Abstract

Credit risk modelling is important for assessing the probability of borrower default and for lending decisions. Even with recent advances in predictive algorithms, however, there are obstacles to implementing transparent, robust, and reliable models that can adapt to uncertain inputs. This paper examines how combining XAI and UQ can improve interpretability and confidence in credit risk predictions. Three modelling methods, logistic regression, Random Forest and XGBoost, were compared with respect to the Home Equity (HMEQ) dataset based on predictive accuracy, probability calibration, interpretability, and uncertainty management. Ensemble methods showed better predictive performance, with over 98% accuracy and AUC values >0.999, while logistic regression showed poor performance. A disparity between accuracy and probabilistic reliability was found through calibration analysis. Random Forest yielded more accurate results with less well calibrated estimates (ECE = 0.0475). Nevertheless, XGBoost had strong prediction accuracy and trustworthy confidence estimates (ECE = 0.0117). Entropy-based uncertainty quantification found that the model’s predictions were quite uncertain in some cases, yet it was able to mark challenging problems accurately. SHAP and LIME consistently found DELINQ, DEROG, and DEBTINC the driving variables of default risk, which was in line with accepted financial risk rationale. Utilising SHAP, LIME and entropy-based uncertainty quantification, the paper develops a framework to improve interpretation, regulatory compliance and trust in automated loan systems. It highlights the need for assurance in addition to prediction.

IPC Classification

Keywords