Start with the test that costs nothing
04.05.04 built a thousand candlestick boards from a random walk and found that 732 of them earned a pattern name. The same test can be run on chart patterns, and it is worth running because chart patterns are supposed to be the larger, slower, more meaningful object.
So: one random walk of a thousand sessions, containing no information whatsoever. Then a rule, stated in advance — two swing highs, each higher than the five closes either side; the second within 2% of the first; ten to ninety sessions apart; with a decline of at least 5% between them.
That procedure finds twenty double tops. Nothing caused any of them. There was no distribution, no large seller, no exhausted demand. There was a random number generator and a shape that a reasonable rule was willing to call a double top.
Then change the rule slightly, and watch
Every number in that procedure is a choice, and none of them is standard. Hold the data completely fixed — the same thousand sessions — and move one choice at a time.
Widen the tolerance on how near the same price from 2% to 5% and the count goes from 20 to 58. Narrow it to 1% and it falls to 12. Change how a swing high is defined, from higher than five neighbours to higher than three, and it becomes 54. Require the decline between the highs to be 10% rather than 5% and it drops to 6.
The same chart contains between six and sixty double tops, and which number you get is decided entirely by definitions that no book agrees on and most do not state.
- The stated rule
- 20 double tops
- Tolerance 1% instead of 2%
- 12
- Tolerance 3% instead of 2%
- 33
- Tolerance 5% instead of 2%
- 58
- Swing high = higher than 3 neighbours, not 5
- 54
- Swing high = higher than 8 neighbours, not 5
- 11
- Required decline 3% instead of 5%
- 60
- Required decline 10% instead of 5%
- 6
Before you can count how often a pattern works, somebody has to decide what one is. That decision moves the answer by a factor of ten.
Why this is worse for chart patterns than for candles
04.05.04's five checks all still apply — what counted as success, by when, on what, compared with what, and net of costs. Three further problems belong to chart patterns specifically, and they are the reason this page exists separately.
Identification is not a rule, it is a person. A candlestick pattern has four numbers and a threshold; `candle-rules.js` will grade it the same way every time. A double top requires somebody to pick which highs are highs, decide how near counts as near, and choose a window. Two careful analysts reading the same chart produce different samples, and a study's result is a property of the sample.
The outcome is often defined by the pattern itself. The usual measure of success is whether price travelled the measured move — the pattern's own height, projected from its boundary. So a tall pattern is graded against a distant target and a short one against a near target. Two patterns with the same success rate have succeeded at different things, and a claim that bigger patterns work better may be a claim about arithmetic.
The base rate is harder to state. For candles you can ask what the market did after all the other days. For a chart pattern you have to ask what it did after all the other stretches of forty sessions, which overlap each other and are not independent. That is a real statistical problem, not a quibble, and it makes a clean comparison difficult even for somebody trying.
The counting that has actually been done
It would be unfair to leave the impression that nobody has tried. The best-known effort is Thomas Bulkowski's, who identified chart patterns across tens of thousands of instances by hand and published tables of what followed each one. That is a serious, laborious piece of work and it is more than most of the people quoting it have done.
What it is, read honestly, is a record of what followed one person's identifications, in one market, over particular decades, under one definition of success. Those are exactly the four things 04.05.04 said a number must state, and to his credit they are stated.
What it is not is a property of the shape. Somebody applying a different tolerance, in a different market, over different years, measuring a different outcome, would get a different table — and the experiment above shows the first of those alone is worth a factor of five.
So the number is not wrong. It is narrower than the sentence it usually appears in. The double top is 78% reliable is a claim about patterns. The underlying work supports something more like these identifications, in this market, over these years, reached this target this often. Those are different sentences and only one of them was measured.
What Granthara will and will not say
Will not: publish a success rate for any chart pattern, repeat one from elsewhere as though it settled the question, or rank patterns by reliability. 07.02.03 already promised that in as many words and this page keeps it.
Will: state the geometry exactly, name the conventions it chose, show where the thresholds bite, and hand you the means to check. `node scripts/noise-test.js` reproduces every figure on this page from the repository, and changing the procedure in it will change them — which is the lesson rather than a defect.
And the practical form of all this is short. When you meet a reliability figure, ask the five questions from 04.05.04 and then one more: who decided what counted as an instance, and would I have drawn the same ones? If the answer is not in the source, the figure is a number about somebody's eye.