The famous ChatGPT headline study, updated: the edge shrank as AI spread
The 2023 paper that made "ChatGPT can predict stocks" a headline has been revised several times. The latest version keeps the core result, adds 12 models, and documents something traders should notice: the strategy's returns fell as AI adoption rose.
Few AI finance papers have traveled as far as "Can ChatGPT Forecast Stock Price Movements?" by Alejandro Lopez-Lira and Yuehua Tang of the University of Florida. First posted in April 2023, it was revised for the sixth time on arXiv on October 28, 2025. The revision matters, because the version most people quote on social media is not the current one.
What the authors did
The authors collected news headlines about 4,123 U.S. common stocks from October 2021 to May 2024. The start date was chosen on purpose: the ChatGPT version they evaluated was trained on data up to September 2021, so every headline in the test came after the model's knowledge cutoff. That is the main defense against the model simply remembering what happened next.
For each headline, the model was asked whether the news was good, bad or neutral for the company's stock. The answers became a score, and the score drove a simple daily strategy: buy stocks with positive news, short stocks with negative news.
The latest version compares 12 language models, from small early models up to GPT-4.
The three findings
1. GPT-4 reads the initial reaction very well. GPT-4 achieves a daily portfolio hit rate of approximately 90% for the direction of the initial price response (portfolio-days over October 2021 to May 2024; no interval disclosed). The authors label this reaction non-tradable: it happens as the news hits, before an ordinary investor could act on a score. Context: n: portfolio-days not stated · Oct 2021 to May 2024 · Lopez-Lira and Tang, arXiv:2304.07619 v6 · margin: not disclosed
2. Some of the move continues afterward, and GPT-4 helps predict it. After the first reaction, prices tend to keep drifting in the same direction for one to two trading days. A long-short strategy built on GPT-4's reading of overnight news earned an average of 34 basis points per day before transaction costs. Its annualized Sharpe ratio was 2.97 over the sample period, against 1.66 for GPT-3.5, 1.26 for DistilBART-MNLI, and negative Sharpe ratios for most of the basic models. Context: n=headlines on 4,123 U.S. stocks · Oct 2021 to May 2024 · Lopez-Lira and Tang, arXiv:2304.07619 v6 · margin: not disclosed
3. The drift is strongest where trading is hardest. The underreaction is most pronounced in smaller stocks and after negative news. The authors' model explains why: in small, illiquid stocks fewer sophisticated traders bother to participate, and shorting on bad news is costly and restricted. When they removed stocks priced at $5 or less and those below the 20th percentile of NYSE market cap, the strategy's cumulative return before costs was still above 300% over the sample period. Context: n=stocks above $5 and above the NYSE 20th size percentile · Oct 2021 to May 2024 · arXiv:2304.07619 v6 · margin: not disclosed
The finding the headlines missed
The revised paper reports what happened to the strategy over time. Its annualized Sharpe ratio was:
| Period | Annualized Sharpe ratio |
|---|---|
| Q4 2021 | 6.54 |
| 2022 | 3.68 |
| 2023 | 2.33 |
| January to May 2024 | 1.22 |
That decline happened while GPT models' capabilities and adoption, in the authors' words, "skyrocketed." The authors call this suggestive evidence, not proof, that AI adoption is reducing information-processing frictions in real time. As more investors use language models to read news, the easy part of the signal gets priced faster.
The paper also argues that only models above a certain level of sophistication can profit at all. Below that threshold, the signal is too noisy to survive market frictions. That is consistent with the ranking by model size: the bigger the model, the better the hit rate and the drift-strategy Sharpe ratio.
The fine print on costs
Every return figure above is before transaction costs. The authors say directly that exploiting these patterns in practice is only feasible for participants whose trading costs are low enough, such as market makers. A daily-rebalanced long-short book in small stocks, with shorts on bad news, is close to the most expensive strategy a retail trader can run.
One more comparison is worth knowing. The authors pit GPT-4's zero-shot reading against supervised models trained on embeddings and actual returns over rolling six-month windows. On intraday news, with 26,109 observations, the embedding model's Sharpe ratio dropped to 1.71, against 2.53 on overnight news with 105,742 observations. GPT-4 held a Sharpe ratio of 2.63 on intraday news. The point: the language model did not need a large training set to read headlines sensibly. Context: n=26,109 intraday and 105,742 overnight observations · Oct 2021 to May 2024 · arXiv:2304.07619 v6 · margin: not disclosed
How it fits with later research
Independent work has since probed the weak spot of this kind of study: models that have read the future. A 2023 Columbia paper by Paul Glasserman and Caden Lin re-ran a simplified version of the strategy and found its out-of-sample profitability broadly aligned with the original, while reporting lower long-short returns. Our explainer on lookahead bias in AI backtests covers why the post-cutoff design used here matters so much.
What it shows / What it doesn't
It shows
- A large language model can turn plain news headlines into a signal that lines up with the direction of next-day price moves, on headlines published after its training cutoff.
- Bigger models did better: GPT-4 beat GPT-3.5 and smaller models on hit rate and on the strategy's Sharpe ratio.
- The strategy's risk-adjusted returns fell steadily from late 2021 to mid-2024 as AI use spread.
It doesn't show
- That a retail trader can capture these returns. All figures are before costs, and the authors say only very low-cost traders such as market makers could exploit them.
- That the signal still works today. The sample ends in May 2024, and the trend at the end of the sample was downward.
- That the roughly 90% hit rate is tradable. It refers to the initial reaction, which the authors describe as non-tradable. Context: n: portfolio-days not stated · Oct 2021 to May 2024 · Lopez-Lira and Tang, arXiv:2304.07619 v6 · margin: not disclosed
- Proof that AI adoption caused the decline. The authors call the link suggestive.
Why it matters
This is the study most often cited to argue that AI can read the news faster than the market. Read in full, it says something more useful: a language model's reading of headlines carried real information, the value sat mostly in small stocks and bad news where trading is costly, and the edge was shrinking as more people used the same tools. When a tool is widely available, the signal it produces tends to get priced in. The AI can do the reading. Whether anything is left to trade is the question to ask before acting.
Primary sources
- VerifiedCan ChatGPT Forecast Stock Price Movements? Return Predictability and Large Language Models (arXiv:2304.07619, v6) (arXiv / University of Florida)Primary source. Version 6 posted October 28, 2025. Sample, hit rate, Sharpe ratios, 34 bps per day and the cost caveat checked in the HTML version (https://arxiv.org/html/2304.07619v6) on October 14, 2026.
- LinkedSSRN version of the same paper (SSRN)Earlier circulation, linked from arXiv; not re-checked.
- VerifiedAssessing Look-Ahead Bias in Stock Return Predictions Generated By GPT Sentiment Analysis (arXiv:2309.17322) (arXiv / Columbia University)Independent check by Glasserman and Lin; comparison with Lopez-Lira and Tang checked October 14, 2026.
In our directories
How we used AI: AI tools helped research and draft this article. Every figure was then checked against the primary sources listed above before publication.
Not financial advice. For information and education only. AI model outputs and research results can be wrong and do not predict future returns. Nothing here is a recommendation to buy or sell any asset. See our disclaimer and methodology.