Quantifying uncertainty in credit rating model development

Banks estimate the creditworthiness of their counterparties by using credit rating models. These models are developed internally, using statistical methods. The input for the final model are certain characteristics (factors) of a counterparty, based on which the model returns a credit rating for this counterparty. During the model development, regression is performed on the factors to obtain the appropriate factor weights. In this thesis, we develop a method to quantify the uncertainty in these estimated weights. After the weights are fully determined, the performance of the model is checked, and – in particular – its discriminatory power. For this, the measure powerstat is used. Also for this measure we provide a method to quantify its uncertainty. Both methods are tested numerically and examples are generated for illustration purposes. Where possible, datasets from real model developments are used.