12 八月 2026
Operations Research and Operations Management
Prof. Gavin Feng
Published in Journal of Financial Markets, June 2026
Can business news help identify firms at risk of bankruptcy before traditional financial indicators do? In a recent study, Professor Gavin FENG from the Department of Decision Analytics and Operations and the Department of Economics and Finance explores how artificial intelligence can unlock valuable predictive signals from real-time business news. Using full-text articles from the Dow Jones Newswires database, the research demonstrates that AI-generated news indicators, particularly those derived using ChatGPT, can significantly improve the prediction of corporate bankruptcy. The findings highlight the growing role of generative AI in financial analytics and risk management, offering practical implications for investors, lenders, regulators, and corporate decision-makers.
WHY: Predicting firm bankruptcy is critical for investors, creditors, and policymakers, yet traditional prediction models rely heavily on financial statements that may not fully capture rapidly changing business conditions. This study investigates whether real-time business news can provide timely and valuable information for anticipating bankruptcy risk.
WHAT: The research examines the predictive value of firm-level news indicators generated from full-text news articles using ChatGPT. It compares ChatGPT-based measures with those produced by FinBERT and traditional dictionary-based approaches, while also evaluating the added contribution of news-derived information beyond conventional financial variables.
WHERE: The study contributes to the fields of financial risk prediction, artificial intelligence, and business analytics by demonstrating how large language models can extract meaningful signals from unstructured textual data. It also provides insights relevant to financial institutions, investors, and regulators seeking more effective early-warning systems.
HOW: Using daily full-text articles from the Dow Jones Newswires database, the research generates firm-level news-based predictors and tests their ability to forecast bankruptcy. The results show that ChatGPT-based variables outperform alternative methods, with sentiment indicators exhibiting strong predictive power across multiple forecasting horizons. The study further finds that full-text articles are substantially more informative than headlines alone and that news-based measures capture timely macroeconomic information, including market volatility (VIX), real GDP growth, and recession probability, enhancing the accuracy of bankruptcy prediction models.