Forecasting China bond default with severe class-imbalanced data: A simple learning model with causal inference

We develop a parsimonious machine-learning model to forecast bond defaults in China, addressing class imbalance and endogeneity issues widely overlooked in similar research. Using data from 2014–2023, we construct over 70 potential predictors and refine the standard ensemble method by training models on class-balanced sub-samples generated by bootstrapping before aggregating them for the final prediction. Besides superior model performance, the study’s contribution lies in transcending the associative nature of comparable studies by conducting sensitivity tests and causal inference to improve interpretability and robustness. One economic insight is that China’s institutional constraints, like ownership type (and the associated implicit guarantee), create “common causes” which introduce bias and distortions in traditional models. An important policy implication is that the risk of state-owned firms, whose connections afford them to assume a higher debt ratio, are well underestimated, particularly during economic downturns or deregulation, whereas a bailout is less likely.