Can Generative AI Really Predict the Stock Market?

Key Takeaways

• Advanced generative AI models can predict the market’s initial reaction to news surprisingly well – often above human forecasts and common technical signals.
• The edge appears to fade as adoption rises: GPT-4 accuracy drops from about 62% (Q1 2024) to about 51% (Q4 2025), close to a coin-flip baseline.
• AI looks most “tradable” where markets are least efficient: small caps, negative news, and complex information types (like insider transactions).
• Transaction costs are the reality check: strong backtests can still fail in live trading once slippage, spreads, and constraints are included.

So… can it actually predict the market?

The short answer: it can often predict directional movement right after news hits, but that’s not the same thing as a plug-and-play trading strategy. This report evaluates directional accuracy across multiple models using 159,137 firm-headline-date observations (Oct 2021–May 2024). The charts below focus on what tends to get cited and linked: model accuracy, how it changes over time, and where performance is strongest.

1) Generative AI Market Prediction Accuracy

This table summarizes reported directional accuracy by model/method, alongside strategy-level risk metrics where available. It’s the clean “reference table” people typically cite when comparing AI approaches.

Here’s the same information in a cleaner visual. For methods reported as ranges (like GPT-4 and human analysts), the horizontal line shows the full range and the dot marks the midpoint.

The takeaway is separation: several AI approaches outperform common baselines on directional calls. But trading profit depends on execution – timing, liquidity, and costs.

2) AI Prediction Accuracy: 12-Month Trend

Does the edge persist – or fade?

This time-series chart is the reality check. It suggests that as AI tools become widely used, the predictive edge can compress. GPT-4 accuracy declines from roughly 62% (Q1 2024) to roughly 51% (Q4 2025), near a 50% baseline.

That pattern makes intuitive sense: if a signal becomes widely available, it can get priced in faster – turning an advantage into “just another input.”

Accelerating adoption eliminates information advantage: The most significant drops occurred during 2024, when ChatGPT reached peak adoption among retail and institutional traders. As AI-generated insights became universally accessible, markets incorporated this information into prices almost instantly, transforming what was once a valuable signal into merely another factor already reflected in market efficiency.

3) AI vs traditional signals (what it beats – and what it doesn’t)

A practical way to read this research is as a benchmark test: how do AI-driven headline strategies compare to the tools traders already use (sentiment feeds, RSI, moving averages, analyst calls, and buy-and-hold)?

Accuracy alone isn’t enough, so the next chart shows risk-adjusted performance (Sharpe ratio). This is where transaction costs and real-world constraints usually show up as the difference between “promising” and “tradeable.”

4) AI Accuracy by Market Condition – Where AI performs best

Performance isn’t uniform. The report shows stronger drift opportunities in places where markets are harder to arbitrage – especially small caps, negative news, and complex information types.

For context, here’s initial-reaction accuracy across the same segments. Notice how some categories can have modest initial accuracy but still show meaningful drift – which suggests slower price discovery rather than immediate overreaction.

Conclusion

Generative AI can predict the market’s first move after news better than many traditional baselines – at least in the tested windows. But the edge is not guaranteed: the time-series trend suggests compression as adoption rises, and transaction costs can erase much of the paper advantage. In practice, these tools tend to shine as decision support – summarizing news, stress-testing theses, and surfacing risk – more than as standalone trade signals.

FAQ

Does high accuracy mean you can profitably trade it? Not necessarily. Accuracy doesn’t include transaction costs, slippage, shorting constraints, or how fast the market prices in the news.

Why does accuracy decline over time in the trend chart? As more market participants use similar AI tools, the market can incorporate that information faster – reducing any informational advantage.

Where is AI most useful according to these results? Small caps, negative-news setups, and complex information categories – where underreaction appears more likely.

How should a typical investor use these tools? As an analysis layer: summarizing news, mapping scenarios, comparing catalysts, and spotting risks – rather than treating the output as a trade instruction.

Sources

Can ChatGPT Forecast Stock Price Movements? Return Predictability and Large Language Models (2025)

AI-Based Stock Trading: Which Gen AI Tool Is Better (2025)

2025’s Highest Profit Factor: The Top 3 AI Trading Agents (2025)

Does generative AI facilitate investor Trading? Early evidence from ChatGPT outages (2025)