Can AI predict the market?
No. A language model predicts the next token in a sequence of text, which is a different operation from predicting the next move in a market — and no amount of model quality closes that gap.
The useful version of this question is not whether AI can forecast prices. It is what a model is actually doing when it produces market commentary, and which parts of a trading workflow that is genuinely good for.
What a model is doing when it comments on a chart
It is generating the text that most plausibly follows your prompt, given patterns in its training data and whatever context you supplied.
That process produces good market commentary in the same way it produces good essays: by reproducing the shape of how such things are written. Plausible commentary and accurate forecasting are different targets, and the model is optimised for the first.
Three consequences follow, and all of them matter more than model quality:
Confidence is not calibrated to accuracy. Fluent, specific language is produced regardless of whether the underlying reasoning is sound. There is no internal signal distinguishing “I have identified a real pattern” from “this is how analysis is usually phrased.”
It cannot know what it cannot see. If your tools expose spot balances but not futures positions, its answers about your exposure are confidently incomplete — and it has no sense of the shape of the hole. Same for delayed data: delayed data looks exactly like current data, and nothing downstream can tell.
Prices are not a language. Text has statistical regularities that persist. Markets are adversarial and adaptive — a persistent, publicly learnable regularity in prices attracts capital until it stops being persistent.
Where the claim usually goes wrong
“The model was right about that move.” Over enough calls, some are right. The question is whether they are right more often than chance, over a sample large enough to mean anything, which requires the kind of record-keeping almost nobody does. See expectancy vs win rate for why a handful of correct calls is uninformative.
“It was trained on financial data.” Training on text about markets teaches the model how market analysis is written. That is not the same as learning what markets do next.
“Newer models are better at this.” They are better at reasoning, instruction-following and not making arithmetic errors — all real improvements that matter for the uses below. None of them make an adaptive system predictable.
What it is genuinely good at
This is the part the prediction framing crowds out, and it is where the value actually is.
Summarising state. Exposure across positions, what is at risk, how concentrated you are. Your exchange UI is usually bad at this and a model is good at it.
Explaining rather than predicting. “Why is this position behaving like this” is a question about mechanics — funding, margin, correlation — and models answer mechanical questions well.
Catching what you did not check. Fees, funding accrual, how close liquidation is, whether a position drifted past your caps. These are things you would find by looking and often do not.
Drafting structure. Turning a vague idea into a concrete plan with an entry, a stop, a size and a stated rationale — which is useful precisely because it forces the parts you were leaving implicit into the open.
Reviewing afterwards. Reading your own trade log back with a model is better than reading it alone, mostly because it asks about the trades you would skip.
None of this requires the model to know what happens next. All of it is available with read-only access, which carries almost no risk — see can Claude trade for me.
The gap that matters
A model producing a trade plan is doing something quite different from a model predicting a price. The plan is a structured artefact you can evaluate: is the stop at a level where the idea is wrong, is the size within your limits, is the reasoning something you agree with.
That artefact is useful even when the model’s directional view is worthless, because the evaluation is yours. What is not useful is treating its conclusion as information about the future.
The distinction has a practical form: a proposal you approve versus an action it takes. The reason that boundary matters, and why it has to be enforced structurally rather than intended, is in why MCP needs an approval layer.
What we do not claim
This site publishes no price predictions, no signals and no strategy performance. Not because we are being cautious for its own sake, but because the honest position is that nobody has demonstrated a reliable way for a language model to forecast prices, and claiming otherwise would be selling something we cannot support.
The things we do claim are testable: exchange APIs behave in specific ways, the arithmetic of position sizing works out as shown, and risk limits enforced in code behave differently from limits written in a prompt.
FAQ
Can ChatGPT or Claude predict stock prices?
No. These are language models — they generate the text most likely to follow their input, which is a different task from forecasting an adaptive system. They produce fluent, confident market commentary because that is what such text looks like, and fluency is not evidence of accuracy. Nothing about model quality changes this.
Is AI useful for trading at all?
Yes, for reading rather than predicting. Summarising exposure, explaining why a position is behaving oddly, checking costs and margin you did not look at, turning a vague idea into a structured plan, and reviewing your own decisions afterwards. All of it works with read-only access and none of it requires the model to know the future.
Why do AI trading tools claim high accuracy?
Because accuracy over a short or selectively-reported period is easy to produce and hard for a reader to check. The questions worth asking are the sample size, whether the record includes every call rather than the memorable ones, and whether performance is net of fees, funding and slippage. Those three usually account for the difference between a claim and a result.
Will better models change this?
They will improve the things models are already good at — reasoning about mechanics, following instructions, not miscalculating. They will not make an adaptive, competitive system predictable, because that is a property of the system rather than a limitation of the tool.