An Interpretable and Fair Machine Learning Framework for Credit Risk Assessment: Predicting Repayment Probability and Timing

This paper presents a unified machine learning framework for credit risk assessment that integrates predictive modelling, interpretability, fairness evaluation, and survival analysis. Using the Lending Club and UCI Credit datasets, the study applies XGBoost to estimate repayment probabilities, SHAP values to interpret model decisions, and a Cox Proportional Hazards model to analyse repayment timing. The framework evaluates fairness across demographic groups using metrics such as selection rate, disparate impact, ROC-AUC, and calibration, ensuring responsible and transparent decision-making. Experimental results show that XGBoost achieves strong predictive performance, with an accuracy of 94.94% and an AUC of 0.987, while maintaining well-calibrated probability estimates. The Cox model achieves a concordance index of 0.92, demonstrating strong capability in predicting repayment timelines. The study highlights the importance of integrating explainability and fairness into machine learning systems used in financial decision-making. By combining probabilistic prediction, feature-level interpretation, and temporal modelling, this work contributes to the development of more transparent, reliable, and equitable credit risk assessment systems.