Interpretable Machine Learning Comparison for Credit Card Default Prediction

Credit card default has become an increasingly urgent issue in the financial field, causing huge economic losses and systemic risks. Traditional statistical methods, are no longer sufficient to handle the complexity and scale of modern financial data, especially their ability to manage nonlinear relationships and class imbalances is also very limited. The research aims to enhance credit card default prediction through interpretable machine learning. Three models - logistic regression (LR), random forest (RF), and eXtreme Gradient Boosting (XGBoost) - were evaluated using real-world credit card datasets from Taiwan. Optimize the model using grid search and validate the model through cross-validation. Performance is evaluated using Area Under Curve(AUC), precision, recall rate, f1 score and accuracy. SHapley Additive exPlanation (SHAP) is used to explain feature contributions and model decisions. The results show that the ensemble methods (RF and XGBoost) are significantly superior to LR, especially in dealing with imbalanced data. Repayment Status in the Most Recent Month (PAY_0), Credit Limit (LIMIT_BAL) and Repayment Status 2 Months Before the Most Recent Month (PAY_2) are the most influential predictors. On this basis, a dynamic analysis framework is proposed to help financial institutions identify high-risk customers and take preemptive measures. The research highlights the potential of explainable machine learning in credit risk analysis and provides actionable insights for financial decision-making.