Do chart patterns beat the market's own baseline?

Three sources, one pattern, three different truths.

Search for the success rate of a bearish engulfing and you will be told 57%, 75.76%, and 79%. The numbers disagree because they are answers to under-specified questions: measured over which universe, which timeframe, how many bars forward, with what separation between overlapping occurrences, and — the omission that matters most — compared against what?

A directional hit rate cannot be read on its own. Over the ten trading days after any arbitrary moment, with no pattern present at all, US stocks and ETFs in this corpus closed higher 57.8% of the time (n=46,038). A bullish pattern that resolves upward 58% of the time has therefore done nothing. Against a 50% reference it looks like an eight-point edge; against what the market delivered anyway it is +0.2 percentage points.

This page reports what happens when that comparison is made consistently, for every pattern, timeframe and confidence band we publish. It is not an argument that technical analysis is worthless — it is one measurement, with its universe, horizon and limits stated, offered so it can be checked.

How the numbers were produced.

The same harness produces both tables; only the corpus and the data source differ.

The forward window

A pattern's outcome is measured over the 10 bars after the pattern becomes knowable — the window opens at detection, never at the first bar of the formation. There is no lookahead. Occurrences closer than 10 bars are dropped so that no two measured windows overlap, which keeps a single sustained move from being counted as several independent successes.

That distinction is worth more than it sounds. A chart pattern only becomes knowable two pivot-confirming bars after its low or high prints: nobody can act on the second trough of a double bottom until the bars that confirm it as a trough have closed. Starting the forward window at the pivot itself — the convention in a great many published backtests — leaks those confirming bars into the result and inflates the measured hit rate by roughly 10 percentage points. The rates on this page start the clock at confirmation, which is a large part of why they are lower than the figures in circulation.

The baseline

For each market, timeframe and horizon, the harness measures the unconditional outcome: over all non-overlapping 10-bar forward windows in the same corpus, what fraction closed up, down, or exactly flat. A bullish pattern is scored against the up rate and a bearish one against the down rate, so the comparison is direction-matched. Lift is the hit rate minus that baseline, in percentage points.

The interval

A textbook interval treats every occurrence as independent. It is not: 117 tickers trade simultaneously and a market-wide day moves most of them together, so 117 observations sharing a date carry far less than 117 observations' worth of information. Both the pattern sample and the baseline sample are therefore aggregated into UTC calendar-day clusters, and the variance is taken across cluster totals, which allows observations inside a day to correlate arbitrarily. At the median this widens the interval by 1.4×. Both intervals are published; every verdict on this page is decided on the clustered one.

Day clusters absorb the cross-sectional market factor. They do not absorb serial correlation across days, so even the clustered interval should be read as a floor rather than a ceiling.

What survives, and why we do not report it as a finding.

Before clustering, 14 of the 105 categories showed a lift whose interval excluded zero. After clustering on the calendar day, 3 remain. All 3 are crypto categories, and 2 of them rest on fewer than 200 occurrences. They are listed here in full rather than left out.

Two reasons not to call these discoveries. First, 3 is fewer than chance predicts. Running 105 tests at the 5% level yields about 5.3 exclusions even when every true effect is zero; finding 3 is what a null result looks like, not evidence against one. Second, none survives correction for having run 105 tests. Widening each interval by the Bonferroni factor (z from 1.96 to 3.49) leaves all 3 covering zero.

Note also that the two daily crypto rows point in opposite directions on mirror-image patterns: the double bottom trails its baseline by 12.2pp while the double top beats its own by 9.8pp, on 143 and 149 occurrences. A pair of small samples disagreeing that neatly is the signature of noise.

What this measurement cannot tell you.

Each of these could move the result, and none is fixed by the analysis above.

• Windows overlap across tickers. The 10-bar separation rule removes overlap within a symbol, not across them: on any given day most of the 117 tickers are inside someone's forward window. Day clustering is a response to this, but it absorbs the cross-sectional factor only — serial correlation from one day to the next is untreated, so the true intervals are wider than the ones published here.
• The ticker universe is survivorship-biased. The corpus is liquid US large-caps and ETFs as constituted today. Companies that were delisted, acquired or collapsed are absent. This inflates the baseline — the drift measured here is the drift of the survivors. It biases the baseline upward, which if anything makes a bullish pattern's lift look worse and a bearish one's look better, so it cannot explain away a null result in both directions at once.
• Hourly history is capped. Yahoo serves roughly 730 days of 60-minute bars, so the 1h and 4h equity categories cover a much shorter and more recent span than the daily and weekly ones — a single regime rather than several.
• One horizon, one exit. Everything is measured at exactly 10 bars, close to close. A pattern that resolves in three bars or in forty is invisible here, and so is any effect that a stop, a target or a trailing exit would surface. Nothing in this study rules those out.
• Gross and directional. Returns carry no fees, no slippage, no spread and no borrow cost. Direction is all that is counted: a category can hit its direction more often and still lose money if the wins are small and the losses large. The distribution is measured separately and shipped with the API, but it is not what is tested here.
• Detector-specific. These are our geometric definitions of the six patterns at our confidence bands. A different detector draws different occurrences. The finding is about these definitions on this corpus — it does not generalise to every implementation of a head and shoulders.
• One snapshot. Everything is as of 2026-07-18. The tables are recomputed periodically and the numbers will move.

If patterns carry no lift, why sell a pattern API?

Fair question, and the honest answer is that the null result is the product rather than a problem for it.

A language model asked what a head and shoulders implies does not abstain. It answers from what the open web says, and the open web says 57%, 75.76% or 79% — confident figures with no baseline, no interval and no sample size behind them. An agent that acts on those has been handed a hallucination with a decimal point on it. Returning the measured hit rate together with its baseline, its cluster-robust interval and its n is what makes the model's uncertainty match reality. Knowing a signal is noise is worth paying for, because the alternative is a model that confidently believes otherwise.

The detection itself also remains useful as description rather than prediction: where the shape sits, what it spans, which level it is pressing against. And this result is narrow. It says nothing about trend and regime, support and resistance levels or interpreted indicators — all of which are measurements of the current state rather than directional claims about the next ten bars, and none of which this study touches.