Can AI Predict the Market? What the Research Says

The short answer

On the published research we have read in the source itself: not reliably. Models can be fitted to historical market data and will produce forecasts. What the literature we could obtain does not show is that those forecasts are dependably good enough, on data the model has never seen, to justify the confidence with which they are sold.

That is a narrower statement than either side of this argument usually makes, and it is deliberately narrow. Below are the two studies, what each actually measured, and what neither can tell you.

Why there is no percentage on this page

You will find pages quoting a single figure for how accurate machine-learning market prediction is. We are not going to, and the reason is worth stating plainly: no such figure has been read in a primary source by anyone writing this.

An earlier draft of our research notes carried a directional-accuracy number that had reached us through a search-result summary of a review we were never able to open. We could not confirm it in the review itself, so it does not appear on this site at all — not hedged, not attributed, not "around" anything. A number you cannot trace is not evidence, and repeating it would make this page exactly like the ones it is trying to be an alternative to.

What follows is only what we read.

Study one: median performance, and an explicit caution

Mokhtari, Yen and Liu applied two kinds of machine learning to the prediction problem: regression models on historical price data, and classification models on public tweets — roughly, the technical and the sentiment approaches. Their conclusion, in their own words:

"The results show a median performance, implying that with the current technology of AI, it is too soon to claim AI can beat the stock markets."

Two things to hold onto. "Median performance" is the authors' own characterisation of what they measured — not weak, not strong, middling. And "too soon to claim" is a statement about the state of the evidence, not a prediction that it will never happen.

This is one study, published in 2021, and we attribute it that way. It is not a survey of the field, and anyone citing it as proof that "AI doesn't work" is overreaching as badly as a vendor citing a backtest.

Study two: fifteen models, and what happened on new data

The more pointed finding comes from a 2023 benchmark. The authors examined fifteen state-of-the-art deep-learning models for stock price trend prediction built on limit-order-book data, and — importantly — built an open framework so that every model was trained and evaluated the same way, rather than each being reported by its own authors under its own conditions.

"Our extensive experiments reveal that all models exhibit a significant performance drop when exposed to new data, thereby raising questions about their real-world market applicability."

All fifteen. Not most, not the weaker ones. That is the single most useful result we found, because it addresses the question that actually matters to someone buying a tool: not "can a model fit the past?" but "does the fit survive contact with data nobody tuned it on?"

Keep the scope honest. This is short-horizon trend prediction from order-book data. It is not a verdict on every model, asset class and time horizon, and we are not going to inflate it into one. But it is a controlled, like-for-like comparison, which is rarer and more informative than a longer list of individually-reported successes.

What we could not get, and why that matters

We tried to read two systematic reviews that would have let us say something about the literature as a whole. One returned a permissions error; the other sat behind a login. We could not read either, so neither is cited here and neither shaped a word of this page.

That gap is worth being open about. It means this article describes two studies, not a field. If you see a page confidently summarising "what the research shows" about AI market prediction, a fair question is which of it the author actually opened.

The question to ask instead

"Can AI predict the market?" is not answerable as posed, partly because "AI" covers several unrelated technologies — we separate them in the three machines sold as AI. A more answerable version, and the one worth putting to any vendor:

What was this measured on, and had the model seen that data before? The fifteen-model result above is the reason. Performance on the data a model was developed against tells you about the fitting process. Performance on genuinely unseen data tells you about the model. Only the second is a reason to act.

A related trap is the strategy that looks excellent because many variants were tried until one did — that mechanism is covered in why good backtests fail live, and it applies to fitted models just as much as to hand-written rules.

How we source and record claims like these is set out in our editorial policy, and the risk disclosure covers what this publication does and does not offer.