Optimization of Machine-Learning-Based Enterprise Loan Default Prediction Models: A Comparative Study with Traditional Scoring Methods

Abstract

This study leverages public datasets such as CSMAR and Wind to conduct a comparative analysis between machine learning models (XGBoost, LightGBM) and traditional credit scoring models (Logistic Regression, Altman Z-score) for optimizing enterprise loan default prediction models. The research addresses three critical challenges: data imbalance, integration of non-financial information, and model interpretability. By employing the SMOTE oversampling technique and SHAP value analysis, the model performance was significantly enhanced. Experimental results demonstrate that the XGBoost model achieved an AUC of 0.85, markedly outperforming the traditional Logistic Regression model (AUC=0.72). Furthermore, incorporating sentiment data improved the recall rate by 15%. The contributions of this study are threefold: first, it systematically compares the performance differences between machine learning and traditional models in enterprise loan default prediction; second, it proposes a dynamic risk assessment framework integrating financial and non-financial features, enhancing the model’s timeliness and adaptability; third, it improves model transparency through interpretable AI techniques (SHAP analysis), aligning with the regulatory requirements of Basel III and providing theoretical and practical support for risk management in commercial banks.