deutsche Bank Research. Frankel, J. A., & Rose, A. K. (1996). Currency crashes in emerging markets: An empirical treatment. Journal of International Economics, 41(3-4), 351-366. VIKALPA • VOLUME 38 • NO 4 • OCTOBER - DECEMBER 2013 Glick R., & Moreno, R. (1999). Money and credit, competitiveness, and currency crises in Asia and Latin America. Center for Pacific Basin Money and Economic Studies FRB of San Francisco WP 99-01. Goldstein, M., Kaminsky, G., & Reinhart, C. (2000). Assessing financial vulnerability: An early warning system for emerging markets. Institute of International Economics, Washington D.C. Jha, R., & Murthy, K. (2006). Environmental degradation index: A survey of composite indices measuring country performance: 2006 update. A UNDP/ODS Working Paper, By Romina Bandura With Carlos Martin del Campo, Office of Development Studies, United Nations Development Programme, New York, 35-36. Kaminsky, G., Lizondo, S., & Reinhart, C. M. (1998). Leading indicators of currency crises. IMF Staff Papers. 45(1), 1-48. Lewis-Beck, M. (1994). Factor analysis and related techniques. Sage: New Delhi. 35 Malik, Anjala (2008). Measurement and analysis of international currency crises: Lessons for India. Unpublished Ph.D. Thesis, University of Delhi. Milesi-Ferritti, G.M., & Razin, A. (1998). Current account reversals and currency crises: Empirical regularities. IMF WP 98/99. Moreno, Ramon (1995). Macroeconomic behavior during periods of speculative pressure or realignment: Evidence from Pacific Basin economies. Economic Review, Federal Reserve Bank of San Francisco, 3-16. K V Bhanu Murthy is a Professor at the Department of Commerce, at Delhi School of Economics, Delhi University. He is a Ph.D. in Economics from the Department of Economics, Delhi School of Economics. His recent contributions have been in the areas of banking and finance, environmental economics, agricultural markets, international business, knowledge management, corporate social responsibility, and business ethics. Sixteen of his papers have been rated in the Top Ten list of the Social Science Research Network Library (SSRN) since its inception. His overall ranking at SSRN (amongst 2,41,000 economists, in the world) is at 99 percentile. Anjala Kalsie is currently an Assistant Professor in the Faculty of Management Studies, University of Delhi. She has a doctorate from the Department of Commerce, Delhi School of Economics, Delhi University, and an M.Phil and M.Com. from the same university. She is also a Fellow Member of the Institute of Company Secretaries of India. She has published 10 empirical papers and has also presented five papers out of which two were in international conferences. She has 14 years of experience out of which four were in the industry. The areas of her interest are financial economics, currency crises, capital markets, and financial modeling. e-mail: bhanumurthykv@yahoo.com e- mail: kalsieanjala@gmail.com 36 MEASUREMENT OF INTERNATIONAL CURRENCY CRISES: A PANEL DATA APPROACH ...