سیستم های هشدار دهنده اولیه برای بحران بدهی های مستقل: نقش عدم تجانس

Sovereign default models that differ in their treatment of unobservable country, regional and time heterogeneities are systematically compared. The analysis is based on annual data over the 1983–2002 period for 96 developing economies. Inference-based criteria and parameter plausibility overwhelmingly favour more complex models that allow the link between the probability response and the fundamentals to vary over time and across countries. However, out-of-sample forecast evaluation using several loss functions and equal-predictive-ability tests suggests that simplicity beats complexity. Parsimonious pooled logit models produce the most accurate sovereign default forecasts and outperform the naive benchmarks.

The sovereign debt crises of recent decades have emphasized the importance of credit risk prediction. Financial institutions use default probabilities to price loans and bonds, to determine adequate concentration limits and as inputs for value-at-risk analyses. Furthermore, the new Basel Accord allows banks to use internal ratings and default probabilities to set their regulatory capital. The available studies can be grouped into three broad types. One group exploits option pricing models to obtain the implied default probability while another seeks to explain the default probability using structural models or panel discrete-choice models. This study belongs to the latter group. A third research area focuses on explicitly generating default probabilities from credit ratings. This is the approach adopted in Fuertes and Kalotychou (2006a) where the finite-sample properties of different credit migration estimators are investigated by bootstrap simulations. A large empirical literature analyses the determinants of sovereign default and the results are quite mixed. Some studies find that liquidity or global business cycle indicators are crucial. The evidence on the importance of structural economic conditions also varies across studies. Most researchers estimate logit models using pooled data for a large number of countries although the validity of the underlying homogeneity assumption has been questioned. McFadden et al. (1985) and Hajivassiliou (1987) argue that the link between debt repayment performance and macroeconomic attributes is likely to vary across countries and over time. For countries with fewer capital controls or more open to global trade, external economic activity signals may have more predictive power. Moreover, qualitative idiosyncrasies (colonial histories, types of government and religious institutions) may explain why country A defaults but country B services its debt while exhibiting similar economic fundamentals and debt structures. As Shleifer (2003; p. 5) succinctly puts it: “Sovereign debt markets could not be more different”. The few studies that control for country heterogeneity adopt either fixed or random effects models ( Detragiache and Spilimbergo, 2001, Li, 1992 and Oral et al., 1992). Evidence from the currency crisis literature suggests that the relevant heterogeneity is regional rather than country specific ( Burkart and Coudert, 2002 and Staikouras, 2004; Kalotychou and Staikouras, 2005). For instance, different key leading indicators of currency crises have been identified for Asia and Latin America. In sovereign default studies, regional differences have been captured using dummy variables ( Feder et al., 1981). Time heterogeneity may reflect changing world conditions such as the business cycle and the development of international capital markets or the changing nature of emerging markets themselves. To control for such time effects that are assumed to be common across countries, some studies include year-dummy variables ( Aylward and Thorne, 1998) whereas others use global macroeconomic indicators such as OECD growth ( Lee, 1991 and Detragiache and Spilimbergo, 2001). The issue of whether controlling for country, regional or time heterogeneity helps to improve the forecast power of sovereign default models is relevant to regulators, practitioners and the rating agencies, all of which are mostly interested in the when rather than the why question of default. Regulators rely on default forecasts to monitor the financial health of banks, pension funds and other financial institutions that include sovereign debt in their portfolios. Practitioners feed such forecasts into theoretical or simulation models for pricing sovereign debt. Moreover, default probability predictions are used to test various hypotheses such as that country risk is priced in stock returns and borrowing costs. Finally, Early Warning Systems (EWSs) are recognized as a potentially fruitful complement to the broader analysis and judgement of decision-makers for identifying looming debt crises ( Berg et al., 1999 and Fuertes and Kalotychou, 2006b). Against this background, it is surprising that forecasting issues have received scant attention in the literature. Most studies compare sovereign default models on the basis of their in-sample forecasts (Hajivassiliou, 1987 and Detragiache and Spilimbergo, 2001). A few studies conduct out-of-sample evaluation but the forecasts are based on parameters estimated once and are limited to a 1- or 2-year holdout period. The forecast accuracy metrics typically used are Type I, Type II and overall error rates (Feder et al., 1981, Manasse et al., 2003, Oka, 2003 and Peter, 2002). Furthermore, the few studies that provide out-of-sample predictions do not confront them with simple benchmarks such as those from random walk type models. This is particularly important in the present context due to the persistence in debt-servicing behaviour. This paper contributes to the literature in two respects. First, to investigate the importance of unobserved heterogeneities in modelling sovereign default, it considers a wide range of logit models that differ in how they treat country, regional and time effects. The models’ ability to describe the data generating process is gauged on the basis of statistical tests, information criteria and parameter plausibility. By characterizing the heterogeneities in different ways we seek to assess whether these unobserved effects are genuine or merely an artefact of misspecification. Some of the parameterizations, such as the random coefficients model with time-dependent and country-specific slopes and the models that allow for region-specific or time-specific slopes, have not yet been utilized in the present context. In addition, three novel global indicators are included to control for time effects in the likelihood of sovereign default. Second, a comprehensive forecasting analysis is conducted. A 12-year window is rolled forward to generate out-of-sample forecasts sequentially over 5 years. A battery of tests is run to assess the importance of heterogeneity in forecasting. For this purpose, several forecast accuracy measures are adopted: probability scores as well as metrics that allow for asymmetric misclassification costs. The tests are applied not only over the entire 5-year holdout sample but also over a positive-directional-change subset in order to gauge the models’ ability to predict new (as opposed to ongoing) defaults. Various uninformative benchmarks are considered including random walk type models and the naive models implicit in the Pesaran and Timmermann (1992) and Donkers and Melenberg (2002) tests. The statistical tests and model selection criteria indicate that the more complex specifications describe the data better. Unobserved heterogeneity across countries, regions and time is significant in explaining sovereign default. By contrast, the forecast horse race suggests that the parsimonious pooled logit model that imposes full homogeneity appears capable of yielding relatively good out-of-sample predictions and beating the benchmark models. Hence, our findings corroborate in this novel context the well-known limited relationship between in-sample fit and out-of-sample forecast performance. Simple variants of the pooled logit model that allow for fixed regional effects or time effects also beat the uninformative benchmarks but, interestingly, are unable to improve significantly upon the baseline pooled logit forecasts. Complex random coefficients models that allow for country and time variation in the link between default risk and the country fundamentals yield poorer forecasts than the simple pooled logit and generally underperform the naive benchmarks. The paper is structured as follows. Section 2 describes the data and the endogenous default indicator. Section 3 outlines the models and the inference-based metrics. Section 4 discusses the forecast framework and Section 5 analyses the empirical results. A final section concludes.

While the empirical literature on sovereign debt crises is vast, only a few studies explicitly focus on specification issues related to the prediction of country default. Model-based early warning systems (EWS) are one approach among many others used for country risk monitoring. Such quantitative tools can usefully complement the sound judgement and broader analysis of decision-makers by yielding objective measures of country vulnerability to debt crises. Concerns have been flagged about the challenges that country heterogeneity and time variation pose with regard to sovereign default modelling. For instance, it has been pointed out in the context of arrears to the IMF that “Temporal stability and country homogeneity that are assumed under probit estimation using panel data might be problematic” (Oka, 2003; p. 33). In a similar vein, Berg et al. (1999) provide theoretical reasons for using a fairly homogeneous group of countries and a stable sample period in the design of EWS for financial crises. Despite such concerns, an investigation of the importance of controlling for unobservable differences across countries and through time is lacking in the EWS literature. This paper seeks to fill the gap. We formulate a large number of logit specifications ranging from a simple pooled regression (full homogeneity) to a random coefficients model where the impact of the macroeconomic indicators on the probability of a debt crisis is both country and time heterogeneous. The analysis is based on a sample of 96 emerging/developing economies over the 1983–2002 period. The observable ratios that emerge as robust leading indicators of sovereign default are trade to GDP, external debt to GDP, official debt to total debt, IMF credit to exports and credit to private sector over GDP. A systematic comparison of the models is conducted from two distinct angles. First, model adequacy to describe the data in-sample is evaluated using statistical hypothesis tests, information criteria and theoretical judgements on the plausibility of the parameter estimates. The analysis corroborates the importance of controlling for country, regional and time effects in sovereign default models. Second, an extensive out-of-sample forecast competition is carried out. Forecasts are generated over a 5-year holdout period on the basis of a rolling estimation window. The forecast comparison includes naive benchmark models and is based on several loss functions and predictive ability tests. We find that by exploiting pooled data across a large number of countries in a recursive modelling approach, it is possible to develop an effective EWS of sovereign default that outperforms various naive predictors. Although the theoretical literature stresses the vulnerability of developing economies to changes in market sentiment and the global environment, the inclusion of either year-dummy variables or proxies for market volatility and risk aversion does not help forecasting. Moreover, accounting for parameter instability stemming from, say, changes over time in the level of economic integration and in market structures, does not enhance predictive performance either. Restricting the pooling of countries to regions does not yield significant forecast gains over a broader (world) pooling approach. Poor forecasts are obtained from models that allow for both country and time variation in the impact of the macroeconomic fundamentals on the probability of default. Most pertinently, this study documents a very weak association between in-sample fit and out-of-sample forecast performance in the novel context of EWS for sovereign default. Simple models outperform more complex models in terms of forecast accuracy although the latter provide a better description of sovereign debt default data. Our findings indirectly corroborate the established wisdom that forecast performance is not necessarily a good guide to model validity.