طراحی بهینه سیستم های هشدار دهنده برای بحران بدهی های مستقل

This paper tackles the design of an optimal early warning system (EWS) for sovereign default from two distinct angles: the choice of the econometric methodology and the evaluation of the EWS itself. It compares K-means clustering of macrodata, a logit regression for macrodata, a logit regression for credit ratings, and the combined forecasts from all three methods. The optimal choice of forecast method is shown to depend on the desired trade-off between missed defaults and false alarms. Hence, it is crucial to account for the decision-maker's preferences which are characterized through a loss function and risk-aversion parameter. Recursive forecast combining generally yields a better balance of type I and type II errors than any of the individual forecasting methods, and outperforms the naïve predictions.

The financial turmoil that hit emerging markets in recent decades has triggered the need for accurate country risk assessment. A number of studies have focused on the development of empirical models for explaining and predicting banking and currency crises (Berg and Pattillo, 1999, Frankel and Rose, 1996, Kaminsky and Reinhart, 1999 and Kumar et al., 2003). As more countries move toward flexible exchange rates, twin crises are becoming less frequent. But sovereign debt crises remain a matter of concern for international financial markets and economic policymakers. The process of building an Early Warning System (EWS) can be broadly divided into four decision stages: the sample (country and time span), the input variables, the econometric approach, and the evaluation of the EWS in relation to its end use by the decision-maker. The first two have by now received extensive attention in the sovereign default literature. Most studies have focused on identifying the nature–region, country, or period specific–of debt crises, or their main determinants among domestic fundamentals and indicators of the international business-cycle and market sentiment. For this purpose, different classification techniques have been used. However, the empirical literature on EWSs with an explicit forecasting objective is relatively young. Scant attention has been paid to forecasting issues and to the design and validation of an EWS tailored to the decision-maker's preferences. The aim of this paper is to contribute to filling this gap. Several studies have applied discriminant analysis (Frank and Cline, 1971 and Taffler and Abassi, 1984), whereas more recent research has been based on panel logit models (Peter, 2002). Non-parametric classification techniques such as clustering and recursive tree analysis, albeit popular in other areas, have received little attention in this context. There is evidence that country credit ratings have predictive power regarding sovereign debt crises and that they Granger-cause sovereign bond spreads (Cantor and Packer, 1996, Reinhart, 2001 and Rojas-Suárez, 2001). Moreover, the New Basel Accord allows banks to use internal ratings for calculating capital requirements. The Institutional Investor ratings can be regarded as consensus internal ratings from major international banks. The upshot is that it is unclear which method and information set one should adopt in developing an EWS for sovereign default. In this respect, forecast combining may be fruitful. This paper presents a novel framework for the optimal design of an EWS focusing on methodological issues. The contribution is twofold. First, it assesses alternative forecasting techniques in the light of the decision-maker's degree of risk-aversion towards default. These are: (i) a multivariate logit model based on macrodata, (ii) a univariate logit model based on the Institutional Investor ratings, (iii) K-means clustering of macrovariables, and (iv) a combination of the above three forecasting methods (or classifiers) using a parametric regression. In the present context, clustering has not been utilized as yet and issues of forecast combination have barely been addressed. The analysis is based on a sample of 75 emerging/developing economies over the 1983–2000 period. Second, the paper explores the evaluation of an EWS in relation to the decision-maker's objective function. We show how the latter can be taken into account to choose the classifier and its embedded parameters. The classifiers are shown to have different strengths in terms of missed defaults and false alarms. Furthermore, their forecast ranking is unstable over the holdout years. On the one hand, these findings imply that the user's loss function and degree of risk-aversion are critical inputs in the assessment of an EWS. On the other hand, they motivate forecast combining. It is shown that a relatively better balance of missed default and false alarms is achieved by combining the classifiers. Our framework can be easily adapted to distinct classifiers and loss functions. Finally, as a by-product of our analysis, some lessons emerge for practitioners in the area of sovereign default prediction. First, optimal recursive in-sample calibration of the classifiers is worthwhile. Second, given the persistence of sovereign default events, it seems sensible to gear the out-of-sample assessment of forecast ability toward default entries rather than continuing defaults. Section 2 outlines the background literature. Section 3 describes the methodology and Section 4 introduces the data. Section 5 illustrates several issues regarding the optimal, recursive calibration of classifiers. The forecast combining analysis is presented in Section 6 before concluding.

An early warning system (EWS) for sovereign default provides a complementary tool to the analysis of decision-makers by facilitating objective measures of vulnerability. This paper investigates the optimal design of an EWS focusing on two aspects: the choice of the econometric methodology and the evaluation of the EWS itself. These two problems raise important issues which are seldom tackled in the literature. The analysis is based on a sample of 75 emerging and developing economies 1983–2000. Forecasts are obtained from a logit regression (LOGIT-M) and K-means clustering, both based on macrovariables, and a logit regression based on internal-bank ratings (LOGIT-R). Clustering has not been used in this context to date. The study has two main components. First, it incorporates the decision-maker's preferences (captured by a loss function and risk-aversion parameter) into the optimal calibration of the classifiers and the assessment of their out-of-sample forecasting properties. Second, it investigates forecast combining issues. For this purpose, a regression framework is adopted that exploits the distinct in-sample forecast ability of the individual methods. The issues of the objective function and forecast combination have received scant attention in the literature. The results suggest that the decision-maker's preferences influence the choice of forecast methodology and its optimal calibration. LOGIT-M outperforms the non-parametric (clustering) and judgmental (LOGIT-R) classifiers by issuing fewer false alarms. But the latter two classifiers dominate LOGIT-M in missing fewer defaults. Moreover, the out-of-sample forecast ranking of the individual classifiers is unstable. These findings vindicate forecast combining. Both individual and naïve forecasts are outperformed by the combined forecasts for a range of risk-aversions. In this paper, the decision-maker's preferences have been formalized using loss functions that simply seek to reflect the desired trade-off between missed defaults and false alarms. Future work could extend this framework by aiming to better capture the economic or utility-based value of debt-crisis predictability.