The Next Generation of Default Prediction Models: Incorporating Signal Strength and Dependency

Abstract

The financial market crisis has exposed a weakness in predicting defaults. A new wave of models must be more powerful and should be more successful in predicting multiple defaults. I argue that such models incorporate signal strength, cross and serial dependencies. The conditional default probability is the most powerful predictor and I demonstrate that heteroscedastic probits with random effects represent a flexible framework to estimate the conditional probability. The lower the variance the higher is the signal strength of the underlying credit score. I derive an approximate likelihood procedure, including asymptotic standard errors, and apply it successfully to a dataset.