Threshold accepting for credit risk assessment and validation

According to the latest Basel framework of Banking Supervision, financial institutions should internally assign their borrowers into a number of homogeneous groups. Each group is assigned a probability of default that distinguishes it from other groups. This study aims at determining the optimal number and size of groups that allow for statistical ex post validation of the efficiency of the credit risk assignment system. Our credit risk assignment approach is based on Threshold Accepting, a local search optimization technique, which has recently performed reliably in credit risk clustering especially when considering several realistic constraints. Using a relatively large real-world retail credit portfolio, we propose a new technique to validate ex post the precision of the grading system.