Can ChatGPT Forecast Stock Price Movements?

GPT-4 can interpret corporate news well enough to predict both immediate stock-price reactions and a short-lived post-announcement drift, but realistic trading frictions and wider AI adoption sharply reduce the opportunity.

• The study covers 159,137 firm-headline-date observations for 4,123 U.S. companies from October 2021 to May 2024, using a GPT-4 version with a September 2021 knowledge cutoff to limit look-ahead bias.
• Given only a company name and headline, GPT-4 correctly identifies the direction of the immediate reaction with hit rates of 93.3% for overnight news and 88.8% for intraday news; positive overnight headlines are followed by an average 1.27% opening gain, while negative ones precede a 1.79% decline.
• Prices continue drifting in the predicted direction for another one to two trading days, producing pre-cost daily returns of 34 basis points for overnight long-short portfolios and 50 basis points for intraday portfolios, with Sharpe ratios of 2.97 and 2.63 respectively.
• Predictability is stronger after negative news and among smaller stocks: the overnight short portfolio earns 26 basis points per day with a 2.01 Sharpe ratio, versus 8 basis points and a 0.78 Sharpe ratio for the long portfolio.
• Transaction costs materially weaken the strategy: the overnight portfolio's cumulative return falls from about 700% before costs to above 300% at 5 basis points and above 100% at 10 basis points, but becomes unprofitable at 20 basis points; its Sharpe ratio also declines from 6.54 in late 2021 to 1.22 through May 2024.

Takeaway: LLMs appear most useful as research assistants for processing complex corporate disclosures rather than as automatic trading systems, since the strongest initial signal is difficult to trade, turnover and market impact are substantial, and the advantage is likely to decay as adoption improves price discovery.