Together or apart? The relationship between currency and banking crises

ARTICLE IN PRESS

JID: JBF

[m5G;September 30, 2019;23:47]

Journal of Banking and Finance xxx (xxxx) xxx

Contents lists available at ScienceDirect

Journal of Banking and Finance journal homepage: www.elsevier.com/locate/jbf

Together or apart? The relationship between currency and banking crises Sylvester C.W. Eijffinger a,c, Bilge Karatas¸ b,∗ a

CentER and the Department of Economics, Tilburg University, PO Box 90153 5000 LE Tilburg, the Netherlands CentER and the Department of Finance, Tilburg University, PO Box 90153 5000 LE Tilburg, the Netherlands c CEPR, 77 Bastwick Street, London EC1V 3PZ, UK b

a r t i c l e

i n f o

Article history: Received 29 September 2015 Accepted 4 September 2019 Available online xxx JEL classification: F30 F31 F41 G01 G21 E44

a b s t r a c t The purpose of this study is to provide empirical evidence on the links between currency and banking crises. Panel data probit and bivariate probit models are estimated to a sample of 21 developed and developing countries having monthly observations between the years 1985 and 2010. The findings indicate that banking crises precede currency crises, and vice versa. Currency crises also indirectly influence future banking crises probability through external shocks, liberalized financial markets, or highly-leveraged banking sectors. The study also finds evidence of contemporaneous correlation between the two crises. The results not only confirm the theoretical links between banking and currency crises, but also underline the importance of higher frequency data in analyzing the relationship between various financial crises. © 2019 Elsevier B.V. All rights reserved.

Keywords: Banking crisis Currency crisis Twin crisis

1. Introduction Twin – currency and banking – crises have led to huge losses of economic welfare in countries experienced them. These crises tend to hit harder than banking or currency crisis alone and the economic recovery takes years. This is probably the most important reason why the academic literature continuously focuses on the types, causes and impacts of twin crises. The occurrence of banking and currency crises in close time intervals has started to attract the attention of the financial crisis literature after the late 1980s with the large costs they bring to countries experiencing these crises. Although there is empirical evidence on twin crises in the economic literature, the main question is not fully answered: Which macroeconomic factors cause currency and banking crises to occur jointly? Theoretical studies to date that focus on the causal links of currency crashes and bank runs point out that currency crises can lead to banking crises, banking problems may cause a crisis in exchange rates, or both crises might occur simultaneously. The mod-

∗

Corresponding author. E-mail addresses: S.C.W.Eijffi[email protected] [email protected] (B. Karatas¸ ).

(S.C.W.

Eijffinger),

els focusing on a banking crisis leading to a currency crisis are part of the third generation currency crisis family. In these models, currency crises are the result of mismatches in the balance sheets of the private – financial and non-financial – sector caused by foreign currency borrowing and domestic currency lending. Some of the important works of this field are: Diaz-Alejandro (1985), Velasco (1987), Calvo (1998), Chang and Velasco (1999), and Goldfajn and Valdes (1997). The models suggesting the reverse causality, from a currency crisis to a banking crisis, require the banking sector to be already highly indebted in foreign currency (Miller, 1996; Mishkin, 1996). Shocks like domestic interest rate hikes (Obstfeld, 1994) or foreign interest rate hikes (Stoker, 1994) spread currency crisis to the banking sector. Finally, some studies like McKinnon and Pill (1998) model the joint occurrence of currency and banking crises due to the “overborrowing syndrome” caused by financial deregulation and international financial liberalization. These dynamic links between currency and banking crises, occurring in both directions, are not reported in the empirical works. Empirical studies in this field focus on binary choice estimations using annual data, apart from Kaminsky and Reinhart (1999) with monthly frequency, and Falcetti and Tudela (2008) with quarterly data. Most of these empirical studies find that banking crises precede currency crises (Kaminsky and Reinhart, 1999; Rossi, 1999;

https://doi.org/10.1016/j.jbankfin.2019.105631 0378-4266/© 2019 Elsevier B.V. All rights reserved.

Please cite this article as: S.C.W. Eijffinger and B. Karatas¸ , Together or apart? The relationship between currency and banking crises, Journal of Banking and Finance, https://doi.org/10.1016/j.jbankfin.2019.105631

JID: JBF 2

ARTICLE IN PRESS

[m5G;September 30, 2019;23:47]

S.C.W. Eijffinger and B. Karatas¸ / Journal of Banking and Finance xxx (xxxx) xxx

Glick and Hutchison, 2001). Some find that currency crises lead to banking crises (Von Hagen and Ho, 2009). Among these, the studies that take into account the simultaneity of currency and banking crises (Falcetti and Tudela, 2008; Glick and Hutchison, 2001; Von Hagen and Ho, 2009) also discover contemporaneous occurrence of these two crises. The asymmetric relationship between currency and banking crises established in the empirical literature might be due to the reliance on annual data which gives limited insight especially in discovering the relationship between twin crises happening in the same year. A spread from one crisis to the other occurring in several months requires an analysis with higher frequency data to empirically determine the relationship between the two crises. Our study attempts to overcome this shortcoming in the empirical literature by using monthly data between 1985 and 2010 for a sample of 21 emerging and developed countries1 to discover the ties between currency and banking crises. Although annual data decreases the concerns about correctly dating the onsets of the crises and gives the opportunity of including more countries in the sample, by relying on monthly data we capture the sudden nature of crises and their influence on each other. With this study, we fill an important gap in the literature since Kaminsky and Reinhart (1999) by providing evidence on twin crises with high frequency data. Our study contributes to the financial crisis literature in several ways: First, and foremost, with the help of the recent systemic banking crises database by Laeven and Valencia (2008, 2012) using monthly frequency data, we analyze the probability of a joint occurrence of currency and banking crises using bivariate probit estimation where we assume both crises are endogenously determined caused by common unobservable factors. Secondly, we examine not only the direct links, but also the indirect links between currency and banking crises. The interaction effects of crises variables with macroeconomic indicators are introduced to empirically test the linkages between banking and currency crises. For the currency crisis model, our hypothesis is that an initial banking crisis increases the likelihood of future currency crises through bank bailouts. This hypothesis follows the macroeconomic trilemma that, in the absence of capital controls, if central banks finance these costly bailouts by creating money they cannot maintain the pegged exchange rate regimes. Although a lot of countries have changed their regimes into much more flexible regimes and use inflation targeting as their monetary policy targets, as documented by Ilzetzki et al. (2017) 80% of the countries in the world use de facto regimes with limited flexibility. Therefore, we expect a stronger link from banking to currency crises in countries engaged in expansionary monetary policies. Similarly for the banking crisis model, our hypotheses are that a currency crisis increases the probability of future banking crisis if a country is exposed to an external shock, the financial sector is internationally liberalized, or the banking sector has high external debts. If a currency crisis is caused by increased international interest rates, there is a high risk that this crisis also damages the domestic financial system by means of capital flights. Of course, the damage to the domestic financial system depends on how open the financial system is to the international markets; in the absence of capital controls, a crisis in the exchange rate market might quickly spread to the banking sector. Finally, since a currency crisis increases the domestic currency price of the banking sector’s foreign liabilities, the result might be a banking crisis if these liabilities represent a large portion of the balance sheets of the banks.

1 Country number is constrained by the limited availability of monthly starting dates of banking crises from Laeven and Valencia (2008, 2012).

This paper is organized as follows: Section 2 describes the methodology and data used in the analyses of the linkages between banking and currency crises, Section 3 presents the insample results of the estimations, Section 4 summarizes the outof-sample results, and Section 5 concludes. 2. Methodology and data 2.1. Starting months of currency and banking crises The determination of the starting month of a currency crisis is relatively straightforward. Although there are various approaches exist in the literature in defining a currency crisis, according to Reinhart and Rogoff (2011) focusing exclusively on the depreciation of the exchange rates gives the most parsimonious currency crisis dates. Therefore in this study, we adopt the definition of a currency crisis used in Eijffinger and Karatas¸ (2012) where a large depreciation of the exchange rate occurs following moderately stable exchange rates.2 The detailed description of the crisis definition can be found in Eijffinger and Karatas (2012) which relies on the thresholds defined by Kraay (2003). Compared to the methodologies for determining currency crises, determination of the banking crises is not well established. Although there are various indexes defined in the literature, most of the starting dates are determined through event-based analyses. Here, we use the database developed by Laeven and Valencia (2008, 2012) for two reasons: Firstly they provide an updated, corrected, and expanded version of the banking crises database of Caprio and Klingebiel (1996), and Caprio et al. (2005) on which most of the empirical work on banking crises relies. Secondly, they provide the starting months of the banking crises for a small subset of countries which indicates a deviation from the previous literature where the onset of the banking crises is identified on an annual basis.3 In their database, Laeven and Valencia (2008, 2012) require the banking crisis to be ‘systemic’ and they exclude banking system distress events that affect isolated banks but are not systemic in nature. Two requirements should be met in diagnosing a systemic banking crisis: The first requirement is ‘financial distress’ designated by bank runs, losses in the banking system and/ or liquidations of banks. Secondly, there should be significant actions taken by the policy makers in response to high losses in the banking system. More explanation of their methodology can be found in Laeven and Valencia (2012: 4–5). These methodologies provide 25 banking and 46 currency crises dates for the 21 countries used in our study between January 1985 and December 2010 (Table A1, Appendix A) . Since almost

2 Kaminsky and Reinhart (1999) also limit their sample to the countries having exchange rate regimes that are fixed or pegged. We require stable exchange rates prior to currency crises to detect the theoretical links between banking and currency crises. The majority of the theoretical studies on the causal link from a currency crash to a banking crisis assume fixed exchange rates and that the crash of the currency is sudden and large. Allowing currency crises where there is no commitment on the part of the policy makers to keep the exchange rate pegged prior to currency crises might also weaken the link between a prior currency crash and the likelihood of banking crisis. As a result, with this definition of currency crisis, we expect a strong link from currency crisis to banking crisis. Robustness of the results to an alternative currency crisis definition (the exchange market pressure index) is checked in the sensitivity analyses of our study. 3 If banking crises arise the from liability side, an event-based approach works well in marking the onset; however if the problems arise from the asset side, eventbased studies might be too late in correctly diagnosing a banking crisis. In this case, non-performing loans could be used to date the onsets. However, as Reinhart and Rogoff (2009) discuss, these data are not available in high frequency for many countries and, also, they might not be accurate, since banks tend to delay as long as possible the disclosure of their problems to the public. Therefore, acknowledging its caveats, we rely on an event-based approach to date banking crises in this study.

Please cite this article as: S.C.W. Eijffinger and B. Karatas¸ , Together or apart? The relationship between currency and banking crises, Journal of Banking and Finance, https://doi.org/10.1016/j.jbankfin.2019.105631

ARTICLE IN PRESS

JID: JBF

[m5G;September 30, 2019;23:47]

S.C.W. Eijffinger and B. Karatas¸ / Journal of Banking and Finance xxx (xxxx) xxx

3

all countries have experienced at least one crisis during the sample period, our sample might suffer selection bias. Table A2 in Appendix A presents a comparison of our crisis dates with the crisis dates of the previous literature (Kaminsky and Reinhart, 1999; Glick and Hutchison, 2001; Von Hagen and Ho, 2009). Analogous to the approach by Kaminsky and Reinhart (1999), the probabilities of each crisis occurring conditioned on the existence of the other crisis are calculated. In our sample, 48% of banking crises are followed by a currency crisis in the twelve months after the onset of a banking crisis, and 32% of banking crises are followed by a currency crisis within the three months after the banking crisis onset. The conditional probability of a banking crisis following a currency crash is somewhat smaller: 15% of currency crises are followed by a banking crisis within the twelve months after the initial currency crash and the conditional probability of a currency crisis followed by a banking crisis within the three months after a currency crisis is only 9%.

are the interactions of the lagged currency and banking crises with the possible control variables that are suggested to spread one crisis to the other in theoretical studies. Eqs. (1) and (2) are estimated by probit model where the cumulative distribution function is assumed to follow the standard normal distribution. Error terms ε i,t and μi,t are assumed to be independent and identically distributed, and normal. It is also assumed that for both banking and currency crisis equations, the explanatory variables are uncorrelated with the corresponding error terms in each time period.4 Lastly the idiosyncratic shocks that affect Ci,t might be correlated with the idiosyncratic shocks that affect Bi,t . Possible correlation between the error terms of models (1) and (2) is handled in Section 3.2 by estimating these equations jointly using the bivariate probit model.

2.2. The model

The choice of macroeconomic indicators in explaining banking and currency crises onsets is based on previous relevant empirical evidence (i.e. Lestano et al., 2003; Kaminsky and Reinhart, 1999; Demirgüç-Kunt and Detragiache, 1997; Kaminsky, 2006): the gross central government debt to GDP, the inflation rate, the percentage change in stock prices and the growth rate of GDP are included as indicators of the domestic real and public sector; the international debt of the banking sector to GDP, the money supply to foreign exchange reserves, the percentage change in the real domestic interest rates, and domestic credit to the private sector over GDP represent the financial sector; the deviation of the real exchange rate from trend, the ratio of the current account balance to foreign exchange reserves, and the percentage change in the real international interest rates are included as external sector indicators. These macroeconomic variables are taken from the IMF’s International Financial Statistics, the World Bank’s World Development Indicators and the Bank for International Settlements. The recent empirical literature underlines the importance of political and institutional factors in predicting currency and banking crises. Therefore monthly institutional indexes from the International Country Risk Guide (ICRG), the parliamentary and presidential election dates from the Consortium for Elections and Political Process Strengthening (CEPPS), and the openness of a country’s capital accounts measured by the Chinn – Ito Financial Openness Index (Chinn and Ito, 2006) are used as institutional and political control variables. From the ICRG, apart from the commonly used law and order index (see Demirgüç-Kunt and Detragiache, 1997, 1998; Rossi, 1999; Eichengreen and Arteta, 20 0 0), other measures of political stability (government stability, bureaucracy quality, democratic accountability and investment profile), financial quality (the ability of a country to finance its debt) and economic quality (a country’s economic weaknesses and strengths) are included to measure institutional stability. The constructions

The onset of currency and banking crises are represented by two binary choice models in a panel data setting (i.e. let i denote country and t denote month; i = 1, ……, N; t = 1,…….,T). The first model measures the effect of a previous banking crisis on the probability of a currency crisis controlling for the effects of macroeconomic and institutional variables. The second model measures the predictive power of prior currency crisis on the occurrence of a banking crisis taking into account the specific macroeconomic and institutional factors influencing the probability of a banking crisis. The models are defined by the following equations:

Ci,t ∗ = β0 + β1 Bi,t −1tot −3 + β2 Zi,t−k + β3 Bi,t −1tot −3 Zi,t−k + μi,t Bi,t ∗ = α0 + α1Ci,t −1tot −3 +

(1)

α2 Xi,t−k + α3Ci,t −1tot −3  Xi,t−k + εi,t (2)

and k = 1,2,3,…… Ci,t ∗ and Bi,t ∗ are unobservable latent random variables representing the onset of currency and banking crises, respectively. However, the discrete dependent variables Ci,t and Bi,t are observable such that

Ci,t = 1 if Ci,t ∗ > 0 and 0 otherwise Bi,t = 1 if Bi,t ∗ > 0 and 0 otherwise The dependent variable Ci,t takes the value 1 if a currency crisis is observed in country i in month t, and zero otherwise. Similarly, the dependent variable Bi,t takes the value 1 if a banking crisis is observed in country i in month t, and zero otherwise. In order to differentiate the beginning of the crisis from the continuation of the same crisis, we apply windows. For currency crisis, the observations in the twelve months after the initial depreciation are considered as the same crisis and excluded from the sample. For banking crisis, each crisis episode has different length. Therefore we use the end dates from Laeven and Valencia (2012) and delete the observations following the onset until the end of banking crisis from our sample. Ci,t -1 to t -3 and Bi,t -1 to t -3 , represent the composite lagged currency and banking crises, respectively. These dummies take the value 1 if a currency/banking crisis occurs in the previous threemonth period. These composite lagged crises dummies give the opportunity of reducing multicollinearity concerns resulting from including crises dummies with various lags into the estimations. The vectors Xi,t-k and Zi,t-k include a set of macroeconomic and institutional variables playing a crucial role in the financial crisis literature in predicting banking and currency crises, respectively. The last term in each equation – Ci,t -1 to t -3 ´Xi,t-k and Bi,t -1 to t -3 ´Zi,t-k –

2.3. Data

4 Note that this is a rather strong assumption given the nature of the banking and currency crises. For instance, economic agents might anticipate the occurrence of a crisis if they follow an indicator which might be linked to the crisis. On the other hand, instrumental variables technique is not possible in the absence of strictly exogenous instruments, especially for banking and currency crisis dummies. Generalized Method of Moments estimation might generate consistent, asymptotically normal, and efficient estimators in the presence of endogeneity. However the method is not applicable for maximum likelihood estimation. Switching to linear estimation techniques with binary dependent variables might lead to misestimated magnitudes and unjustified hypothesis tests for independent variables. Therefore this method is not applied in our study. The endogeneity issues are apparent in other studies that successfully predict currency and banking crises (such as Berg and Pattillo (1999), Rossi (1999), Von Hagen and Ho (2007), and Komulainen and Lukkarila (2003) amongst others). Taking this into account, the correlations between economic indicators, and currency and banking crises in this study should not be interpreted as causal relationships.

Please cite this article as: S.C.W. Eijffinger and B. Karatas¸ , Together or apart? The relationship between currency and banking crises, Journal of Banking and Finance, https://doi.org/10.1016/j.jbankfin.2019.105631

ARTICLE IN PRESS

JID: JBF 4

[m5G;September 30, 2019;23:47]

S.C.W. Eijffinger and B. Karatas¸ / Journal of Banking and Finance xxx (xxxx) xxx Table 1 Descriptive statistics. Variable

Obs.

Mean

Std. Dev.

Min

Max

Unit of account

Currency Crisis Banking Crisis ࢞ Public Debt Real International Interest Rate Real Domestic Interest Rate Exchange Rate Overvaluation Current Account Position GDP Growth Inflation Stock Prices Money Supply ࢞ Domestic Credit by Banking Sector ࢞ Domestic Credit to Private Sector ࢞ Banking Sector International Debt Capital Account Openness Election Political Environment Market Environment

4531 4071 4531 4531 4125 4531 4531 4531 4531 4529 3327 4531 4531 3660 4531 4531 4531 4531

0.008 0.004 −0.016 0.018 0.066 −0.015 −0.004 0.372 0.010 0.018 10.084 0.104 0.087 0.036 0.292 0.035 0.213 0.197

0.088 0.064 0.646 0.546 1.837 0.092 0.090 0.306 0.023 0.108 15.333 0.855 0.678 0.189 1.482 0.184 0.971 0.989

0 0 −4.650 −3.390 −35 −0.714 −0.770 −1.244 −0.097 −0.780 0.862 −7.956 −5.506 −0.737 −1.904 0 −2.083 −3.115

1 1 3.758 5.041 59 0.634 0.625 1.524 0.474 2.129 105.597 4.913 4.186 2.191 2.374 1 2.229 2.609

Dummy Dummy Ratio (% of Percentage Percentage Percentage Ratio (% of Percentage Percentage Percentage Ratio (% of Ratio (% of Ratio (% of Ratio (% of Index Dummy Index Index

and sources of all the variables used in the study can be found in Table A3 in Appendix A. The seven institutional variables from the ICRG might capture similar effects and create multicollinearity in the estimations. Therefore, to construct fewer uncorrelated variables to include in our estimations, we use “factor analysis” (Kim and Mueller, 1978; Torres-Reyna, 2012). This technique generates unobserved, uncorrelated random variables – factors – that explain the variability among several observed and correlated random variables. We use Kaiser Criterion5 to condense the seven indexes into two factors which explain 67% of the total variance of these seven indexes. Table B1 in Appendix B displays the rotated6 loadings of the two factors and indicates the weight of each index in the factors. The higher the load of an index in a factor, the more relevant it is in defining the factor’s dimensionality. We name the factor having higher loadings on political indexes as “political environment” and the factor having loadings on financial, economic and investment quality as “market environment”. The loadings of the indexes on both factors are positive indicating that higher scores of these factors correspond to a higher institutional quality. Our unbalanced sample includes monthly observations starting from January 1985 until December 2010 for 21 developed and developing countries. For each country we have at most 312 monthly observations which raises concerns about the non-stationarity of the data. After conducting the Im-Pesaran-Shin (2003) test which allows for heterogeneity in the unbalanced panel, we find that the variables public debt over GDP, domestic credit by the banking sector over GDP, and the banking sector foreign debt over GDP have unit roots. Therefore we apply first-differencing transformation for these variables.7 After the transformations, in order to derive out-of-sample forecasts for assessing the performance of the models, we randomly remove one emerging and one developed country from the sample. Therefore we estimate our equations with 19 countries excluding Ecuador and Norway. These in-sample estimation results are used for out-of-sample predictions of the crises occurred in these two countries.8 Table 1 represents the descriptive statistics of the data used for the in-sample estimations.

5 This criterion retains the factors having eigenvalues – the total variance accounted by each factor – greater than or equal to one. 6 Orthogonal varimax rotation is applied to generate uncorrelated loadings for the factors by maximizing the variance of the squared loadings within factors. 7 First-differencing the indicators puts the emphasis on the short-term effects of these variables on our dependent variables. Although this is not fully in-line with

GDP) Change Change Deviation Reserves) Change Change Change Reserves) GDP) GDP) GDP)

3. In-sample results We initially estimate Eqs. (1) and (2) by fixed-effects and random-effects models in order to discover if unobserved heterogeneity needs to be controlled for. The additive unobserved fixedeffects in the probit estimations9 indicate that the fixed country effects are jointly insignificant. Following this, we estimate random effects probit model and check if the results differ from the pooled probit estimations. The likelihood ratio test fails to reject the null that there is no unobserved individual level heterogeneity in the models. Therefore pooled probit estimation method is preferred for the estimations of the two crises models. 3.1. Single equation estimation results In this section we ignore the simultaneity between two crisis models (i.e. (Corr (ε i,t, μi,t ) = 0), and estimate them separately. For every specification, the estimations are done by clustering the robust standard errors by country in order to correct for serial correlation in the error terms.10 In order to minimize the simultaneity concerns, we lag all explanatory variables at least one month. We use general-to-specific approach to determine the appropriate lag structure of the regressors in each equation. Initially, we include up to twelve lags of each variable into the estimations, and, in every stage, we exclude the statistically insignificant lags. The presented specifications include only the first significant lags of the explanatory variables. Each table presents the estimated coefficients, z-statistics, and marginal effects per regressor for various specifications. We include marginal effects, since in probit models the coefficients do not give the change in the conditional mean of the dependent variable with the change in each independent variable. For the interaction terms, marginal effects change for each observation. Thus

the theoretical considerations which, emphasize long-term relationships, not treating the non-stationary variables would create estimates that cannot be trusted. 8 We also estimate the models including all 21 countries. The unpublished results remain robust. 9 The sample with large T (each country has at least 90 observations) reduces the incidental parameters bias of the fixed-effects probit estimator. 10 In order to account for the correlation in the standard errors among different countries in the same month and across time in the same country, the models are also estimated by clustering the error terms in two dimensions (i.e. across time and across countries. See Peterson (2009) for details of the method.). The estimated standard errors, available upon request, are similar to the results presented in the paper and do not lead to major changes in the significance of the variables.

Please cite this article as: S.C.W. Eijffinger and B. Karatas¸ , Together or apart? The relationship between currency and banking crises, Journal of Banking and Finance, https://doi.org/10.1016/j.jbankfin.2019.105631

JID: JBF

ARTICLE IN PRESS

[m5G;September 30, 2019;23:47]

S.C.W. Eijffinger and B. Karatas¸ / Journal of Banking and Finance xxx (xxxx) xxx

we calculate the minimum, maximum and mean marginal effects, and their z-statistics for all interaction terms and present them in Appendix B. Tables also include goodness of fit measures as loglikelihood, pseudo R-squared and the percentage of correctly classified observations. The thresholds for the correct predictions of crisis and non-crisis observations are chosen to increase the correct classification of the crisis observations without causing a large increase in type 1 errors. Taking into account the rarity of crisis observations in the sample, relatively lower thresholds are chosen to correctly predict crisis observations. To classify an observation as ‘crisis’ we set two cut-off values: 1% and 10%. Increasing the thresholds beyond these values cause a significant decrease in the correctly classified crisis observations without any gain in the correct classification of non-crisis observations. 3.1.1. Estimations of the currency crisis model The estimation results of the currency crisis model with the set of macroeconomic, political, and institutional variables are represented in Table 2. The three-month composite lagged banking crisis dummy is included in the estimations from the second specification onwards. Since data for the ratio of money supply to foreign exchange reserves is not available for the entire sample, this variable is included in column 3 and its interaction with the lagged banking crisis dummy is represented in column 4. This interaction term seeks the contribution of monetary expansion during a banking crisis on the likelihood of a currency crisis.11 The results indicate that the onset of banking crises significantly increases the probability of future currency crises. A banking crisis occurring in the three-month period prior to a currency crisis onset increases the currency crisis likelihood by around 4%, an effect that is economically crucial. In the existing literature the magnitude of this effect changes depending on the sample period and the frequency of the data used in the studies. For instance, Glick and Hutchison (2001) find for all countries in their sample that previous banking crises increase the currency crises probability by around 5%. On the other hand, Falcetti and Tudela (2008) find an insignificant effect of a banking crisis on the currency crisis likelihood of, only, 0.6% in the following quarter. Other studies, like Von Hagen and Ho (2009) and Rossi (1999), do not present the marginal effects of their estimations. In order to identify the different effects of banking crisis indicator on an average developing and developed country, we also calculate the marginal effect of our lagged banking crisis indicator by taking the mean values of the other indicators for developing and developed country subsamples using the estimation results in column 2 of Table 2. The results suggest an initial banking crisis increases future currency crisis likelihood, on average, by 6.4% in developed countries, and by 3.6% for developing countries. This result is interesting since taking the vulnerabilities of the developing countries and the fragility of their exchange rates to the investors’ and foreign creditors’ opinions, the expectation would be that an initial banking crisis contributes more to the currency crisis probability for an average developing country. However, in developed countries, the higher transparency of information and more advanced banking system might actually make the relationship from banking to a currency crisis stronger. Since we have only 4 de-

11 Theoretical studies emphasizing this relationship are Diaz-Alejandro (1985), Velasco (1987) and Calvo (1998). Following theoretical studies, we also consider interacting lagged banking crisis dummy with foreign monetary policy and the international financial liberalization indicator. However, the interaction terms do not significantly contribute to the likelihood of currency crises. Throughout the study, the interaction terms should be interpreted carefully, since there might be potential endogeneity problems.

5

veloped countries in our sample, this effect should be interpreted with caution. Among macroeconomic controls, the change in stock prices suggests that a fall in stock prices by 1% in a given month leads to an increase in the probability of a currency crisis occurring in the next three months by around 2%. The fall in the GDP growth, overvalued real exchange rates, and the expansion of the banking sector domestic credit significantly increase the likelihood of future currency crises. These results confirm the findings of previous studies such as Frankel and Rose (1996), Berg and Pattillo (1999), Komulainen and Lukkarila (2003), and Goldstein et al. (20 0 0). Although the factor representing the political environment is not significant, the factor market environment enters significant in the last two specifications of Table 2. This indicates that with lower financial and economic quality a currency crisis is more likely. This result is in line with the expectations since when a country’s ability to repay its debts decreases, it becomes a less desirable place for the foreign investors. This result partly contradicts the findings of Rossi (1999) since he finds no significant contribution of the institutional indicators in predicting currency crises. We also fail to establish any relationship between the ratio of public debt to GDP (confirming the result of Kaminsky and Reinhart (1999), real foreign interest rates, current account balances, international financial liberalizations, presidential/parliamentary elections and currency crises. According to Diaz-Alejandro (1985) and Velasco (1987), central bank bailouts of troubled banks as a result of a banking crisis lead to excessive money creation. This can trigger a currency crisis if the majority of short-term obligations of the banks are in foreign currency. In order to analyze this effect, we include money supply to foreign exchange reserves and it’s interaction with the banking crisis indicator in columns 3 and 4 of Table 2. However, the variable as well as the interaction term do not contribute significantly to the currency crisis probability. The marginal effect of the interaction term is also insignificant for all of the observations (see Table B3 in Appendix B). Although, we expected a strong link through expansionary policies from banking to currency crisis, especially during the Global Financial Crisis (GFC) this link loosened. The reason might be that the cause of the recent currency crashes was not linked to the monetary policies, but to the GFC started in the US which was an exogenous shock for most of the countries in our sample. Additionally, although many countries allow limited flexibility in their exchange rates (Ilzetzki et al., 2017), central banks are not committed in maintaining fixed exchange rate regimes, which also explains why we fail to confirm the link between money creation and currency crises. The model explains 16 to 21% of the variation in the dependent variable. The inclusion of the lagged banking crisis dummy and the money supply indicator increases the performance of the model. This is similar to Von Hagen and Ho (2009), and Glick and Hutchison (2001) which are able to explain up to 30% of the variance in the currency crisis onset. With a 10% threshold for defining an observation as ‘crisis’, our model successfully calls 9 to 15% of the crisis periods, while predicting, on average, 99% of the observations correctly. Compared to other studies, our model has a better performance in predicting non-crisis observations. For example, with the same cut-off value the model of Von Hagen and Ho (2009) predicts 30% of crisis and 89% of non-crisis observations. Glick and Hutchison (2001) explain 80% of the crisis and, only, 47% of the non-crisis observations. If we lower the cut-off value to 1% in defining a ‘crisis’, 64 to 80% of the crisis observations are correctly called, while 83 to 87% of the total observations are correctly classified. Hence, our model is quite successful in predicting both crisis and non-crisis observations with a better fit to the non-crisis periods.

Please cite this article as: S.C.W. Eijffinger and B. Karatas¸ , Together or apart? The relationship between currency and banking crises, Journal of Banking and Finance, https://doi.org/10.1016/j.jbankfin.2019.105631

ARTICLE IN PRESS

JID: JBF 6

[m5G;September 30, 2019;23:47]

S.C.W. Eijffinger and B. Karatas¸ / Journal of Banking and Finance xxx (xxxx) xxx Table 2 Single equation pooled probit estimation results of currency crisis. Dependent variable: currency crisis onset

(1) Estimates (z-stats) elasticity

(2) Estimates (z-stats) elasticity

(3) Estimates (z-stats) elasticity

(4) Estimates (z-stats) elasticity

−0.009 (−0.11) −0.0001 −2.603∗ ∗ ∗ (−3.73) −0.024 −0.865 (−1.38) −0.008 −2.989∗ ∗ ∗ (−3.97) −0.028 −0.003 (−0.05) −0.00003 0.086 (0.59) 0.001 −0.493∗ ∗ (−2.29) −0.005 0.245∗ ∗ ∗ (3.70) 0.002 0.218 (0.76) 0.003 0.045 (0.46) 0.0004 −0.081 (−0.85) −0.001

1.092∗ ∗ ∗ (3.88) 0.042 −0.022 (−0.20) −0.0002 −2.071∗ ∗ ∗ (−3.14) −0.017 −0.636 (−0.67) −0.005 −2.511∗ ∗ ∗ (−3.31) −0.021 0.066 (0.92) 0.001 0.080 (0.41) 0.001 −0.475∗ ∗ (−1.97) −0.004 0.217∗ ∗ ∗ (2.74) 0.002 0.355 (1.25) 0.005 −0.008 (−0.07) −0.0001 −0.138 (−1.52) −0.001

0.942∗ ∗ ∗ (3.81) 0.027 0.067 (0.78) 0.001 −2.117∗ ∗ ∗ (−2.98) −0.016 −0.217 (−0.25) −0.002 −4.040∗ ∗ ∗ (−3.74) −0.030 0.121 (1.41) 0.001 0.091 (0.42) 0.001 −0.442∗ (−1.66) −0.003 0.108 (1.13) 0.001 0.216 (0.54) 0.002 −0.176 (−1.29) −0.001 −0.219∗ ∗ (−2.17) −0.002 0.004 (0.68) 0.00003

Pseudo-R2 Number of Observations Log-Likelihood

0.163 4531 −171.562

0.188 4094 −135.961

0.207 3028 −95.468

1.098∗ ∗ ∗ (4.64) 0.025 0.067 (0.77) −0.001 −2.120∗ ∗ ∗ (−2.99) −0.005 −0.362 (−0.42) 0.001 −4.053∗ ∗ ∗ (−3.76) −0.006 0.126 (1.44) 0.0004 0.099 (0.45) 0.001 −0.420 (−1.63) −0.001 0.111 (1.18) 0.0001 0.223 (0.57) 0.003 −0.175 (−1.28) −0.001 −0.215∗ ∗ (−2.12) −0.001 0.005 (0.87) 0.00003 −0.012 (−1.57) −0.001 0.209 3028 −95.178

Goodness of fit (10% cutoff) % of correctly predicted observations % of correctly predicted crises % of correctly predicted non-crises

98.77 8.57 99.44

98.93 14.29 99.51

98.98 15.00 99.53

98.15 15.00 99.47

Goodness of fit (1% cutoff) % of correctly predicted observations % of correctly predicted crises % of correctly predicted non-crises

82.79 80.00 82.81

85.78 64.29 85.93

86.82 70.00 86.93

86.49 70.00 86.60

Variables Banking Crisis

t -1 to t -3

Real International Interest Rate

Exchange Rate Overvaluation

Current Account Position

Stock Prices

t -1

t -1

t -3

Capital Account Openness

 Public Debt GDP Growth

t -1

t -1

t -1

t -1

 Domestic Credit by Banking Sectort -1 Election

t -1

Political Environmentt -1

Market Environmentt -1

Money Supplyt -1

BCt -1 to

t -3

X Money Supplyt -1

Notes: Robust standard errors are clustered by country. The significance levels of the variables are indicated by ∗ (10%), ∗ ∗ (5%) and ∗ ∗ ∗ (1%). Counter intuitively signed coefficients are represented in italics. Highly significant coefficients with anticipated signs are represented in bold. The marginal effects are evaluated at the sample mean for continuous variables and for change from zero to one for dummy variables holding all other variables at their mean. In order to convert the marginal effects into percentages they should be multiplied by 100.

3.1.2. Estimations of the banking crisis model Following the currency crisis model, we estimate the banking crisis model with the macroeconomic, political, and institutional variables, and present the results in Table 3. We include the dummy representing the three-month composite lagged onset of currency crisis from the second specification onwards. The international indebtedness of the banking sector is involved in the specifications from column 3 onwards since this variable is not available for the entire sample. In order to find out the indirect effects

of currency crises on future banking crises probability, we include the interaction terms of lagged currency crisis with real foreign interest rates, capital account openness and banking sector international debt in columns 4 through 6. Our estimation results indicate that the lagged currency crisis also help in explaining the onset of a banking crisis. The statistical and economic significance of this effect, however, is somewhat lower than their reverse relationship. On average, a currency crisis increases the likelihood of a banking crisis by 0.8% within the fol-

Please cite this article as: S.C.W. Eijffinger and B. Karatas¸ , Together or apart? The relationship between currency and banking crises, Journal of Banking and Finance, https://doi.org/10.1016/j.jbankfin.2019.105631

ARTICLE IN PRESS

JID: JBF

[m5G;September 30, 2019;23:47]

S.C.W. Eijffinger and B. Karatas¸ / Journal of Banking and Finance xxx (xxxx) xxx

7

Table 3 Single equation pooled probit estimation results of banking crisis. Dependent variable: banking crisis onset

Variables Currency Crisis

t -1

Capital Account Openness

Current Account Position

t -1

t -2

 Public Debt GDP Growth

Election

t -1

t -6

Stock Prices

(2) Estimates (z-stats) elasticity

(3) Estimates (z-stats) elasticity

(4) Estimates (z-stats) elasticity

(5) Estimates (z-stats) elasticity

(6) Estimates (z-stats) elasticity

−2.341∗ ∗ ∗ (−3.31) −0.011 −0.169 (−1.62) −0.001 −2.153∗ ∗ ∗ (−4.10) −0.010 7.223∗ ∗ ∗ (2.94) 0.033 −2.534∗ ∗ ∗ (−3.04) −0.011 0.316∗ ∗ (2.53) 0.001 −0.905∗ ∗ ∗ (−3.47) −0.004 0.714∗ ∗ ∗ (2.61) 0.009 0.069∗ ∗ (1.98) 0.0003 −0.033∗ ∗ (−2.01) −0.0001 0.161 (1.27) 0.001 0.122 (0.90) 0.001 0.374∗ ∗ ∗ (2.71) 0.002

0.676∗ (1.77) 0.007 −2.529∗ ∗ ∗ (−3.22) −0.009 −0.179∗ (−1.82) −0.001 −2.451∗ ∗ ∗ (−3.32) −0.009 8.095∗ ∗ ∗ (2.84) 0.029 −2.533∗ ∗ ∗ (−2.80) −0.009 0.321∗ ∗ ∗ (2.77) 0.001 −0.816∗ ∗ ∗ (−2.95) −0.003 0.814∗ ∗ ∗ (3.05) 0.010 0.106∗ ∗ (2.03) 0.0004 −0.041∗ ∗ (−2.17) −0.0001 0.201 (1.54) 0.001 0.086 (0.75) 0.0003 0.410∗ ∗ (2.41) 0.001

0.723∗ (1.80) 0.008 −2.256∗ ∗ (−2.53) −0.008 −0.234∗ (−1.88) −0.001 −1.692 (−1.41) −0.006 8.107∗ ∗ ∗ (2.73) 0.027 −2.538∗ ∗ ∗ (−2.70) −0.009 0.430∗ ∗ ∗ (3.00) 0.001 −0.765∗ ∗ (−2.47) −0.003 0.929∗ ∗ ∗ (3.42) 0.013 0.052 (0.29) 0.0002 −0.046∗ ∗ (−2.34) −0.0002 0.192 (1.11) 0.001 0.061 (0.54) 0.0002 0.318 (1.62) 0.001 0.650 (1.39) 0.002

0.870∗ ∗ (2.21) 0.012 −2.242∗ ∗ (−2.55) −0.008 −0.238∗ (−1.88) −0.001 −1.656 (−1.37) −0.006 8.113∗ ∗ ∗ (2.74) 0.027 −2.528∗ ∗ ∗ (−2.71) −0.009 0.420∗ ∗ ∗ (2.97) 0.001 −0.807∗ ∗ (−2.52) −0.003 0.924∗ ∗ ∗ (3.45) 0.013 0.030 (0.17) 0.0001 −0.046∗ ∗ (−2.36) −0.0002 0.195 (1.14) 0.001 0.060 (0.52) 0.0002 0.318 (1.62) 0.001 0.655 (1.39) 0.002 5.745∗ ∗ ∗ (2.90) 0.393

−1.553 (−1.16) −0.001 −2.049∗ ∗ (−2.18) −0.006 −0.269∗ ∗ (−1.98) −0.001 −1.666 (−1.37) 0.00003 8.107∗ ∗ ∗ (2.78) 0.023 −2.561∗ ∗ ∗ (−2.72) −0.008 0.403∗ ∗ ∗ (2.89) 0.001 −0.854∗ ∗ ∗ (−2.57) −0.003 0.935∗ ∗ ∗ (3.46) 0.012 0.042 (0.24) 0.001 −0.047∗ ∗ (−2.35) −0.0002 0.201 (1.20) 0.001 0.089 (0.76) 0.0003 0.327 (1.63) 0.001 0.656 (1.41) 0.003

−0.745 (−1.00) −0.001 1.909∗ (−1.94) −0.006 −0.282∗ ∗ (−2.03) −0.001 −1.689 (−1.41) −0.0001 8.034∗ ∗ ∗ (2.79) 0.025 −2.458∗ ∗ (−2.51) −0.008 0.371∗ ∗ (2.54) 0.001 −0.894∗ ∗ (−2.51) −0.003 0.939∗ ∗ ∗ (3.53) 0.013 0.051 (0.29) 0.001 −0.047∗ ∗ (−2.38) −0.0002 0.205 (1.28) 0.001 0.121 (0.97) 0.001 0.330 (1.65) 0.001 0.599 (1.27) 0.003

t -1 to t -3

Exch. Rate Overvaluation

Inflation

(1) Estimates (z-stats) elasticity

t -6

t -1

t -1

Real International Interest Rate

Real Domestic Interest Rate

t -2

t -1

 Dom. Credit to Private Sector Political Environment

Market Environment

t -1

t -1

t -1

 Banking Sector Foreign Debt

t -1

t -2

CCt -1 to

t -3

X Real Int.. Int. Rate

CCt -1 to

t -3

X Cap. Acc. Openness

t -1

CCt -1 to

t -3

X Bank. Foreign Debt

t -1

1.523∗ ∗ (2.10) 0.078 6.984∗ ∗ (2.51) 0.155

Pseudo-R2 Number of Observations Log-Likelihood

0.209 3900 −99.168

0.235 3697 −87.067

0.266 3000 −73.188

0.268 3000 −72.994

0.276 3000 −72.201

0.276 3000 −72.164

Goodness of fit (10% cutoff) % of observations correctly predicted % of crises correctly predicted % of non-crises correctly predicted

99.31 20.00 99.72

99.22 22.22 99.59

98.90 25.00 99.30

98.90 25.00 99.30

98.93 25.00 99.33

99.13 31.25 99.50

Goodness of fit (1% cutoff) % of observations correctly predicted % of crises correctly predicted % of non-crises correctly predicted

88.95 70.00 89.05

89.99 72.22 90.08

90.53 75.00 90.62

90.60 75.00 90.68

90.90 75.00 90.99

90.80 75.00 90.88

Notes: Robust standard errors are clustered by country. The significance levels of the variables are indicated by ∗ (10%), ∗ ∗ (5%) and ∗ ∗ ∗ (1%). Counter intuitively signed coefficients are represented in italics. Highly significant coefficients with anticipated signs are represented in bold. The marginal effects are evaluated at the sample mean for continuous variables and for change from zero to one for dummy variables holding all other variables at their mean. In order to convert the marginal effects into percentages they should be multiplied by 100.

Please cite this article as: S.C.W. Eijffinger and B. Karatas¸ , Together or apart? The relationship between currency and banking crises, Journal of Banking and Finance, https://doi.org/10.1016/j.jbankfin.2019.105631

JID: JBF 8

ARTICLE IN PRESS

[m5G;September 30, 2019;23:47]

S.C.W. Eijffinger and B. Karatas¸ / Journal of Banking and Finance xxx (xxxx) xxx

lowing three months. Glick and Hutchison (2001) also discover a lower marginal effect (though insignificant) – around 1% – of the lagged currency crisis on the banking crisis probability. If we take the developed and developing country averages of the other indicators by using the specification in column 2 of Table 3 and calculate the marginal effects of the initial currency crisis on the probability of a banking crisis, we see that for developed countries this effect is 0.69%, and for developing countries it is 0.68%. Hence we do not observe a significantly different effect of currency crisis on banking crisis between developing and developed countries. The overvaluation of real exchange rates enters highly significant confirming Goldstein et al. (20 0 0), amongst others, that it stands as one of the best predictors of banking crises. The variable indicates that a 1% increase in the real exchange rate misspecification increases the future banking crisis likelihood by around 6 – 11%. This effect is similar to the results of Falcetti and Tudela (2008). The financial liberalization indicators, capital account openness and domestic real interest rates12 indicate that the international and domestic liberalizations of the financial sector decrease the likelihood of experiencing a banking crisis. The results seem to contradict the conventional view that financial liberalization is linked with higher banking crises probability. On the other hand, the association of the falling domestic interest rates with higher banking crisis risk supports the claim of Calvo et al. (1994) that with the fall in the interest rates, banking system might become illiquid due to the large withdrawals of depositors from the domestic banking system. The international openness decreasing banking crises likelihood supports Eichengreen and Arteta (20 0 0), and Bekaert et al. (2011). They also find that banking crises are more likely when capital controls are present. However, the effect is insignificant in Bekaert et al. (2011). Additionally, the fall in the net worth of the firms, slowing economic activity, rise in the inflation rate, rising public sector indebtedness compared to GDP, and elections are all significant in increasing the likelihood of future banking crises. On average, banking crisis probability increases by 1% following an election. Although not significant in the currency crisis model, the rise in the US policy rates and worsening of the current account position play roles in predicting a banking crisis onset. On the other hand, similar to Demirgüç-Kunt and Detragiache (1997), private sector credit to GDP does not help in predicting banking crises. Banking sector international debt also fail to predict banking crises. The insignificant institutional factors – political and market environment – confirm the results of Rossi (1999). In column 4 of Table 3, we include the interaction of the currency crisis dummy with the change in the real foreign interest rates. According to Stoker (1994), an increase in the real US interest rates during a currency crisis might weaken the banking sector through capital flight. If the banking sector is highly exposed to foreign finances, this situation might cause a banking crisis. In the estimations, both the coefficient and the mean marginal effect of the interaction term are positive, and for the majority of the observations the marginal effect is significant (see Table B4 in Appendix B). In order to analyze the true influence of real US interest rates on the banking crisis likelihood, we calculate its marginal effect when there is no currency crisis and compare it to its effect during a currency crisis. The much higher marginal effect of the real US interest rates during a currency crisis suggests that a 1 percentage point increase in US interest rate increases the banking crisis probability by 39%. This economically large effect is in

12 In the financial crisis literature, the rise in the real domestic interest rates is associated with the liberalization of the financial sector (Lestano et al., 2003).

line with the theory that an external shock during currency crises increases the likelihood of banking crises. The theoretical studies of McKinnon and Pill (1996, 1998) suggest that the liberalization of the capital account increases the foreign borrowing of the banking sector and its exposure to currency risk. Although we find that international financial liberalization decreases the banking crisis probability, during a currency crisis the absence of capital controls might trigger a banking crisis. Foreign and domestic investors can withdraw from the domestic banks making the financial sector illiquid. In order to analyze this effect, we use the interaction of capital account openness indicator with the currency crisis dummy. The result indicates a positive and significant coefficient for the interaction term represented in column 5. For most of the observations, the term’s marginal effect is also positive and significant, as indicated in Table B4. The sign of the marginal effect of the term is negative if there is no currency crisis compared to a positive effect when there is a currency crisis. This result indicates that although lower restrictions of a country to the international financial markets improve the banking system and decrease the banking crisis likelihood, a currency crash reverses the positive effect of international liberalization and increases the banking crisis probability through capital flights. Lastly, Mishkin (1996) suggests that large foreign liabilities of the banks put pressure on the financial sector in the course of a currency crisis since devaluations worsen the balance sheets of the banks. This situation increases the risk of insolvency, and hence the risk of a banking crisis. In the last column, we try to control this by including an interaction term between the currency crisis dummy and foreign indebtedness of the banking sector. The term has strong influence on the probability of a banking crisis – for most of the observations the marginal effect is positive and significant – showing that an already fragile banking system might fall into crisis with the sudden currency crash. This result is in line with the argument of Demirgüç-Kunt and Detragiache (1997) that in the case of a speculative attack against the domestic currency, foreign liabilities risk the profitability of the banking sector and might lead to a banking crisis. The model explains 21 to 28% of the variance in the banking crisis onset which is a fairly satisfactory fit and similar to other studies. The model of Von Hagen and Ho (2009) has 19% explanatory power for their whole sample, and in Glick and Hutchison (2001) the pseudo R-squared ranges from 20% for all countries to 26% for the emerging economies. The correctly predicted crisis observations ranges from 20 to 31% when the 10% cut-off is applied for defining a ‘crisis’ observation. The inclusion of the lagged currency crisis dummy and its interactions with macroeconomic variables decreases the false signals in the predictions. Compared to other studies, similar to our currency crisis model, the predictions perform better for the overall observations. Glick and Hutchison (2001) predict 85% of the observations correctly while their model calls 50% of the crisis observations correctly. The main reason for a larger type 2 error in our models than that of comparable studies is due to the monthly frequency of the sample data. Therefore we see a significant increase in the correct predictions of the actual crisis observations if we decrease the threshold to 1% in defining an observation as ‘crisis’. In this case, the model predicts 70 to 75% of the crisis observations correctly which represents a successful fit compared to similar studies in the literature. 3.2. Joint estimation of banking and currency crises The single equation pooled probit estimations assume that the error terms of the banking and currency crises equations are independent. Therefore, the estimations do not take into account that the currency and banking crises occur contemporaneously and that they are driven by common unobservable factors. In this section,

Please cite this article as: S.C.W. Eijffinger and B. Karatas¸ , Together or apart? The relationship between currency and banking crises, Journal of Banking and Finance, https://doi.org/10.1016/j.jbankfin.2019.105631

ARTICLE IN PRESS

JID: JBF

[m5G;September 30, 2019;23:47]

S.C.W. Eijffinger and B. Karatas¸ / Journal of Banking and Finance xxx (xxxx) xxx

9

Table 4 Bivariate probit estimation results. Dependent variable: currency crisis onset Variables

Real International Interest Rate

Exchange Rate Overvaluation

Current Account Position

Stock Prices

t -1

t -1

t -3

Capital Account Openness

 Public Debt GDP Growth

t -1

t -1

t -1

t -1

 Domestic Credit by Banking Sector Election

t -1

Political Environment

Market Environment

t -1

t -1

t -1

Dependent variable: banking crisis onset Estimates (z-stats) elasticity

Variables

Estimates (z-stats) elasticity

−0.041 (−0.40) −0.0004 −1.869∗ ∗ ∗ (−2.92) −0.019 −0.656 (−0.70) −0.007 −2.105∗ ∗ ∗ (−2.94) −0.021 0.077 (1.12) 0.001 0.121 (0.81) 0.001 −0.473∗ ∗ (−2.09) −0.005 0.186∗ ∗ (2.27) 0.002 0.347 (1.33) 0.006 −0.053 (−0.48) −0.001 −0.134 (−1.64) −0.001

Exchange Rate Overvaluation

Capital Account Openness

Current Account Position

Inflation

t -1

t -1

t -6

Stock Prices

t -2

 Public Debt GDP Growth

Election

t -1

t -6

t -1

t -1

Real International Interest Rate

Real Domestic Interest Rate

t -2

t -1

 Domestic Credit to Private Sector Political Environment

Market Environment

t -1

t -1

ρ

t -1

−2.396∗ ∗ ∗ (−2.62) −0.007 −0.348∗ ∗ (−2.11) −0.001 −2.442∗ ∗ ∗ (−3.08) −0.008 7.549∗ ∗ ∗ (2.73) 0.024 −3.107∗ ∗ ∗ (−3.56) −0.010 0.254∗ ∗ (2.05) 0.001 −0.915∗ ∗ ∗ (−2.96) −0.003 0.850∗ ∗ ∗ (3.22) 0.010 0.104∗ ∗ (1.99) 0.0003 −0.039∗ (−1.84) −0.0001 0.192 (1.53) 0.001 0.176 (1.39) 0.001 0.448∗ ∗ (2.44) 0.001

0.454 Wald test of ρ = 0: chi2 (1) = 4.549 Prob > chi2 = 0.033 Log-Likelihood −206.150 Number of Observations 3631

Notes: The estimation is conducted using the “biprobit” command for STATA 13. Robust standard errors are clustered by country. The significance levels of the variables are indicated by ∗ (10%), ∗ ∗ (5%) and ∗ ∗ ∗ (1%). Counter intuitively signed coefficients are represented in italics. Highly significant coefficients with anticipated signs are represented in bold. The marginal effects are evaluated at the sample mean for continuous variables and for change from zero to one for dummy variables holding all other variables at their mean. In order to convert the marginal effects into percentages they should be multiplied by 100.

we jointly estimate the two crises equations by bivariate probit model using maximum likelihood estimation. This model estimates an extra parameter, ρ , that measures the correlation between the error terms of the two crises equations (Corr (ε i,t, μi,t ) = ρ ), and accounts for the endogeneity of the crises models. The bivariate probit model estimation for the specifications in the first columns of Tables 2 and 3 for the currency and banking crises equations, respectively, is presented in Table 4. Accounting for the possibility of the correlation in the error terms of currency and banking crises equations do not radically change the results of the coefficients of the explanatory variables. The contemporaneous correlation between the error terms of the two crises equations indicates a positive correlation between the currency and banking crises, i.e. ρ = 0.454. The likelihood ratio test of the significance of ρ rejects the null hypothesis that the correlation coefficient between the two equations is zero. This cru-

cial result is robust to different specifications and indicates that the two crises are endogenously determined caused by common unobservable factors. 3.3. Sensitivity analyses Crises are rare events, and with the monthly frequency of our sample data, the percentage of the crisis observations in the overall sample becomes quite low. This might lead to a bias in our estimation results and should be corrected. In order to account for this problem, we estimate the currency and banking crises models with the Rare Events Logistic Model (King and Zeng, 2001a). According to King and Zeng (2001a) if not accounted for, the sample having rare events underestimates the event probabilities, and for this reason they suggest decreasing the rareness of the event in the full sample by selecting a small sample of non-crisis periods

Please cite this article as: S.C.W. Eijffinger and B. Karatas¸ , Together or apart? The relationship between currency and banking crises, Journal of Banking and Finance, https://doi.org/10.1016/j.jbankfin.2019.105631

JID: JBF 10

ARTICLE IN PRESS

[m5G;September 30, 2019;23:47]

S.C.W. Eijffinger and B. Karatas¸ / Journal of Banking and Finance xxx (xxxx) xxx

to increase the efficiency of the estimator. Hence with Rare Events Logistic Estimator we estimate the corrected sample with unbiased coefficients having lower variance.13 The results of the rare events logit estimations, available upon request, are similar to the pooled probit estimations. Similarly, in the currency crisis equation, the lagged onset of a banking crisis has a high explanatory power. For the banking crisis equation, the prior onset of currency crisis is, again, among the important determinants of a banking crisis. Following the robustness checks with the change in the estimation technique, to compare our results with an alternative definition of a currency crisis, we construct the exchange market pressure index (EMP) which is developed by Eichengreen et al. (1996). This widely used index is a broader definition of currency crises as it also takes into account the unsuccessful attacks to the currency where the attack only alters the foreign exchange reserves without any significant change in the domestic currency price. Additionally, this definition of currency crisis brings the advantage of analyzing speculative attacks under both fixed and flexible exchange rate regimes. We calculate the index following Kaminsky and Reinhart (1999) by taking the weighted average of changes in the exchange rates and foreign exchange reserves. The weights are applied to make sure that the two components of the index have the same sample volatilities. For each country, we consider the months that the index is above its mean by two standard deviations as crisis months. Since we calculate standard deviations separately for each country, the crisis threshold changes per country. Appendix C provides further explanation on the construction of the index. Using the EMP index definition, we determine 82 currency crisis observations between 1985 and 2010 for the countries in our sample. The conditional probabilities suggest that 60% of banking crises are followed by a currency crisis in the following twelve months, while only 16% of the currency crises are followed by a banking crisis in the twelve-month period after a currency crisis onset. We re-estimate the specifications of the currency crisis model in Table 2, the banking crisis model in Table 3, and the bivariate probit estimation in Table 4 by using our new currency crisis indicator. We apply, once more, a twelve-month window after each currency crisis onset and exclude these observations from the sample. The results of our in-sample estimations, are presented in Tables C1 to C5 in Appendix C. The estimation results of the currency crisis model represent, once more, a high predictive power of lagged banking crises on the likelihood of currency crises. A banking crisis increases the probability of a currency crisis within the following three months by an average of 6%.14 However, currency crises do not help in predicting future banking crises once the definition of a currency crisis allows for the unsuccessful speculative attacks against currency. This result supports the findings of Kaminsky and Reinhart (1999), and Glick and Hutchison (2001) where currency crises tend to follow rather than precede banking crises. Both studies also use the EMP index in defining a currency crisis. In the currency crisis model, presented in Table C1, most of the previous results are confirmed. Additionally, the link between currency crises and current account deficits become stronger. In the banking crisis equation, presented in Table C3, the inflation rate and capital account openness are no longer significant; GDP growth, real international and domestic interest rates have lower significance levels. Domestic credit to private sector, on the other

13 The readers can refer to King and Zeng (2001a) for technical explanations, and King and Zeng (2001b) for applied examples. 14 The separate marginal effects of banking crisis for developed and developing countries do not differ significantly; for developed countries the marginal effect is 6.5%, and for developing countries it is 5.3%.

hand, has higher significance. This can be explained by the change in the currency crisis indicator. Since with the change in the currency crisis indicator the sample also changes, it is not clear if the changes in currency crisis onsets or the changes in the observations leading to the highly significant coefficient of the private sector domestic credit. Demirgüç-Kunt and Detragiache (1997) also find changing significance of this variable in explaining banking crisis with the changes in regressors. The coefficients of the other variables are not affected by the change in the currency crisis definition. The interactions of lagged currency crisis with the macroeconomic indicators also have significant marginal effects, yet smaller magnitudes compared to the main results. Therefore, although the direct effect of currency crises on the likelihood of banking crises is insignificant, the indirect effects are robust to the change in the definition of the currency crisis variable. The bivariate probit estimation results, presented in Table C5, suggest a positive correlation coefficient between the error terms of the two equations. However, the likelihood ratio test fails to reject the null of no correlation between the error terms at 5% significance level. Thus, once the currency crises definition is changed to allow for currency fluctuations prior to the crisis and unsuccessful attacks on currency, we confirm that banking crises increases the future currency crises likelihood. However, we find that the contemporaneous correlation between currency and banking crises weakens as well as the predictive power of previous currency crises on the banking crises probability. Our preference of using monthly data in establishing the close relationship of twin crises comes with a trade-off. On the one hand we establish the sudden and dynamic relationship between the two crises, on the other hand we might suffer a measurement error if we fail to establish the exact starting month of the crises. This is especially true with the banking crises onsets since we rely on event-based methodologies. Therefore, in order to confirm our results with the application of lower frequency crisis dating, we use the methodologies of Goldstein et al. (20 0 0), and Bussiere and Fratzscher (2006) in widening our window for currency and banking crises onsets. Since macroeconomic variables deteriorate before the actual onset of the crisis, we define our currency crisis variable equals to 1 if there is a currency crisis in the next 12 months. For the banking crisis onset, we define the crisis window from 12 months prior to the crisis month extending to 12 months after the crisis month. This choice of the window for banking crisis comes from the main finding of Kaminsky and Reinhart (1999) that the peak of the banking crisis comes after the onset of the crisis. In order to compare the results with our main findings, we use the specifications of Tables 2–4 with our new crisis indicators. The in-sample estimation results are presented in Appendix D in Tables D1 to D5. Table D1 representing the currency crisis estimations with our 12-month window currency crisis variable confirms our main findings. Banking crisis indicator – lagged 1 month – enters highly significant together with overvalued exchange rates, change in stock prices, and domestic credit. GDP growth has lower significance, and real US interest rates have significant, but negative coefficient. As for the banking crisis equation, we include the composite currency crisis indicator lagged by 12 months as the main regressor. The variable enters highly significant to the estimations. This indicates that the spread from currency to banking crisis can take longer than 3-month period and applying a wider window makes this link stronger. The other results confirm our previous findings, except for the lower significance of public indebtedness, election dates, real domestic interest rates, and inflation rates. Banking sector foreign indebtedness enters significant to the estimations. As for the interaction terms, although their coefficients are insignificant most of the marginal effects are significant confirming our main findings. The wider crises windows changes the goodness of fits of our two crises estimations that our models

Please cite this article as: S.C.W. Eijffinger and B. Karatas¸ , Together or apart? The relationship between currency and banking crises, Journal of Banking and Finance, https://doi.org/10.1016/j.jbankfin.2019.105631

ARTICLE IN PRESS

JID: JBF

[m5G;September 30, 2019;23:47]

S.C.W. Eijffinger and B. Karatas¸ / Journal of Banking and Finance xxx (xxxx) xxx

11

Table 5 Out of sample predictions. Currency crises Goodness of fit (10% cutoff) % of observations correctly predicted % of crises correctly predicted % of non-crises correctly predicted

98.81 20.00 99.76

Goodness of fit (1% cutoff) % of observations correctly predicted % of crises correctly predicted % of non-crises correctly predicted

85.27 60.00 85.58

99.75 50.00 100.00

Goodness of fit (1% cutoff) % of observations correctly predicted % of crises correctly predicted % of non-crises correctly predicted

91.88 50.00 92.09

Banking crises Goodness of fit (10% cutoff) % of observations correctly predicted % of crises correctly predicted % of non-crises correctly predicted

predict crisis observations much better compared to the main results. Table D5 shows the results of the bivariate probit estimations. The correlation coefficient of the error terms of our two crises equations is similar to our previous finding, i.e. ρ = 0.523, and also highly significant. This result is reasonable because of the wider windows of our crises variables. Therefore, we can conclude that our results are robust to the application of wider crises windows. As a final sensitivity test, we check if our results are robust to the alternative banking crises dates in the literature. Therefore we estimate our equations using the banking crises starting dates of Kaminsky and Reinhart (1999), Von Hagen and Ho (2009), and Glick and Hutchison (2001) presented in Table A2 in Appendix A. For the monthly banking crises dates of Kaminsky and Reinhart (1999), based on their work, we set a 12-month window and delete the observations following the banking crisis onset. For the quarterly banking crisis dates of Von Hagen and Ho (2009), we take the 3 months corresponding to their quarterly dating of the banking crisis onset and delete 24 months (8 quarters) after the onset of crisis. For the yearly dating of Glick and Hutchison (2001), we set 12 months as banking crisis corresponding to their crisis years and set the window according to their end dates. Sample size can get very small depending on the sample period and country selection of each study. Therefore we estimate only the second specifications of Tables 2 and 3 for currency and banking crisis models, respectively, and do not include the interaction terms. We estimate these specifications separately, as well as jointly using bivariate probit estimation. The results are presented in Tables E1 to E4 in Appendix E. The results with the banking crisis dates of Kaminsky and Reinhart (1999), presented in Table E1, indicate that the lagged banking crisis contributes to future currency crisis likelihood, but not vice versa. On the other hand, in Table E2, with the banking crises dates of Von Hagen and Ho (2009), we find currency crises help in predicting future banking crises, but we do not find a significant converse relationship. Finally, with the banking crises dates of Glick and Hutchison (2001), we establish the symmetric result that both lagged crises types contributing to each other’s likelihood. A result that is in line with our main findings. The bivariate probit estimations, presented in Table E4, establish a positive and significant correlation coefficient of the error terms of our two crises equations for the dates of Von Hagen and Ho (2009), and Glick and Hutchison (2001). However, for the dates of Kaminsky and Reinhart (1999), although the results indicate a perfect correlation15 between the error terms, it is not significant. These results, partially confirm our main findings. However, the asymmetric results in the single equation estimations with the banking crisis dates of

15 The boundary value of ρ might be the result of the poor fit of our models. This might be caused by too many zeros in our dependent variables that some episodes contain fewer 1’s compared to others.

Kaminsky and Reinhart (1999), and Von Hagen and Ho (2009) indicate that banking crisis dating might be crucial in establishing the links between the two crisis models. However, the different sample periods and selection of countries might also be the reason behind these asymmetric results. 4. Out-of-sample forecasts In this section, we test the performances of the currency and banking crises models by conducting out-of-sample predictions for two countries that are excluded from the in-sample estimations: Norway and Ecuador. For currency crises, the estimation results of the specification in column 1 of Table 2 are used to generate the prediction of the currency crisis variable. For banking crises, estimation results in column 1 of Table 3 are used. The success of these models in predicting the observations of Norway and Ecuador are checked for two different cut-off values – 10 and 1% – of crisis classification (Table 5). For the currency crisis model, the out-of-sample predictions cannot predict the two currency crises that occurred in Ecuador since the data for Ecuador is only available after 1995, which does not include the currency crises dates. For Norway, if the crisis classification threshold is set to 10%, the model predicts one out of five crisis observations correctly while only sending 0.24% false signals. If we set the threshold to 1% to define an observation as ‘crisis’, the model correctly predicts March 1991, November 1992 and October 2008 crises, but misses July 1988 and May 2010 crises, although the false alarms are in the run up to crisis. For the banking crisis predictions, the model correctly predicts Ecuador’s 1998 crises under both thresholds. However, Norway’s 1991 banking crisis could not be predicted even under the low threshold value. We also conduct out-of-sample predictions for the alternative currency crisis definition using the EMP index. Under this definition, Ecuador experiences three currency crises (12/1985, 09/1988, 03/1999). Since the data for Ecuador is available after 1995, our model predicts the 1999 crisis under the low threshold. For Norway, we establish eight crisis dates using the EMP index: 05/1986, 07/1988, 03/1991, 11/1992, 12/1996, 06/1999, 09/2008, 05/2010. The model, fails to predict any of these crises with the 10% threshold. If the threshold is set to 1%, the model predicts six crises correctly, and fails to predict crises occurred on 12/1996 and 06/1999. 5. Conclusion In this study, we analyze the likelihood of banking and currency crises emphasizing the theoretical connections between the two crises by controlling for their lagged and contemporaneous impacts on each other. The sample contains 21 countries having monthly observations between January 1985 and December 2010. The estimations are conducted by pooled probit and bivariate probit models that use the method of maximum likelihood estimation.

Please cite this article as: S.C.W. Eijffinger and B. Karatas¸ , Together or apart? The relationship between currency and banking crises, Journal of Banking and Finance, https://doi.org/10.1016/j.jbankfin.2019.105631

JID: JBF 12

ARTICLE IN PRESS

[m5G;September 30, 2019;23:47]

S.C.W. Eijffinger and B. Karatas¸ / Journal of Banking and Finance xxx (xxxx) xxx

Interesting conclusions can be derived from the results of our study: banking crises tend to precede devaluations and, to a lower extent, currency crises help in predicting future banking crises. In our main results, we also find evidence on the contemporaneous correlation between the two crises. This important result shows that these two crises are closely intertwined which is proved in a monthly setting. If we define currency crises to allow for unsuccessful attacks against currency and depreciations in flexible exchange rate regimes (the EMP index), the predictive power of currency crises on the future and contemporaneous banking crises decreases. Under this definition of currency crisis, speculative attacks tend to follow banking crises, and there is no evidence on the common stochastic factors causing the contemporaneous occurrence of the two crises in the same month. These results support other studies using the EMP index in defining a currency crisis, like Kaminsky and Reinhart (1999), and Glick and Hutchison (2001), although Glick and Hutchison (2001) reports contemporaneous correlation for their emerging economy subsample. The result of banking crises preceding currency crises is robust under two currency crisis definitions. In this sense, our results are in line with those of Kaminsky and Reinhart (1999). Their claim is that currency crises tend to follow the onset of banking crises and deepen the banking crises by creating a vicious circle. However, there are two major conclusions to add to the early work of Kaminsky and Reinhart (1999). The first one is that by taking into account the possibility of common factors leading two crises occurring jointly in the same month, we establish the endogeneity between these two crises even in a monthly setting. This important result proves that currency and banking crises are intertwined, and shows the importance of dynamic analysis in establishing the links between them. Secondly, different from Kaminsky and Reinhart (1999) who only use the EMP index to define currency crisis, we find that under a narrower currency crisis definition where large fluctuations in the exchange rates are not allowed prior to domestic currency depreciations, currency crises also have a strong leading effect on the probability of banking crises. Our narrow currency crisis definition only focuses on national currency depreciation and does not allow for the volatility of exchange rate prior to currency crash. Hence, under this definition currency crisis is rather unexpected and publicly unknown. Currency depreciation (successful attacks on exchange rates) actually increases the foreign debt obligations of the banking system and also is a clearer signal of balance of payments problems. Since under the narrow definition we do not allow depreciations under flexible regimes, our result supports the argument of Obstfeld and Rogoff (1995) that countries operating under fixed exchange rate regimes are more prone to banking sector problems once their currency is under attack. This strong relationship loosens under EMP index definition; once the currency crises observations include unsuccessful attacks on the currency and depreciations under flexible regimes. This difference in currency crisis definition provides crucial insights to the policy makers regarding exchange rate management.

Mainly we confirm the results of earlier studies on early warning indicators of currency and banking crises that most of the macroeconomic variables deemed to be significant determinants of banking and currency crises are successful in explaining the likelihood of these two crises. Conversely, even though a variety of financial crises follow financial liberalizations, we establish that both domestic and international liberalization decrease the banking crisis likelihood. This result might seem unconventional but confirms the literature on the benefits of financial openness on the development of financial system. Levine (2001) discusses that international liberalization increases the variety and availability of capital, and that it leads to the improvement of the financial infrastructure. Schmuckler (2004) argues that the domestic financial system benefits both from external and internal liberalization because they improve the infrastructure and the governance of the domestic financial system. As for the institutional indicators, while we find some evidence that they help to predict future currency crises, we fail to support the similar argument for the banking crisis model which contradicts the results of Demirgüç-Kunt and Detragiache (1997), but confirms those of Rossi (1999). Lastly, we find indirect links from currency to banking crises and they are robust to the alternative currency crises definitions. The findings indicate that the banking crisis probability increases if a country in the wake of a currency crisis faces rising foreign real interest rates, international liberalization of the financial sector, or rising international obligations of the banking sector. Note that even though international liberalization by itself decreases the banking crisis likelihood, during a currency crisis it increases the risk of future banking crisis. The currency crisis changes the positive influence of liberalization to negative, suggesting that even though liberalization improves the banking system, it has the opposite effect with the easiness of capital flight during a currency crisis. On the whole, these indirect links confirm the theoretical literature linking banking and currency crises. Our results underline the importance of analyzing twin crises with monthly frequency data, since the time period between these crises is rather narrow – a few months –, annual analysis results in a serious loss of information in analyzing the lead-lag relationship of these two events. We also establish that the asymmetric results of the previous empirical literature are highly related to the differences in banking and currency crises definitions, that deserves much more attention in the twin crises analysis. In this respect the empirical evidence of our study contributes significantly to the existing twin crises literature. Acknowledgements We would like to thank the editor, Geert Bekaert, the two anonymous referees, Benedikt Goderis, Jakob de Haan, Harry Huizinga, Clemens Kool and Wolf Wagner for their valuable comments and suggestions. We also thank the seminar participants at the Central Bank of the Republic of Turkey and at the IFABS 2015 International Conference for their helpful discussions.

Please cite this article as: S.C.W. Eijffinger and B. Karatas¸ , Together or apart? The relationship between currency and banking crises, Journal of Banking and Finance, https://doi.org/10.1016/j.jbankfin.2019.105631

ARTICLE IN PRESS

JID: JBF

[m5G;September 30, 2019;23:47]

S.C.W. Eijffinger and B. Karatas¸ / Journal of Banking and Finance xxx (xxxx) xxx

Appendix A. Crisis dates and data sources

13

.

Table A1–A3

Table A1 Currency and banking crisis dates.

1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21

Country

Banking crisis

Currency crisis

Argentina Brazil Chile China Colombia Ecuador Finland India Indonesia Japan Korea Malaysia Mexico Norway Philippines Russia Sweden Thailand Turkey United Kingdom Venezuela

12.1989, 01.1995, 11.2001 02.1990, 12.1994

01.2002 01.1999, 10.2008 07.1986, 09.2008 12.1985, 07.1988, 07.1991 09.1986, 03.1991 11.1997, 12.1997 12.1994, 07.1988, 09.1997 09.1998, 03.1991, 07.1997 02.2001, 10.1992, 12.1986,

11.1998 06.1998 08.1998 09.1991 09.1993a 11.1997 11.1997 08.1997 07.1997 12.1994 10.1991 07.1997 08.1998, 09.2008 09.1991, 09.2008 07.1997 11.2000 09.2007 01.1994

10.2002, 09.2008 01.1994 09.1992 03.1991, 09.1992, 10.2008, 05.2010 08.1997, 11.2008 10.2008 09.1998, 10.2008 03.1991, 11.1992, 10.2008, 05.2010 01.2009 11.1992, 09.2008 10.2008 08.2008 12.1995, 02.2002, 01.2010

a The starting month is taken from Khan (2011) as the forced merger between New Bank of India and Punjab National Bank due to increased problems of New Bank of India.

Table A2 Overview of currency and banking crises dates in previous literature. Country

Our study Sample Period: 1985–2010

Kaminsky and Reinhart (1999) Sample Period: 1970–1995

Glick and Hutchison (2001) Sample Period: 1975–1997

Von Hagen and Ho (2009) Sample Period: 1980–2004

Banking Crisis

Currency Crisis

Banking Crisis

Currency Crisis

Banking Crisis

Currency Crisis

Banking Crisis

Currency Crisis

01/02

03/80, 05/85, 12/94 11/85, 12/94

02/81, 09/86, 02/90 11/86, 10/91

80–82, 89–90, 95–97 90, 94–97

Q2/89

Q4/87, Q1/90

Q1/91, Q1/99

Chile China Colombia Ecuador

11/98 06/98 08/98

09/81 Not in Sample 07/82 Not in Sample

08/82 Not in Sample 03/83 Not in Sample

76, 81–83 82–86 82–87 80–82, 96–97

75–76, 82–83, 89–91 82–83, 87, 90–91, 95 85

Q2/89, Q3/02

Brazil

12/89, 11/95, 12/01 02/90, 12/94

Q4/84 Not in Sample Not in Sample Q2/84

Q3/82, Q4/84 Not in Sample Not in Sample Q3/82, Q1/99

Finland

09/91

09/91

11/91

91–94

Q4/89

Q1/83, Q3/92

India Indonesia

09/93 11/97

Not in Sample 11/92

Not in Sample 09/86, 08/97

93–97 94, 97

76, 91, 93, 95 78, 83, 86, 97

Q4/99 Q3/84, Q1/98

Q4/81, Q3/91 Q1/98

Japan Korea Malaysia Mexico

11/97 08/97 07/97 11/94

Not in Not in 07/85, 09/82,

Not in Not in 07/75, 12/82,

92–97 97 85–88, 97 81–91, 95–97

Q4/89 Q4/97 Not in Sample Q3/82, Q1/95

10/91

79, 89–90 80, 97 86, 97 76, 82, 85, 94–95 78, 86, 97

Q3/98 Q4/81, Q1/98 Not in Sample Q2/89, Q2/95

Norway

Not in Sample

Not in Sample

Philippines Russia Sweden

07/97 08/98, 09/08 09/91, 09/08

Not in Sample Not in Sample Q3/92

07/97

83–84, 86, 97 Not in Sample 77, 81–82, 92–93 81,84, 97

Not in Sample Not in Sample Q3/92

Thailand

07/88, 09/92, 05/10 07/91 09/86, 11/08 03/91 11/97, 12/97 12/94, 10/08 07/88, 11/92, 05/10 09/97 09/98, 03/91, 09/08 07/97

Q3/97

Q1/81, Q3/97

Turkey

11/00

78–80, 94

Q1/01

Q2/94, Q1/01

United Kingdom

Not in Sample

Not in Sample

Venezuela

76, 79, 81–82, 86, 92 84,86, 94–96

Q4/97

Q1/89, Q2/96

Argentina

01/99, 10/02, 09/08 10/08 07/86, 01/94 09/08 12/85, 09/92 03/91, 10/08,

08/97,

10/08 09/98, 03/91, 10/08,

Sample Sample 09/97 10/92

Sample Sample 08/97 12/94

11/88

05/86

87–93

01/81, 07/97 Not in Sample 11/91

10/83, 07/97 Not in Sample 11/92

81–87, 97 Not in Sample 90–93

11/78, 11/84, 07/97 03/94

83–87, 97

02/01, 10/08

03/79, 10/83, 05/96 01/91

09/07

10/92, 08/08

Not in Sample

Not in Sample

82–85, 91, 94–95 75–76, 84

01/94

12/86, 12/95, 02/02, 1/10

10/93

05/94

78–86, 94–97

01/09 11/92,

85 82–83, 85–86, 88 77–78, 82, 91–93

Please cite this article as: S.C.W. Eijffinger and B. Karatas¸ , Together or apart? The relationship between currency and banking crises, Journal of Banking and Finance, https://doi.org/10.1016/j.jbankfin.2019.105631

JID: JBF 14

ARTICLE IN PRESS

[m5G;September 30, 2019;23:47]

S.C.W. Eijffinger and B. Karatas¸ / Journal of Banking and Finance xxx (xxxx) xxx

Table A3 Data descriptions. Explanatory variables

Construction and sources

Frequency

Public Debt Real International Interest Rate

Gross Central Government Debt divided by GDP (Reinhart and Rogoff, 2009a) The percentage change in the US Federal Funds rate subtracted by the US inflation rate (IMF IFSb lines 60B and 64) The percentage change in the money market interest rate subtracted by the inflation rate (IMF IFSb lines 60B and 64) The deviation of the real exchange rate from the trend which is calculated using Hodrick-Prescott filter with a parameter of 129,000 (IMF IFSb lines 60B and 64) The ratio of the difference of a country’s exports and imports expressed in US dollars to non-gold reserves. (IMF IFSb lines 70D, 71D, RF and 64) The Chinn-Ito Index measuring country’s degree of financial openness (Chinn and Ito, 2006)c The percentage change in the Consumer Price Index (IMF IFSb line)64 The percentage change in share prices (IMF IFSb line 62) The percentage change in the nominal GDP expressed in local currency. (World Bank, WDId ) The ratio of domestic credit provided by the banking sector to GDP. (World Bank, WDId ) The ratio of domestic credit to private sector to GDP. (World Bank, WDId ) Parliamentary and presidential election dummy. (Election Guide, Consortium for Elections and Political Process Strengthening) The ratio of money plus quasi-money, converted in US dollars, to non-gold reserves. (IMF, IFSb lines 59 MB, RF and 1LD) International debt securities outstanding by financial institutions divided by GDP (BISe and World Bank, WDId ) Risk rating that assesses the country’s ability to finance its official, commercial and trade debt obligations. (PRS Group, ICRGf ) Risk rating that assesses the country’s currency economic strengths and weaknesses. (PRS Group, ICRGf ) Risk rating that assesses the government’s ability to carry out its declared programs and its ability to stay in the office. (PRS Group, ICRGf ) Risk rating that assesses the strength and quality of the bureaucracy in the political system. (PRS Group, ICRGf ) Risk rating that assesses the strength of the legal system and observance of law. (PRS Group, ICRGf ) Risk rating that assesses how responsive the government is towards its people. (PRS Group, ICRGf ) Risk rating that assesses the risk to investment. (PRS Group, ICRGf )

Annual, linear interpolation Monthly

Real Domestic Interest Rate Exchange Rate Overvaluation Current Account Position Capital Account Openness Inflation Stock Prices GDP Growth Domestic Credit by Banking Sector Domestic Credit to Private Sector Election Money Supply Banking Sector International Debt Financial Quality Economic Quality Government Stability Bureaucracy Quality Law and Order Democratic Accountability Investment Profile a b c d e f

Monthly Monthly Monthly Annual Monthly Monthly Annual, linear interpolation Annual, linear interpolation Annual, linear interpolation Monthly Monthly Annual, linear interpolation Monthly Monthly Monthly Monthly Monthly Monthly Monthly

Extracted from: International Monetary Fund, International Financial Statistics Extracted from: World Development Indicators Bank of International Settlements. Political Risk Services Group, International Country Risk Guide Rating.

Appendix B. Statistics Table B1–B4.

Table B1 Rotated factor loadings of institutional variables. Factor 1 Political Environment

Factor 2 Market Environment

Uniqueness

Financial Quality Economic Quality Government Stability Investment Profile Bureaucracy Quality Law and Order Democratic Accountability

0.308 0.443 −0.073 0.409 0.859 0.758 0.835

0.743 0.701 0.810 0.618 0.211 0.288 0.039

0.353 0.312 0.338 0.452 0.218 0.343 0.301

Variance Explained

0.353

0.316

Notes: The factor analysis is conducted in STATA 13 based on principal-component method. All institutional variables are lagged one-month. Varimax rotation is implemented to generate uncorrelated factor loadings. The relevant variable per factor is indicated in bold. The two factors explain 67% of the total variance in the indicators.

Please cite this article as: S.C.W. Eijffinger and B. Karatas¸ , Together or apart? The relationship between currency and banking crises, Journal of Banking and Finance, https://doi.org/10.1016/j.jbankfin.2019.105631

ARTICLE IN PRESS

JID: JBF

[m5G;September 30, 2019;23:47]

S.C.W. Eijffinger and B. Karatas¸ / Journal of Banking and Finance xxx (xxxx) xxx

15

Table B2 Correlation coefficients.

Currency Crisis Banking Crisis Real Int. Interest Rate ER Overvaluation CA Position Inflation Stock Prices ࢞Public Debt GDP Growth ࢞Domestic Credit Election Money Supply KA Openness Real Dom. Interest Rate ࢞ Priv. Sector Dom. Cr. ࢞Bank Int. Debt Political Environment. Market Environment

࢞Domestic Credit Election Money Supply KA Openness Real Dom. Interest Rate ࢞Priv. Sector Dom. Cr. ࢞Bank Int. Debt Political Environment. Market Environment

Currency Crisis

Banking Crisis

Real Int. Int. Rate

ER Overval.

CA Position

Inflation

Stock Prices

࢞Public Debt

GDP Growth

1.000 −0.004 −0.014 −0.030 −0.021 0.002 −0.034 0.047 −0.067 −0.004 0.028 0.050 0.001 0.000 −0.030 −0.010 −0.023 −0.035

1.000 −0.003 −0.069 −0.029 0.014 −0.029 0.024 −0.051 −0.043 0.018 0.017 −0.045 −0.005 −0.037 0.019 −0.012 −0.011

1.000 0.034 0.012 −0.025 −0.021 −0.015 0.033 −0.060 −0.002 −0.011 0.009 0.013 −0.045 0.016 0.009 0.026

1.000 0.056 0.066 0.026 −0.048 0.035 −0.035 −0.004 −0.002 0.019 −0.003 0.011 −0.000 0.062 −0.038

1.000 −0.254 −0.015 −0.130 0.062 −0.077 0.017 −0.427 0.090 −0.000 −0.049 −0.271 −0.149 0.282

1.000 0.137 −0.001 −0.029 −0.079 0.016 −0.175 −0.336 −0.013 −0.105 −0.100 −0.185 −0.378

1.000 −0.021 0.022 −0.011 −0.014 −0.067 −0.079 −0.030 0.002 −0.049 −0.065 −0.091

1.000 −0.366 0.082 0.018 0.150 0.014 −0.020 −0.087 0.024 0.138 −0.061

1.000 −0.130 −0.008 −0.202 −0.247 0.016 0.043 −0.147 −0.251 0.092

࢞Dom. Credit

Election

Money Supply

KA Open

Real dom Int. Rate

࢞Prv. Dom. Cr.

࢞Bank Int. Debt

Political Env.

Market Env.

1.000 −0.018 0.152 0.135 0.009 0.859 0.092 0.182 0.094

1.000 0.016 −0.034 0.001 −0.026 0.031 −0.007 −0.040

1.000 0.378 −0.019 0.156 0.447 0.498 −0.053

1.000 −0.043 0.171 0.263 0.568 0.230

1.000 0.005 0.002 −0.042 0.019

1.000 0.120 0.181 0.152

1.000 0.268 0.147

1.000 −0.206

1.000

Table B3 Marginal Effects of the Interaction Term in Currency Crisis Equation. Interaction of banking crisis with

Mean

Min

Max

Money Supply (z-stats) Money Supply if BC Money Supply if BC

−0.001 (−0.91) 0.00007 −0.0005

−0.004 (−1.85)

−0.000 (−0.17)

t -1 to t -3 = 0 t -1 to t -3 = 1

Notes: The marginal effects of the change in the interacted terms are calculated with the “inteff” command (Norton et al., 2004) in STATA 13. The marginal effect of the continuous indicator for two values of dummy indicator is calculated with the “margins” command in STATA 13. Table B4 Marginal effects of the interaction terms in banking crisis equation. Interaction of currency crisis with

Mean

Min

Max

Real International Interest Rate (z-stats) Real International Interest Rate if CC Real International Interest Rate if CC

0.393 (1.02) 0.0003 0.399

−0.002 (−0.19)

2.304 (7.90)

0.078 (1.12) −0.003 0.071

0.000 (0.06)

0.545 (4.47)

0.155 (0.12) 0.007 0.142

−0.121 (−1.65)

3.018 (6.65)

Capital Account Openness (z-stats) Capital Account Openness if CC Capital Account Openness if CC

t -1 to t -3 = 0 t -1 to t -3 = 1

t -1 to t -3 = 0 t -1 to t -3 = 1

Banking Sector International Debt (z-stats) Banking Sector International Debt if CC Banking Sector International Debt if CC

t -1 to t -3 = 0 t -1 to t -3 = 1

Notes: The marginal effects of the change in the interacted terms are calculated with the “inteff” command (Norton et al., 2004) in STATA 13. The marginal effect of the continuous indicator for two values of dummy indicator is calculated with the “margins” command in STATA 13.

Appendix C. Exchange market pressure index estimations Table C1–C5. Definition: Exchange market pressure index (EMP) The exchange market pressure index for country i in period t is defined as follows:

EMP Ii,t = ei,t /ei,t − −(σe /σr ) (ri,t /ri,t )

(C1)

where (࢞ei,t / ei,t ) and (࢞ri,t / ri,t ) are the rate of change in exchange rates and foreign exchange reserves, respectively. In order to have equal variances of the two elements of the index, the standard deviation of the change in the exchange rate σ e is divided by σ r which is the standard deviation of the change in the reserves. The periods that are two standard deviations above the mean of the index are classified as crisis periods. In order to avoid misdiagnosing a currency crisis during hyperinflation periods, the sample is divided for the periods where in the previous six months

Please cite this article as: S.C.W. Eijffinger and B. Karatas¸ , Together or apart? The relationship between currency and banking crises, Journal of Banking and Finance, https://doi.org/10.1016/j.jbankfin.2019.105631

ARTICLE IN PRESS

JID: JBF 16

[m5G;September 30, 2019;23:47]

S.C.W. Eijffinger and B. Karatas¸ / Journal of Banking and Finance xxx (xxxx) xxx Table C1 Single equation pooled probit estimation results of currency crisis with alternative currency crisis definition. Dependent variable: currency crisis onset

Variables Banking Crisis

Exchange Rate Overvaluation

Current Account Position

t -1

t -1

Capital Account Openness

GDP Growth

t -1

t -3

 Public Debt

t -1

t -1

t -1

 Domestic Credit by Banking Sector Election

Political Environment

Market Environment

BCt -1 to

t -1

t -1

Money Supply

t -3

(2) Estimates (z-stats) elasticity

(3) Estimates (z-stats) elasticity

(4) Estimates (z-stats) elasticity

−0.096∗ (−1.83) −0.002 −1.864∗ ∗ (−2.23) −0.044 −0.702∗ (−1.82) −0.017 0.095 (0.15) 0.002 −0.064 (−1.63) −0.002 0.004 (0.04) 0.0001 −0.688∗ ∗ ∗ (−2.98) −0.017 0.128 (1.19) 0.003 −0.107 (−0.35) −0.002 0.072 (1.17) 0.002 −0.151∗ ∗ (−1.96) −0.004

0.915∗ ∗ ∗ (2.73) 0.056 −0.081 (−1.13) −0.002 −2.169∗ ∗ (−2.27) −0.044 −1.119∗ ∗ ∗ (−2.60) −0.023 −1.203∗ ∗ (−2.18) −0.025 −0.011 (−0.34) −0.0002 0.049 (0.44) 0.001 −0.373 (−1.26) −0.008 0.206∗ ∗ ∗ (3.07) 0.004 0.003 (0.01) 0.0001 0.022 (0.37) 0.001 −0.147∗ (−1.88) −0.003

0.737∗ ∗ (1.99) 0.033 −0.054 (−0.61) −0.001 −3.236∗ ∗ ∗ (−3.03) −0.059 −1.498∗ (−1.88) −0.027 −2.023∗ ∗ ∗ (−2.77) −0.037 −0.006 (−0.12) −0.0001 −0.044 (−0.46) −0.001 −0.550∗ ∗ (−2.06) −0.010 0.213∗ ∗ (2.06) 0.004

1.005∗ ∗ ∗ (2.63) 0.061 −0.056 (−0.61) −0.001 −3.283∗ ∗ ∗ (−3.10) −0.058 −1.655∗ ∗ (−2.19) −0.029 −1.966∗ ∗ ∗ (−2.64) −0.035 −0.003 (−0.07) −0.0001 −0.048 (−0.49) −0.001 −0.525∗ ∗ (−1.98) −0.009 0.218∗ ∗ (2.53) 0.004

−0.091 (−1.35) −0.002 −0.159∗ (−1.91) −0.003 0.001 (0.18) 0.00002

−277.793 0.093 4217

−224.728 0.129 3828

−154.540 0.154 2719

−0.096 (−1.52) −0.002 −0.157∗ (−1.93) −0.003 0.003 (0.39) 0.0001 −0.017∗ ∗ ∗ (−3.03) −0.002 −153.815 0.158 2719

98.06 6.90 99.33

98.07 12.50 99.15

98.01 17.65 99.03

97.98 14.71 99.03

55.63 82.76 55.25

65.46 81.25 65.26

68.26 79.41 68.12

68.78 79.41 68.64

t -1 to t -3

Real International Interest Rate

Stock Prices

(1) Estimates (z-stats) elasticity

t -1

t -1

t -1

X Money Supply

t -1

Log-Likelihood Pseudo-R2 Number of Observations Goodness of fit (10% cutoff) % of correctly predicted observations % of correctly predicted crises % of correctly predicted non-crises Goodness of fit (1% cutoff) % of correctly predicted observations % of correctly predicted crises % of correctly predicted non-crises

Notes: Robust standard errors are clustered by country. The significance level of the variables are indicated by ∗ (10%), ∗ ∗ (5%) and ∗ ∗ ∗ (1%). Counter intuitively signed coefficients are represented in italics. Highly significant coefficients with anticipated signs are represented in bold. The marginal effects are evaluated at the sample mean for continuous variables and for change from zero to one for dummy variables holding all other variables at their mean. In order to convert the marginal effects into percentages they should be multiplied by 100. Election dummy drops in the specifications (3) and (4) due to multicollinearity. Table C2 Marginal effects of the interaction term in currency crisis equation. Interaction of banking crisis with

Mean

Min

Max

Money Supply (z-stats) Money Supply if BC Money Supply if BC

−0.002 (−1.08) 0.00007 −0.002

−0.006 (−3.02)

−0.000 (−0.25)

t -1 to t -3 = 0 t -1 to t -3 = 1

Notes: The marginal effects of the change in the interacted terms are calculated with the “inteff” command (Norton et al., 2004) in STATA 13. The marginal effect of the continuous indicator for two values of dummy indicator is calculated with the “margins” command in STATA 13.

Please cite this article as: S.C.W. Eijffinger and B. Karatas¸ , Together or apart? The relationship between currency and banking crises, Journal of Banking and Finance, https://doi.org/10.1016/j.jbankfin.2019.105631

ARTICLE IN PRESS

JID: JBF

[m5G;September 30, 2019;23:47]

S.C.W. Eijffinger and B. Karatas¸ / Journal of Banking and Finance xxx (xxxx) xxx

17

Table C3 Single equation pooled probit estimation results of banking crisis with alternative currency crisis definition. Dependent variable: banking crisis onset

Variables Currency Crisis

t -1 to t -3

Exch. Rate Overvaluation

t -1

Capital Account Openness

Current Account Position

Inflation

t -1

t -6

Stock Prices

t -2

 Public Debt GDP Growth

Election

t -1

t -6

t -1

t -1

Real International Interest Rate

Real Domestic Interest Rate

t -2

t -1

 Dom. Credit to Private Sector Political Environment

Market Environment

t -1

t -1

t -1

 Banking Sector Foreign Debt

(1) Estimates (z-stats) elasticity

(2) Estimates (z-stats) elasticity

(3) Estimates (z-stats) elasticity

(4) Estimates (z-stats) elasticity

(5) Estimates (z-stats) elasticity

0.538 (0.69) 0.001 −2.982∗ ∗ ∗ (−3.58) −0.006 −0.150 (−1.48) −0.0003 −2.721∗ ∗ ∗ (−3.70) −0.005 2.664 (0.61) 0.005 −2.793∗ ∗ ∗ (−3.07) −0.006 0.528∗ ∗ ∗ (4.48) 0.001 −0.708∗ ∗ (−2.30) −0.001 0.775∗ ∗ (2.48) 0.006 0.107∗ ∗ (2.00) 0.0002 −0.038 (−1.29) −0.0001 0.417∗ ∗ ∗ (4.69) 0.001 0.052 (0.46) 0.0001 0.474 (2.55) 0.001

0.495 (0.81) 0.002 −3.012∗ ∗ ∗ (−2.99) −0.005 −0.207 (−1.60) −0.0003 −2.094∗ (−1.87) −0.003 0.694 (0.14) 0.001 −2.420∗ ∗ (−2.07) −0.004 0.771∗ ∗ ∗ (4.93) 0.001 −0.410 (−1.19) −0.001 0.929∗ ∗ ∗ (2.99) 0.008 −0.078 (−0.30) −0.0001 −0.048∗ (−1.66) −0.0001 0.538∗ ∗ ∗ (4.50) 0.001 0.014 (0.16) 0.00002 0.324 (1.41) 0.001 0.450 (1.09) 0.001

0.497 (0.78) 0.002 −3.011∗ ∗ ∗ (−2.98) −0.005 −0.207 (−1.60) −0.0003 −2.094∗ (−1.87) −0.003 0.697 (0.14) 0.001 −2.419∗ ∗ (−2.06) −0.004 0.771∗ ∗ ∗ (4.93) 0.001 −0.411 (−1.18) −0.001 0.929∗ ∗ ∗ (3.00) 0.008 −0.079 (−0.31) −0.0001 −0.048∗ (−1.66) −0.0001 0.538∗ ∗ ∗ (4.48) 0.001 0.014 (0.16) 0.00002 0.324 (1.41) 0.001 0.450 (1.09) 0.001 0.060 (0.04) 0.0003

−0.493 (−0.77) −0.0003 −3.284∗ ∗ ∗ (−2.94) −0.004 −0.242 (−1.60) −0.0003 −2.148∗ (−1.78) −0.003 0.869 (0.18) 0.001 −2.619∗ ∗ (−2.33) −0.003 0.777∗ ∗ ∗ (4.76) 0.001 −0.529 (−1.34) −0.001 0.906∗ ∗ ∗ (2.82) 0.006 0.009 (0.04) 0.00001 −0.052∗ (−1.70) −0.0001 0.558∗ ∗ ∗ (4.48) 0.001 −0.007 (−0.08) −0.00001 0.321 (1.36) 0.0004 0.447 (1.06) 0.001

−0.195 (−0.28) −0.0002 −3.205∗ ∗ ∗ (−2.99) −0.005 −0.224 (−1.60) −0.0003 −2.123∗ (−1.86) −0.003 0.871 (0.18) 0.001 −2.457∗ ∗ (−2.11) −0.004 0.762∗ ∗ ∗ (4.81) 0.001 −0.489 (−1.31) −0.001 0.930∗ ∗ ∗ (2.92) 0.007 −0.009 (−0.04) −0.00001 −0.051∗ (−1.69) −0.0001 0.565∗ ∗ ∗ (4.36) 0.001 −0.008 (−0.10) −0.00001 0.307 (1.35) 0.0004 0.402 (1.02) 0.001

t -1

CCt -1 to

t -3

X Real Int. Int. Rate

CCt -1 to

t -3

X Cap. Acc. Openness

t -1

CCt -1 to

t -3

X Bank. Foreign Debt

t -1

t -2

0.824∗ ∗ ∗ (2.94) 0.018 3.243∗ ∗ ∗ (3.22) 0.060

Log-Likelihood Pseudo-R2 Number of Observations

−68.606 0.249 3513

−54.935 0.292 2846

−54.935 0.292 2846

−53.623 0.309 2846

−54.072 0.303 2846

Goodness of fit (10% cutoff) % of correctly predicted observations % of correctly predicted crises % of correctly predicted non-crises

99.37 21.43 99.69

99.19 25.00 99.51

99.19 25.00 99.51

99.26 33.33 99.54

99.19 33.33 99.47

Goodness of fit (1% cutoff) % of correctly predicted observations % of correctly predicted crises % of correctly predicted non-crises

91.29 64.29 91.40

91.95 83.33 91.99

91.95 83.33 91.99

92.09 83.33 92.13

92.06 83.33 92.59

Notes: Robust standard errors are clustered by country. The significance levels of the variables are indicated by ∗ (10%), ∗ ∗ (5%) and ∗ ∗ ∗ (1%). Counter intuitively signed coefficients are represented in italics. Highly significant coefficients with anticipated signs are represented in bold. The marginal effects are evaluated at the sample mean for continuous variables and for change from zero to one for dummy variables holding all other variables at their mean. In order to convert the marginal effects into percentages they should be multiplied by 100.

Please cite this article as: S.C.W. Eijffinger and B. Karatas¸ , Together or apart? The relationship between currency and banking crises, Journal of Banking and Finance, https://doi.org/10.1016/j.jbankfin.2019.105631

ARTICLE IN PRESS

JID: JBF 18

[m5G;September 30, 2019;23:47]

S.C.W. Eijffinger and B. Karatas¸ / Journal of Banking and Finance xxx (xxxx) xxx

Table C4 Marginal effects of the interaction terms in banking crisis equation. Interaction of currency crisis with

Mean

Min

Max

Real International Interest Rate (z-stats) Real International Interest Rate if CC t -1 to t -3 = 0 Real International Interest Rate if CC t -1 to t -3 = 1 Capital Account Openness (z-stats) Capital Account Openness if CC t -1 to t -3 = 0 Capital Account Openness if CC t -1 to t -3 = 1 Banking Sector International Debt (z-stats) Banking Sector International Debt if CC t -1 to t -3 = 0 Banking Sector International Debt if CC t -1 to t -3 = 1

0.0003 (0.00) −0.0007 −0.0005 0.018 (0.86) −0.002 0.015 0.060 (0.35) 0.004 0.060

−0.0001 (−0.01)

0.021 (0.04)

0.000 (0.05)

0.289 (3.13)

−0.065 (−0.84)

1.452 (4.32)

Notes: The marginal effects of the change in the interacted terms are calculated with the “inteff” command (Norton et al., 2004) in STATA 13. The marginal effect of the continuous indicator for two values of dummy indicator is calculated with the “margins” command in STATA 13. Table C5 Bivariate probit estimation results with alternative currency crisis definition. Dependent variable: currency crisis onset Variables

Real International Interest Rate

Exchange Rate Overvaluation

Current Account Position

Stock Prices

t -1

t -1

t -3

Capital Account Openness

 Public Debt GDP Growth

t -1

t -1

t -1

t -1

 Domestic Credit by Banking Sector Election

t -1

Political Environment

Market Environment

t -1

t -1

t -1

Dependent variable: banking crisis onset Estimates (z-stats) elasticity

Variables

Estimates (z-stats) elasticity

−0.081 (−1.13) −0.002 −2.031∗ ∗ (−2.09) −0.045 −1.219∗ ∗ ∗ (−3.45) −0.027 −1.168∗ ∗ (−2.23) −0.026 −0.039 (−0.82) −0.001 −0.022 (−0.13) −0.001 −0.479∗ (−1.85) −0.011 0.168∗ ∗ (2.38) 0.004 0.048 (0.14) 0.001 0.035 (0.56) 0.001 −0.121∗ ∗ (−2.04) −0.003

Exchange Rate Overvaluation

Capital Account Openness

Current Account Position

Inflation

t -1

t -2

 Public Debt GDP Growth

t -6

t -1

t -1

Real International Interest Rate

Real Domestic Interest Rate

t -2

t -1

 Domestic Credit to Private Sector Political Environment

Market Environment

ρ

t -1

t -6

Stock Prices

Election

−2.830∗ ∗ ∗ (−2.88) −0.005 −0.268∗ (−1.90) −0.001 −2.984∗ ∗ ∗ (−3.47) −0.005 3.030 (0.65) 0.005 −3.083∗ ∗ ∗ (−3.44) −0.006 0.423∗ ∗ ∗ (3.95) 0.001 −0.919∗ ∗ (−2.19) −0.002 0.801∗ ∗ ∗ (2.61) 0.006

t -1

t -1

t -1

t -1

(1.91) 0.0002 −0.038 (−1.07) −0.0001 0.405∗ ∗ ∗ (4.58) 0.001 0.193 (1.36) 0.0003 0.528∗ ∗ (2.13) 0.001

0.360 Wald test of ρ = 0: chi2 (1) = 3.266 Prob > chi2 = 0.071

Log-Likelihood Number of Observations

−268.089 3411

Notes: The estimation is conducted using the “biprobit” command for STATA 13. Robust standard errors are clustered by country. The significance levels of the variables are indicated by ∗ (10%), ∗ ∗ (5%) and ∗ ∗ ∗ (1%). Counter intuitively signed coefficients are represented in italics. Highly significant coefficients with anticipated sings are represented in bold. The marginal effects are evaluated at the sample mean for continuous variables and for change from zero to one for dummy variables holding all other variables at their mean. In order to convert the marginal effects into percentages they should be multiplied by 100.

Please cite this article as: S.C.W. Eijffinger and B. Karatas¸ , Together or apart? The relationship between currency and banking crises, Journal of Banking and Finance, https://doi.org/10.1016/j.jbankfin.2019.105631

ARTICLE IN PRESS

JID: JBF

[m5G;September 30, 2019;23:47]

S.C.W. Eijffinger and B. Karatas¸ / Journal of Banking and Finance xxx (xxxx) xxx

the inflation rate is higher than 150% and the index is calculated separately for these subsamples. The original index developed by Eichengreen et al. (1996) also includes interest rate changes, but we stick to the calculation of Kaminsky and Reinhart (1999) since the interest rate data is not available for every country in our sample. Further explanations can be found in Kaminsky and Reinhart (1999).

19

Table D2 Marginal effects of the interaction term in currency crisis equation. Interaction of banking crisis with

Mean

Min

Max

Money Supply (z-stats) Money Supply if BC Money Supply if BC

−0.0004 (−0.09) 0.0009 0.0006

−0.003 (−0.77)

0.0001 (0.27)

t -1 t -1

=0 =1

Notes: The marginal effects of the change in the interacted terms are calculated with the “inteff” command (Norton et al., 2004) in STATA 13. The marginal effect of the continuous indicator for two values of dummy indicator is calculated with the “margins” command in STATA 13.

Appendix D. Estimations with wider crises windows Table D1–D5. Table D1 Single equation pooled probit estimation results of currency crisis. Dependent variable: currency crisis onset

(1) Estimates (z-stats) elasticity

(2) Estimates (z-stats) elasticity

(3) Estimates (z-stats) elasticity

(4) Estimates (z-stats) elasticity

−0.111∗ ∗ ∗ (−5.74) −0.013 −5.144∗ ∗ ∗ (−3.10) −0.605 −1.012 (−1.21) −0.119 −1.323∗ ∗ ∗ (−2.97) −0.156 −0.037 (−0.42) −0.004 0.077 (0.41) 0.009 −0.580 (−1.63) −0.068 0.293∗ ∗ ∗ (3.37) 0.034 −0.121 (−0.74) −0.013 0.039 (0.52) 0.005 0.062 (0.52) 0.007

1.146∗ ∗ ∗ (4.45) 0.255 −0.122∗ ∗ ∗ (−4.67) −0.014 −4.238∗ ∗ ∗ (−2.92) −0.497 −0.671 (−0.67) −0.079 −0.929∗ ∗ ∗ (−2.62) −0.109 0.056 (0.62) 0.007 0.059 (0.24) 0.007 −0.286 (−0.81) −0.034 0.228∗ ∗ (2.52) 0.027 −0.102 (0.58) −0.011 0.022 (0.16) 0.003 −0.012 (−0.11) −0.001

0.950∗ ∗ ∗ (3.56) 0.186 −0.134∗ ∗ ∗ (−3.65) −0.015 −6.382∗ ∗ ∗ (−3.64) −0.717 0.149 (0.17) 0.017 −1.432∗ ∗ ∗ (−2.57) −0.161 0.067 (0.80) 0.008 0.046 (0.18) 0.005 −0.254 (−0.68) −0.029 0.186∗ (1.78) 0.021 −0.044 (−0.22) −0.005 −0.163 (−0.93) −0.018 −0.036 (−0.30) −0.004 0.007 (0.87) 0.001

1.035∗ ∗ ∗ (3.04) 0.210 −0.137∗ ∗ ∗ (−3.70) −0.015 −6.395∗ ∗ ∗ (−3.58) −0.713 −0.041 (−0.04) −0.005 −1.429∗ ∗ ∗ (−2.62) −0.159 0.070 (0.81) 0.008 0.047 (0.18) 0.005 −0.244 (−0.65) −0.027 0.193∗ (1.85) 0.022 −0.060 (−0.28) −0.006 −0.163 (−0.93) −0.018 −0.032 (−0.26) −0.004 0.008 (0.89) 0.001 −0.006 (−0.54) −0.0004

Log-Likelihood Pseudo-R2 Number of Observations

−1137.405 0.187 4533

−994.852 0.241 4184

−720.280 0.261 3104

−719.245 0.262 3104

Goodness of fit (10% cutoff) % of correctly predicted observations % of correctly predicted crises % of correctly predicted non-crises

75.35 76.19 75.26

80.35 71.46 81.28

78.77 73.22 79.35

78.45 73.90 78.92

Goodness of fit (1% cutoff) % of correctly predicted observations % of correctly predicted crises % of correctly predicted non-crises

15.93 97.14 7.64

14.58 97.98 5.86

18.36 95.93 10.22

19.07 95.93 11.00

variables Banking Crisis

t -1

Real International Interest Rate

Exchange Rate Overvaluation

Current Account Position

Stock Prices

t -1

t -1

t -3

Capital Account Openness

 Public Debt GDP Growth

t -1

t -1

t -1

t -1

 Domestic Credit by Banking Sector Election

t -1

Political Environment

Market Environment

Money Supply

t -1

t -1

t -3

t -1

BCt -1 X Money Supply

t -1

Notes: Robust standard errors are clustered by country. The significance level of the variables are indicated by ∗ (10%), ∗ ∗ (5%) and ∗ ∗ ∗ (1%). Counter intuitively signed coefficients are represented in italics. Highly significant coefficients with anticipated signs are represented in bold. The marginal effects are evaluated at the sample mean for continuous variables and for change from zero to one for dummy variables holding all other variables at their mean. In order to convert the marginal effects into percentages they should be multiplied by 100.

Please cite this article as: S.C.W. Eijffinger and B. Karatas¸ , Together or apart? The relationship between currency and banking crises, Journal of Banking and Finance, https://doi.org/10.1016/j.jbankfin.2019.105631

ARTICLE IN PRESS

JID: JBF 20

[m5G;September 30, 2019;23:47]

S.C.W. Eijffinger and B. Karatas¸ / Journal of Banking and Finance xxx (xxxx) xxx

Table D3 Single equation pooled probit estimation results of banking crisis. Dependent variable: banking crisis onset

Variables Currency Crisis

t -2

Capital Account Openness

Current Account Position

t -1

t -2

 Public Debt GDP Growth

Election

t -1

t -6

Stock Prices

(2) Estimates (z-stats) elasticity

(3) Estimates (z-stats) elasticity

(4) Estimates (z-stats) elasticity

(5) Estimates (z-stats) elasticity

(6) Estimates (z-stats) elasticity

−1.284∗ ∗ (−2.40) −0.196 −0.277∗ ∗ (−2.38) −0.042 −0.515 (−0.55) −0.079 6.739∗ ∗ (2.40) 1.023 −0.744∗ ∗ (−2.19) −0.114 0.079 (0.71) 0.012 −1.728∗ ∗ ∗ (−5.75) −0.264 0.101 (0.81) 0.017 0.018∗ ∗ ∗ (3.03) 0.003 −0.009 (−1.55) −0.001 0.160 (0.85) 0.024 0.105 (0.56) 0.016 (2) 0.243 (1.93) 0.037

0.961∗ ∗ ∗ (3.27) 0.201 −1.146∗ ∗ (−2.04) −0.159 −0.274∗ ∗ (−2.18) −0.038 −0.816 (−0.79) −0.113 5.442∗ (1.84) 0.756 −0.581∗ (−1.86) −0.081 0.078 (0.80) 0.011 −1.359∗ ∗ ∗ (−5.38) −0.189 0.135 (0.83) 0.021 0.033∗ ∗ ∗ (3.29) 0.005 −0.006 (−0.80) −0.001 0.052 (0.25) 0.007 0.106 (0.57) 0.015 (1) 0.256 (1.84) 0.036

1.090∗ ∗ ∗ (2.97) 0.257 −1.455∗ ∗ (−2.50) −0.212 −0.364∗ ∗ (−2.26) −0.053 −0.144 (−0.20) −0.021 4.592 (1.49) 0.669 −0.605∗ (−1.94) −0.088 0.099 (0.82) 0.015 −1.569∗ ∗ ∗ (−4.18) −0.229 0.061 (0.30) 0.009 −0.038 (−0.53) −0.006 0.008 (0.51) 0.001 0.034 (0.14) 0.005 0.134 (0.65) 0.020 (2) 0.215 (1.33) 0.031 0.898∗ ∗ (2.44) 0.131

1.098∗ ∗ ∗ (2.98) 0.259 −1.469∗ ∗ (−2.53) −0.214 −0.364∗ ∗ (−2.26) −0.053 −0.152 (−0.21) −0.022 4.549 (1.47) 0.661 −0.599∗ (−1.92) −0.087 0.102 (0.83) 0.015 −1.578∗ ∗ ∗ (−4.25) −0.229 0.057 (0.28) 0.009 −0.126 (−1.52) −0.018 0.008 (0.51) 0.001 0.030 (0.12) 0.004 0.135 (0.69) 0.020 (3) 0.215 (1.33) 0.031 0.893∗ ∗ (2.44) 0.139 0.809 (1.40) 0.232

1.048∗ ∗ ∗ (2.67) 0.241 −1.397∗ ∗ (−2.54) −0.200 −0.393∗ ∗ (−2.06) −0.056 −0.100 (−0.14) −0.014 4.794 (1.60) 0.688 −0.587∗ (−1.89) −0.084 0.109 (0.89) 0.016 −1.579∗ ∗ ∗ (−4.25) −0.227 0.070 (0.35) 0.010 −0.031 (−0.46) −0.004 0.008 (0.53) 0.001 0.051 (0.51) 0.007 0.133 (0.68) 0.019 (4) 0.210 (1.30) 0.030 0.940∗ ∗ (2.42) 0.135

0.983∗ ∗ (2.48) 0.222 −1.399∗ ∗ (−2.33) −0.205 −0.365∗ ∗ (−2.31) −0.054 −0.218 (−0.30) −0.032 4.579 (1.51) 0.672 −0.587∗ (−1.95) −0.086 0.103 (0.88) 0.015 −1.553∗ ∗ ∗ (−4.20) −0.228 0.061 (0.31) 0.009 −0.046 (−0.66) −0.007 0.007 (0.49) 0.001 0.029 (0.12) 0.004 0.147 (0.75) 0.022 (5) 0.215 (1.34) 0.031 0.708∗ ∗ (2.05) 0.104

t -1 to t -12

Exch. Rate Overvaluation

Inflation

(1) Estimates (z-stats) elasticity

t -6

t -1

t -1

Real International Interest Rate

Real Domestic Interest Rate

t -1

 Dom. Credit to Private Sector Political Environment

t -2

t -1

t -1

Dependent variable: banking crisis onset Market Environment t -1

 Banking Sector Foreign Debt

t -1

t -2

CCt -1 to

t -12

X Real Int. Int. Rate

CCt -1 to

t -12

X Cap. Acc. Openness

t -1

CCt -1 to

t -12

X Bank. Foreign Debt

t -1

0.122 (0.52) −0.034 1.512 (1.18) 0.566

Log-Likelihood Pseudo-R2 Number of Observations

−1228.604 0.175 4119

−1123.306 0.245 4113

−865.495 0.300 3295

−894.580 0.301 3295

−893.353 0.302 3295

−890.843 0.304 3295

Goodness of fit (10% cutoff) % of correctly predicted observations % of correctly predicted crises % of correctly predicted non-crises

69.19 78.88 67.90

75.22 71.84 75.67

73.35 75.69 73.00

73.38 75.93 73.00

73.26 76.85 72.72

74.29 75.46 74.12

Goodness of fit (1% cutoff) % of correctly predicted observations % of correctly predicted crises % of correctly predicted non-crises

15.39 100.00 4.15

17.02 99.79 6.01

20.42 98.84 8.59

20.70 98.61 8.94

21.58 98.15 10.02

19.91 99.07 7.96

Notes: Robust standard errors are clustered by country. The significance levels of the variables are indicated by ∗ (10%), ∗ ∗ (5%) and ∗∗∗ (1%). Counter intuitively signed coefficients are represented in italics. Highly significant coefficients with anticipated signs are represented in bold. The marginal effects are evaluated at the sample mean for continuous variables and for change from zero to one for dummy variables holding all other variables at their mean. In order to convert the marginal effects into percentages they should be multiplied by 100.

Please cite this article as: S.C.W. Eijffinger and B. Karatas¸ , Together or apart? The relationship between currency and banking crises, Journal of Banking and Finance, https://doi.org/10.1016/j.jbankfin.2019.105631

ARTICLE IN PRESS

JID: JBF

[m5G;September 30, 2019;23:47]

S.C.W. Eijffinger and B. Karatas¸ / Journal of Banking and Finance xxx (xxxx) xxx

21

Table D4 Marginal effects of the interaction terms in banking crisis equation. Interaction of currency crisis with

Mean

Min

Max

Real International Interest Rate (z-stats) Real International Interest Rate if CC t -1 to t -12 = 0 Real International Interest Rate if CC t -1 to t -12 = 1 Capital Account Openness (z-stats) Capital Account Openness if CC t -1 to t -12 = 0 Capital Account Openness if CC t -1 to t -12 = 1 Banking Sector International Debt (z-stats) Banking Sector International Debt if CC t -1 to t -12 = 0 Banking Sector International Debt if CC t -1 to t -12 = 1

0.232 (1.17) −0.017 0.213 −0.034 (−0.62) −0.052 −0.086 0.566 (1.45) 0.095 0.656

0.002 (0.37)

0.309 (4.30)

−0.067 (−2.02)

0.121 (2.34)

−0.279 (−1.77)

0.836 (3.76)

Notes: The marginal effects of the change in the interacted terms are calculated with the “inteff” command (Norton et al., 2004) in STATA 13. The marginal effect of the continuous indicator for two values of dummy indicator is calculated with the “margins” command in STATA 13. Table D5 Bivariate probit estimation results. Dependent variable: currency crisis onset Variables

Real International Interest Rate

Exchange Rate Overvaluation

Current Account Position

Stock Prices

t -1

t -1

t -3

Capital Account Openness

 Public Debt GDP Growth

t -1

t -1

t -1

t -1

 Domestic Credit by Banking Sector Election

t -1

Political Environment

Market Environment

t-1

t -1

t -1

Dependent variable: banking crisis onset Estimates (z-stats) elasticity

Variables

Estimates (z-stats) elasticity

−0.115∗ ∗ ∗ (−4.37) −0.016 −4.540∗ ∗ ∗ (−3.00) −0.646 −0.838 (−1.13) −0.119 −0.872∗ ∗ ∗ (−2.60) −0.124 −0.062 (−0.71) −0.009 0.067 (0.29) 0.010 −0.696∗ (−1.83) −0.099 0.221∗ ∗ ∗ (2.86) 0.031 −0.095 (−0.60) −0.013 0.065 (0.55) 0.009 0.060 (0.51) 0.009

Exchange Rate Overvaluation

Capital Account Openness

Current Account Position

Inflation

t -1

t -2

 Public Debt GDP Growth

t -6

t -1

t -1

Real International Interest Rate

Real Domestic Interest Rate

t -2

t -1

 Domestic Credit to Private Sector Political Environment

Market Environment

ρ

t -1

t -6

Stock Prices

Election

t -1

t-1

t -1

t -1

−2.579∗ ∗ (−2.09) −0.339 −0.250∗ ∗ (−2.06) −0.033 −1.333 (−1.25) −0.175 5.361∗ (1.84) 0.705 −0.594∗ (−1.74) −0.078 0.005 (0.03) 0.001 −1.464∗ ∗ ∗ (−5.45) −0.192 0.032 (0.18) 0.004 0.040∗ ∗ ∗ (3.10) 0.005 −0.008 (−0.96) −0.001 0.076 (0.36) 0.010 0.088 (0.57) 0.011 0.254∗ (1.76) 0.033

0.523 Wald test of ρ = 0: chi2 (1) = 13.688 Prob > chi2 = 0.0002

Log-Likelihood Number of Observations

−1984.154 3748

Notes: The estimation is conducted using the “biprobit” command for STATA 13. Robust standard errors are clustered by country. The significance levels of the variables are indicated by ∗ (10%), ∗ ∗ (5%) and ∗ ∗ ∗ (1%). Counter intuitively signed coefficients are represented in italics. Highly significant coefficients with anticipated sings are represented in bold. The marginal effects are evaluated at the sample mean for continuous variables and for change from zero to one for dummy variables holding all other variables at their mean. In order to convert the marginal effects into percentages they should be multiplied by 100.

Please cite this article as: S.C.W. Eijffinger and B. Karatas¸ , Together or apart? The relationship between currency and banking crises, Journal of Banking and Finance, https://doi.org/10.1016/j.jbankfin.2019.105631

ARTICLE IN PRESS

JID: JBF 22

[m5G;September 30, 2019;23:47]

S.C.W. Eijffinger and B. Karatas¸ / Journal of Banking and Finance xxx (xxxx) xxx

Appendix E. Estimations with alternative banking crises dates Table E1–E4.

Table E1 Kaminsky and Reinhart (1999): Single equation pooled probit estimation results. Dependent variable: currency crisis onset Variables

Banking Crisis

t -1 to t -3

Real International Interest Rate

Exchange Rate Overvaluation

Current Account Position

t -1

t -1

t-1

Stock Pricest-3

Capital Account Openness

 Public Debt GDP Growth

t -1

t -1

t -1

 Domestic Credit by Banking Sector Election

t -1

Political Environment

Market Environment

t -1

t -1

t -1

Dependent variable: banking crisis onset Estimates (z-stats) elasticity

Variablesa

2.709∗ ∗ ∗ (4.64) 0.005 5.125∗ ∗ (2.13) 0.000002 −15.563∗ ∗ ∗ (−3.18) −0.00001 3.256 (1.47) 0.000001 2.104∗ (1.75) 0.000001 −0.465∗ ∗ ∗ (−5.32) −0.0000002 0.093 (1.35) 0.00000004 −0.622 (−0.91) −0.0000003 1.099∗ ∗ ∗ (2.83) 0.000001 −0.051 (−0.08) −0.00000002 1.473∗ ∗ ∗ (2.83) 0.000001 0.351∗ ∗ (2.19) 0.0000001

Currency Crisis

Estimates (z-stats) elasticity t -1 to t -6

Exchange Rate Overvaluation

Capital Account Openness

Current Account Position

t -1

t -1

t -1

Inflationt -6

Stock Prices

t -2

 Public Debt

t -6

GDP Growtht -1

Real International Interest Rate

Real Domestic Interest Rate

t -2

t -1

 Domestic Credit to Private Sector Political Environment

Market Environment

t -1

t -1

t -1

0.147 (0.29) 0.002 −2.659 (−1.52) −0.026 0.072 (0.55) 0.0007 0.636 (0.68) 0.006 6.993∗ ∗ (2.20) 0.068 −3.988∗ ∗ ∗ (−5.05) −0.039 0.044 (0.26) 0.0004 −0.503 (−1.62) −0.005 0.200 (0.08) 0.002 −1.094 (−1.55) −0.011 0.169 (0.89) 0.002 0.014 (0.07) 0.0001 −0.015 (−0.07) −0.0001

Pseudo-R2 Number of Observations Log-Likelihood

0.529 1043 −24.362

Pseudo-R2 Number of Observations Log-Likelihood

0.193 766 −35.896

Goodness of fit (10% cutoff) % of correctly predicted observations % of correctly predicted crises % of correctly predicted non-crises

97.60 55.56 97.97

Goodness of fit (10% cutoff) % of correctly predicted observations % of correctly predicted crises % of correctly predicted non-crises

98.30 12.50 99.21

Goodness of fit (1% cutoff) % of correctly predicted observations % of correctly predicted crises % of correctly predicted non-crises

91.95 100.00 91.88

Goodness of fit (1% cutoff) % of correctly predicted observations % of correctly predicted crises % of correctly predicted non-crises

75.46 87.50 75.33

Notes: Robust standard errors are clustered by country. The significance levels of the variables are indicated by ∗ (10%), ∗∗ (5%) and ∗ ∗ ∗ (1%). Counter intuitively signed coefficients are represented in italics. Highly significant coefficients with anticipated signs are represented in bold. The marginal effects are evaluated at the sample mean for continuous variables and for change from zero to one for dummy variables holding all other variables at their mean. In order to convert the marginal effects into percentages they should be multiplied by 100. Sample period runs from 1985 to 1995; for Indonesia, Malaysia, Philippines, and Thailand from 1985 to 1997. a The composite currency crisis indicator is lagged by 6 months due to the separation problem, and Electiont-1 drops from the estimation due to multicollinearity.

Please cite this article as: S.C.W. Eijffinger and B. Karatas¸ , Together or apart? The relationship between currency and banking crises, Journal of Banking and Finance, https://doi.org/10.1016/j.jbankfin.2019.105631

ARTICLE IN PRESS

JID: JBF

[m5G;September 30, 2019;23:47]

S.C.W. Eijffinger and B. Karatas¸ / Journal of Banking and Finance xxx (xxxx) xxx

23

Table E2 Von Hagen and Ho (2009): Single equation pooled probit estimation results. Dependent variable: currency crisis onset Variablesa

Banking Crisis

t -3

Real International Interest Rate

Exchange Rate Overvaluation

Current Account Position

Stock Prices

t -1

t -1

t -3

Capital Account Openness

 Public Debt GDP Growth

t -1

t -1

t -1

t -1

 Domestic Credit by Banking Sector Election

t -1

Political Environment

Market Environment

t -1

t -1

t -1

Dependent variable: banking crisis onset Estimates (z-stats) elasticity

Variables

Estimates (z-stats) elasticity

0.747 (1.51) 0.019 −0.034 (−0.11) −0.0003 −2.216∗ ∗ ∗ (−2.91) −0.019 −0.923 (−0.62) −0.008 −1.205∗ ∗ (−2.53) −0.010 0.052 (0.60) 0.0004 0.227∗ ∗ (2.12) 0.002 −0.878∗ ∗ ∗ (−3.72) −0.008 0.236∗ ∗ (2.12) 0.002 0.267 (0.58) 0.003 −0.097 (−0.65) −0.001 −0.213∗ ∗ (−2.08) −0.002

Currency Crisis

t -1 to t -3

Exchange Rate Overvaluation

Capital Account Openness

Current Account Position

Inflation

t -1

t -1

t -6

Stock Prices

t -2

 Public Debt GDP Growth

Election

t -1

t -6

t -1

t -1

Real International Interest Rate

Real Domestic Interest Rate

t-2

t -1

 Domestic Credit to Private Sector Political Environment

Market Environment

t -1

t -1

t -1

1.109∗ ∗ ∗ (2.82) 0.058 0.175 (0.16) 0.002 −0.004 (−0.03) −0.0001 −2.489∗ ∗ (−2.13) −0.032 3.153 (1.11) 0.040 −2.391∗ ∗ ∗ (−4.31) −0.031 −0.064 (−0.20) −0.0008 −1.007∗ ∗ (−2.39) −0.013 0.461 (1.57) 0.010 −0.098 (−0.48) −0.001 0.040∗ ∗ (2.17) 0.0005 0.260∗ (1.91) 0.003 0.127 (0.53) 0.002 0.202 (1.28) 0.003

Pseudo-R2 Number of Observations Log-Likelihood

0.198 2050 −78.894

Pseudo-R2 Number of Observations Log-Likelihood

0.204 1839 −87.806

Goodness of fit (10% cutoff) % of correctly predicted observations % of correctly predicted crises % of correctly predicted non-crises

98.78 23.53 99.41

Goodness of fit (10% cutoff) % of correctly predicted observations % of correctly predicted crises % of correctly predicted non-crises

97.82 35.00 98.52

Goodness of fit (1% cutoff) % of correctly predicted observations % of correctly predicted crises % of correctly predicted non-crises

82.29 70.59 82.39

Goodness of fit (1% cutoff) % of correctly predicted observations % of correctly predicted crises % of correctly predicted non-crises

76.73 70.00 76.80

Notes: Robust standard errors are clustered by country. The significance levels of the variables are indicated by ∗ (10%), ∗∗ (5%) and ∗ ∗ ∗ (1%). Counter intuitively signed coefficients are represented in italics. Highly significant coefficients with anticipated signs are represented in bold. The marginal effects are evaluated at the sample mean for continuous variables and for change from zero to one for dummy variables holding all other variables at their mean. In order to convert the marginal effects into percentages they should be multiplied by 100. Sample period runs from 1985 to 2004. a The banking crisis indicator is lagged by one quarter (three months).

Please cite this article as: S.C.W. Eijffinger and B. Karatas¸ , Together or apart? The relationship between currency and banking crises, Journal of Banking and Finance, https://doi.org/10.1016/j.jbankfin.2019.105631

ARTICLE IN PRESS

JID: JBF 24

[m5G;September 30, 2019;23:47]

S.C.W. Eijffinger and B. Karatas¸ / Journal of Banking and Finance xxx (xxxx) xxx

Table E3 Glick and Hutchison (2001): Single equation pooled probit estimation results. Dependent variable: currency crisis onset Variablesa

Banking Crisis

t -3 to t -12

Real International Interest Rate

Exchange Rate Overvaluation

Current Account Position

Stock Prices

t -1

t -1

t -3

Capital Account Openness

 Public Debt GDP Growth

t -1

t -1

t -1

t -1

 Domestic Credit by Banking Sector Election

t -1

Political Environment

Market Environment

t -1

t -1

t -1

Dependent variable: banking crisis onset Estimates (z-stats) Elasticity

Variables

Estimates (z-stats) Elasticity

0.710∗ ∗ ∗ (3.06) 0.006 1.349 (0.82) 0.005 −5.922∗ ∗ ∗ (−4.24) −0.020 −2.262∗ ∗ ∗ (−2.73) −0.008 −0.786 (−0.74) −0.003 −0.063 (−1.11) −0.0002 0.142∗ ∗ (2.45) 0.0005 −0.079 (−0.18) −0.0003 0.332∗ ∗ (2.40) 0.001 0.021 (0.05) 0.0001 0.531∗ ∗ ∗ (3.07) 0.002 −0.075 (0.55) −0.0003

Currency Crisis

t -1 t o t -3

Exchange Rate Overvaluation

Capital Account Openness

Current Account Position

Inflation

t -1

t -1

t -6

Stock Prices

t -2

 Public Debt GDP Growth

Election

t -1

t -6

t -1

t -1

Real International Interest Rate

Real Domestic Interest Rate

t -2

t -1

 Domestic Credit to Private Sector Political Environment

Market Environment

t -1

t -1

t -1

1.245∗ ∗ (2.40) 0.314 −1.423 (−0.89) −0.173 −0.216 (−1.17) −0.026 0.874 (0.71) 0.106 6.842∗ ∗ (2.14) 0.830 −0.529 (−0.96) −0.064 −0.391∗ ∗ ∗ (3.37) −0.047 −2.526∗ ∗ ∗ (−4.85) −0.307 0.199 (1.16) 0.028 1.950∗ ∗ (1.96) 0.237 −0.180 (−1.00) −0.022 0.252 (1.18) 0.031 −0.041 (−0.16) −0.005 0.322∗ (1.81) 0.039

Pseudo-R2 Number of Observations Log-Likelihood

0.287 1710 −61.300

Pseudo-R2 Number of Observations Log-Likelihood

0.262 1374 −360.171

Goodness of fit (10% cutoff) % of correctly predicted observations % of correctly predicted crises % of correctly predicted non-crises

98.42 40.00 98.94

Goodness of fit (10% cutoff) % of correctly predicted observations % of correctly predicted crises % of correctly predicted non-crises

71.62 83.44 70.09

Goodness of fit (1% cutoff) % of correctly predicted observations % of correctly predicted crises % of correctly predicted non-crises

83.92 80.00 83.95

Goodness of fit (1% cutoff) % of correctly predicted observations % of correctly predicted crises % of correctly predicted non-crises

26.71 99.36 17.34

Notes: Robust standard errors are clustered by country. The significance levels of the variables are indicated by ∗ (10%), ∗∗ (5%) and ∗ ∗ ∗ (1%). Counter intuitively signed coefficients are represented in italics. Highly significant coefficients with anticipated signs are represented in bold. The marginal effects are evaluated at the sample mean for continuous variables and for change from zero to one for dummy variables holding all other variables at their mean. In order to convert the marginal effects into percentages they should be multiplied by 100. Sample period runs from 1985 to 1997. a The composite banking crisis indicator is lagged by three months extending to twelve months following the crisis year.

Please cite this article as: S.C.W. Eijffinger and B. Karatas¸ , Together or apart? The relationship between currency and banking crises, Journal of Banking and Finance, https://doi.org/10.1016/j.jbankfin.2019.105631

ARTICLE IN PRESS

JID: JBF

[m5G;September 30, 2019;23:47]

S.C.W. Eijffinger and B. Karatas¸ / Journal of Banking and Finance xxx (xxxx) xxx

25

Table E4 Bivariate probit estimation results.

Variables

K&R Estimates (z-stats) elasticity

H&H Estimates (z-stats) elasticity

G&H Estimates (z-stats) elasticity

3.979 (1.42) 0.00003 −14.423∗ ∗ ∗ (−4.13) −0.0001 3.953 (1.19) 0.00003 3.096∗ ∗ ∗ (4.79) 0.00002 −0.198∗ (−1.75) −0.000001 0.098 (0.86) 0.000001 −1.095∗ (−1.68) −0.00001 0.658∗ ∗ ∗ (2.74) 0.000004 0.165 (0.25) 0.000002 1.117∗ ∗ ∗ (3.21) 0.00001 0.157 (0.68) 0.000001

−0.038 (−0.14) −0.0004 −1.786∗ ∗ (−2.17) −0.020 −0.949 (−0.63) −0.011 −0.737 (−1.33) −0.008 0.067 (0.80) 0.0008 0.245∗ ∗ ∗ (2.70) 0.003 −0.696∗ ∗ ∗ (−2.90) −0.008 0.229∗ ∗ (2.35) 0.003 0.270 (0.62) 0.004 −0.148 (−1.03) −0.002 −0.206∗ (−1.91) −0.002

1.054 (0.70) 0.007 −3.966∗ ∗ (−2.80) −0.026 −2.361∗ ∗ ∗ (−3.72) −0.016 −0.717 (−0.55) −0.005 −0.047 (−0.54) −0.0003 0.238∗ ∗ ∗ (3.01) 0.002 −0.226 (−0.51) −0.002 0.432∗ ∗ ∗ (2.89) 0.003 0.189 (0.53) 0.002 0.487∗ ∗ (2.35) 0.003 −0.052 (−0.43) −0.0003

−2.442 (−1.34) −0.011 0.064 (0.47) 0.0003 1.239 (1.39) 0.005 6.608∗ (1.95) 0.029 −4.875∗ ∗ ∗ (−5.44) −0.022 0.108 (0.72) 0.0005 −0.160 (−0.44) −0.0007 −5.374∗ ∗ ∗ (−6.66) −0.002 −0.157 (−0.07) −0.0007 −1.463∗ ∗ (−1.97) −0.006 0.095 (0.45) 0.0006

−0.263 (−0.21) −0.003 0.153 (1.01) 0.002 −2.260 (−1.72) −0.028 2.448 (0.68) 0.031 −2.745∗ ∗ ∗ (−4.76) −0.034 −0.032 (−0.10) −0.0004 −0.796∗ ∗ (−2.08) −0.010 0.174 (0.47) 0.003 −0.133 (−0.71) −0.002 0.028∗ ∗ ∗ (3.60) 0.0004 0.211∗ (1.82) 0.003

−1.193 (−0.72) −0.137 −0.208 (−1.13) −0.024 1.250 (0.94) 0.144 6.114∗ ∗ (1.94) 0.704 −0.564 (−0.87) −0.065 −0.486∗ ∗ ∗ (−4.13) −0.056 −2.511∗ ∗ ∗ (−4.70) −0.289 0.130 (0.67) 0.016 2.166∗ ∗ (2.18) 0.249 0.270 (1.30) −0.032 0.270 (1.30) 0.031

Dependent variable: currency crisis onset Real International Interest Rate

Exchange Rate Overvaluation

Current Account Position

Stock Prices

t -1

t -1

t -3

Capital Account Openness

 Public Debt GDP Growth

t -1

t -1

t -1

t -1

 Domestic Credit by Banking Sector Election

t -1

t -1

Political Environment

Market Environment

t -1

t -1

Dependent variable: banking crisis onset Exchange Rate Overvaluation

Capital Account Openness

Current Account Position

Inflation

t -1

t -1

t -6

Stock Prices

t -2

 Public Debt GDP Growth

Election

t -1

t -6

t -1

t -1

Real International Interest Rate

Real Domestic Interest Rate

t -2

t -1

 Dom. Credit to Private Sector

t -1

(continued on next page)

Please cite this article as: S.C.W. Eijffinger and B. Karatas¸ , Together or apart? The relationship between currency and banking crises, Journal of Banking and Finance, https://doi.org/10.1016/j.jbankfin.2019.105631

ARTICLE IN PRESS

JID: JBF 26

[m5G;September 30, 2019;23:47]

S.C.W. Eijffinger and B. Karatas¸ / Journal of Banking and Finance xxx (xxxx) xxx Table E4 (continued)

Variables

Political Environment

Market Environment

t -1

t -1

ρ

Wald test of ρ = 0: chi2 (1) Prob > chi2 Log-Likelihood Number of Observations

K&R Estimates (z-stats) elasticity

H&H Estimates (z-stats) elasticity

G&H Estimates (z-stats) elasticity

−0.019 (−0.12) −0.00001 0.012 (0.07) 0.00005 1.000

0.027 (0.12) 0.0003 0.028 (0.20) 0.0003 0.478

−0.024 (−0.09) −0.003 0.263 (1.48) 0.030 0.532

0.215 0.643 −50.323 749

4.854 0.028 −147.971 1803

14.684 0.0001 −396.649 1344

Notes: The estimation is conducted using the “biprobit” command for STATA 13. Robust standard errors are clustered by country. The significance levels of the variables are indicated by ∗ (10%), ∗ ∗ (5%) and ∗ ∗ ∗ (1%). Counter intuitively signed coefficients are represented in italics. Highly significant coefficients with anticipated signs are represented in bold. The marginal effects are evaluated at the sample mean for continuous variables and for change from zero to one for dummy variables holding all other variables at their mean. In order to convert the marginal effects into percentages they should be multiplied by 100.

References Bekaert, G., Harvery, C.R., Lundblad, C., 2011. Financial openness and productivity. World Dev. 39, 1–19. Berg, A., Pattillo, C., 1999. Predicting currency crises: the indicators approach and alternative. J. Int. Money Finance 18, 561–586. Bussiere, M., Fratzscher, M., 2006. Towards a new early warning system of financial crises. J. Int. Money Finance 25, 953–973. Calvo, G.A., 1998. Varieties of capital-market crises. In: Calvo, C., King, M. (Eds.), The Debt Burden and Its Consequences for Monetary Policy. MacMillan Press, London. Calvo, G.A., Leiderman, L., Reinhart, C.M., 1994. The capital inflows problem: concepts and issues. Contemp. Econ. Policy West. Econ. Assoc. Int. 12 (3), 54–66. Caprio, G., Klingebiel, D., 1996. Bank Insolvencies: Cross-Country Experience. World Bank Policy Research Working Paper No. 1620. Caprio, G., Klingebiel, D., Laeven, L., Noguera, G., 2005. Appendix: banking crisis database. In: Honohan, P., Laeven, L. (Eds.), Systemic Financial Crises: Containment and Resolution. Cambridge University Press. Chang, R., Velasco, A., 1999. Liquidity Crises in Emerging Markets: Theory and Policy. NBER Working Paper Series No. 7272. Chinn, M.D., Ito, H., 2006. What matters for financial development? Capital controls, institutions, and interactions. J. Dev. Econ. 81, 163–192. Demirgüç-Kunt, A., Detragiache, E., 1997. The Determinants of Banking Crises: Evidence from Developing and Developed Countries. IMF Working Paper No. 97/106. Demirgüç-Kunt, A., Detragiache, E., 1998. Financial Liberalization and Financial Fragility. IMF Working Paper No. 98/83. Diaz-Alejandro, C.F., 1985. Good-bye financial repression, hello financial crash. J. Dev. Econ. 19 (1–2), 1–24. Eichengreen, B., Arteta, C.O., 20 0 0. Banking Crises in Emerging Markets: Presumptions and Evidence. Mimeo. University of California, Berkeley. Eichengreen, B., Rose, A.K., Wyplosz, C., 1996. Contagious Currency Crises. Centre for Economic Policy Research Discussion Paper No. 1453. Eijffinger, S.C.W., Karatas¸ , B., 2012. Currency crises and monetary policy: a study on advanced and emerging economies. J. Int. Money Finance 31, 948– 974. Falcetti, E., Tudela, M., 2008. What do twins share? A joint probit estimation of banking and currency crises. Economica 72, 199–221. Frankel, J., Rose, A., 1996. Currency crashes in emerging markets: an empirical treatment. J. Int. Econ. 41, 351–366. Glick, R., Hutchison, M., 2001. Banking and currency crises: how common are twins? In: Glick, R., Spiegel, M.M. (Eds.), Financial Crises in Emerging Markets. Cambridge University Press, Cambridge. Goldfajn, I., Valdes, O.R., 1997. Capital Flows and the Twin Crises: the Role of Liquidity. IMF Working Paper No. 97/87. Goldstein, M., Kaminsky, G., Reinhart, C., 20 0 0. Assessing Financial Vulnerability, an Early Warning System for Emerging Markets: Introduction. University Library of Munich, Germany MPRA Paper No. 13629. Ilzetzki, E., Reinhart, C.M., Rogoff, K.S., 2017. Exchange Arrangements Entering the 21st Century: Which Anchor Will Hold?. NBER Working Paper Series No. 23134. Im, K.S., Pesaran, M.H., Shin, Y., 2003. Testing for unit roots in heterogeneous panels. J. Econom. 115, 53–74.

Kaminsky, G.L., 2006. Currency crises: are they all the same? J. Int. Money Finance 25, 503–527. Kaminsky, G.L., Reinhart, C., 1999. The twin crises: the causes of banking and balance-of- payments problems. Am. Econ. Rev. 89 (3), 473–500. Khan, A.A., 2011. Merger and acquisitions (M&As) in the Indian banking sector in post liberalization regime. Int. J. Contemp. Bus. Stud. 2 (11), 2156–7506. Kim, J., Mueller, C.W., 1978. Introduction to Factor Analysis: What it is and How to Do it. Sage Publications. King, G., Zeng, L., 2001a. Logistic regression in rare events data. Political Anal. 9, 137–163. King, G., Zeng, L., 2001b. Explaining rare events in international relations. Int. Organ. 55 (3), 693–715 Summer. Komulainen, T., Lukkarila, J., 2003. What drives financial crises in emerging markets? Emerg. Mark. Rev. 4, 248–272. Kraay, A., 2003. Do high interest rates defend currencies during speculative attacks? J. Int. Econ. 59, 297–321. Laeven, L., Valencia, F., 2008. Systemic Baking crisesCrises: A new New Database. IMF Working Paper No. 224. Laeven, L., Valencia, F., 2012. Systemic Banking Crises Database: An Update. IMF Working Paper No. 163. Lestano, N.V., Jacobs, J., Kuper, G.H., 2003. Indicators of Financial Crises Do Work! An Early-Warning System for Six Asian Countries. University of Groningen CCSO Working Papers Vol. 200313. Levine, R., 2001. International financial liberalization and economic growth. Rev. Int. Econ. 9 (4), 688–702. McKinnon, R.I., Pill, H., 1998. International Overborrowing: A Decomposition of Credit and Currency Risks. Stanford University, Palo Alto, California Working paper. McKinnon, R.I., Pill, H., 1996. Credible liberalizations and international capital flows: the ‘overborrowing syndrome’. In: Ito, T., Krueger, A.O. (Eds.), Financial Deregulation and Integration in East Asia. University of Chicago Press, pp. 7–42. Miller, V., 1996. Speculative currency attacks with endogenously induced commercial bank crises. J. Int. Money Finance 15, 385–403. Mishkin, F.S., 1996. Understanding financial crises: a developing country perspective. In: Bruno, M., Pleskovic, B. (Eds.), Annual World Bank Conference on Development Economics. World Bank, pp. 29–62. Norton, E.C., Wang, H., Ai, C., 2004. Computing interaction effects and standard errors in logit and probit models. Stata J. 4 (2), 154–167. Obstfeld, M., 1994. The logic of currency crises. Cahiers Economiques et Monétaires Bank of France 43, 189–213. Obstfeld, M., Rogoff, K., 1995. The mirage of fixed exchange rates. J. Econ. Perspect. 9 (4), 73–96. Peterson, M.A., 2009. Estimating standard errors in finance panel data sets: comparing approaches. Rev. Financ Stud. 22 (1), 435–480. Reinhart, C., Rogoff, K., 2009. This Time is Different: Eight Centuries of Financial Folly. Princeton University Press, New Jersey. Reinhart, C., Rogoff, K., 2011. From financial crash to debt crisis. Am. Econ. Rev. 101, 1676–1706. Rossi, M., 1999. Financial Fragility and Economic Performance in Developing Countries. IMF Working Papers No. 99/66. Schmukler, S., 2004. Financial Globalization: Gain and Pain for Developing Countries. Federal Reserve Bank of Atlanta Economic Review Second Quarter, pp. 39–66.

Please cite this article as: S.C.W. Eijffinger and B. Karatas¸ , Together or apart? The relationship between currency and banking crises, Journal of Banking and Finance, https://doi.org/10.1016/j.jbankfin.2019.105631

JID: JBF

ARTICLE IN PRESS S.C.W. Eijffinger and B. Karatas¸ / Journal of Banking and Finance xxx (xxxx) xxx

Stoker, J., 1994. Intermediation and the Business Cycle Under a Specie standard: The Role of the Gold Standard in English Financial Crises, 1790–1850. Mimeo, University of Chicago. Torres-Reyna, O., 2012. Getting Started in Factor Analysis (using STATA 10). Princeton University. Velasco, A., 1987. Financial crises and balance of payments crises: a simple model of the southern corn experience. J. Dev. Econ. 27.

[m5G;September 30, 2019;23:47] 27

Von Hagen, J., Ho, T.-.K., 2007. Money market pressure and the determinants of banking crises. J. Money Credit Bank. 39 (5), 1037–1066. Von Hagen, J., Ho, T.-.K., 2009. Twin crises: An Examination of the Empirical Links. Bundesbank Working Paper.

Please cite this article as: S.C.W. Eijffinger and B. Karatas¸ , Together or apart? The relationship between currency and banking crises, Journal of Banking and Finance, https://doi.org/10.1016/j.jbankfin.2019.105631