Can ChatGPT Forecast Stock Price Movements? Return Predictability and Large Language Models

Abstract

We examine the potential of ChatGPT and other large language models (LLMs) to predict stock market returns using news. Categorizing headlines with ChatGPT as positive, negative, or neutral for companies' stock prices, we document a significant correlation between ChatGPT scores and subsequent daily stock returns, outperforming traditional methods. Basic models like GPT-1 and BERT cannot accurately forecast returns, indicating return forecasting is an emerging capacity of more complex LLMs, which deliver higher Sharpe ratios. We explain these puzzling return predictability patterns by testing implications from economic theories involving information diffusion frictions, limits to arbitrage, and investor sophistication. Predictability strengthens among smaller stocks and following negative news, consistent with these theories. Only advanced LLMs maintain accuracy when interpreting complex news and press releases. Finally, we present an interpretability technique to evaluate LLMs' reasoning. Overall, incorporating advanced language models into investment decisions can improve prediction accuracy and trading performance.

Keywords: Natural Language Processing (NLP), Generative Pre-training Transformer (GPT), Return Predictability, Large Language Models, ChatGPT

JEL Classification: C53, G10, G11, G12, G14, G17