Modeling and forecasting international credit risk : the case of sovereign loans
This thesis investigates the relative merits of econometric modeling, statistical and judgmental techniques for predicting debt crises and assessing the risk of credit migration. The increased reliance on econometric or statistical approaches and credit rating systems in risk management has intensified the need for more rigorous analysis of their finite sample properties. A better understanding of the available tools has implications for credit risk management, regulation and policy decision-making. The thesis contributes to the extant sovereign risk literature in three areas. First, it addresses the question of whether controlling for unobserved heterogeneity is important for predicting debt crises and explores a pervasive inference problem in Early Warning Systems (EWSs). Second, it addresses the development of an `optimal' EWS for sovereign debt crises that accommodates the decision maker's preferences. Third, it considers the measurement of sovereign credit migration matrices using different estimators and explores non Markov effects in the rating dynamics. Chapter 2 confronts competing models of sovereign default that differ in how country-, region- and time-specific effects are treated. Statistical tests and information criteria overwhelmingly favour more complex models with country heterogeneity that possibly changes over time. However, simplicity beats complexity in terms of forecasting. Simple pooled logit parameterization, that control either for regional heterogeneity or for time effects produce the most accurate forecasts and outperform several naive predictors. Chapter 3 investigates the severity of the autocorrelation problem in EWS of sovereign default. This stems from seeking to provide crisis warnings over a horizon that is longer than the frequency at which the forecasts are updated and from the sluggishness of the typical exogenous indicators. Neglecting residual serial autocorrelation in such models is shown to be far from innocuous. Inferences are overturned when using a correction. This phenomenon is generally clearer for the macroeconomic ratios that are more persistent. Chapter 4 combines three fundamentally different classification techniques - econometric, statistical and judgmental- to produce an EWS for sovereign default. The optimal choice of crucial EWS elements is shown to depend on the decision-makers' preferences. The forecast ranking of classifiers is found to be unstable and overall the classifiers appear to have different strengths. Payoffs from forecast combination are documented and the combining scheme is shown to depend on the decision-makers' loss function. Chapter 5 turns to the estimation of sovereign transition probability matrices and evaluates the popular discrete multinomial estimator against two continuous hazard rate methods that differ in their treatment of time-heterogeneity. Bootstrap simulations of the rating generating process reveal interesting insights. Hazard rate estimators yield more reliable default probabilities. Efficiency is further enhanced upon relaxing homogeneity. Downgrade momentum and duration effects are found to be present in the rating process.