Can ChatGPT Forecast Stock Price Movements?

Key Takeaways

• GPT‑4 predicted overnight market direction 93.3% of days
• Long‑short strategy would have earned ~700% return pre‑costs
• Drift signal strongest for small‑cap stocks and negative news
• Only top‑tier models captured post‑announcement drift; FinBERT failed
• Sharpe dropped to 1.22 by early 2024 as AI spread

Pulse Analysis

The October 2025 study by Lopez‑Lira and Tang leverages a clean, out‑of‑sample dataset that sidesteps look‑ahead bias, feeding GPT‑4 real‑time headlines it had never seen. By prompting the model to act as a financial expert, the researchers turned natural‑language understanding into a binary sentiment signal. The result—a 93% hit rate on initial price moves—demonstrates that the newest generation of large language models can parse nuanced corporate news with a precision previously reserved for specialized quant tools.

Performance analysis reveals a stark hierarchy among language models. While GPT‑4 and its close cousin GPT‑3.5 produced positive Sharpe ratios for post‑announcement drift, older architectures such as BERT, DistilBART‑MNLI, and even FinBERT—despite its financial fine‑tuning—failed to generate any edge. The drift effect was most pronounced for small‑cap stocks and negative news, where market participants are slower to incorporate information. A hypothetical long‑short portfolio that bought stocks flagged as "good" and shorted those marked "bad" would have amassed roughly 700% cumulative return before transaction costs, underscoring the practical profit potential of LLM‑driven signals.

However, the study also documents rapid erosion of that advantage. As AI‑driven news‑analysis tools proliferated, the strategy’s Sharpe ratio fell from 6.54 in late 2021 to just 1.22 by early 2024. High turnover—about 190% daily—and modest transaction costs render the approach viable only for institutions with ultra‑low‑latency execution. The broader lesson is that while LLMs can unlock hidden informational asymmetries, those gaps close quickly, reinforcing the need for continuous innovation in data processing and execution speed.