Scholarlink

<p>To address the challenges of class imbalance, insufficient model generalization ability, and lack of explainability in loan default prediction, this research proposes a three-stage loan default prediction (TSLDP) model. Unlike prior works, this study combines hybrid sampling, ensemble learning, and explainability analysis to form a predictive framework that balances accuracy and explainability. First, we design a hybrid sampling strategy integrating wasserstein generative adversarial network and local outlier factor to address the class imbalance problem. Second, we propose a stacking-based ensemble learning method that adaptively adjusts the weights of base classifiers using an improved Newton-Raphson-based optimizer strategy, thereby improving the model’s predictive accuracy and robustness. Finally, we employ the SHapley Additive exPlanations model to identify the key features that influence the probability of default. Experimental results on two real-world credit datasets demonstrate that TSLDP has competitive predictive performance and explainable results compared to the state-of-the-art methods. This study offers financial institutions a precise and interpretable tool for predicting loan default risks, thereby mitigating financial losses and safeguarding broader economic stability. </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

A Three-Stage Loan Default Prediction and Explainability Analysis Method Using WGAN-Based Sampling, Ensemble Learning, and SHAP

  • 期刊论文

  • 发表日期：

    2026年6月20日

查看全文
  • Feifei Jin,
  • Dandan Zhang,
  • Pingfan Xia,
  • Jinpei Liu

摘要

To address the challenges of class imbalance, insufficient model generalization ability, and lack of explainability in loan default prediction, this research proposes a three-stage loan default prediction (TSLDP) model. Unlike prior works, this study combines hybrid sampling, ensemble learning, and explainability analysis to form a predictive framework that balances accuracy and explainability. First, we design a hybrid sampling strategy integrating wasserstein generative adversarial network and local outlier factor to address the class imbalance problem. Second, we propose a stacking-based ensemble learning method that adaptively adjusts the weights of base classifiers using an improved Newton-Raphson-based optimizer strategy, thereby improving the model’s predictive accuracy and robustness. Finally, we employ the SHapley Additive exPlanations model to identify the key features that influence the probability of default. Experimental results on two real-world credit datasets demonstrate that TSLDP has competitive predictive performance and explainable results compared to the state-of-the-art methods. This study offers financial institutions a precise and interpretable tool for predicting loan default risks, thereby mitigating financial losses and safeguarding broader economic stability.