Published Paper Details:

While AI is transforming retail banking operations from loan application to credit scoring, fraud prevention to customer segmentation, there is a high level of opacity in the high-performing models that is harming customer confidence and making regulatory compliance more difficult. This paper examines how Explainable Artificial Intelligence (XAI) can help achieve the balance between prediction accuracy and explainability in retail banking machine learning platforms. We analyze the comparative behaviors of four popular explanation methods: SHapley Additive exPlanations (SHAP), Local Interpretable Model-agnostic Explanations (LIME), ELI5 and Counterfactual Explanations between interpretable models (Logistic Regression, Decision Tree) and black-box models (Random Forest, XGBoost, LightGBM, CatBoost, Neural Network). We propose an end-to-end novel framework called TrustBank-XAI that consists of Data Preprocessing, Feature Engineering, Prediction Engine, Explainability Module, Trust Assessment Module and Decision Support System for creating human explainable, auditable and fairness aware credit and fraud decisions. It is assessed conceptually on four public benchmark datasets: Give Me Some Credit, Bank Marketing, Home Credit Default Risk and German Credit, with predictive metrics (Accuracy, Precision, Recall, F1-score, ROC-AUC, MCC) and explainability metrics ( Trust Score, SHAP Consistency, Explanation Fidelity, Explanation Stability, Fairness Score). Results show that the accuracy of the ensemble black-box models is slightly better than that of the interpretable ones, and the explanations generated by SHAP show the most consistency and stability across folds. The resulting trade-off between prediction and explainability, future directions on real-time explainable banking systems that are aligned with regulatory compliance and fairness, as well as the implications of regulation fairness and compliance are discussed.