Multi-Period Credit Default Prediction : A Survival Analysis Approach

The book deals with the problem to estimate credit default probabilities under a flexible multi-period prediction horizon. Multi-period predictions are naturally desirable because the maturity of loans usually spans several periods. However, single-period models largely prevail in the literature so far due to their simplicity. Predicting over multiple periods indeed entails certain challenges that do not arise within a single-period view. Among the main contributions of this work to the literature is to show that there are relatively simple solutions to these challenges available.From a methodological point of view, a survival analysis approach is used. In a survival analysis context, the time until default (or lifetime) is the central variable under investigation as opposed to the traditional approach of reducing the information to a binary variable representing the default event. Modeling the time until default has the advantage that both the timing of default events and censored data are utilized. Since both issues gain importance as the prediction horizon grows it is no coincidence that a survival analysis approach is selected for the multi-period prediction problem.The main results of the work are the following. First, a new index for measuring the predictive accuracy of default predictions is proposed and its advantages over commonly used indices are shown both theoretically and by an empirical analysis. The second chapter further includes new methods of statistical inference for the new index. In the third chapter, default prediction models for the case of panel datasets with time-varying covariates are dealt with. A new approach is developed that is simpler than the models available in the literature so far. In an empirical study concerning North American public firms, we further provide evidence that the proposed approach delivers more accurate predictions than its competitors as well. In the final chapter, the problem of assigning default probability estimates to given rating grades is examined. If default events are rare, standard approaches have certain drawbacks. As an alternative, an empirical Bayes approach is presented that mitigates the effects of data sparseness. The new estimator is applied to a comprehensive sample of sovereign bonds. Among the main findings of the empirical part is that capital requirements for sovereign bonds are likely to be underestimated by using standard approaches but not when using the empirical Bayes estimator.