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

Overview

• This paper explores the capability of large language models (LLMs) like ChatGPT to predict stock price movements using news headlines, without direct financial training.
• The researchers find that ChatGPT's scores significantly predict out-of-sample daily stock returns, outperforming traditional methods.
• The predictability is stronger among smaller stocks and following negative news.
• The researchers develop a theoretical model to explain these findings, incorporating information capacity constraints, underreaction, limits-to-arbitrage, and LLMs.
• The model makes several key predictions, which are empirically tested.
• The paper also introduces an interpretability framework to evaluate LLMs' reasoning, contributing to AI transparency and economic decision-making.

Plain English Explanation

The researchers wanted to see if large language models (LLMs) like ChatGPT could predict how stock prices will change, even without being specifically trained on financial data. They found that ChatGPT's scores, which measure how it interprets news headlines, can actually predict future daily stock returns better than traditional methods.

This predictability was stronger for smaller companies and when the news was negative. To explain these findings, the researchers created a theoretical model that takes into account things like the limits on how much information LLMs can process, how investors sometimes underreact to news, and the constraints on being able to take advantage of these predictions.

The model made some key predictions, which the researchers then tested with data. For example, the model showed that there is a critical level of AI capability needed to make profitable predictions, and that only advanced LLMs can effectively interpret complex information like news articles.

The researchers also introduced a way to better understand how the LLMs are reasoning, which helps make these AI systems more transparent and useful for economic decision-making. Overall, the study suggests that sophisticated AI systems can now forecast stock returns in a way that was previously difficult, and this could change how information spreads and decisions are made in financial markets.

Technical Explanation

The researchers first demonstrate that ChatGPT's scores, which measure its interpretation of news headlines, can significantly predict out-of-sample daily stock returns. This predictability is stronger among smaller stocks and following negative news.

To explain these findings, the researchers develop a theoretical model that incorporates information capacity constraints, underreaction, limits-to-arbitrage, and the capabilities of LLMs. The model makes several key predictions:

1. It establishes a critical threshold in AI capabilities necessary for profitable predictions. Below this threshold, AI systems cannot effectively interpret complex information.

2. It shows that only advanced LLMs like ChatGPT have the necessary information processing power to extract predictive signals from news.

3. It predicts that widespread LLM adoption can enhance market efficiency by improving information diffusion and decision-making.

The researchers empirically test these model predictions using data. They also introduce an interpretability framework to evaluate how LLMs reason, contributing to AI transparency and economic decision-making.

Critical Analysis

The paper provides a comprehensive theoretical and empirical analysis of LLMs' ability to forecast stock returns. However, there are a few potential limitations and areas for further research:

• The study focuses on ChatGPT, which may have unique capabilities compared to other LLMs. Evaluating a broader range of models could strengthen the generalizability of the findings.

• The paper does not explore the potential risks or downsides of widespread LLM adoption in financial markets, such as increased volatility or the concentration of power in the hands of a few AI-driven traders.

• The interpretability framework is a useful first step, but more work is needed to fully understand the "black box" decision-making of these complex models and ensure their transparency and accountability.

Despite these caveats, the research presents a compelling case for the emerging capabilities of AI systems in sophisticated financial forecasting and the potential for these technologies to transform information diffusion and decision-making in markets.

Conclusion

This paper documents the impressive ability of large language models like ChatGPT to predict stock price movements using news headlines, even without direct financial training. The researchers develop a theoretical model to explain these findings, highlighting the critical thresholds in AI capabilities, the advantages of advanced LLMs, and the potential for widespread LLM adoption to enhance market efficiency.

By introducing an interpretability framework, the study also contributes to the transparency and economic decision-making applications of these AI systems. Overall, the research suggests that sophisticated return forecasting is an emerging capability of AI, with significant implications for how information flows and decisions are made in financial markets.