A1194
Title: ChatGPT-extracted news sentiment and the cross-section of U.S. stock returns
Authors: Lukas Petrasek - Charles University Prague (Czech Republic) [presenting]
Abstract: The purpose is to analyze more than one hundred thousand news headlines to investigate their impact on U.S. equity prices, leveraging ChatGPT to extract sentiment, novelty, importance, and other key linguistic metrics. For every day between 1st January 2004 and 30th November 2024, an index ranging between -1 (bad) and 1 (good) is constructed, describing the overall sentiment of news published on that given day. Individual U.S. stock returns are then regressed on the sentiment index to determine their sensitivity to the sentiment of the news headlines. Based on these exposures, a portfolio-sorting approach is implemented to analyze the cross-sectional effects of news-driven sentiment on stock performance. Findings reveal that stocks that have been highly exposed to news sentiment in the previous year earn 3.7\% higher annualized returns in the following month compared to stocks that have been less exposed to the sentiment index. The contribution is to AI-driven asset pricing models and offers practical implications for investors incorporating news-based factors into trading strategies.
