Explainable AI for Credit Risk and Customer Segmentation in Subprime Lending: A Comprehensive Framework with Implementation Protocol

Abstract

Subprime lending sets a conflict of operations between predictive accuracy and regulatory transparency. Although advanced machine learning methods are more efficient in default prediction, they have low interpretability, which limits their use in regulated credit systems. The paper presents a combined explainable artificial intelligence system that builds on eXtreme Gradient Boosting (XGBoost) and SHAP (SHapley Additive exPlanations) to assist in transparent credit risk measurements and explanation-based customer segmentation. The framework comes up with three contributions, namely maintaining high predictive performance and allowing instance-level explanations, clustering of borrowers around explanation vectors instead of risk scores, and generating regulation-consistent adverse action notices automatically. Assessment using a simulated dataset of 125,000 subprime loan applications, under realistically simulated conditions, shows that the proposed algorithm has a competitive predictive accuracy, a silhouette score of 0.61 in the explanation-based segmentation, a 12.7% reduction in the default, and an 8.9% increase in the approval of credit-worthy applicants. These results suggest that explainability is an operational capability that can be adopted to improve regulatory compliance and lending in subprime credit markets.