Get early access

AI forecasting5 min read

Can AI predict the stock market? What the research actually shows

AI can read the news and predict short-term stock moves, but the edge is small, shrinking and easy to fake in a backtest. What the research shows.

By Kristian FagerlieFounder & CEO, Sikt Intelligence

Short answer: partly, briefly, and less over time. Research finds that large language models (LLMs) such as GPT-4 can read financial news and predict how stocks will react, better than older methods. But most of that reaction happens before anyone can trade on it, the rest is concentrated in small stocks, and it shrinks as more investors use AI. Many impressive "AI beats the market" backtests are also flawed, because the AI already knew what happened. AI is on much firmer ground forecasting the events that move markets than the prices themselves.

The landmark study: ChatGPT reads the headlines

The best-known research is by Alejandro Lopez-Lira and Yuehua Tang, now published in the Journal of Financial Economics (arXiv, journal). They gave LLMs company news headlines and asked whether each was good or bad news for the stock. Their findings:

  • Using only headlines from after the model's training cutoff, GPT-4 captured the initial market response, with about 90% portfolio-day hit rates for that first reaction. But that initial move is, in the authors' words, "non-tradable": it happens before an investor could act.
  • GPT-4's scores also "significantly predict the subsequent drift" in prices, especially for small stocks and negative news.
  • Forecasting ability generally grows with model size: older, smaller models could not do it.
  • Strategy returns decline as LLM adoption rises, "consistent with improved price efficiency."

That last finding is the important one. The more people use the same tool, the faster prices absorb the news, and the smaller the opportunity becomes.

Why a stock-price edge shrinks

Stock prices are adversarial. If a signal predicts returns and many investors act on it, their trading moves prices until the signal stops working. This is the core idea of efficient markets, for which Eugene Fama shared the 2013 Nobel Prize in economics. It does not mean prices are always right, but it does mean that easy, widely known edges rarely last.

That is also why, as we wrote in can AI predict the future?, short-term market moves are much closer to random than most real-world events.

The trap: lookahead bias

An LLM learns from text up to a training cutoff date. If you test it on news from before that date, it may already have read what happened next. The backtest then looks brilliant, but only because the model is remembering, not forecasting.

A December 2025 study measured this directly (Gao, Jiang and Yan). Their "lookahead propensity" statistic estimates how likely a model is to have internalized the real outcome. It is "materially positive throughout the in-sample period and collapses essentially to zero right after the training-data cutoff." In two tasks, predicting stock returns from headlines and capital spending from earnings calls, the LLM's predictive power was amplified exactly where it was most likely to know the answer, and that effect lost significance on data from after the cutoff.

A five-point check for any "AI beats the market" claim

  1. Was it tested after the model's training cutoff? If not, assume lookahead bias.
  2. Was the signal tradable? A prediction of the first reaction is not a strategy.
  3. What about costs? Small stocks are expensive to trade in size.
  4. Is there a live record? Forecasts made in real time beat any backtest.
  5. Is there a fair baseline? Compare with simple methods and the market itself, using proper scores such as the Brier score.

Where AI forecasting does work: events

AI has made clear progress on a different task: putting probabilities on real-world events, such as elections, central bank decisions and geopolitical developments. On ForecastBench, a benchmark of such questions, several AI systems are now "statistically indistinguishable from superforecaster-level accuracy," the Forecasting Research Institute reported in July 2026. The researchers add important caveats: the results are "more consistent with superforecaster parity than with outperformance," and the comparison relies on superforecaster forecasts last collected in 2024 (FRI). We cover the details in AI vs. superforecasters.

Events suit AI better than prices for three reasons:

  • They resolve clearly. The Fed hikes or it does not. There is no ambiguity about whether the forecast was right.
  • They can be researched. Evidence, history and base rates actually inform the answer.
  • Your forecast does not change the outcome. A widely used stock signal gets traded away; a good forecast of an election does not make the election less predictable.

And events move markets. Rate decisions, inflation prints, elections and conflicts regularly move prices, which is why professional investors increasingly track event probabilities. See how hedge funds and quant firms use prediction markets.

Where Sikt Intelligence fits

Sikt Intelligence does not try to predict stock prices. Our AI superforecaster forecasts the events behind them: rate decisions, elections, geopolitical flashpoints and technology milestones, questions with clear answers that can be researched and checked. Each forecast sits next to the market's price, like our live Fed rate odds.

We hold ourselves to the standard in this article: forecasts are judged only on questions whose outcomes the AI could not have known, and every forecast is scored against what actually happens. Sikt is in research; join the waitlist for early access. Nothing here is financial or investment advice.

Key takeaways

  • LLMs such as GPT-4 can read news and predict short-term stock reactions better than older methods.
  • Most of that reaction is not tradable, the rest is concentrated in small stocks and bad news, and returns decline as AI adoption rises.
  • Lookahead bias makes many AI backtests look far better than they are: test only after the model's training cutoff.
  • AI forecasting is strongest on real-world events, where top systems now roughly match superforecasters on ForecastBench.
  • Forecasting the events that move markets is a sounder use of AI than forecasting prices directly.

FAQ

Can ChatGPT predict stock prices?

In research, GPT-4 predicted short-term stock reactions to news headlines better than older methods, especially for small stocks and negative news. But most of the reaction happens too fast to trade, and strategy returns fall as more investors use AI.

Why do AI trading backtests look so good?

Often because of lookahead bias: when tested on data from before its training cutoff, the model may already know what happened. Studies show the predictive power largely disappears on data from after the cutoff.

Is AI better at forecasting events or prices?

Events. Real-world events resolve clearly, can be researched and do not adapt to your forecast, and top AI systems now roughly match superforecasters on event-forecasting benchmarks. Stock prices adapt quickly once a signal is widely used.

Sources

Keep reading