Internal credit rating systems: methodology and economic value

Abstract

In this paper we propose a methodology for translating credit risk scores into empirical probabilities of default (PD) and then segment these measures into rating classes to obtain predefined target default rates for each class. The translation of risk scores into empirical PD values is done using a piecewise multiperiod logistic regression model based on observed credit risk scores with associated historical default/non-default behavior. The segmentation is based on an optimization algorithm for creating “obligor risk ratings” in such a way that within each class the average empirical PD is closest to a predefined fixed target PD. Taken together, these two algorithms allow one to fix a PD-based rating scale, and then map an unlimited number and variety of credit risk scores onto that common rating scale in a way that minimizes the expected deviation of empirical PDs from their target values. We use an economic model to quantify the potential value-add to an organization implementing this more advanced methodology as opposed to using the basic cohort methodor standard regression models. An example using a proprietary dataset is provided.