Comparative Analysis of Ensemble Learning Models for Credit Scoring: Challenges, Improvements, and Explainability

Credit scoring is an important tool in modern financial systems, helping lenders evaluate creditworthiness and manage risks. Traditional models like logistic regression are commonly used but face difficulties with complex, highdimensional, and imbalanced data. This study researches the use of ensemble learning methods, including Bagging, Boosting (such as AdaBoost, Gradient Boosting, XGBoost), and Stacking, to improve the model performance of credit scoring models. Using datasets such as the Taiwan Credit Card dataset, German Credit Data, and Lending Club Data, we compare these methods with traditional approaches. The results show that ensemble models perform better than logistic regression in terms of accuracy and AUC-ROC. However, issues like low recall for default predictions and the lack of interpretability remain challenges. To address these problems, we discuss strategies like data balancing, advanced feature selection, and using tools such as SHAP to improve model transparency. Our findings provide guidance for building more accurate and interpretable credit scoring models, making them suitable for real-world financial applications.