How AI Trading Signals Are Generated

The short answer

A trading signal is a suggestion to take a position, and a well-formed one has four parts: the instrument, an entry zone, an invalidation — the point at which the idea is wrong — and a rationale saying why it exists at all.

How it was generated matters less than whether those four parts are present. A signal that tells you what to buy but never where it would be wrong cannot be evaluated, by you or by anyone else. That is this publication's standard, and we apply it regardless of whether a model, an indicator or a person produced the signal.

Where signals come from

In practice a signal service is doing one of a few things, and the label "AI" is used across all of them.

A rule firing. Conditions are computed from price and volume, and when they coincide, a signal is emitted. This is the most common and the most transparent: the logic is fixed and someone wrote it.

A model's output crossing a threshold. A fitted model scores the current state and emits a signal when the score passes some cut-off. Here the rule is not written but learned, which means nobody can fully state why this particular signal appeared.

A person, with software assistance. An analyst forms a view and uses tools to frame it.

A copy of someone else's position. No analysis is performed for you at all — another account's trades are mirrored into yours.

These are genuinely different products, and only the last one has a regulator's attention in quite the way described below. The underlying distinction between a written rule and a learned one is the same one running through how the underlying tools work.

The part that separates a signal from a tip

Invalidation. It is the only component that makes the rest checkable.

A signal with a stated invalidation is a falsifiable claim: it says, in advance, what would have to happen for this idea to be wrong. You can hold it to that afterwards. A signal without one can never be wrong — it can only be early, and "it was early" is the most expensive sentence in trading.

The rationale does similar work more softly. If the reason is given, you can judge whether it still holds when conditions change. If there is no reason, you are left holding a position you cannot reason about, which means you cannot decide when to abandon it either.

To be clear about what this is: our definition, adopted because a reader can actually check it. It is not a regulatory requirement, and it is not a research finding. We are not telling you that signals with invalidations are profitable — no source we have read supports any claim about signal profitability, and we would need one before saying so.

When following signals becomes something else legally

Here the ground shifts from editorial standards to a regulator's framework. In March 2023, ESMA published a supervisory briefing on firms offering copy trading services, addressing exactly the question of what such a service is in law.

On an arrangement where a firm executes orders from signal providers the client has chosen, without further intervention from the client, the briefing records at paragraph 28:

"The answer concluded that the described arrangement requires authorisation in relation to portfolio management."

And where the client has to act each time, at paragraph 29:

"when client action is required prior to each transaction, the activity performed by the firm should not be qualified as portfolio management"

That does not mean such a service is unregulated — it means it may be something else:

"Depending on the interaction with the client, other investment services may still be relevant, notably investment advice in the case of personal recommendations and RTO or execution of orders (when no personal recommendation is provided to the client). Such assessment needs to be made on a case-by-case basis."

Three cautions about reading that. It addresses firms offering copy trading services under MiFID II in the EU — not every vendor who publishes signals you choose to act on manually. It is a supervisory briefing, which sets out expectations for supervisors and firms, rather than a rule aimed at you. And ESMA itself insists the assessment is case-by-case, so anyone converting this into a tidy two-box test is going further than the regulator did.

What it does give you is a sharp practical question: does acting on this service require me to do anything each time? If the answer is no, you have handed over discretion, and a regulator considers that a materially different arrangement from reading a suggestion.

What to ask of any signal service

• Where is the invalidation? If it is absent, you are being given a tip.
• Why does this signal exist? A service that will not say is asking for trust it has not earned.
• Who executes, and do I act each time? This is the question above, and it decides what you are actually buying.
• What happens when it is wrong? Every method is wrong regularly. A service that presents only its good outcomes is selecting what you see, which is the same problem that makes backtests so misleading.

We do not publish signals, rate signal services or repeat anyone's performance figures — see the editorial policy for why, and the risk disclosure for the limits of everything here.