Predicting Bond Defaults in China: A Comparative and Interpretable Machine Learning Approach Based on the IFind Database

Abstract

Since 2014, the Chinese bond market has seen a large increase in default, whose shortcomings the conventional credit risk models reveal. The paper presents a systematic machine learning model to forecast bond defaults with a new dataset in the IFind database (2014-2024). We compare eight machine learning models, such as Logistic Regression, Random Forest, XGBoost, and Deep Neural Networks. The findings indicate that tree-based ensemble models, especially XGBoost (F1-score: 0.58, AUC: 0.92), are much better than the traditional linear models. Most importantly, we use SHAP (SHapley Additive exPlanations) to give model interpretability, which indicates that profitability (ROE), leverage (Debt-to-Asset Ratio), and liquidity (Quick Ratio) are the most significant predictors, whereas the predictive power of official credit ratings is surprisingly low. A suggested hybrid KMV-XGBoost model is similarly accurate but has theoretically based interpretability. This paper presents a testable, readable early warning mechanism of the Chinese bond market with direct risk management and regulatory policy implications.