AI crypto trading signals: what they assert, and how to judge one

Published 2026-08-02Updated 2026-08-02

A signal claims something is worth acting on; an analysis claims something is present. Only the second can be checked before the outcome. The four things a provider must publish, and why almost none of them do.

An AI crypto trading signal and an AI chart analysis are different products that look almost identical from the outside. One tells you what to do. The other tells you what is there. Confusing them is the most expensive mistake a beginner makes in this category, and the marketing of both is designed to make the confusion easy.

This is not an argument that signals are worthless. It is a description of what a signal is actually asserting, what a provider has to publish before that assertion can be evaluated, and why almost none of them publish it.

Chart Intel is not a signal service and is structurally barred from becoming one — the output policy prohibits directional calls in every locale, and the request schema has no field for holdings, capital or risk tolerance, so there is nowhere to put the inputs a personalised signal would need. That makes this a description from outside the category rather than a competitor's complaint about it.

What a signal asserts, and what an analysis asserts

SignalAnalysis
Core claimSomething is likely enough to act onSomething is present on the chart
Falsifiable byThe outcome of the tradeInspecting the chart
Needs to know about youYes, to be usefulNo
Can be checked before the outcomeNoYes
Failure modeLossBeing wrong about geometry, visibly

The last row is the important one. If an analysis says an ascending triangle is present, anyone can look at the pivots and disagree. The claim is checkable immediately, by inspection, at no cost.

A signal cannot be checked before its outcome. And a single outcome tells you almost nothing, because any individual call can be right or wrong by luck. Evaluating a signal provider requires a run of them, which means the evaluation takes months and costs money to perform. That asymmetry is the entire economics of the signal business.

The four things a provider must publish to be evaluable

Very few do. Ask for all four before anything else.

1. Every call, timestamped before its outcome

Not a selection. Not screenshots. A complete, timestamped record published at the time each call was made, including the ones that went badly.

The failure this guards against is not usually fraud. It is the ordinary human process by which the memorable wins stay in the marketing and the forgettable losses do not. A provider with an incomplete record is not necessarily dishonest, but their track record is not evidence.

2. The base rate they are being compared against

A signal that is right 60% of the time sounds good. In a market that rose over the sample period, a rule as naive as "up, every time" also scores well above 50% — our own nine-year record shows exactly this effect, and it is the most likely explanation for why a double top resolved in its textbook bearish direction only 29.1% of the time at ten bars.

So 60% against what? The honest benchmark is not a coin flip. It is the simplest possible constant rule over the same window and the same instruments. A hit rate published without that comparison is not interpretable.

3. Magnitude, not just direction

Direction alone is a broken metric, and it is broken in a way that flatters the publisher.

Our own record shows a descending triangle resolving in its textbook direction 57.5% of the time at sixty bars, with a median return of −0.28%. More than half the time it moved the "right" way, and the typical move was roughly a quarter of one percent. A provider quoting only direction on results like these is quoting a number that is technically true and practically empty.

Ask for the distribution, not the hit rate: median, quartiles, and the share of outcomes that finished more than a standard deviation from flat.

4. Costs, explicitly

Fees, funding and slippage are excluded from most published results, including the raw figures in our own record — which is why we say so wherever those figures appear.

For any claimed effect under roughly one percent per event, this single exclusion decides whether the result exists at all. Round-trip costs on a leveraged crypto position routinely exceed a quarter of a percent, and a strategy whose edge is a quarter of a percent does not have an edge.

The reproducibility test

There is one question that separates signal providers into two groups faster than any track record, and it costs nothing to ask:

Does the same market state produce the same call, every time?

If the answer is no — because a human is exercising judgement, or because a language model is being asked what it sees — then no published rate about the service is checkable. Not because anyone is lying, but because there is no fixed population of calls to measure. Run the same chart twice, get two different calls, and "the share of calls that worked" has no stable denominator.

This is the same structural point that applies to chart tools built on a model reading an image. A probabilistic front end forfeits reproducibility, and reproducibility is a precondition for measurement rather than a nice property to have.

It also explains something otherwise puzzling about this industry: why the providers with the most confident statistics are frequently the least specific about their method. Vagueness is not always concealment. Sometimes the method genuinely cannot be pinned down, and the statistic was never well-defined to begin with.

What pattern-based signals inherit

Any signal derived from chart formations inherits the limits of the underlying formation data, and those limits are more severe than the confident tone of most signal marketing suggests.

Replaying strict detectors over 182,341 bars of ten major crypto pairs across nine years produced 661 completed formations. From that:

  • Only nine of eighteen formation types reached thirty instances at any horizon. The rest cannot be scored at all.
  • Head and shoulders, the most quoted shape in technical analysis, was found fifteen times in nine years.
  • Triple top and triple bottom never appeared once.
  • Double top resolved in its textbook direction 29.1% of the time at ten bars over 55 samples.
  • Only double bottom and rising wedge stayed above a coin flip at every horizon measured.

A provider issuing frequent, confident calls on head-and-shoulders formations is working from a base of evidence that, measured honestly, contains fifteen examples. That does not make the calls wrong. It does mean the statistical justification a reader imagines is behind them is not there.

And on noise: forty independent 400-bar random walks run through the same detectors produced a nameable formation in fourteen of them. Detectors find structure in randomness. The rate at which they do so is a function of a threshold somebody chose, and if that threshold is not published, the false-positive rate is unknown.

Incentives, stated plainly

Two revenue models dominate, and they fail differently.

Subscription. The provider is paid whether or not the calls work, and their incentive is retention. Retention is served by frequency and confidence — a channel that goes quiet for three weeks loses subscribers, regardless of whether three quiet weeks was the correct read. This is a structural pressure toward more calls than the evidence supports.

Exchange affiliate revenue. The provider is paid a share of trading fees, often for the lifetime of the referred account. Their incentive is volume. Every major exchange offers this, commonly at 30–50% revenue share, and it is the dominant business model behind free crypto analysis of every kind.

The second one is ours, and stating that is part of the answer rather than a disclaimer at the bottom. This site is paid when readers trade, while publishing analysis to inexperienced people. That is a real conflict of interest and good intentions do not dissolve it. The controls that do something about it are narrow, structural and checkable:

  1. Disclose wherever a link appears, visibly, not in a footer.
  2. Never tune the engine toward more activity — no sensitivity setting, no "more actionable" mode. Determinism and version stamping are what make this auditable rather than a promise.
  3. Avoid manufactured urgency. No "opportunity", no "don't miss", no "now is the time".
  4. Do not rank exchanges by what they pay, or say so on the page if that ever changes.

When you evaluate any provider, the question of who is paid for which reader behaviour belongs beside the question of whether the analysis is good. A provider whose revenue rises with your activity has a reason to find something today. A tool that most days reports no structural change is, at minimum, not acting on that reason.

Why this product refuses the category

The refusal is enforced in code rather than in editorial policy, which is the only version of such a commitment that means anything.

The output policy permanently prohibits three capabilities in every locale: directional calls, entry and exit levels, and personalised sizing. In the Japanese locale it additionally prohibits measured targets, following the 2026 amendment bringing crypto under Japan's financial instruments law. The request schema is strict, and it contains no field for holdings, capital or risk tolerance — there is physically nowhere to put the information a personalised call would require.

The consequence is a narrower product. It names what is on the chart, shows the confidence breakdown that produced that name, and attaches what historically followed the same shape with its sample count. It does not tell anyone what to do, and it cannot be quietly extended to.

A checklist for any signal service

  1. Is every call published before its outcome is known? If not, the record is not evidence.
  2. What benchmark is the hit rate measured against? If it is a coin flip rather than a naive constant rule over the same period, the number is inflated by drift.
  3. Is magnitude published alongside direction? Direction-only results hide the rows where the typical move is a fraction of a percent.
  4. Are costs included? Below about one percent per event, this decides everything.
  5. Does the same state produce the same call? If not, no rate about the service is checkable.
  6. What are the detection thresholds? Without them, the false-positive rate on noise is unknown.
  7. How many historical instances support this pattern? Fifteen is a common answer for famous shapes.
  8. Who pays the provider, and for what behaviour? Subscription rewards frequency; affiliate rewards volume.

In short

  • A signal asserts something worth acting on; an analysis asserts something is present. Only the second can be checked before the outcome.
  • Four things make a provider evaluable: a complete timestamped record, a real benchmark, magnitude alongside direction, and costs.
  • Reproducibility is a precondition for any published rate. A provider whose calls vary run to run has no measurable denominator.
  • Pattern-based signals inherit thin evidence: 661 formations in nine years, nine of eighteen shapes unscoreable, head and shoulders found fifteen times, and 14 of 40 random walks producing a formation anyway.
  • Subscription revenue rewards frequency; affiliate revenue rewards volume. This site runs on the second, which is why it says so here.
  • Nothing on this page is investment advice, and Chart Intel is not registered to provide any.

Frequently asked questions

What is the difference between an AI trading signal and AI chart analysis?
A signal asserts that something is likely enough to act on; an analysis asserts that something is present on the chart. The analysis can be checked immediately by inspecting the pivots, at no cost. A signal cannot be checked until its outcome is known, and one outcome tells you almost nothing.
Are AI trading signals reliable?
That cannot be answered in general, only per provider, and only if the provider publishes four things: every call timestamped before its outcome, the benchmark the hit rate is measured against, magnitude alongside direction, and whether costs are included. Most publish none of them.
Why is a 60% hit rate not necessarily good?
Because the benchmark matters. In a sample period that mostly rose, a rule as naive as answering "up" every time already scores well above 50%. The meaningful comparison is a constant-direction rule over the same window and instruments, not a coin flip.
How can I tell whether a provider is reproducible?
Ask whether the same market state produces the same call every time. If a human is exercising judgement, or a language model is being asked what it sees, then there is no fixed population of calls and no published rate about the service can be checked.
How many chart formations are actually usable as a basis for a signal?
On our nine-year record, two of eighteen clear all three bars at once — enough instances, direction holding above chance at every horizon, and a median move large enough to survive costs. Those are double bottom and rising wedge. Nine of eighteen never reach thirty instances at all.
Why should I be cautious about free signals?
Because the cost of running the service is recovered somewhere. The usual routes are exchange affiliate revenue, a funnel into a paid tier, paid coin promotion, or advertising. The third is the most distorting and the hardest to detect from outside when it is not disclosed.
Does Chart Intel provide trading signals?
No, and it cannot be extended to. The output policy permanently prohibits directional calls, entry and exit levels and personalised sizing in every locale, and the request schema has no field for holdings, capital or risk tolerance, so there is nowhere to put the inputs a personalised call would need.
What changed for Japanese readers in 2026?
The 15 July 2026 amendment brought crypto assets under Japan’s financial instruments law, so providing individual trading advice requires registration. One visible consequence here is that the Japanese locale does not output measured targets while the English locale does — same engine, same chart, different output.

Read a live chart

Figures on this page come from the same deterministic engine the tool runs on. The method is documented, and the thresholds used are printed alongside the numbers.

Everything shown here is produced by software that mechanically computes and charts publicly available market data. It is general information published identically to every user and is not personalised to your circumstances, objectives, financial situation or holdings.

Nothing here is a recommendation to buy, sell or hold any crypto-asset, and no entry, exit, stop-loss or position-size guidance is provided. We are not a registered investment adviser and we do not provide personal recommendations within the meaning of applicable investment-advice rules.

Historical patterns and statistics describe the past and do not indicate or guarantee future results. Crypto-asset prices are highly volatile and you may lose your entire investment.

Any decision you take is your own. Consider seeking advice from a licensed professional in your jurisdiction before acting.

All articles

AI crypto trading signals: what they assert, and how to judge one | Chart Intel