Beyond Accuracy: A Multi-Dimensional Explainability and Model-Risk Framework for AI-Based Credit Decisions in Banking

Beyond predictive discrimination, selecting AI models for mortgage lending requires examining explainability, fairness, calibration, stability, and governance. This study proposes a Beyond-Accuracy, Interpretability and Model-Risk Scoring (BAIMRS) framework and analyzes recorded decisions for 9,360,942 applications in the 2025 Home Mortgage Disclosure Act (HMDA) data. These labels describe historical lender denials, not borrower default; public HMDA data do not include credit scores. Four algorithms (Logistic Regression, Random Forest, XGBoost, and LightGBM) were compared with a monotone LightGBM benchmark on a locked test cohort of 1,211,242 applications and three same-year institution holdouts. LightGBM had the highest ROC-AUC (0.8729) and precision-recall AUC (0.7696), narrowly ahead of XGBoost (0.8727, 0.7688); their paired AUC difference was 0.000179 (95% bootstrap interval -0.000149 to 0.000499). However, LightGBM AUC decreased to 0.7530 at Institution A and its calibration ECE10 reached 0.1129 at Institution C. Following correction of numeric-noise robustness tests, Random Forest and XGBoost had close nominal provisional technical means of 64.362 and 63.209, respectively; XGBoost ranked first under a 0.75 anchor-tolerance scale. Demographic-parity gaps were substantial, and expanded SHAP checks used 300 validation cases per model. Formal deployment rankings remained unavailable owing to incomplete independent governance and auditability evidence. BAIMRS provides a transparent descriptive comparison, not lending approval or regulatory certification.