A stacking ensemble model with SMOTE for improved imbalanced classification on credit data

This research is based on a significant problem in credit risk analysis in the banking sector caused by class imbalance. We face the problem of the model’s inability to accurately identify risks in the ‘‘Charged Off’’ class. As a solution, we propose a stacked ensemble approach that utilizes synthetic minority over-sampling technique (SMOTE) to balance the class distribution. Experiments were conducted by applying SMOTE to the training data before training the credit model using gradient boosting (XGBoost) and random forest (RF) algorithms in a single ensemble. The results show significant improvements in precision, recall, and F1-score after applying SMOTE on the unbalanced classes. The updated model achieved a striking accuracy rate of 0,97 on resampled training data. This re-search clearly identifies the problem of class imbalance as a major challenge in credit risk analysis. The application of SMOTE in a stacked ensemble was found to be effective in improving model performance, making a valuable contribution to the development of more reliable credit models for better risk management and revenue generation in financial institutions.

本研究聚焦银行业信贷风险分析中由类别不平衡引发的突出问题。当前模型存在的核心缺陷在于，无法精准识别“核销类（Charged Off）”样本中的风险。为此，我们提出一种堆叠集成学习方案，该方案借助合成少数类过采样技术（SMOTE）平衡类别分布。实验过程中，我们先对训练数据应用SMOTE进行重采样平衡，随后基于梯度提升树（XGBoost）与随机森林（RF）两种算法搭建单集成信贷风险模型。实验结果显示，在不平衡类别上应用SMOTE后，模型的精确率、召回率与F1分数均得到显著改善；经优化的模型在重采样训练集上的准确率高达0.97，表现亮眼。本研究明确证实，类别不平衡问题是信贷风险分析中的核心挑战。实践证明，将SMOTE应用于堆叠集成学习框架可有效提升模型性能，为金融机构开发更可靠的信贷风险模型以优化风险管理与营收增长贡献了宝贵的技术参考价值。