A Comparative Study on Loan Default Classification with Imbalanced Data Processing

Credit risk default classification is a cornerstone of modern financial risk management, enabling institutions to optimize lending, allocate capital efficiently, and mitigate losses, with accurate predictions directly impacting financial system stability amid economic volatility. A critical challenge is data imbalance: default samples typically make up only 5–15% of datasets, biasing models toward the majority class and harming recall, the key metric for minimizing losses. This study compares four models (Logistic Regression, Extreme Gradient Boosting (XGBoost), Light Gradient Boosting Machine (LightGBM), Random Forest) combined with four imbalance-handling methods, using Accuracy, Recall, and F1 score as metrics. Results show tree-based models outperform Logistic Regression across metrics. For Logistic Regression, class weighting effectively improves recall; for tree-based models, class weighting boosts recall but slightly reduces F1, while Synthetic Minority Over-sampling Technique (SMOTE) and Adaptive Synthetic Sampling (ADASYN) enhance F1 but risk noise. These findings highlight optimal strategies, with future work needed on ensemble methods and interpretability to refine credit risk assessment.