A Latent Variable Approach to Validate Credit Rating Systems

We suggest a new parametric framework to assess the accuracy of estimated default probabilities (PDs). Whereas the traditional methods to validate credit rating systems focus primarily on the discriminatory power, recent advances in credit risk management and banking regulation has shifted the focus of rating validation to the accuracy of PD estimates. In this paper we introduce bank and obligor specific error terms which are additive to a suitably transformed latent variable, which in our approach has the interpretation of the "true", unobservable PD of the underlying obligor. Provided with rating information from different sources (e.g., banks, rating tools, or rating agencies) and an appropriate model for the distribution of the latent PDs the parameters of the errors distribution can be obtained. In a specific application we assume the error terms to be jointly normal and estimate this model for a data set provided by the Austrian central bank where we observe PD estimates for 2090 obligors provided by 13 banks