Large Language Models and Financial Market Sentiment

Abstract

We investigate the predictive capabilities of large language models (LLMs) in the context of forecasting aggregate stock market returns. We use ChatGPT to analyse daily news summaries of the U.S. stock market, from which we construct a market sentiment indicator for the S&P 500 Index. Our findings reveal a noteworthy negative correlation between this sentiment indicator and short-term market returns. Notably, LLMs outperform conventional sentiment classifiers, with ChatGPT exhibiting a slight edge in out-of-sample performance. This analysis underscores the substantial potential of LLMs in text analysis-a relatively underexplored data sourcefor gaining insights into asset markets.

Keywords: ChatGPT, GPT, BARD, large language model, LLM, Natural Language Processing (NLP), sentiment, behavioural finance, market return predictability, asset pricing

JEL Classification: G10, G14, G17, G41, C53