This research paper investigates the application of large language models (LLMs) in sentiment analysis of U.S. financial news to predict stock market returns, comparing advanced LLMs with traditional methods. The study analyzes a comprehensive dataset of 965,375 news articles from 2010 to 2023, evaluating models including BERT, OPT, FINBERT, and the Loughran-McDonald dictionary. It finds that LLMs, particularly OPT (a GPT-3 based model), significantly outperform traditional approaches in sentiment prediction accuracy and their association with subsequent daily stock returns.

The methodology involves sentiment scoring of news articles using the various models, followed by regression analyses to quantify the impact on returns and portfolio performance assessments through long-short strategies. Results show OPT achieving the highest accuracy (74.4%) and the strongest predictive power in regressions, with coefficients indicating a positive effect on next-day returns. In portfolio management, OPT-based strategies yield a superior Sharpe ratio of 3.05, highlighting its practical utility. These findings challenge the dominance of dictionary-based methods like Loughran-McDonald, which show minimal effectiveness, and underscore a shift towards LLMs in financial analysis, with implications for market prediction, regulation, and policy.