Rank-order statistics for validating discriminative power of credit risk models

This paper provides practical insights into common statistical measures used to validate a model’s discriminatory power for the probability of default (PD), loss liven default (LGD) and exposure at default (EAD). The review of available rank-order statistics is not based on analysing empirical data. Thus, the study has more of an informative value without delivering empirical evidence. When there is an alternative model available for comparison, this paper proposes to use the cumulative accuracy curve and the accuracy ratio to assess the rank-order ability for PD models given their popularity in practice. When there is no model available for comparison, due to the limited techniques in this area, this paper proposes to compare the confidence intervals in order to prove that a rating system has any discriminative power. For the LGD/EAD/slotting models, this paper recommends using a graph to check the rank-order ability. No statistical test is recommended. Focusing on enhancing practical implications for the financial industry, this paper advises banks on the existing CRR self-attestation requirements.