Do chart patterns work in crypto? Nine years of evidence, shape by shape

Short answer: on this evidence, mostly not — and for two of the eighteen shapes there is no evidence at all.

Across nine years of ten major crypto pairs, deliberately strict detectors found 661 completed formations. Of the eighteen shapes the engine recognises, nine never reached a usable sample, two of those never occurred once, and of the nine that did, only two stayed above a coin flip at every horizon measured.

That is the whole answer. The rest of this page is the working, shape by shape, with the sample count next to every claim so you can see which lines are evidence and which are arithmetic.

What was measured

A hit rate here is the share of completed formations that moved in the direction classical technical analysis associates with the shape. It says nothing about how far, which is why the median change sits next to it, and nothing about reliability on its own, which is why the sample count does too.

This matters because the published figures elsewhere are not crypto figures. The most widely reproduced chart-pattern statistics come from Thomas Bulkowski's work on the stock market, and the tools that reprint them say so — one of the leading crypto pattern scanners states plainly that "no equivalent crypto-wide statistical study exists". Other platforms publish rates in the eighty-percent range with no sample count, no period, no instrument set and no detection threshold attached. A rate quoted without those cannot be checked against anything.

The answer, shape by shape

Each line is the pooled record. Follow a link for the full four-horizon table, the split by interval, and how that shape is detected.

The two that held up

• Double bottom — 50 examples. 63.3% at ten bars, still 57.5% at sixty, and the median change rises with the horizon (+8.19% at twenty, +19.47% at sixty). The only shape where direction and magnitude agree at every horizon.
• Rising wedge — 34 examples. 61.8% at five bars, 64.7% at ten, 61.8% at twenty, with a median change that is negative throughout. Smaller sample, but internally consistent.

The ones that resolved at or below chance

• Double top — 55 examples, and 29.1% at ten bars. Price was higher ten bars later in roughly seven cases out of ten, which is the opposite of the classical reading.
• Falling wedge — 66 examples. Decays with time: 57.6% at five bars, 53.0% at ten, 47.0% at twenty, 45.5% at sixty. Whatever it picks up appears to be short-lived.
• Ascending triangle — 43 examples. 44.2% at five bars and 48.8% at sixty, with a median change of +0.35% at twenty.

The ones that are technically above chance and practically empty

• Descending triangle — 74 examples, the second most common shape found. 57.5% at sixty bars, with a median change of −0.28%. More than half moved the "right" way, by roughly a quarter of one percent.
• Bull pennant — 42 examples, but the rate swings from 40.5% at ten bars to 66.7% at sixty. A sample of this size produces exactly that kind of movement.
• Bear pennant — 32 examples, sitting between 51.6% and 56.3% across all four horizons. Close enough to a coin flip that the difference is not meaningful.

The ones that make no directional claim

• Symmetrical triangle — 211 examples, by far the most common formation in the dataset, nearly a third of everything found. It indicates a move is coming, not which way, so no hit rate is published. Judge it by the spread of outcomes.
• Rectangle — same reasoning, and found twice.

The nine with nothing to say

Head and shoulders deserves a line of its own, because it is the most recognisable formation in technical analysis and the one most often quoted with a rate in the eighties. Fifteen examples in nine years cannot distinguish a reliable shape from a coin flip, and no honest procedure produces a percentage from it.

Triple top and triple bottom were found zero times. That is a statement about the detector, not a claim the shape does not exist: a triple top must clear the confidence and height thresholds three times inside one structure, and across 182,341 bars no candidate did. It is also the kind of row a marketing page has no incentive to show you.

Why the "yes" answers you find elsewhere look so different

Four things vary between any two measurements of the same question, and none of them are usually stated:

1. Detector strictness. A permissive detector finds many marginal formations; a strict one finds few clean ones. The populations are not comparable.
2. The definition of success. "Reached a projected level" can be satisfied by a brief spike. "Higher or lower after N bars" cannot. The first produces much prettier numbers.
3. Sample period and market. Nine years of crypto majors is not thirty years of equities, and the drift is not the same.
4. Whether under-sampled shapes are reported at all. A table that silently drops everything below thirty examples looks dramatically more consistent than one that shows them.

That last one is the largest single reason our table looks worse than the ones you will find on vendor sites. Half our rows say "too few examples". Those rows did not disappear from other people's data — they were never printed.

Direction is not the same as profit

Even the shapes that resolve in their textbook direction more often than not may be worth nothing after costs.

All returns here are close-to-close and exclude fees, funding and slippage. Round-trip costs on a leveraged crypto position routinely exceed a quarter of one percent, so any effect smaller than about one percent per event should be assumed gone once execution is real.

The interval nobody mentions

The pooled table above combines the daily and 4-hour intervals. Split them and something uncomfortable appears: on the daily interval only one shape reaches thirty examples — symmetrical triangle, at exactly 30. Falling wedge is found 16 times, double bottom 15, double top 6, head and shoulders once in nine years.

On the 4-hour interval, eight shapes clear the floor. So every rate on this page is, in practice, a 4-hour statistic. Double top at ten bars measured on the 4-hour alone is 28.6% over 49 samples against 29.1% pooled — the daily sample is too thin to move the combined number.

The practical consequence runs against a common belief. Longer intervals are usually treated as more reliable, and they may well carry less noise, but they carry far less evidence. Nine years is only about 2,900 daily bars, and a strict detector over 2,900 bars leaves a dozen formations.

What this does not prove

1. 661 formations is a small sample, and half the shapes never reach a usable one.
2. The ten instruments are correlated. They rise and fall together, so the effective sample is smaller than the count suggests — closer to a handful of independent market regimes than to 661 independent trials.
3. The period is one and a half market cycles, all of it post-2017 crypto. The double-top result is most plausibly explained by the drift of a sample that mostly rose, and if drift explains it then it does not carry into a period with different drift.
4. Detectors find patterns in noise. Forty independent 400-bar random walks run through the same detectors produced a nameable formation in fourteen of them.
5. Nothing here is forward-looking, and nothing here is investment advice.

How to check this yourself

Every figure on this page is reproducible, and that is the point of publishing the thresholds rather than only the results.

The engine is deterministic: the same candles produce the same output every time, on any machine, with no model call anywhere in the path that computes a level, a swing or a formation. Language models are used only to put an already-computed reading into sentences, and cannot alter a number.

So the claim "double top resolved in its textbook direction 29.1% of the time at ten bars, n=55, thresholds 0.72 confidence and 2.5 ATR, on ten USDT majors between 2017-08-17 and 2026-07-25" is a checkable statement. Anyone with the same public candle data and the same thresholds arrives at the same figure, or demonstrates that we did not.

That property also rules out a whole category of tool from ever making this kind of claim. A system that answers by passing a chart to a language model returns different readings of the same chart on different days, so there is no stable population of detections to measure and no rate about it can be reproduced. It is not a criticism of the model — it is a structural fact about what can be measured.

So should anyone look at chart formations at all?

The defensible version of the answer is narrower than either the "patterns work" or the "patterns are useless" camps would like.

What the record supports: formations are a way of describing what a chart has done, against a fixed standard, so that two people looking at the same chart can agree on what is present. That is worth something, and it is checkable.

What the record does not support: treating a named formation as a reason to expect a particular outcome. On these instruments, over this period, the directional information in most shapes is small enough to be indistinguishable from the drift of the market they were measured in.

The honest use is as a base rate with its sample count attached — "this shape has occurred 55 times and resolved this way 29.1% of the time" — rather than as a signal. That is the form these figures are published in here, and it is why the rows with nothing to say are still on the page.