Scholarlink

<p>To avoid lending losses due to default, properly modeling a propensity for a borrower to default is a critical part of underwriting and naturally relies on a good predictive model. Loan defaults, as a rare event, are usually a small minority of historical outcomes result in highly imbalanced data. In this study, we contrast twenty-six separate predictive modeling strategies on a real loan data set. Modeling strategies include resampling to remove imbalances, model optimization with a cost-sensitive objective, and ensemble modeling approaches. We tested the performance of the model on four subsets of the original data (10%, 20%, 30%, and 40%) using six evaluation metrics: precision, recall, F1-Score, F2-Score, G-Mean and accuracy. Among the ensemble models, LightGBM and stacking demonstrated superior performance across all metrics, followed closely by CatBoost and XGBoost. For resampling approaches, the decision tree achieved the best results when combined with synthetic oversampling (SMOTE). Similarly, in cost-sensitive modeling, the decision tree also outperformed other models. Overall, ensemble methods, particularly LightGBM and stacking, consistently delivered the best performance across all dataset splits, highlighting their effectiveness in addressing imbalanced loan default prediction.</p>

错误：搜索内容不能为空，请输入英文关键词
错误：关键词超出字数限制，请精简
高级检索

An Experimental Study of Binary Classification Approaches for Imbalanced Data on Loan Default Prediction

  • 期刊论文

  • 发表日期：

    2026年4月11日

查看全文
  • Donglin Wang,
  • Russell Zaretzki

摘要

To avoid lending losses due to default, properly modeling a propensity for a borrower to default is a critical part of underwriting and naturally relies on a good predictive model. Loan defaults, as a rare event, are usually a small minority of historical outcomes result in highly imbalanced data. In this study, we contrast twenty-six separate predictive modeling strategies on a real loan data set. Modeling strategies include resampling to remove imbalances, model optimization with a cost-sensitive objective, and ensemble modeling approaches. We tested the performance of the model on four subsets of the original data (10%, 20%, 30%, and 40%) using six evaluation metrics: precision, recall, F1-Score, F2-Score, G-Mean and accuracy. Among the ensemble models, LightGBM and stacking demonstrated superior performance across all metrics, followed closely by CatBoost and XGBoost. For resampling approaches, the decision tree achieved the best results when combined with synthetic oversampling (SMOTE). Similarly, in cost-sensitive modeling, the decision tree also outperformed other models. Overall, ensemble methods, particularly LightGBM and stacking, consistently delivered the best performance across all dataset splits, highlighting their effectiveness in addressing imbalanced loan default prediction.