ABSTRACT

The evaluation of credit portfolio risk models is an important issue for both banks and regulators. It is impeded by the scarcity of credit events, long forecast horizons, and data limitations. To make efficient use of available information, the evaluation can be based on a model’s density forecasts, instead of examining only the accuracy of point forecasts such as value-at-risk. We suggest the Berkowitz (2001) procedure, which relies on standard likelihood ratio tests performed on transformed loss data. We simulate the power of this approach to detect misspecified parameters in asset value models, focusing on asset correlations. Monte Carlo simulations show that a loss history of ten years can be sufficient to resolve uncertainties currently present in credit risk modeling. The power is better for two-state models than for multi-state models, and it can be improved by incorporating cross-sectional information.