This paper investigates the potential of large language models (LLMs), specifically ChatGPT and GPT-4, to forecast stock price movements based on news headlines, despite lacking direct financial training. By analyzing post-knowledge-cutoff headlines, the study finds that GPT-4 can accurately capture initial market reactions, achieving hit rates around 90% for non-tradable responses, and significantly predicts subsequent price drifts, especially in cases involving small stocks and negative news. The research highlights that forecasting performance generally improves with model size, suggesting that financial reasoning emerges as a capability in more complex LLMs, rather than requiring specialized training.

To contextualize these empirical findings, the authors develop a theoretical model that incorporates LLM technology, information-processing capacity constraints, underreaction phenomena, and limits to arbitrage. This model helps rationalize how LLMs can enhance market efficiency by processing public information more effectively, leading to declining returns for strategies based on LLM predictions as adoption increases. The study contributes to quantitative finance by demonstrating the practical applications of LLMs in financial forecasting and offering insights into their impact on market dynamics and efficiency.