Can AI Detect Credit Card Risk? A Predictive Accuracy Comparison of Machine Learning Models for Credit Card Default Prediction

Credit card default prediction is a critical task for financial institutions to manage credit risk and reduce potential losses. This study evaluates the predictive accuracy of 15 machine learning models using the Taiwanese credit card default dataset and analyzes feature relevance through chi-square tests. The models cover six major machine learning categories, including linear models (Logistic Regression and Support Vector Machine), distance-based models (k-Nearest Neighbors), Naive Bayes models (Categorical and Multinomial Naive Bayes), tree-based models (Decision Tree, Random Forest, Extra Trees, and Bagging), boosting models (AdaBoost, Histogram-based Gradient Boosting, XGBoost, LightGBM, and CatBoost), and neural models (Multi-Layer Perceptron). For feature importance ranking for credit card default, our results show that recent repayment behavior, particularly payment delays, emerges as the most significant predictor of default, followed by credit limit and recent payment amounts. In contrast, demographic characteristics and billing amount features contribute minimally to predictive power. For machine learning model performance comparison, Categorical Naive Bayes achieved the highest F1-score of 0.5120, while gradient boosting models demonstrated superior performance, with CatBoost and HistGradientBoosting attaining the highest ROC AUC scores of 0.7841 and 0.7827, respectively, along with a top accuracy of 0.8215. The findings indicate that boosting algorithms offer superior performance for risk ranking and probability calibration, making them suitable for tiered intervention strategies. In contrast, Naive Bayes models excel in recall-focused applications, where identifying potential defaulters takes precedence over false positive rates. Compared to prior studies, our models achieved improved accuracy, F1-score, and AUC metrics, providing a robust framework for operational credit risk modeling. In summary, this study provided an integrated assessment of feature relevance and model performance in credit card default prediction. The findings provide practical insights that strike a balance between predictive accuracy and deployment considerations, thereby supporting more effective credit risk management strategies in the financial services industry.