Demand and Supply Drivers of Foreign Currency Loans in CEECs: A Meta-Analysis
←
→
Page content transcription
If your browser does not render page correctly, please read the page content below
Demand and Supply Drivers of Foreign Currency Loans in CEECs: A Meta-Analysis September 2013 Abstract We present a meta-analysis of the determinants of foreign currency loans in Central and Eastern Europe. We base our inferences on the results of 21 studies that provide around 800 estimated coefficients for seven determinants of foreign currency loan demand. Our results indicate that, on average, supply factors (foreign currency deposits and the minimum variance portfolio ratio) appear to play a more significant role than demand factors (interest rate differentials) of foreign currency loan. Moreover, we show that the estimates reported in literature tend to be influenced by selected study characteristics such as the econometric methodology and their regional focus. JEL: E51, F31, C11. Keywords: Foreign currency loans, meta-analysis, publication bias, meta-regression, Bayesian model averaging.
1 Introduction As part of their financial integration and catching-up process, several Central and Eastern European Countries (CEECs) have built up high stocks of assets and liabilities denominated in foreign currency (mainly in euro or Swiss franc). As a result, high shares of unhedged foreign currency loans have fuelled fears concerning systemic risks with regard to financial vulnerability (see for example Beckmann et al., 2012, EBRD Transition Report 2011).The analysis of the exposure of households and firms to foreign currency risks in CEECs has been gaining importance in the last few years and numerous studies have been published on this topic. Despite a large number of publications on this issue, especially in the recent years, there is no unique agreement in the literature concerning the main driving factors of foreign currency loans in the region (Crespo Cuaresma et al., 2011, Zettelmeyer et al., 2010,). Thus, this contribution presents a meta-analysis of foreign currency loan determinants in CEECs, including publication bias analysis and the explicit assessment of model uncertainty in the specification of meta-analytic regressions. To the best of our knowledge, these issues have hardly been analyzed in the previous literature. In addition, besides applying conventional estimation approaches for meta-analysis (e.g. the random effects model, random effects maximum likelihood model, and weighted least squares), we use Bayesian model averaging (BMA) methods to assess the degree of model uncertainty attached to the meta-regressions for foreign currency loans. The presence of foreign currency loans dominates the private sector indebtedness in some CEECs, while foreign currency loans are basically nonexistent in others. For instance, in the Czech Republic, at one side of the spectrum, the share of foreign currency-denominated private debt has remained low and relatively constant over the past years, being only present in the corporate sector while almost absent in the household sector (Fidrmuc et al., 2012). On the contrary, the shares of foreign currency loans have increased rapidly to about two thirds of total loan stocks in South East European countries (e.g. Croatia and Serbia). Moreover, the phenomenon of foreign currency loans quickly extended from corporate lending to household loans (especially in Hungary and Romania). The previous empirical and theoretical literature has identified several determinants of foreign currency loans in CEECs. The determinants which are most often cited in the studies surveyed include the interest rate differential, the inflation rate, the exchange rate changes, the volatilities of inflation and the exchange rate and their ratio (the so called minimum variance 2
portfolio ratio), as well as the degree of bank funding in foreign currency.1 These determinants reflect demand and supply factors, as well as the interaction of both of them. Therefore, the impact of some determinants may be ambiguous, depending on whether supply or demand motives dominate. While it is difficult to assess the actual importance of demand and supply factors for the individual determinants, we can expect that demand factors dominate the role of the interest rate differential. In contrast, the asset structure of banks can be largely viewed as a supply factor. One interesting finding of our contribution is that the banks’ deposits in foreign currency and the portfolio factors (MVP-ratio) are highly robust determinants of differences in foreign currency loans, while the demand-driven factors yield a much less clear-cut picture. This can be interpreted as an indication that supply factors have actually played a more important role in foreign currency loans than generally argued in the literature (see e.g. Brown and De Haas, 2012, Brown et al., 2010 and 2011, Fidrmuc et al., 2013). Additionally, empirical studies related to foreign currency loans in the CEECs are inherently characterized by significant data problems, a characteristic which is common to many other fields in the natural and social sciences. Meta-analyses of existing studies on a common topic have been suggested as a potentially fruitful way of overcoming such a problem and gaining more powerful results (see e.g. Lipsey and Willson, 2001) by extending the analysis beyond standard literature surveys. In addition to providing a more precise aggregate view on the subject, this approach allows for the analysis of possible factors which may influence the results (e.g. methodology used, data definition, characteristics of the authors). In the past decade, the use of meta-analysis has become a popular research tool in economics (see e.g. Stanley, 2001, Stanley and Jarrell, 2005), in particular in monetary economics (see Knell and Stix, 2005 and Rose and Stanley, 2005, ), and in international economics (see Havránek et al., 2013, Iršová and Havránek, 2013, Hanousek et al., 2011). Transition economics, being especially restricted by data problems, has also become a field of meta-applications (see Campos and Coricelli, 2002, Fidrmuc and Korhonen, 2006). Inference related to the determinants of foreign currency loans is a particularly appropriate field for a meta-analysis. On the one hand, the recent financial crisis has fuelled a general interest for the determinants of foreign currency loans, in particular with the aim of improving the design of economic policy actions. Hence, the number of studies based both on aggregate 1 We do not cover some possible special aspects as for example the perspective of EU accession and euro adoption (see for example Rosenberg and Tirpák, 2009) due to their seldom inclusion in the empirical studies and thus the low number of coefficients. 3
and microeconomic data has mushroomed in the recent years. On the other hand, the robustness of results reported in the literature may be questionable because of the reasons sketched out above. In fact, a central finding of our meta-analysis is that previous studies on foreign currency loans succeed in explaining only a minor fraction of foreign currency lending developments. The majority of coefficients reported for the most common determinants of foreign currency loans are statistically insignificant even at the marginal level of 10%. Does this mean that standard economic theory fails to explain the credit behavior in euroized countries? We try to answer this question and show that only supply factors are relatively robust determinants of foreign currency loans. The paper is structured as follows. Section 2 presents a literature review on the determinants of foreign currency loans in CEECs with the aim of identifying the most common explanatory factors. Section 3 then describes the meta-analysis framework. Section 4 provides the first descriptive evidence of the effects of the selected determinants of foreign currency loans and section 5 continues with the analysis of the publication bias. Section 6 discusses the results of the meta-regression and the robustness of the effects. This analysis is extended by a Bayesian model averaging exercise in section 7. The last section concludes. 2 The Determinants of Foreign Currency Loans: A Review While loan dollarization emerged initially in Latin American countries and early studies analyzed the determinants of the borrowing and lending behavior in foreign currency in this region (e.g. Barajas and Morales, 2003), the number of studies on private sector dollarization in CEECs has substantially increased during the past few years. While an increasing number of empirical studies in recent years are based on survey-based (micro) data, the majority of studies use aggregate data. Therefore they mainly concentrate on the effects of macro-level determinants such as inflation, exchange rate depreciation and their volatilities as well as foreign currency deposits and the interest rate differential. The literature on loan dollarization in the private sector discusses three major factors. The traditional demand side determinants in particular include interest rate differential, while the supply side determinants tend to include the degree of deposit dollarization. This indicator is often explained by the minimum variance portfolio (MVP) approach following Ize and Levy-Yeyati (2003). Finally the last major factor is the ambiguous effects which exchange rates and inflation can exert on foreign currency loans, as these variables are able to reflect both demand and supply side factors (Kapounek and Lacina, 2011). 4
Turning first to the determinants on the demand side; the differential between domestic and euro area loan interest rates is included in most of the empirical studies as an explanatory variable. On the one hand, it reflects the relative price of foreign currency loans. Hence, a higher interest rate differential would induce more borrowing in foreign currency. On the other hand, the real interest rate differential is influenced by macroeconomic stability and its significance could be a result of the trade-off between currency risk (in the case of a large devaluation of the domestic currency) and real interest rate risk (in the case of a lower-than- expected inflation rate). Such a positive effect has been documented for economies of Latin America (e.g. Barajas and Morales, 2003), while the findings for the CEECs are rather mixed. Égert et al. (2007) document that borrowers in countries with a credible peg or currency board regime react more strongly to differentials between local and euro interest rates than borrowers in countries with flexible exchange rate regimes. Similarly, Rosenberg and Tirpák (2009) analyze the determinants of private sector loans in euro and Swiss francs in the new EU Member States and in Croatia and find that the interest rate differential is a robust determinant of foreign currency loans. In contrast, using bank-level data, Brown and De Haas (2010) do not find a significant impact of the interest rate differential (as compared to the euro area) on foreign currency loans. In turn, the supply side factors are usually proxied by the banks’ funding in foreign currency. Basso et al. (2011) argue that currency matching plays a major role in the choice of the currency denomination by borrowers. Not only do banks try to match their open foreign currency positions as a result of regulatory measures (see e.g. Luca and Petrova, 2008), the borrowers also want to match their foreign currency deposits and lending (as shown by Brown and De Haas, 2010). In addition, the large share of remittances in some CEECs (e.g. Bosnia and Herzegovina, Albania, and Romania) also strengthens the matching willingness of the borrowers. Moreover, the MVP-ratio (Ize and Levy-Yeyati, 2003) can also be considered as a predominant supply-side determinant of foreign currency loans. The MVP approach possibly receives the most attention in the empirical analysis of foreign currency loans. The concept derives the optimal share of foreign assets as the ratio of the inflation volatility and real exchange rate volatility,2 2 s MVP 2 , (1) ss2 2 s 2 The importance of the link between inflation and exchange rate volatility for Central and Eastern European economies has been shown for example by Kočenda and Poghosyan (2009). 5
where σ is the variance or covariance of inflation, , and changes in the real exchange rate, s. The MVP-ratio is based on expectations about future developments of the particular factors, although the empirical literature usually substitutes realized data for the respective expectations of inflation and real exchange rate volatility. Basso et al. (2011) confirm that a higher MVP induces a statistically significant higher degree of both deposit and loan dollarization in CEECs, although the effect is negligible in economic terms. By contrast, Neanidis and Savva (2009) find no relationship between loan dollarization and the MVP indicator in the short run. Fidrmuc et al. (2013) use survey data on households’ portfolio behavior, and find MVP to be a significant and robust determinant of households’ foreign currency loans. Finally, the remaining determinants of foreign currency loans could simultaneously express demand and as well as supply side factors. For instance, higher inflation volatility induces more borrowing in foreign currency due to the fact that it can be associated with more stable real interest rates than borrowing in local currency. Hence, the impact of this factor on foreign currency loans depends on the trade-off from the borrower’s and lender’s perspective between currency and real interest rate risks. The empirical literature has shown that the problem of high inflation is less dominant in EU countries. Studies based on aggregate data and survey- based studies generally show a positive effect on loan dollarization (e.g. Zettelmeyer et al., 2010), although some studies also show a significant negative effect from inflation (e.g. Steiner, 2011). Several empirical studies on foreign currency lending in CEECs have also included the real exchange rate depreciation in the domestic currency and its volatility, which can also be determined by demand and supply side factors. The exchange rate is an important factor for output dynamics (Kočenda et al., 2012), thus often fueling lending booms. Moreover, the theoretical impact of exchange rate depreciation may also be ambiguous, as it can have a different impact on the behavior of lenders and borrowers. Lenders may try to shift the exchange rate risks to the borrowers, who in turn would try to avoid it. Empirical evidence suggests that a negative impact reflects the credit default risk of unhedged loans and a positive influence could emerge from the expected stability of repayment rates.3 Barajas and Morales (2003) provide evidence for Latin American countries in which exchange rate volatility tends to reduce credit dollarization in the short run. Luca and Petrova (2008) confirm this result for a large set consisting of 21 transition countries. Rosenberg and Tirpák (2009) find that exchange rate volatility has negative but small effects on the share of foreign 3 The exact effect depends on the elasticity of the euro loan default probability and the rate of depreciation. 6
currency loans in the new EU Member States and Croatia. In contrast to this, past exchange rate volatility is not found to play a significant role, which has been explained by the increase of the perceived stability of the exchange rate due to the EU membership making the economic agents more willing to accept currency risks. 3 A Meta-Analysis of the Determinants of Foreign Currency Loans Empirical studies on the determinants of foreign currency loans tend to build upon linear regression models of the following type, FCLit X it it , (2) where FCL stands for the corresponding share of foreign currency loans or the change in this share, X is a vector of variables that explains the differences in the dependent variable and is an error term assumed to fulfill the usual assumptions of the standard linear regression model. Usually equation (2) is estimated for sectors in one or more countries, which are indexed by i, while the time period is indexed by t. Similarly, microeconomic (survey) studies, where the dependent variable is a dummy which measures whether a given individual borrower (firm or household) has taken a foreign currency loan, estimate usually the following limited dependent variable model P(FCLit = 1 | X ) = F ( + X it ), (3) where F(.) is a nonlinear function, usually the cumulative normal distribution function for probit models or the logistic function for logit models. Similarly to the previous estimation, equation (3) is estimated for individual borrowers (possibly in several countries) indexed by i over time t. Although the comparability of micro econometric results with macroeconomic studies is questionable, it is to be kept in mind that all the reviewed studies using specifications such as those in equation (3) report marginal effects whose interpretation is similar to the elasticities reported in a standard OLS regression. Moreover, the conditional expectation of the dependent variable in equation (3), E( FCLit | X it ) , can be interpreted as the share of foreign currency loans held by individuals (households) or a firms in the whole population of individual firms or households. Using the corresponding parameter estimates from the 21 studies (see Table A.1 in appendix) that dealt with the determinants of foreign currency loans in CEECs, we estimate meta- regression models which aim to explain the differences in the estimated coefficients. A typical meta-regression is therefore given by ̂ lm Dlm ulm (4) 7
where ˆ is the estimate corresponding to variable l in study m and D is a vector containing variables reflecting various characteristics of the study. It is further assumed that u is the regression error term, which may have a different distribution for each of the analyzed studies. The vector D includes both continuous and dummy variables, which summarizes information related to data definition, data structure, estimation method, publication and included control variables. As control variables in our meta-regressions we include information on the sample used and the characteristics of the study (for the exact definitions of the control variables, see table A.2 in the appendix). The year of publication of the study shows whether there is a trend in the overall analysis of foreign currency loans. This could correspond to actual structural changes (e.g. the increasing role of foreign currency loans, changes in the strategies of foreign banks acting in CEECs) or to generally accepted views on determinants of foreign currency loans. Next, we differentiate between publications using aggregate data (base category) and micro- econometric datasets. We define dummy variables for models with fixed effects (such as country, region or firm fixed effects)4 and special factors that are not frequently included in the estimation. Such factors are for example the foreign currency income of the borrowers (from own export activities in the case of firms or remittances in the case of households), as well as an EU enlargement variable, which indicates whether a study accounts for the EU accession or the expectation of euro adoption. We also include a dummy variable for publications after the beginning of the recent financial crisis. Further variables describe the geographic focus of the paper, e.g. data including Latin American and former Soviet Union countries, and the exclusion of currency board countries (Bulgaria, Bosnia-Herzegovina, and the Baltic states) from the sample. With the exception of the publication year variable, which is measured as the deviation from the mean publication year in our sample, all other variables are binary covariates. Since not all regression models reported in the studies in our sample include information concerning these covariates, our meta-regression specifications do not include all these variables for each of the parameters of interest. We present several estimates based on the alternative specifications of the equation (4). First, we start with a random effects model that does not explicitly take into account the relative precision of individual estimates (i.e. their significance). We consider possible heteroscedasticity in the meta-regression and report standard errors clustered at the level of individual studies. In addition, we employ a weighted least squares (WLS) estimation by 4 Particularly studies using survey-level data (e.g. Brown et al., 2009 and 2011) have underlined the importance of including country fixed effects for the statistical significance of the coefficient estimates. 8
using the precision of each parameter estimate (measured by the inverse of their standard errors or standard deviation) as a weight in the regression. This approach is consistent, for instance, with Knell and Stix (2005), although the controversy of the weighting approach has been acknowledged by various authors (e.g. Krueger, 2003). While the application of WLS can account for the relative precision of the estimates within the specification given by (4), it cannot deal with the potential heterogeneity in estimates across studies. In particular, if we assume that the true value of 1 can only be imperfectly approximated by Dm , so that l Dlm l , where l is a normally distributed random variable with zero mean and variance 2 equal to the standard error reported for ̂ lm in individual studies, then (4) can be written as ̂ lm Dlm l ulm . (5) Under the assumption that l and u lm are uncorrelated, the model in equation (5) can be estimated by using random maximum likelihood methods (see e.g. Thompson and Sharp, 1999). This specification is thus able to account for the between-study variance, 2 , and the individual variance of the estimate when accounting for the relative precision across the observed values of ̂ lm . 4 Meta-Statistics For our meta-analysis we employ estimates from 21 papers which conduct empirical work with the use of data on foreign currency loans in CEECs.5 Together, they provide nearly 800 estimates for seven determinants, with the interest rate differential being most often included in the studies. Most of these studies rely on aggregate data (estimating the share or the change in the share of foreign currency loans in the private sector as the main variable of interest) although a growing number of studies are being based on survey-level data of firms or households. Actually, an advantage of our regional focus on CEECs as opposed to Latin 5 We searched EconLit Database using keywords “foreign currency” and “loans”, “lending” or “borrowing”, which were reviewed to check whether they presented regression analysis of determinants of foreign currency loans in CEECs. This search identified about 20 studies. Several papers were published first as working papers and then as journal articles. Both versions were surveyed and included in the meta-regressions unless the journal article was fully identical to the working paper version. We completed the meta-analysis by few studies (PhD thesis and forthcoming papers) which were quoted by other sources. The exclusion of these studies has no influence on the results. The literature search was carried out in September 2012. 9
American countries is the availability of results from survey data (mostly data on firms; two papers use bank-level data and two papers include household survey data). Table 1 reports some descriptive statistics for the coefficients of the determinants in the studies surveyed. The coefficients of the determinants of foreign currency loans seem to be rather similar despite differences in data types or the dependent variable (aggregate share of loans or share of respondents with a foreign currency loan). We can draw three important conclusions from the descriptive evidence given in Table 1. First, the average estimated coefficient for the interest rate differential over all available estimates is surprisingly low. The corresponding t-tests fail to reject the null hypothesis that the reported coefficients are zero not only for the interest rate differential, but also for exchange rate volatility and inflation. Second, there is a substantial variation for all analyzed determinants of foreign currency loans. Third, an intriguing result from the descriptive evidence is that the share of significant coefficients in the studies surveyed is mostly around 50% or even lower for all determinants (with the exception of foreign currency deposits, where 76% of the coefficients are significant). However, this descriptive discussion does not take the differences between studies explicitly and systematically into consideration. As the variation across subsamples of coefficients indicates, it may well be that the coefficients are influenced by certain specific characteristics of the individual studies and that correcting these particularities would lead to a more coherent picture. 5 Publication Bias Analysis The analysis of economic policy issues may often be the subject of general expectations on findings. This can result in an unintended publication bias if authors, reviewers and publishers follow their preference for statistically significant and theoretically sound results. Publication bias or publication selection is a term often used also for other types of selection biases that lead to estimates which are asymmetrically distributed around the true effect. Such a phenomenon can be detected by a so called funnel plot, which is a scatter diagram displaying the precision (e.g. the inverse of the standard error) against the estimated effect. If publication bias is insignificant, the funnel plot should look like an inverted funnel and the estimates should vary symmetrically around the true effect. The estimates which are close to the true effect should be characterized by the highest precision. Similarly, the least precise estimates should be located in the lower part of the chart. In contrast, if publication selection leads to an 10
overrepresentation of significant results in the sample, the funnel plot becomes hollow and unduly wide. However, we have to keep in mind that the funnel plot is a subjective tool for the detection of publication bias. Stanley (2005) notes that modeling issues (for example omitted variables, estimation techniques) can likewise be a source of misspecification bias that can be wrongly attributed to the publication bias. Bearing in mind these limitations, we examine the funnel plots for the parameter estimates corresponding to our covariates of interest, which are displayed in Figure 1. The precision (y- axis) is defined as the inverse of the standard error. Evidence for publication bias can be seen for the parameters associated to several of the analyzed variables, especially for the MVP- ratio and the interest rate differential. Other variables, particularly exchange rate depreciation and volatility (and to a lesser degree inflation volatility) do not exhibit the expected funnel form, underlining the heterogeneity of these results. Nevertheless, a visual examination of the funnel plots is often not conclusive in the detection of asymmetry. To test for symmetry in a more formal manner, we employ the funnel asymmetry test (FAT). It is based on the simple meta-regression of available effects and the corresponding standard errors (Card and Krueger, 1995; Ashenfelter et al., 1999) ˆki = k SEki + k + ki (6) where ̂ ki denotes the reported estimates of the particular kth coefficient, that is assumed to vary around the so called “true” effect , while stands for the so called publication bias. If the estimates are distributed symmetrically around the true effect, , then the coefficient should not be significantly different from zero. If, however, there is a tendency to report certain parameter values or significant results, would be non-zero and significant and the estimate would be proportional to the standard error. Thus, the publication selection bias can be detected through the correlation of reported effects and their standard errors. To deal with the differences in quality across studies, it is recommended to use inverse standard errors as weights, which gives more weight to precise than to imprecise effects. This means that equation (6) is rewritten as ˆ ki 1 t ki k k ki (7) SEki SE ki This equation puts the t-statistic for significance in relationship with the inverse standard error. Following Egger et al. (1997), we test for 0 , using standard and weighted versions of the FAT test, whose rejection confirms the presence of publication bias (presence of 11
asymmetry). We use OLS with clustered standard errors and random effects for the individual studies. Similar to the funnel plots, Table 2 shows that all test specifications confirm publication bias for the MVP-ratio, and nearly all do so as well for the interest rate differential. There is only weak evidence (see OLS results) that publication bias exists also for foreign currency deposits and for both of the exchange rate determinants. Finally, no publication bias is robustly revealed for the inflation variables. 6 Meta-Regression Analysis Meta-regressions can identify the adjusted effect of the individual determinants of foreign currency loans, which is shown by the corresponding intercept in the specification. In addition, we also discuss the impact of the included control variables as described in the appendix on the performance of the meta-regression. Tables 3 to 5 present the results of the meta-regression model given by equation (4) for our seven most common determinants of foreign currency loans using random effects (RE), weighted least squares (WLS), and random effects maximum likelihood (REML). For each of the seven determinants, we perform separate meta-regressions. We use robust standard errors clustered at the level of individual studies. In the case of WLS, we apply the inverse standard errors of the coefficient estimates as weights. Thus, more precise coefficients (i.e coefficients with smaller standard errors) receive a larger weight in the regression (table 4). Even though the results are largely similar, the REML estimates can be expected to provide the most reliable findings as they account for the variance both within and between studies. The interest rate differential is commonly viewed as the main explanatory factor of foreign currency loans on the demand side and is included in almost all studies (apart from a few microeconomic estimations). Due to the larger number of observations of this coefficient (288), it can be expected that the results are more clear-cut than for other determinants of foreign currency loans. However, the intercept through the meta-regression appears mostly insignificant with the exception of the REML estimates. It is even negative for the RE specification, although it is counteracted by the large and positive (but insignificant) coefficients for the main results. Additionally, the interest rate differential appears to have played a more important role in earlier years, as shown by a negative coefficient for the last year of the sample period. In spite of the latter result, the dummy for the financial crisis is positive and significant, hence implying that the effects of the financial crisis are rather unclear, because the crisis dummy and time trend (expressed by the last year of observation) have counteracting effects. In general, micro-econometric studies report lower interest rate semi-elasticities as compared to estimates from macroeconomic analyses. This implies that, 12
even though the variable has been widely emphasized in literature, interest rate differentials do not appear to play a major role in the demand for foreign currency. Moreover, this result is irrespective of the applied meta-approach. This finding casts some doubt on whether demand factors are actually a main determinant of foreign currency loans in CEECs. A large part of the reviewed studies likewise include the MVP-ratio according to Ize and Levy-Yeyati (2003), as defined by equation (1). Our meta-analysis confirms that the intercept is highly significant in the MVP estimations. Though interestingly, as the CIS dummy shows, the MVP ratio might play a smaller role in former Soviet Union countries, corresponding to the less developed financial markets of this region. Several authors use inflation and exchange rate volatility separately (e.g. Brown and De Haas, 2010; Zettelmeyer et al., 2010). The estimations confirm the importance of inflation volatility, which is in line with the lack of a monetary credibility argument as proposed by Fidrmuc et al. (2013). In turn, exchange rate volatility exerts a less robust influence on the foreign currency loans. Overall, the meta-regression results detect inflation volatility as the more important driver of foreign currency loans. Supply factors are often proxied for by the share of foreign currency deposits in total deposits. Banks with high deposits in foreign currency have an interest in shifting currency risk towards their customers. In fact, the MVP indicator is usually found to be highly correlated with the degree of deposits in foreign currency. Our meta-analysis confirms that foreign currency deposits are, on average, a significant determinant of foreign currency loans. The meta-regression analysis shows that the determinants of foreign currency loans which are related mainly to supply factors (foreign currency deposits and the MVP-ratio) keep on average the expected signs, while demand side factors (the interest rate differential) show on average rather ambiguous results. In addition to the meta-regressions based on the whole sample of estimates in each study, we also used the subsamples of results which were reported as “preferred” or “baseline” specifications in the analyzed studies. However, one has to be aware of the smaller number of coefficients in this case. The results of the meta-regressions for these subsamples of estimates confirm the important role of the MVP-ratio, exchange rate volatility and foreign currency deposits, while the other determinants are insignificant.6 It is rather difficult to assess the overall effect of all explanatory variables given a best- practice approach (Havránek and Iršová, 2011), because different authors may also have different preferences regarding the choice of an appropriate specification. Nevertheless, there is a general agreement that the estimations should use methods which address the endogeneity 6 Detailed results are available upon request. 13
problems. Similarly, panel estimations (fixed effects) are likewise superior to cross-sections and time-series analysis. Finally, micro econometric approaches have recently gained in importance and popularity, but the impact of macroeconomic variables (e.g. interest rate differentials) is often difficult to estimate using micro-econometric data. Correspondingly, it may be more difficult to compare studies using macroeconomic and survey data. While an overall evaluation should reflect the preferences of a reader, a comparison of several approaches shows that by and large the effects remain similar to the base specification. In general, variables related to methodology do not seem to have large and significant effects. Moreover, they often work in different directions. More importantly, results based on post- crisis samples strengthen the effects found with data prior to the crisis. 7 Bayesian model averaging In addition to the estimation of the single specification in the previous section, we extend our analysis using Bayesian model averaging of specifications in the model space spanned by specifications of the type given by equation (4). Assuming that there are K different potential explanatory covariates and that all (linear) combinations of them can form a model of the type described in (4), 2K specifications can be considered in order to assess the role played by the characteristics of the analysis on the estimates. The lack of theoretical guidance concerning the choice of independent variables implies that the specification choice may be an important source of uncertainty when it comes to quantifying how the methods and samples used in empirical studies affect the elasticities of foreign currency loans in CEECs to the variables of interest. Some recent studies have paid attention to model uncertainty in the framework of meta- regression analysis (Moeltner and Woodward, 2009). As in Iršová and Havránek (2012), we use Bayesian model averaging (BMA) methods to assess the degree of model uncertainty attached to the meta-regressions for foreign currency loans. In particular, following the standard BMA literature (see Doppelhofer, 2008, for a review on model averaging methods in economics) we obtain weighted averages of parameter estimates in meta-regressions in which the weights are given by the posterior model probabilities of the corresponding specifications. Let the posterior model probability of a model, Mi, be given by P(Mi|Y). Using Bayes theorem, such a posterior probability is proportional to the product of the prior probability of the model, P(Mi) and its marginal likelihood, P(Y| Mi). For a given choice of priors over the model space and over the parameters of each individual model, the posterior model probability can be obtained using standard Bayesian methods (see Koop, 2003). Following the 14
recent literature on model averaging in econometrics, we use a g-prior over the parameters of individual models (see Fernández et al., 2001a and 2001b) and a beta-binomial prior on the inclusion of individual covariates (see Ley and Steel, 2009). Using these prior settings, we obtain model-averaged estimates of the parameters in (4) and compute the posterior inclusion probability of each covariate, which is defined as the sum of posterior model probabilities of specifications including that variable. This statistic thus summarizes the importance of the covariate when it comes to explaining differences in estimates, ˆ , across studies. The results of our BMA exercise are presented in Table 6. We obtained BMA estimates for specifications based on the coefficients of the interest rate differential, exchange rate depreciation, exchange rate volatility, inflation, inflation volatility, minimum variance portfolio and foreign currency deposits. As above, the set of covariates used for the BMA analysis differs between coefficients depending on the number of available observations of the covariates of the meta-regressions, as well as on the existence of perfect multicollinearity across explanatory variables. Since all models estimated include a constant, the posterior inclusion probability of the intercept in equation (4) is equal to one in all cases. 7 Table 6 indicates that the power of explanatory variables in terms of discriminating robustly across the results of different studies is strongly dependent on the elasticity that we are trying to explain. While none of the variables used in the analysis appears robust in explaining differences in the coefficient of interest rates and inflation across empirical studies, some characteristics of the analysis are able to discriminate across coefficient estimates for the rest of the estimates analyzed. For the remaining determinants, the last year of the sample, the dummy for the financial crisis and methodological variables appear to be robust covariates in the meta-regressions. Using the BMA results, we concentrate on the median probability model (proposed by Barbieri and Berger, 2004), that is, the specification which includes those variables which have a posterior inclusion probability above 0.5 as covariates for each dependent variable. Through the use of the median probability model, we are able to obtain the predicted estimate for each coefficient that corresponds with the median values of the explanatory variable. We dub such an estimate the median probability meta-estimate and report it in the last row of Table 6, together with the corresponding standard deviation. The high degree of uncertainty 7 We used Markov Chain Monte Carlo (MCMC) methods to explore the model space. The results presented are based on 100.000 models visited by the chain. Computing the correlation between the analytical and simulated posterior density over models for selected subsamples of the visited set of specifications indicates that the chain converged. 15
surrounding the estimates is supported by the fact that only a half of the median probability meta-estimates is significantly different from zero. The median probability meta-estimates for the supply side indicators, especially for the degree of deposit dollarization, are particularly close to the mean values of the coefficients. This finding supports the existence of a significant effect of these determinants on foreign currency loans. 8 Conclusions Foreign currency loans have played an important role in the catching-up process of the majority of the CEECs and have gained increasing attention from economic researchers in the past few years. In particular, the financial crisis intensified the attention paid to the lending in foreign currency due to its negative impact on the financial stability in the CEECs. Overall, we were able to find 21 empirical studies on the determinants of foreign currency loans in the CEECs. From these studies, we collected about 300 estimation equations that roughly provided 800 coefficients on the seven most common determinants. Furthermore, while the literature underlines the importance of foreign currency loans, it also provides a highly ambiguous picture of their effective determinants. The behavior of individuals and firms concerning foreign currency loans is determined not only by economic and monetary policies, but also through more general political developments. Additionally, once lenders get used to foreign currency, it may take a relatively long period of time to change their behavior again, which in turn has an indirect impact on foreign currency borrowing. Although several papers have been published by high-ranking journals or working paper series with a strong policy impact, we can see that only roughly half of the published coefficients are actually statistically significant. Moreover, a similarly wide differentiation can be perceived in the economic significance of the coefficients. However, these differences are not so surprising when considering the heterogeneity of the analyzed data and the methods applied. The literature analyzes developments in highly heterogeneous countries, including the EU Member States and the Western Balkan countries or CIS states. Some studies also choose a more general focus and additionally include developing countries, with a tendency to Latin American countries. Similarly, the level of data aggregation is highly diversified. Some papers look at the aggregate share of foreign currency loans, while others focus more on the foreign currency loans of individual firms and households based on survey-level data. Despite these differences, we try to identify common ground, by using meta-statistics and meta-regressions. Our results show that especially supply side indicators and foreign currency 16
deposits play a robust role. The Bayesian model averaging analysis shows these variables to being robust and explanatory, therefore making them significant determinants of the foreign currency loans. Similarly, we also show that several standard determinants play only a minor role in loan euroization. For example, foreign currency loans are often viewed as a result of high domestic interest rates. From this perspective, borrowers take an excessive risk when taking up foreign currency loans and underestimating the danger of exchange rate depreciation. We show that, on average, interest rate differentials do not influence the currency selection for loans. Thus, foreign currency loans do not seem to be only a result of the demand of borrowers in the CEECs. In addition, we document that exchange rate movements do not play a clear role in foreign currency loans. On the one hand, the exchange rate depreciation (not its volatility) does not robustly influence the demand for foreign currency loans. This finding corresponds to the previously mentioned weak results for interest rate differentials. Yet on the other hand, it also shows that borrowers in the CEECs, may well underestimate the potential losses from depreciations, despite their experience with trend appreciation before the financial crisis. Altogether, our findings confirm that supply factors play a more significant role than demand factors for foreign currency loans in CEECs. References Ashenfelter, O., Harmon, C., Oosterbeek, H., 1999. A review of estimates of the schooling/earnings relationship, with tests for publication bias. Labour Economics 6(4), 453-470. Barajas, A., Morales, R.A., 2003. Dollarization of liabilities: Beyond the usual suspects. IMF Working Paper 03/11. Barbieri, M., Berger, J., 2004. Optimal predictive model selection. Annals of Statistics 32,870-897. Basso, H.S., Calvo-Gonzalez, O., Jurgilas, M., 2011. Financial dollarization: The role of foreign-owned banks and interest rates. Journal of Banking and Finance 35(4), 794-806. Beckmann, E., Fidrmuc, J., Stix, H., 2012. Foreign Currency Loans and Loan Arrears of Households in Central and Eastern Europe. Working Paper No. 181, Oesterreichische Nationalbank, Vienna. Brown, M., De Haas, R., 2010. Foreign currency lending in Emerging Europe: Bank-level evidence”. EBRD Working Paper 122/2010. 17
Brown, M., De Haas, R., 2012. Foreign banks and foreign currency lending in Emerging Europe. Economic Policy 69, 57–98. Brown, M., Kirschenmann, K., Ongena, S., 2010. Foreign currency loans - demand or supply driven?. CEPR Discussion Paper 7952. Brown, M., Ongena, S., Yesin, P., 2011. Foreign currency borrowing by small firms. Journal of Financial Intermediation 20(3), 285-302. Campos, N. F., Coricelli, F., 2002. Growth in transition: What we know, what we don't, and what we should. Journal of Economic Literature XL, September, 793-836. Card, D., Krueger, A.B., 1995. Time-series minimum-wage studies: A meta-analysis. American Economic Review 85(2), 238-243. Crespo Cuaresma, J., Fidrmuc, J., Hake, M., 2011. Determinants of foreign currency loans in CESEE countries: A meta-analysis. Focus on European Economic Integration, OesterreichischeNationalbank, No. 4, 69-87. Doppelhofer, G., 2008. Model averaging. The New Palgrave Dictionary of Economics.Second Edition.Eds. Steven N. Durlauf and Lawrence E. Blume. Palgrave Macmillan, 2008. EBRD. 2011. Towards a pan-European banking architecture.Transition Report 2011, Chapter 3, 44-61. Egger, M., Smith, G. D., Scheider, M., Minder, C., 1997. Bias in meta-analysis detected by a simple, graphical test. British Medical Journal 316, 629-634. Égert, B., Backé, P., Zumer, T., 2007. Private-sector credit in Central and Eastern Europe: New (Over) shooting stars? Comparative Economic Studies.49 (2).201-231. Fernández, C., Ley, E., Steel, M.F., 2001a. Benchmark priors for Bayesian model averaging. Journal of Econometrics 100, 381-427. Fernández, C., Ley, E., Steel, M.F., 2001b.Model uncertainty in cross-country growth regressions. Journal of Applied Econometrics 16, 563-576. Fidrmuc, J., Korhonen, I., 2006. Meta-Analysis of the business cycle correlation between the Euro Area and the CEECs. Journal of Comparative Economics 34 (3), 518-537. Fidrmuc, J., M. Hake, Stix, H., 2013. Households’ foreign currency borrowing in Central and Eastern Europe. Journal of Banking and Finance 37(6), 1880–1897. Hanousek, J., Kočenda, E., Maurel, M., 2011. Direct and indirect effects of FDI in emerging European markets: Survey and Meta-analysis. Economic Systems, 35(3), 301-322. Havránek, T., Iršová, Z., 2011. Estimating vertical spillovers from FDI: Why results vary and what the true effect is. Journal of International Economics 85(2), 234-244. 18
Havránek, T., Horváth, R., Rusnák, M.,2013. How to solve the price puzzle? A Meta-analysis. Journal of Money, Credit and Banking 45(1), 37-70. Ize, A., Levy-Yeyati, E., 2003. Financial dollarization. Journal of International Economics 59, 323-347. Iršová, Z., Havránek, T., 2013. Determinants of horizontal spillovers from FDI: Evidence from a large meta-analysis. World Development 42(C), 1-15. Kapounek, S., Lacina, L., 2011. Inflation perceptions and anticipations in the old eurozone member states. Prague Economic Papers 2011(2), 120-139. Knell, M., Stix, H., 2005. The income elasticity of money demand: A Meta analysis of empirical results. Journal of Economic Surveys 19(3), 513-533. Kočenda, E., Poghosyan, T., 2009. Macroeconomic sources of foreign exchange risk in new EU members. Journal of Banking and Finance, 33(11), 2164-2173. Kočenda, E., Maurel, M., Schnabl, G., 2012. Short-Term and Long-Term Growth Effects of Exchange Rate Adjustment.Working Paper Series 4018, CESifo Munich. Koop, G., 2003. Bayesian Econometrics, John Wiley and Sons. Krueger, A.B., 2003. Economic considerations and class size. Economic Journal 113(485), F34–F63. Ley, E., Steel, M.F., 2009. On the effect of prior assumptions in Bayesian Model averaging with applications to growth regression. Journal of Applied Econometrics 24, 651–674. Lipsey, R., Soydan Wilson, D.B., 2001. Practical Meta analysis. Applied Social Research Methods Series. Vol.49. London: SAGE Publications. Luca, A., Petrova, I., 2008. What drives credit dollarization in transition economies? Journal of Banking and Finance, 32, 858-869. Moeltner, K., Woodward, R., 2009. Meta-functional benefit transfer for wetland valuation: Making the most of small samples. Environmental & Resource Economics 42, 89–108. Neanidis, K.C., Savva, C.S., 2009. Financial dollarization: Short-run determinants in transition economies. Journal of Banking and Finance, 33, 1860-1873. Rose, A.K., Stanley, T., 2005. A Meta-analysis of the effect of common currencies on international trade. Journal of Economic Surveys 19(3), 347-365. Rosenberg, C., Tirpák, M., 2008. Determinants of foreign currency borrowing in the new Member states of the EU. IMF Working Paper 173. IMF. Rosenberg, C., Tirpák, M., 2009. Determinants of foreign currency borrowing in the new member states of the EU. Czech Journal of Economics and Finance, 59(3), 216-228. 19
Stanley, T.D., 2001. Wheat from chaff: Meta-analysis as quantitative literature review. Journal of Economic Perspectives 15 (3), 131-150. Stanley, T.D., 2005. Beyond publication bias. Journal of Economic Surveys 19, 309-345. Stanley, T.D., Jarrell, S.B., 2005. Meta-regression analysis: A quantitative method of literature surveys. Journal of Economic Surveys, 19(3), 299-308, Steiner, K., 2011. Households’ exposure to foreign currency loans in CESEE EU member states and Croatia. Focus on European Economic Integration Q1/2011, 6-24. Thompson, S.G., Sharp, S.J., 1999. Explaining heterogeneity in meta-analysis: A comparison of methods. Statistics in Medicine 18, 2693–2708. Zettelmeyer, J., Nagy, P., Jeffrey, S., 2010. Addressing private sector currency mismatches in Emerging Europe. EBRD Working Paper No. 115. 20
Figure 1: Funnel Plots for Selected Determinants Interest Rate Differential MVP-Ratio Exchange Rate Depreciation Interest Rate Minimum Variance Portfolio Exchange Rate Depreciation 100 200 50 80 40 150 60 30 Precision Precision Precision 100 40 20 50 20 10 0 0 0 -5 0 5 10 -5 0 5 10 -2 -1 0 1 2 Coefficient Coefficient Coefficient Exchange Rate Volatility Inflation Volatility Inflation Exchange Rate Volatility Inflation Volatility 30 Inflation 1.5 20 15 20 1 Precision Precision Precision 10 10 .5 5 0 0 0 -.5 0 .5 1 1.5 -10 -5 0 5 10 15 Coefficient Coefficient -10 -5 0 5 Coefficient Foreign Currency Deposits Foreign Currency Deposits 40 30 Precision 20 10 0 -1 -.5 0 .5 1 1.5 Coefficient 21
Table 1: Meta-Statistics No of Share of Variable Mean Std. Dev. t-test Min Max observ. sign. coeff. Interest Rate Differential 288 0.053 0.958 0.930 -2.988 9.300 49.7 Exchange Rate Volatility 76 -0.139 1.273 -0.953 -4.017 7.069 23.7 Exchange Rate Depr. 120 0.129 0.579 2.450** -2.411 1.310 51.7 Minimum var. portfolio 183 0.143 0.772 2.506** -4.898 7.511 54.1 Inflation 88 -0.095 1.611 -0.553 -9.700 5.700 33.0 Inflation Volatility 77 0.969 3.696 2.300** -10.010 12.540 50.6 Foreign Currency Deposits 88 0.272 0.408 6.250*** -1.047 1.329 72.7 Note: a – The t-test tests whether the mean of reported coefficients is significantly different from 0. **, and *** stands for significance at the 5%, and 1% level, respectively. 22
Table 2: Funnel Asymmetry Test Estimation Interest MVP Exch. Rate Exch. Rate Inflation FC Inflation Method Rate Diff. Ratio Depr. Volatility Volatility Deposits OLS 0.238*** 0.361*** 0.330* 0.847 0.264 -0.003 0.000*** (0.029) (0.014) (0.173) (1.004) (0.159) (0.007) (0.000) RE 0.238*** 0.361*** 0.280*** 1.423** 0.264* -0.003 0.000*** (0.029) (0.014) (0.096) (0.599) (0.159) (0.007) (0.000) WLS 3.407** 1.343*** 2.829 -0.274 0.889 0.387 18.742 (1.529) (0.135) (3.333) (0.660) (0.504) (0.289) (18.129) WRE 8.963 1.451*** 13.899 -0.306 0.340 0.852 18.742 (6.403) (0.261) (13.628) (0.529) (0.494) (0.560) (18.129) Observations 288 183 120 76 77 88 88 No of studies 20 9 11 9 8 8 10 Note: We report coefficients and their t-statistics according to equation (6) and (7) for the OLS and random effect (RE) specifications, respectively. WLS and WRE denote the results for weighted tests (weighted by inverse standard errors). Robust standard errors in parentheses. *, **, and *** stands for significance at the 10%, 5%, and 1% level, respectively. 23
Table 3: Meta-Regression, Random Effects Interest MVP Exch. rate Exch. rate Inflation FC Inflation rate diff. ratio depr. volatility volatility deposits intercept -0.073 2.703** 0.499** 5.529*** 1.916 -10.198*** 0.685*** (0.339) (1.280) (0.198) (1.279) (2.284) (1.821) (0.258) main results 0.229 -0.074 -0.044 0.072 -0.765 -1.064 0.086 (0.287) (0.099) (0.052) (0.161) (0.839) (1.301) (0.060) last year -0.201 -0.461*** 0.073* 0.210 0.441* -2.919*** -0.153*** (0.170) (0.085) (0.043) (0.275) (0.247) (0.271) (0.005) micro study -0.404** -0.012 5.234*** -1.045 9.202*** -0.016 (0.203) (0.370) (0.980) (1.433) (0.756) (0.025) fixed effects 0.483** -1.481 -0.171* 0.415** -0.652 0.436 -0.233* (0.213) (1.381) (0.098) (0.179) (0.406) (0.703) (0.136) bias correction 0.071 -0.190 0.172 -1.092 -1.748 16.152*** -0.121** (0.093) (0.190) (0.221) (0.970) (1.436) (0.895) (0.049) mortgage -0.062 -1.539*** -0.048 6.411*** -1.168*** (0.123) (0.297) (0.358) (1.365) (0.240) hedging -0.087 0.161* -0.610*** 1.826 1.100*** -0.308*** (0.101) (0.096) (0.131) (2.150) (0.256) (0.038) EU enlargement 0.212 0.060 -6.502*** (0.142) (0.329) (1.631) post-crisis period 0.351 1.852*** 0.086 -11.140*** -2.948 -0.209 -0.053 (0.829) (0.330) (0.447) (2.623) (2.034) (0.514) (0.039) FSU dummy -0.395 -1.084*** 0.339 -5.402*** -1.708 -0.225 (0.300) (0.175) (0.211) (0.784) (2.043) (0.250) no curr. boards 0.023 0.067 -1.064 0.293 (0.076) (0.161) (1.301) (0.252) Latin America -0.163 -0.331 11.908*** 2.237 (0.863) (0.428) (2.465) (1.418) Observations 288 164 97 71 69 68 77 No of studies 20 8 10 9 7 7 10 Note: *, **, and *** stands for significance at the 10%, 5%, and 1% level, respectively. Robust standard errors in parentheses. 24
Table 4: Meta-Regression, Weighted Least Squares Interest MVP Exch. rate Exch. rate Inflation FC Inflation rate diff. ratio depr. volatility volatility deposits intercept 0.070* 1.521*** 0.705* 2.116*** -0.103 12.429*** 0.459* (0.035) (0.006) (0.384) (0.047) (0.148) (0.313) (0.219) main results 0.000 0.014*** -0.016 0.011 -0.001 0.000 -0.172** (0.000) (0.000) (0.017) (0.018) (0.003) (0.000) (0.064) last year -0.004 -0.585*** 0.037 0.009 0.054 -2.946*** -0.083 (0.028) (0.000) (0.121) (0.015) (0.059) (0.074) (0.092) micro study -0.005 1.163*** -0.101 2.625*** 0.333** -6.823*** 0.455 (0.028) (0.005) (0.448) (0.045) (0.128) (0.165) (0.332) fixed effects 0.021 -0.074*** -0.049 0.010 -0.001 0.017** -0.190** (0.028) (0.004) (0.042) (0.017) (0.003) (0.007) (0.083) bias correction -0.024 0.034*** 0.087 0.520*** -0.052 -0.505 (0.051) (0.008) (0.397) (0.053) (0.116) (0.281) mortgage -0.049 -1.706** -0.476 (0.075) (0.549) (0.334) hedging 0.000 0.064*** -0.628* -0.503 0.321 (0.000) (0.000) (0.335) (0.388) (0.322) EU enlargement -0.032 2.319*** 0.045 -2.547*** -0.898*** (0.038) (0.005) (0.240) (0.082) (0.001) post-crisis period -0.017 0.019 -5.314*** 8.832*** -0.171** (0.100) (0.954) (0.121) (0.222) (0.064) FSU dummy -0.091* -1.246*** 0.052 -2.213*** 0.004 -5.869*** 0.018 (0.044) (0.003) (0.342) (0.027) (0.094) (0.147) (0.081) no curr. boards 0.000 0.088*** -0.160*** 0.995*** (0.000) (0.000) (0.034) (0.269) Latin America 0.071 -0.053 5.322*** 0.219 -8.815*** (0.090) (0.938) (0.117) (0.348) (0.224) Observations 288 164 95 71 66 68 77 Adjusted R2 0.078 0.192 0.773 0.167 1.000 0.644 0.997 Note: *, **, and *** stands for significance at the 10%, 5%, and 1% level, respectively. Robust standard errors in parentheses. 25
You can also read