An ordinary bar goes up 51.73% of the time

Search for the success rate of almost any candlestick pattern and you will find a number in the low fifties, presented as evidence that the pattern does something. Fifty-one percent, fifty-three, sometimes a suspiciously round sixty.

Almost all of them are scored against 50%, on the reasonable-sounding assumption that a coin is the thing to beat.

We scanned 1,372,584 five-minute decision points across 22 US equities and ETFs to check that assumption. An ordinary bar — no pattern, no filter, no context — is followed by a higher close fifteen minutes later 51.73% of the time.

So 50% is not the bar. Not for this universe, this period and this fifteen-minute horizon — the number will be different elsewhere, and that is the point: it has to be measured, not assumed. Where we measured it, a pattern scoring 51% is not slightly ahead of chance. It is behind an unselected bar from the same tape.

What that does to a reported number#

This is not a subtle correction. Many of the pattern success rates published online are edges of exactly this size, so subtracting the drift is the difference between a positive result and a negative one.

The mechanism is ordinary and boring: over this period, this universe and this horizon, price drifted up. Every scan run over it inherits that drift. If you do not subtract it, you are crediting the pattern with something the calendar did.

We wanted to see how much survives once you do subtract it, so we took the pattern that gets written about most and measured it properly.

First problem: nobody agrees what an engulfing candle is#

Before you can measure the pattern you have to decide which bars are the pattern, and that turns out to be the harder half. Published descriptions disagree on at least three axes:

• Bodies or the whole bar? Most wordings say the current body must engulf the previous body. Plenty say the entire candle, wicks included, must swallow the previous one.
• May the boundary touch? Some say the open must be at or beyond the previous close; others insist it must strictly exceed it.
• Is context part of the pattern? Textbooks are firm that an engulfing candle only counts as a reversal when it arrives after a move the other way. Many scans ignore that entirely.

Each reading is defensible. Each selects a different set of bars. So we wrote four:

Then we asked how much the four actually agree.

Of the 159,880 bars the most common wording flags, only 39.6% are also flagged when wicks are included. Read the other way round, 57.5% of the wicks-included set survives the bodies-only rule.

Two articles reporting a success rate for "the bullish engulfing" can be describing largely different bars. Before anyone argues about whether the pattern works, that is worth sitting with.

Second problem: it does not matter, because they all say the same thing#

Here is the part we did not expect. The four definitions disagree profoundly about which bars qualify, and then produce the same answer.

Comparing each one against the unconditional 51.73% would already be an improvement on comparing it against a coin, but it would not be a controlled comparison — pattern bars might cluster in symbols, years, times of day or volatility regimes that drift differently from the average. So each occurrence is instead compared with the non-pattern bars in its own stratum: same symbol, same year, same hour of the session, same volatility tercile within that symbol. That is 1,931 strata, and fewer than four occurrences in total had no match.

Note the two different baselines, because they are easy to conflate: a bullish call is compared with how often a matched bar went up, around 51.7%, and a bearish call with how often one went down, around 48.3%. They are not the same number and the pattern is not being asked the same question in each column.

Eight cells, not one of them positive. The best result in the table is strict's bullish half landing exactly on its matched expectation — a clean null, p = 0.99.

And the largest deviation belongs to the version that follows the textbook most closely. Adding the requirement that the pattern arrive after a move the other way — the rule every source repeats, the one supposed to separate a real reversal from a coincidence — moved it a full point further from reversal, in both directions, with p below 0.001 on each.

One plausible reading is that the context rule selects bars that have just moved, and in this sample those bars appear more likely to continue than to turn. We did not test that mechanism directly, so treat it as a hypothesis. What the measurement supports without interpretation is narrower and strange enough on its own: the filter that is supposed to make the pattern work made it worse.

It is not a volatility signal either#

A fair defence of pattern recognition is that a pattern might not call direction while still marking something worth attention — that the next fifteen minutes will be eventful, for instance.

We checked. On this horizon, the median absolute 15-minute move on a pattern bar is 0.72 to 0.74 times the symbol's own ATR, against 0.743 for the baseline. There is nothing there. The engulfing bar is not followed by a larger move than an average bar; if anything it is marginally smaller.

The number that makes all of this moot#

Suppose you disagree with everything above and believe the one-point gap is real and tradable in the other direction. Fine. How would you confirm that from your own experience?

Detecting a one percentage point deviation from a 51.73% baseline, at 80% power and 5% two-sided, takes about 19,600 occurrences. Half a point takes about 78,400.

Very few individual traders will ever accumulate that many comparable setups, and fewer still will have records clean enough to isolate an effect that size. This is not a claim that traders are careless — it is arithmetic about what a few hundred observations can resolve, and the answer is "nothing of this magnitude". Whatever the confident success rates circulating online are based on, it cannot be personal experience, because personal experience does not have the resolution.

We ran into the same wall building our own statistics panel, from the other side: it is why the app's read on a player stays locked until 85 answers, and why separating two of its situation slices properly takes 392 answers on each side.

What this does not say#

Not that patterns are worthless — this is a directional test at one fixed horizon with no stop, no target and no position management, and a null here is not evidence that no strategy built around the pattern works. Not that the result holds on daily bars, on other instruments, or in another period; it is 22 US names on 5-minute bars from 2022 to 2026.

And emphatically not that you should trade against engulfing candles. A one-point deviation is not something to act on, which is this article's own argument pointed back at itself.

What it does say is narrower, and we think more useful. Check the baseline before you believe a success rate, and check the definition before you compare two of them. We had already found something similar looking at the last candle and at opening-range breakouts, which is why our questions are built around what price actually did next rather than around pattern labels — there is no engulfing tag in the question bank, and this is why.

SwipeTA is a training game and a simulation: no real money, no broker, and it does not provide investment advice. The parameters behind our measurements are on the methodology page.