Interpretable machine learning for default risk prediction in stress testing

Junqi Zhao, Nan Zhou and Zailong Wan

Abstract

Stress testing is essential for testing the resilience of banks’ portfolios against possible future economic conditions. A key component in stress testing is predicting the probability of default within the expected loss framework for credit risk modeling. Traditional probability of default models based on logistic regression (LR) struggle to capture complex, nonlinear relationships and require great efforts in manual feature engineering to satisfy model assumptions and achieve the best model fit. In model validation a benchmark model can serve as a valuable tool to effectively challenge the limitations of an existing model. This study develops a benchmark model – an inherently interpretable machine learning model using the explainable boosting machine (EBM) within a discrete-time survival analysis setting – to predict the forward-looking probability of default of a real-world credit card portfolio. The EBM-based model is compared against a classic LR model, and results show that it achieves comparable predictive accuracy to the LR model and enhanced macrosensitivity with a sample one-fifth of the size of the LR model and without the lengthy manual feature engineering of the LR model. Moreover, the EBM’s learned shape functions make the model inherently interpretable. The study contributes a valuable option for financial institutions seeking to either adopt machine-learning-based methods in order to optimize stress testing model development or enhance benchmarking practices in risk model validation while maintaining compliance with regulatory expectations.