We let a pattern detector read nine years of charts. Here is every hit rate.

A chart pattern hit rate is the share of completed formations that went on to move in the direction classical technical analysis associates with that shape. It is the number every pattern-recognition tool is implicitly claiming when it draws a head and shoulders on your screen, and it is the number almost none of them publish.

This page publishes ours. All eighteen shapes the detector can find, including the ones that resolved against their textbook meaning more often than not, and the ones where there is not enough evidence to say anything at all. Nothing is withheld because it was unflattering, which is the only condition under which a table like this means anything.

What was measured

The figures below come from replaying the detection engine over the full available price history of ten instruments and recording what happened next.

Two of those rows do more work than the rest.

The detector thresholds. Confidence 0.72 and height 2.5 ATR are the library defaults, unmodified. That matters because pattern counts are almost entirely a function of how permissive the detector is. Loosen those two numbers and the same nine years produce several times as many formations, most of them marginal, and every rate on this page changes. A hit rate quoted without its detection thresholds is not a measurement anyone can check.

The sample floor. Below thirty completed instances, a percentage is arithmetic rather than evidence. Three cases that all moved one way is 100%, and it means nothing. Rows under the floor appear in the table with the rate withheld rather than being deleted, because a reader deserves to know the shape was looked for and rarely found.

No look-ahead

A formation's event bar is the bar on which its break became observable. The detectors derive their geometry only from bars at or before their final pivot, and scan forward only far enough to locate the first breaking bar. Forward returns are measured strictly after that bar.

This is the failure that quietly inflates most backtests of chart patterns: identify the shape using the whole swing, including the part that only exists in hindsight, then measure the move you already used to define it. The separation above is what keeps the numbers below from being circular.

The record, pooled across all instruments and timeframes

The warning marker denotes a hit rate at or below 50%. The dagger denotes a shape that makes no directional claim, so no hit rate is published for it.

Read the hit rate and the median return as one figure in two columns. Neither is interpretable without the other, and neither is interpretable without the sample count beside them.

Two shapes finished at or below a coin flip

At the twenty-bar horizon, two of the eight directional patterns with a usable sample resolved in their textbook direction no more often than chance.

Double top — 40.0% over 55 samples at 20 bars, and 29.1% at 10 bars. The classical reading of a double top is a reversal lower. In this dataset, price was higher ten bars later in roughly seven cases out of ten. The median return at that horizon is +3.19%.

Falling wedge — 47.0% at 20 bars and 45.5% at 60, over 66 samples. The interesting property here is not the level but the decay: 57.6% at five bars, 53.0% at ten, then below half. Whatever the shape is picking up appears to be a short-lived effect that has fully dissipated within a few dozen bars.

Why the double top row looks like that

The obvious objection is that the sample is drawn from a period in which these instruments spent much of their time rising. That objection is correct, and it is the most likely explanation. A bearish shape measured against a rising sample inherits the drift of the sample.

The figure is published unadjusted anyway, for two reasons.

The first is that adjusting it would require choosing a benchmark, and every choice of benchmark is an argument rather than a measurement. Subtract the instrument's own drift over the same window and the double-top row moves toward the middle — but so does every other row, and the ranking between them, which is the part a reader actually uses, barely changes.

The second is that the unadjusted number is the honest answer to the question people are really asking. Someone looking at a double top on a chart is not asking whether the shape has predictive power net of market drift. They are asking what tends to happen next. Over the last nine years, on these instruments, what tended to happen next was that price went up.

What it does not mean

It does not mean the double top is an inverted signal worth reading backwards. Fifty-five samples across ten correlated instruments is a small, narrow sample, and a rate of 29.1% at one horizon sitting next to 40.0% at the next is a pattern of numbers consistent with noise. The defensible conclusion is narrower and duller: on this evidence, a completed double top on a major crypto pair does not establish that price is more likely to fall.

Two shapes never appeared at all

Triple top and triple bottom were found zero times in 182,341 bars.

This is a statement about the detector, not a claim that the shape does not occur. A triple top has to clear a 0.72 confidence threshold and a 2.5 ATR height requirement three times inside one structure, with the geometry between the peaks staying within tolerance throughout. Each additional touch multiplies the ways a candidate can fail. Across nine years, none survived.

There is a lesson in it about every screener that offers triple tops in its feature list. A shape can be in the catalogue, be genuinely implemented, and still be something the tool has never once reported. The only way to know which you are looking at is a published count, and a count of zero is exactly the kind of thing a marketing page has no incentive to show you.

Nine of eighteen shapes cannot be scored at all

Only nine pattern kinds reach thirty completed instances at any horizon. The other nine are listed with their rates withheld.

Head and shoulders deserves particular attention, because it is the single most recognisable formation in technical analysis and the one most often quoted with a high success rate. Here it was found fifteen times in nine years. Fifteen is not enough to distinguish a genuinely reliable shape from a coin flip, and no honest procedure produces a percentage from it.

Notice also what the small samples do to the return columns. Bull flag shows a median of +9.32% at five bars — an eye-catching number produced by eight episodes, where the median and the mean are identical because there are too few observations for them to differ meaningfully. Rounding bottom shows a decisive share of 100% at five bars across three cases. Neither figure is an estimate of anything. They are printed for completeness and marked so.

Direction is not size

A hit rate counts which way price went. It says nothing about how far. Several rows in the table pair a respectable hit rate with a median return close to zero, and those rows are worth more suspicion than the ones with a low hit rate.

Descending triangle is the clearest case. It resolves in its textbook direction 57.5% of the time at sixty bars, which sounds usable. Its median return at that horizon is −0.28%. Slightly more than half the cases moved down, and the typical move was a fraction of a percent. Fees and slippage are not included in these figures, and they do not need to be large to consume a move of that size entirely.

Ascending triangle shows the same shape of result from the other side: 53.5% at twenty bars with a median of +0.35%.

Compare those with double bottom, which is the only row in the table where the two columns agree in an unambiguous way — 63.3% at ten bars with a median of +3.57%, rising to a median of +19.47% at sixty bars over 47 remaining samples. Whether that survives the drift objection raised earlier is exactly the question the double-top section raises, and the answer is that it partly does not. But it is the one row where direction and magnitude point the same way at every horizon.

The decisive share

The full distribution table carries one more column: the share of cases that finished more than one standard deviation away from zero. It separates formations followed by a real move in either direction from formations followed by nothing much.

It is the right column to read for the shapes that make no directional claim. A symmetrical triangle does not predict a direction — it is a compression that predicts a move. Scoring it against "did price move at all" would produce a figure close to 100% and would read, to anyone skimming, as a pattern that is always right. The engine returns no hit rate for those shapes on purpose, and the table prints "no directional claim" rather than a number.

For the record, symmetrical triangle is by a wide margin the most common formation in the dataset: 211 of the 661 completed instances, nearly a third of everything found.

How this compares with published success rates elsewhere

Pattern-recognition platforms routinely publish success rates in the eighty percent range. Those figures are not necessarily wrong, and the difference between them and this table is worth stating precisely rather than treating as a dispute.

Four things vary between any two such measurements:

1. Detector strictness. A permissive detector finds many marginal formations; a strict one finds few clean ones. The populations are not the same population.
2. Definition of success. Reaching a projected level is a different test from being higher or lower after a fixed number of bars. The first can be satisfied by a brief spike; the second cannot.
3. Sample period and instruments. 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. A table that silently drops every pattern with fewer than thirty instances looks far more consistent than one that shows them.

The practical consequence is that a success rate published without its sample count and detection thresholds cannot be compared with one that has them. It is not that the higher figures are dishonest; it is that they are unfalsifiable. That is the reason the thresholds sit at the top of this page rather than in a footnote.

Why these numbers can be checked

The engine that produced this table 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 in this product only to put an already-computed reading into sentences, and they cannot alter a number.

That property is what makes a claim like "double top resolved in its textbook direction 29.1% of the time at ten bars, n=55, thresholds 0.72 and 2.5 ATR" a checkable statement rather than an assertion. Anyone with the same public candle data and the same thresholds can arrive at the same figure or demonstrate that we did not.

It also rules out a category of tool from ever making this kind of claim. A system that answers a question by asking a language model to look at a chart returns a different answer to the same chart on different days. There is no version of that architecture in which a published hit rate is reproducible, because there is no fixed procedure to reproduce.

Every published rate is effectively a 4-hour statistic

The table above pools the daily and 4-hour timeframes. Separating them changes how it should be read.

On the daily timeframe, exactly one shape reaches thirty completed instances: symmetrical triangle, at 30. Everything else falls under the floor — falling wedge 16, double bottom 15, bear pennant 8, double top 6, rising wedge 4, and head and shoulders once in nine years. Rectangle, rounding bottom, triple top and triple bottom were never found on the daily at all.

On the 4-hour timeframe, eight shapes clear the floor: symmetrical triangle 181, descending triangle 68, falling wedge 50, double top 49, ascending triangle 41, bull pennant 36, double bottom 35, rising wedge 30.

So every rate printed in the pooled table 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 figure much in either direction.

This has a practical consequence worth stating. Someone looking at a daily chart and recalling "this shape historically did such-and-such" is very likely relying on a statistic counted on a much shorter timeframe. Nine years sounds long, but it is only about 2,900 daily bars, and a strict detector run over 2,900 bars leaves a dozen or so formations. The longer the timeframe, the thinner the evidence behind any claim about it.

Where the AI actually sits

"An AI agent analyses the crypto market" describes completely different systems depending on the implementation. In the one that produced this table, the work is split into six stages, and a language model touches exactly one of them.

1. Fetch candles — the price series for the market and timeframe, from a public exchange API.
2. Compute indicators — moving averages, ATR, volume measures, cross-checked against reference implementations.
3. Detect swings — extract the meaningful turning points from the series.
4. Match formations — test the sequence of turning points against the geometric conditions for eighteen shapes. This is where the 0.72 confidence and 2.5 ATR thresholds apply.
5. Score — combine five weighted lanes (trend, momentum, volatility, volume, pattern) into a summary of the chart's current state.
6. Narrate — and only here does a language model appear, assembling already-fixed numbers and formation names into readable sentences.

Nothing in stages one through five involves a language model. Every level, every formation name and every hit rate on this page comes out of deterministic computation. If the model fails, the system degrades to a deterministic template sentence rather than taking the analysis down with it.

That division has a consequence that bears directly on this page: the same chart returns the same answer, every time. The inverse also holds. A tool that answers by passing a chart image to a language model returns different readings of the same chart on different days, which means it cannot define a hit rate at all, let alone publish one.

The useful question to ask about any "AI market analysis" claim is therefore not how capable the model is. It is which stage the model occupies. On the side that produces the numbers, reproducibility is gone. On the side that turns them into prose, it survives.

Six questions to ask of any pattern-detection claim

There is no need to take the figures above on trust. The better use of them is as a template for interrogating any such claim, this one included.

1. What is the sample count? A percentage without an n cannot be evaluated.
2. What are the detection thresholds? Without a confidence floor and a minimum size, the population is undefined.
3. What period and which instruments? A sample drawn only from a rising market flatters upward-resolving shapes.
4. How is "success" defined? Direction after a fixed number of bars and reaching a projected level are different tests producing different numbers.
5. Are under-sampled shapes shown? A table without them may be one where only the convenient rows survived.
6. Are fees and slippage included? For any shape with a median return under 1%, this single point can reverse the conclusion.

A figure that cannot answer these six is not necessarily wrong. It is unverifiable, and an unverifiable figure is not usable as a basis for anything.

What this evidence does not establish

Stated plainly, because these caveats belong next to the numbers wherever they appear:

1. 661 formations is a small sample. It is what deliberately unforgiving detectors produce over nine years, and it is thin. A few percentage points between two rows is noise, not a ranking.
2. The ten instruments are correlated. They are USDT majors that mostly rise and fall together, so the effective sample is smaller than the raw count suggests — closer to a few independent regimes than to 661 independent trials.
3. Returns are close-to-close. Fees, funding and slippage are excluded. For the rows with sub-1% median returns, that exclusion is larger than the effect being measured.
4. The period is one market cycle and a bit. Everything here is drawn from crypto's post-2017 history. Nothing establishes that these rates carry into a different regime.
5. Nothing here is forward-looking. The table records what followed these shapes historically. It is not a forecast, and it is not investment advice — Chart Intel holds no registration to provide any.

A table that produced a clean, high, uniform set of success rates from this sample would be evidence of a methodological error somewhere, not of a good detector.

What changes when detection is automated

The reason to run this as an engine rather than by eye is not speed. It is that an automated detector can be held to a fixed definition.

A person scanning charts for double tops applies a threshold that moves with what they are hoping to find, cannot recall the instances they skipped, and never counts the ones they never noticed. That is not a criticism of the person; it is the structure of the task. The denominator in a hit rate is the thing human review cannot supply, and without a denominator there is no rate.

An engine with fixed thresholds produces the denominator by construction. It also produces the uncomfortable rows — the shape that never fired, the nine that never reached a usable sample, the famous formation that resolved against its own definition — because it has no mechanism for preferring one outcome over another. Those rows are the actual output of the exercise. A version of this table without them would be a marketing document.

In short

• Across 182,341 bars of ten major crypto pairs, deliberately strict detectors found 661 completed formations in nine years.
• Only nine of eighteen shapes reached thirty instances. The other nine have no published rate, and two of them never appeared at all.
• Double top resolved in its textbook direction 29.1% of the time at ten bars (n=55). Falling wedge decayed from 57.6% at five bars to 45.5% at sixty.
• Double bottom and rising wedge are the only shapes holding above a coin flip at every horizon.
• Hit rate without median return is misleading, and both without a sample count are meaningless.
• The whole table is thinner than it looks, and the honest reading of it is that chart formations on these instruments carry far less directional information than their reputation implies.

You can see the current formation on any of the markets below, with this same record attached to it.