Overcoming Data Imbalance in Risk Management: A Comparative Study of Sampling Methods

Data imbalance is a significant challenge in risk management, especially in classification tasks where critical events—such as loan defaults, employee attrition, or company bankruptcy—occur less frequently than normal cases. This paper presents a comparative study of eight sampling methods—Random Undersampling (RUS), Random Oversampling (ROS), Edited Nearest Neighbor (ENN), One-Sided Selection (OSS), SMOTE, ADASYN, SMOTEENN, and SMOTETomek—across three imbalanced datasets: Taiwanese Bankruptcy Prediction, IBM HR Analytics Employee Attrition, and Loan Prediction. Using eight machine learning classifiers, the study evaluates performance using F1 Score and Negative Predictive Value (NPV), two metrics suited for imbalanced data. The results reveal that ENN achieves the highest F1 scores in high-dimensional and severely imbalanced datasets, while SMOTE-based methods perform best in large-scale datasets with moderate imbalance. Notably, RUS consistently delivers the highest NPV, highlighting its effectiveness in minimizing false negatives and supporting conservative decision-making. The findings underscore the importance of aligning sampling strategies with dataset characteristics and specific risk management objectives.