Research on Default Prediction of Chinese‐Listed Companies Integrating News Headlines

ABSTRACT Accurately predicting the default status of listed companies is crucial for risk management and investment decisions. This study develops a default prediction model for Chinese‐listed companies by combining news headlines and numerical indicators. A CNN–BiLSTM architecture is employed to encode news headline data, and a DNN is used to encode numerical indicators, with a fusion weight applied to integrate the two data types. During training, the weights of the CNN–BiLSTM and DNN encoders are frozen, while the hybrid and output layers are fine‐tuned. This approach improves prediction accuracy and robustness by combining textual and numerical data through weighted information fusion. Additionally, the incorporation of an attention mechanism enhances the model's interpretability by allowing it to focus on the most important features in the data. The main findings of the research are: (1) The proposed model, which combines both news headlines and numerical indicators, outperforms models that rely only on numerical data or only on news headlines. (2) The proposed model outperforms nine baseline models (including BERT series, Transformers, CatBoost, etc.) in terms of G‐mean, AUC and Type II Error. By reducing Type II error, the model helps minimise potential financial losses for institutions. (3) Key numerical indicators, such as ‘Audit opinion type’, ‘Common Stock Profitability Rate’ and ‘Eligibility for Margin Trading and Securities Lending’, are crucial for predicting both default and nondefault samples. In default samples, important news headline terms include ‘Aggressive expansion’, ‘Trustee management’ and ‘Pledged financing’, whereas nondefault samples are characterised by terms such as ‘Wide market potential’, ‘Stabilization and recovery’ and ‘Opening high and rising steadily’.