Predicting Financial Distress With <scp>ESG</scp>‐Driven Deep Learning and Risk‐Based Stratification

ABSTRACT This study contributes to the financial distress prediction literature by integrating Environmental, Social and Governance (ESG) information into predictive modelling alongside traditional financial indicators. Using a comprehensive panel of 6882 firm‐year observations from publicly listed Chinese firms, we assess the incremental predictive value of ESG variables under both raw and deep‐learned representations. We develop a two‐stage modelling framework that first applies supervised deep representation learning to construct a latent ESG risk index, and then stratifies firms into risk regimes for regime‐specific classification. Empirical results demonstrate that incorporating ESG information, jointly with financial data, significantly improves out‐of‐sample prediction accuracy and AUC across multiple machine learning algorithms, under various multicollinearity thresholds and sampling strategies. These findings illustrate the importance of structure‐aware ESG integration in capturing the conditional, non‐linear and forward‐looking aspects of corporate financial vulnerability. By offering a rigorous approach, this study provides new insights for developing advanced early‐warning systems in credit risk and financial stability assessment.