Borderline-SMOTE for Imbalanced Credit Default Prediction Using Malaysian Banking Data
Abstract
Credit risk prediction is challenged by severe class imbalance, where standard classifiers achieve misleadingly high accuracy while failing to detect actual defaults. Most existing studies apply imbalance handling techniques without analyzing whether the chosen technique matches the underlying data structure. This study applies Borderline-SMOTE to credit default prediction using a simulated Malaysian banking dataset of 10,000 borrowers with a 35.63:1 class imbalance ratio. Three machine learning models representing distinct algorithmic paradigms — Logistic Regression, Extreme Gradient Boosting, and Support Vector Machine with RBF kernel — are evaluated under four imbalance conditions: no treatment, Borderline-SMOTE, SMOTE-Tomek, and cost-sensitive learning. A K-Nearest Neighbor (K = 5) analysis reveals that 50.5% of default cases are borderline samples situated near the decision boundary, providing a data-driven basis for Borderline-SMOTE selection. Without imbalance handling, all models exceeded 97% accuracy but achieved recall below 11%, confirming the accuracy paradox. Borderline-SMOTE achieved the highest average F1-Score (0.3127) and Macro-F1 (0.6344) among all techniques, demonstrating the best balance between precision and recall. These results suggest that imbalance handling technique selection should be guided by the minority class structure rather than applied arbitrarily.