Guide · Research method · Updated 2026-09-20

A pattern reliability percentage is a statement about somebody's definition, somebody's universe and somebody's search

Every quantity below was computed from the exchange's daily index close files rather than quoted from anywhere. That is not a formality here. The subject of this page is a class of number that is almost always inherited, and a page that inherited one to make its point would be demonstrating the error instead of correcting it.

A pattern is a sentence until somebody writes the test

Take the most quoted candlestick pattern there is. In words it is a short session followed by a longer session in the opposite direction whose body covers the first. Every part of that sentence is a decision waiting to be taken. Does the second body have to cover the first body, or the first session's whole range including the thin extensions above and below? Must the second session open strictly below the first session's close, or is opening level with it enough? Is there a minimum size below which the first body is too small to be covered by anything? Is a preceding downward move required, and measured how?

None of those questions has a canonical answer, and published pattern research answers them differently from one volume to the next. That is not sloppiness. It is the ordinary condition of a descriptive vocabulary that predates any attempt to count with it. The consequence, though, is exact: two people who both report the reliability of this pattern may have counted disjoint sets of days.

There is an India specific twist that most treatments miss entirely. Two of the four numbers in this definition are opens, and on this exchange the open is not a trade. It is the price discovered by the pre-open call auction, the level at which the most volume clears across all the orders entered in that window. The open is a computed clearing price rather than a transaction, and an index open is a further aggregate of those clearing prices across constituents. A pattern whose definition turns on where the session opened is a pattern defined on an auction outcome.

One tolerance, and the count moves by a factor of 11

So write the test properly and expose the choice. Take the body form: the previous session closed below where it opened, this session closed above where it opened, this session opened at or below the previous close, and this session closed at or above the previous open. Now add a single tolerance, expressed as a fraction of the previous session's body, by which the last two conditions may be relaxed. Zero is the exact textbook form. A negative value demands the previous body be exceeded rather than merely matched. A positive value forgives a near miss.

Two things in that table decide how a reliability claim should be read. The first is the range of the count: 50 occurrences at the strictest setting against 545 at the loosest, on identical sessions, under a name that does not change. The second is the level of the count. At the exact form the pattern occurred 70 times in 13.7 years, about 5.1 times a year. Anybody quoting a reliability percentage for this pattern on one Indian index is quoting a percentage of roughly seventy observations, and a proportion measured on 70 observations carries a standard error of 6.0 points before anything else goes wrong.

The horizon is a second free parameter, and it is rarely stated

Reliability also requires a window. Followed by a higher close when? The same occurrences, measured over five different horizons, produce five different answers, and the base rate they must be compared against moves at the same time because a market that drifts upward is more likely to be higher after twenty sessions than after one.

Notice what happens to the unconditional column as the horizon lengthens. It climbs from 53.7 per cent to 61.8 per cent, because the longer the window the more the market's own upward drift dominates it. A pattern study that reports a high hit rate over a long horizon and no base rate has reported the drift and attributed it to the pattern. This is the single most common way a true number becomes a false claim.

The base rate is the number the claim leaves out

Here is the arithmetic, stated once so it cannot be lost. If a pattern is followed by a rise sixty per cent of the time in a market that rises fifty five per cent of the time, the pattern has added five points, not sixty. The level is almost meaningless; the difference is the entire content. Reporting the level alone is not a rounding error in presentation, it is a change of subject.

And the comparison has to be the right one. Zero is not the alternative. Neither, usually, is the unconditional rate. The correct comparison is the pattern's own shape with the tested part removed, because the shape carries information of its own. This pattern requires a lower close followed by a higher close, which is not a neutral pair of sessions: it is a short reversal, and short reversals have their own behaviour. That matched group occurred on 24.8 per cent of sessions across the panel. Choosing the alternative before seeing the result is the part of this that cannot be repaired afterwards.

Pooled across the panel the pattern occurred 1,842 times, on 2.41 per cent of available sessions, and was followed by a higher close 54.1 per cent of the time. The matched group managed 56.2 per cent and every session managed 55.5 per cent. The difference against the matched group is -2.09 points.

To know whether that is anything, rotate. Slide the outcome series against the pattern flags by a random number of sessions and recompute. Rotation keeps every scrap of serial dependence in the outcomes and every scrap of clustering in the flags, and destroys only the alignment between them, which is exactly the thing under test. Across 4,000 rotations the difference ranged over a band of -3.17 to +3.13 points at the fifth and ninety fifth percentiles, with a standard deviation of 1.94. The measured -2.09 sits comfortably inside it, at p 0.29. Nothing was detected.

What a null result does and does not rule out

A null is only as informative as the test's ability to see. So that was measured too, by injecting a known effect and checking it comes back. Forcing a small share of the flags onto sessions that did go on to rise produced a median difference of +2.8 points, and the rotation test rejected in 7 of 12 trials. A smaller injection, with a median of +1.1 points, dropped that to 2 of 12. Random flag sets of the same size drawn from the matched pool, containing nothing at all, rejected in 1 of 40. The instrument catches an effect near 2.8 points more often than not, catches one near 1.1 in only 2 of 12, and manufactures neither.

That bounds the claim precisely. This measurement would probably have caught a difference near 2.8 points, so it rules that out for this definition on this data. At 1.1 points it caught the effect in only 2 of 12 trials, so it says next to nothing about a difference that size, and no amount of careful wording would change that. On a single index the picture is weaker still: 70 occurrences give a rotation band of -9.5 to +9.6 points, so on one index anything under about nine points either way is indistinguishable from nothing. The sample size that would settle a question is nearly always larger than the sample anybody has.

Twenty three universes is not twenty three tests

The panel table is where a careless reader finds what they came for. 6 of the 23 series show a positive difference, the best at +8.6 points and the worst at -13.7, with a mean of -1.8 and a median of -1.3. Present the best row alone and you have a publishable finding. Present all 23 and you have noise with a spread.

The spread is also smaller evidence than it looks, because these series are not independent. Their mean pairwise daily return correlation is 0.68, running as high as 0.998 between two of the broad size indices which share most of their constituents by weight. Treating correlated series as independent replications and applying the usual correction for the number of them inflates the evidence badly; a standard adjustment puts the effective number of independent series at about 1.4, not 23. A study reporting that a pattern held across two dozen indices has mostly reported that Indian equity indices move together.

How a headline figure gets manufactured without anybody lying

Nothing above requires bad faith, and the mechanism is worth watching in slow motion. Start with the strict definition and a difference of -0.92 points on one index. Now add one reasonable extra condition: require that the previous session's body be at least some fraction of the recent average range, on the entirely sound reasoning that covering a body of nearly nothing should not count.

At the far end of that table the pattern has a hit rate of 75 per cent and a difference of +19.8 points. It also has 4 occurrences in 13.7 years. Every step was defensible, no number was fabricated, and the result is meaningless. It is the same machinery as any other condition added after the results were seen, and it is why the four question filters that pattern teaching commonly applies, location and higher timeframe trend and volume and invalidation, need to be understood for what they are. Each is a sensible condition. Four of them together is a search over a large space of conditions, run on a sample that was already small, and it needs a trial count exactly as much as an automated parameter sweep does.

A search wide enough finds something whether or not anything is there

That claim can be measured rather than asserted. Take the whole grid: 23 index series by 8 tolerances, 184 cells, each with its own matched base rate. Compute every cell. The best cell returns +8.6 points on 72 occurrences, which is the top row of the panel table, the strict setting itself; the worst returns -21.3. 69 of 184 cells are positive and the median cell is -1.0.

Now run the identical grid 1,200 times on randomly rotated outcomes, where by construction there is nothing to find, and record the best cell each time. A single rotation is applied to all series at once, so the cross sectional correlation that makes these series dependent is preserved.

The median best cell from pure noise is +13.3 points. The fifth percentile is +7.9 and the ninety fifth is +20.2. One draw reached +29.4. The real best cell, +8.6, is exceeded by noise in 93.4 of every 100 draws. A search of this width would have produced a headline of thirteen points on data containing no signal whatever, and thirteen points with a hit rate near sixty nine per cent is exactly the sort of figure that gets printed in a pattern table.

This is why the fourth question decides. The best of many tested rules is a statistic about the search, and the correction depends on a number that is never reported: how many definitions, universes, horizons and filters were tried before the reported one. A reader who cannot obtain the trial count can still do what was done here, which is to estimate what a plausible search would have returned from noise and require the claim to beat that instead of zero.

The universe you were shown is the one that survived

Selection does not stop at the search. It is also in the list of instruments. An index constituent list is maintained, and maintenance means removal: instruments leave for poor liquidity, for corporate events, for ceasing to meet the size criterion. A pattern study run on today's constituent list and back-extended is a study of the instruments that lasted, and the ones that failed most dramatically, whose patterns would have failed most dramatically too, are exactly the ones missing. The index series used here sidesteps some of that by being the published index level itself, which carries its own reconstitutions, but it does not escape the point: the series is continuous only because it is maintained.

Two further directions are worth naming because they are invisible. The pattern vocabulary itself is a survivor: the named formations in circulation are the ones that looked like something to somebody, which is a selection on appearance applied over decades of charts. And the published studies are survivors: work that found nothing was mostly not written up, so the visible literature is a biased sample of the work done. Reading any published claim means holding all three in view at once.

The four questions, in the order they decide

The order matters. An unanswerable first question makes the rest moot, because a figure whose definition is unknown cannot be reproduced or applied. A satisfactory first and second question with no third leaves a number that may be entirely accounted for by drift. And a claim that answers the first three and cannot answer the fourth is not false, it is simply unverified, which is a different verdict and should be recorded as one.

Applying this to borrowed figures has a blunt consequence. A reliability percentage measured on another market, in another decade, under a definition you cannot see, adjusted by a factor somebody invented, is not a number about your market at all. There is no correction that rescues it. The only thing to do with a published pattern figure is to treat it as a hypothesis worth testing and then test it, which is a day of work on data that is freely published.

What patterns are actually for

None of this makes the vocabulary worthless, and the dismissal is as lazy as the credulity. A named pattern is a compact description of a short sequence of prices. Its real function is that it makes a hypothesis explicit enough to be tested. Before a pattern has a name, the observation is a feeling about a chart; after it has a name and a written test, it is a statement that can be counted, given a base rate, and refuted. That is a considerable upgrade, and it is the upgrade that pattern atlases genuinely delivered.

What they did not deliver, and what a reader has to supply, is the comparison. The pattern is the hypothesis. The base rate is the experiment. A trader who reads a pattern and then asks what fraction of similar sessions did the same thing anyway is doing the work properly, whether or not any number survives. A trader who reads a pattern, recalls a percentage from a book, and sizes a position on it has skipped the only step that was ever going to tell them anything.

And the discipline generalises past candles. Any claim of the form "when X happens, Y follows Z per cent of the time" is answerable by the same four questions, and most of the trading claims a reader will meet are of exactly that form.

What this measurement is not

Stated plainly, because a page about reading claims carefully has to survive being read carefully. This is one pattern under one definition at one horizon on index level data. It carries no volume condition, no location condition, no higher timeframe trend condition and no invalidation rule, all of which serious pattern work applies and any of which could change the answer. Index open, high, low and close are computed aggregates, nobody transacts at an index level, and every figure here is gross of costs, taxes, spreads and execution effects. The result is a measurement of an index and a written rule, not of a tradable outcome, and it is not a forecast or a recommendation.

What does carry across is the method: define in code, fix the horizon, compute the base rate of the matched group first, count the variants, and estimate what a search of that width returns from noise. Run that on your own market and your own instrument, and the answer belongs to you rather than to somebody else's decade. Publishing it so somebody else can check it is the last step, and the one that separates a result from an anecdote.