An Analytics Process for Forecasting Expected Credit Losses for the Lifetime of Loans : auto loan portfolios

Abstract

The main goal of this study is to conduct a thorough assessment of expected credit risk losses from loan origination throughout the lifetime of loans by utilizing an integrated methodology that combines domain knowledge with advanced analytics. The data used, are extracted from Santander loan applications, consisting of auto loan data from Finnish consumers and companies and Danish borrowers. Following a correlation-based feature selection process, XGBoost is applied as a challenger model to logistic regression to forecast the early performance of loans. A second model that incorporates the early performance as input is trained to predict the expected losses at loan maturity. This nested method enables reforecasting and stress testing the predictions at an early stage of loan admission. The results demonstrate that the analytical process proposed may be adopted by portfolios of other countries and effectively incorporate loan-specific complexities. Furthermore, the XGBoost method shows improvement in the accuracy of predicting default and expected losses in auto loans. Additional insights into the adverse effects of unpredictable events on loan portfolios can be derived by including a systematic shock in credit risk predictions. Finally, model stability and transparency are maintained by including out-of-sample continuous monitoring and SHAP values.