عوامل تعیین کننده بحران ارز: نقش الگوی عدم قطعیت

We tackle explicitly the issue of model uncertainty in the framework of binary variable models of currency crises. Using Bayesian model averaging techniques, we assess the robustness of the explanatory variables proposed in the recent literature for both static and dynamic models. Our results indicate that the variables belonging to the set of macroeconomic fundamentals proposed by the literature are very fragile determinants of the occurrence of currency crises. The results improve if the crisis index identifies a crisis period (defined as the period up to a year before a crisis) instead of a crisis occurrence. In this setting, the extent of real exchange rate misalignment and financial market indicators appear as robust determinants of crisis periods.

Over the course of the last couple of decades several parts of the world have experienced rather harsh financial market crises, sometimes repeatedly, and mostly accompanied by painful real shocks. The very last wave of such turmoils, initially triggered on the US (subprime) mortgage market, has exemplified that financial market turbulences are not confined only to the developing and emerging economies. Moreover, the recent tensions have clearly unveiled challenges financial stability authorities and policy makers have to face in the age of ever deeper and more global markets. Most importantly, diminishing barriers to capital flows and instant information distribution increase the potential sudden evasiveness of capital. As evidenced by the shocking promptness with which the US mortgage malaise extended from one corner of the financial market to another, crises can spread swiftly between different types of markets in geographical and technical terms. One of the most frequent targets of speculators is the currency market and substantial devaluations of the currency under attack generally imply severe consequences for the respective economy. Against this backdrop it is not surprising that both in the academic literature and in the private sector a variety of empirical attempts has been undertaken to predict currency crises. Following the pioneering indicator approach by Kaminsky et al. (1998) a whole plethora of early warning systems for currency crises has been developed. Some of the rather recent approaches employ innovative methodologies such as Markov switching models (see e.g. Abiad, 2003 or Chen, 2005) or financial market tools (see e.g. Malz, 2000 or Crespo Cuaresma and Slacik, 2007) to predict currency attacks. The vast majority of the empirical literature assesses the effect of various potential determinants on the probability of a currency crisis using limited dependent variable – logit or probit – models. The discrete crisis variable is regressed on a set of fundamental indicators, such as, inter alia, current account and government balances, exchange rate overvaluation or liquidity ratios. The choice of regressors is typically inspired by the three generations of theoretical models on balance-of-payment crises. In one of the most recent empirical contributions on this topic Bussière (2007) overhauls the usually static specification, in which, moreover, all regressors tend to enter at the same lag. He thus extends the usual set of explanatory variables by including several lags of the regressors as well as of the dependent binary crisis variable. He finds that there are several variables significantly affecting the probability of a crisis in a dynamic logit model. However, the impact of the indicators ranges between short run (4–6 months) e.g. for the liquidity measures to very long-run (2 years) in case of over-appreciation of the exchange rate. In addition, his results indicate that past crisis episodes increase the probability of a new attack, particularly in the short run. Notwithstanding substantial variations in the literature on early warning systems with respect to methodology, data as well as results, there is one general caveat which applies to all existing binary choice models. Given that there is no unique theoretical framework linking the potential set of determinants with the realizations of currency crises, the issue of model uncertainty surrounding both the choice of variables and the estimates obtained deserves to be treated seriously. Model uncertainty can be explicitly taken into account using Bayesian statistical techniques, in particular with the use of the Bayesian model averaging (BMA) methodology which proposes averaging of the parameter values over all (relevant) alternative models using posterior model probablities as respective weights to evaluate the relative importance of different variables (see Raftery, 1995 for a general discussion and Sala-i-Martin et al., 2004 and Fernández et al., 2001 or Crespo Cuaresma and Doppelhofer, 2007 for applications to economic growth regressions). The different theoretical settings used to explain different crises episodes give rise to alternative sets of potential explanatory variables (with intersections which are not necessarily empty) for the probability of a crisis ocurring. The so called first generation models ( Krugman, 1979 and Flood and Garber, 1984) concentrate on bad economic policy leading to unsustainable developments of some fundamental macroeconomic variables. The abandonment of the fixed exchange rate regime is then precipitated by the eventual exhaustion of the central bank’s foreign reserves. The second generation of currency crises models (see for instance Obstfeld, 1994), explains crises as the consequence of self-fulfilling expectations in theoretical settings with multiple equilibria. In contrast, the third generation of models ( Krugman, 1998) explains the outbreak of a currency run as a symptom of accumulated problems in the banking and financial sector. In the theoretical setting, government guarantees aimed at attracting foreign investment lead to a bubble on the asset market that eventually bursts and creates the crisis. Obviously, given the different theoretical nature of the ultimate cause of the currency crises in the different generations of models, the potential empirical determinants to be included in econometric studies vary strongly depending on the theory used to select covariates. The objective of the present paper is to revisit binary variable models for currency crises based on macroeconomic fundamental data by explicitly taking into account model uncertainty. In particular, we want to work out to what extent model uncertainty puts the robustness of the explanatory variables of the logit models championed in the literature (e.g. Bussière and Fratzscher, 2006 or Bussière, 2007) under strain. On the one hand, our results indicate that the usual macroeconomic variables used in empirical studies of currency crisis are very fragile determinants of the occurrence of such episodes. On the other hand, if we redefine the crisis indicator as to give a signal for observations up to one year prior to the crisis, several variables appear as robust determinants of these crisis periods. Financial market indicators and the deviations of the real exchange rate from a linear trend present very high posterior model inclusion probabilities and thus can be considered robust determinants of crisis periods. The remainder of the paper is structured as follows: Section 2 sketches the Bayesian model averaging procedure. In section 3 the data are described and variables defined. Section 4 presents the results on the extent to which model uncertainty matters, while section 5 concludes.

The dominant majority of early warning mechanisms for currency crises employs some version of fundamental-based binary choice models. To our knowledge, none of the papers on the subject tackels the issue of model uncertainty in currency crisis model explicitly. In the present paper we have explicitly taken into account model uncertainty in the framework of a binary choice model. By means of Bayesian model averaging we estimate the coefficients for each variable as weighted averages over the alternative models from the model space, where the weights correspond to the posterior probability of each model. In order to figure out the value added by this approach as opposed to “standard” logit regressions we have used the same data set as one of the most recent studies on the subject by Bussière (2007). If the discrete dependent variable is constructed so as to predict the exact month in which a crisis may happen our conclusions are twofold. On the one hand, we have found that coefficients mostly have the expected signs coinciding with the benchmark study. On the other hand, however, our principal quality gauge, the posterior inclusion probability (the sum of posterior probabilities of all models containing a particular variable), unveils the lacking robustness of the relationships between regressors and the dependent variable. These results imply that at least in this setting the best model to explain a currency crisis is a mere time and country-unspecific constant. Our results, therefore, indicate that none of the usual macroeconomic fundamental variables is a robust determinant of a currency crisis for the definition and sample used. The results improve considerably if we consider defining “crisis periods” instead of crisis occurrences. Defining crisis periods as observations up to one year prior to the crisis, we find that real exchange rate developments and financial variables are able to robustly explain differences in the probability of a country experiencing such episodes. Since our sample starts in 1994 it could well be that episodes of currency distress included in the sample are crises rather of the second and third generation type. In such a case it would not be surprising that fundamental data show only limited explanatory power. To turn the argument around, the fundamentals should play a much more significant role in a sample covering the first generation type of crises. Exactly along these paths we are planning to conduct our future research. A finer way of testing the different theoretical frameworks proposed by the three generations of currency crises models would imply grouping variables by theory and computing the joint inclusion probability of these groups of variables. The construction of groups of variables by theory could be handled in the BMA framework using the proposal by Brock et al. (2003) of using a hierarchical prior in order to sort variables into theories or thematic indicators (see also the recent contribution by Doppelhofer and Weeks (forthcoming), for the concept of jointness of determinants in the BMA framework). Although we did not follow this approach in the paper, we propose it as a potentially fruitful path of further research. An interesting issue that has not been directly tackled in the paper and that would deserve further scrutiny is the possibility of nonlinear effects in form of interactions among the potential determinants of crises. Developments in some relevant variables may just be responsible for preparing the ground for imbalances that end up a currency crisis when triggered by an unsound development in an additional variable. The use of interaction terms in a BMA setting could assess the importance of this type of effects.