How AI Trading Tools Actually Work

The short answer

"AI trading tool" does not describe one technology. It is a shop window: at least three unrelated machines are sold behind it, they take different inputs, they produce different kinds of output, and — this is the part that matters — they go wrong in different ways.

Most tools marketed as AI for trading are one of three things, or a bundle of them:

1. A rule-based indicator. It computes something from price data. A person decided the formula in advance.
2. A fitted model. It is shown historical data and adjusts its own parameters until it predicts that data well. It then produces a forecast or a classification for new data.
3. A language model. It reads text — news, your question, your trade history — and replies in sentences.

Only the second is making a prediction in the sense most buyers have in mind. The first is doing arithmetic on prices. The third is writing prose. Conflating them is how a tool that summarises news ends up being bought as a tool that forecasts markets.

So the question "does AI trading work?" cannot be answered as asked. You have to know which machine you are being sold first.

Machine one: the indicator that cannot act

An indicator takes price data and computes a number from it. Moving averages, oscillators and the rest all work this way: the formula is fixed, written by a person, and it produces the same output for the same input every time.

Two things follow, and platform documentation states both. MetaQuotes, which publishes MetaTrader 5, describes custom indicators as "intended for analyzing price dynamics" and says they are "only used for analyzing symbol price dynamics". Then it adds the sentence worth keeping: "Indicators cannot trade."

That is a hard architectural line, not a disclaimer. The thing that computes the number and the thing that places the order are separate components. A tool built on indicators is describing the market to you; it is not deciding anything. The instrument list this runs over is similarly plain — MetaQuotes describes the Market Watch window as providing "an overview of price data of financial instruments: quotes, price statistics and tick chart".

If a sales page shows an indicator and implies judgement, the judgement is being supplied by you or by something else in the bundle. The component that can act is a separate program again — on MetaTrader that is an expert advisor, and what one actually does is worth knowing before you let any software near an order.

Machine two: the model that is actually forecasting

This is the one people mean. Rather than a person writing the formula, the formula is fitted: the system is given historical data and tunes its parameters until its outputs match what happened. It is then pointed at data it has not seen.

Because this machine is genuinely making a forecast, it is the one the research addresses — and the research is the part vendors rarely quote. We go through it in detail in what the published studies actually found; the short version follows here.

A 2021 open-access study by Mokhtari, Yen and Liu applied regression machine learning to historical price data and classification machine learning to public tweets. Their own conclusion:

"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."

Read that precisely, because it is easy to over-read in either direction. It is one study, from 2021, not a survey of the field, and Trading Tech Review attributes it as such. It does not say models are useless. It says the measured performance was middling, and that the confident claim is not yet supportable.

Note also what it does not say. It does not establish that a person would have done better. "No demonstrated edge for the model" and "humans are better" are different findings, and only the first has been measured here — a distinction we take apart in where the delegation decision sits.

The output of this machine is also what most signal products are built on. When a model's score crosses a threshold, something has to be emitted, and how that becomes a signal determines whether you can evaluate it at all.

Machine three: the one that writes sentences

A language model reads text and produces text. Ask it why a market moved and you will get a fluent answer. Fluency is the trap: the output is shaped to read like an explanation whether or not the underlying reasoning is sound, and it is produced with no stake in the outcome.

This is where the European Securities and Markets Authority has been direct. In its March 2025 warning to investors on using AI for investing, ESMA writes:

"AI tools can provide valuable support in many areas, but they also come with a set of risks. Despite their innovative potential, AI tools can generate advice that could be inaccurate or misleading and that may result in poor investment decisions and significant financial losses."

That is a regulator stating the failure mode plainly: not that the tools are fraudulent, but that confident, wrong output is a normal thing for them to produce, and the cost of acting on it is yours. What else regulators have published on this — including the point that giving investment advice in the EU requires authorisation — is collected in what regulators say about using AI to invest.

This machine is also what sits behind most products sold as an assistant or a coach, and the gap between what one can do and what it can be accountable for is the subject of what an AI trading assistant does.

How to tell which machine you are looking at

Vendor copy will rarely tell you outright. Three questions usually will.

What goes in? Price and volume only, or text as well? A tool that reads news is doing something categorically different from one that reads a chart.

What kind of thing comes out? A number, a category ("bullish"), or a paragraph? A number is arithmetic. A category is a classification — ask what it was trained on. A paragraph is language, and carries no measurement at all.

Can it act? If it can place or modify orders, you have delegated a decision, not bought an opinion. That is a different purchase and deserves a different level of scrutiny — and if the case for it rests on a backtest, read why those so often fail live first.

A tool that cannot answer these three questions clearly is not necessarily a bad tool, but you cannot judge it, which amounts to the same thing when it is your money.

Where this leaves you

None of the above says do not use these tools. It says the word "AI" on a pricing page tells you almost nothing, and the three machines behind it warrant different levels of trust: the indicator is transparent and limited by design, the fitted model is the one making a real forecast and the one whose published evidence is weakest, and the language model is the most persuasive and the least measured.

What follows from that is a habit rather than a rule. Before trusting any of them, work out which machine you are holding, what it was given, and what it is allowed to do on its own. Turning that into something you can actually put to a vendor is the job of the eight questions we ask of any tool. Our editorial policy sets out how we source claims like the ones above, and the risk disclosure covers what this publication does and does not offer.