Performance Analysis of Credit Card Behavior Score Prediction Using Machine Learning

Credit scoring is vital in the financial risk evaluation in order to inform lending decisions of institutions. Conventional credit scoring approaches examine historical financial records and rule-based model that may not be able to model the non-linear phenomenon of customer behaviors. Machine learning (ML) models present an opportunity to improve risk prediction through data-based discovery. In this study, we examine a machine-learning approach to predict credit behaviour by using the Kaggle-funneled Credit Card Behaviour Score data set. The model we trained had an accuracy of 98.57%, but the low ROC AUC (0.76) values, in addition to the poor recall values across the high-risk population, indicate a serious class imbalance issue. SHAP analysis was utilized to interpret the model decision logic, where onus attributes, bureau inquiry history, and transaction behaviors were more important. These results suggest that although machine learning has potential advantages in credit risk assessment, the issue of class imbalance and model fairness must be overcome for real-world applications in financial risk prediction.