Can news predict firm bankruptcy? 

Abstract

This paper investigates whether rapidly changing business news can predict firm bankruptcy. Using the Dow Jones Newswire database (Wall Street Journal), we create U.S. firm bankruptcy-related news-based variables using ChatGPT and compare their predictive performance with FinBERT and dictionary-based models from 1998 to 2023. ChatGPT-generated variables consistently outperform alternatives, with sentiment scores showing strong predictive power for short-and long-term bankruptcy. Leveraging the full text of news articles significantly improves prediction accuracy compared to using headlines alone. Even after controlling for traditional financial predictors, news data provides incremental value, highlighting the superior capabilities of large language models in bankruptcy prediction.

Keywords: Bankruptcy, ChatGPT, Generative AI, News Data, Sentiment JEL Classification: C53, G11, G17, G33