White Box Finance: Interpreting AI Decisions in Finance through Rules and Language Models

Although loan defaults continue to cause substantial financial losses, this study focuses on improving how AI credit risk models are explained. Beyond developing a predictive model based on the demographics of the borrower, the attributes of the loan, and the credit history, the core contribution lies in introducing and comparing explanation methods. Specifically, we evaluated two ways to provide explanations. One method is a module that integrates SHAP values and GPT-4 to generate human-friendly narratives, a second is a rule-based logic explanation. This approach aims to enhance interpretability and trust, offering a clearer understanding of model predictions than traditional explanation techniques.