Optimal structure of an expected loss credit rating model

According to Basel III, expected loss is regarded as a key indicator of credit risk. In this study, we introduce an innovative expected loss credit rating model (ELCRM) using expected loss as the risk indicator and construct the ELCRM using data from peer-to-peer lending platforms. Additionally, we employ the XGBoost method to build each risk component of the ELCRM, making the XGBoost approach interpretable through the use of SHapley Additive exPlanations values, and further discuss the impact of loan variables on expected losses. We also utilize Kruskal-Wallis tests and Conover's t-tests to demonstrate that, compared to credit rating models based on the probability of default, ELCRM possesses superior capabilities in discriminating between good and bad borrowers. Finally, we show that incorporating prepayment risk into the credit rating model significantly enhances its performance in predicting actual borrower losses.