Patterns, and the hindsight problem

The chapter that connects everything so far to the evidence in chapter 9.

How a pattern comes to exist

A named formation — a shape in price with a label and an expected consequence — enters the vocabulary by a specific route:

1. Someone examines historical charts.
2. They notice that a recognisable shape was often followed by a particular move.
3. They name it and publish it.
4. Readers find it on their own charts.

Every step of that is a search of past data for a shape that preceded an outcome. It is the definition the research literature gives for the most blatant form of the problem: an ex post, in-sample search for profitable rules.

This does not make patterns false. It makes the evidence offered for them the wrong kind of evidence. A shape found by looking for shapes that worked will, with certainty, be a shape that worked.

Four specific defects

Selection after the fact. The pattern is defined by cases where it preceded the expected move. Cases where the same shape preceded nothing are not collected, because nobody was looking for them.

No error rate. Ask how often a pattern fails and there is usually no answer — not a disputed one, none. Chapter 3 noted that support which breaks stops being called support. The same reclassification happens here: a formation that did not work is said not to have been a genuine instance. A claim that cannot record a failure cannot be evaluated, and the absence of a failure count is not a minor gap in the literature; it is the whole gap.

Identification requires judgement. Chapter 4 made this point about candles and it is sharper here. Two experienced readers will disagree about whether a given stretch of price qualifies. A method whose application is contested cannot be applied mechanically, and what cannot be applied mechanically cannot be scored.

The eye manufactures them. Humans find shapes in noise with great reliability — it is among the better-established facts about perception. Generate a random series, plot it, and recognisable formations appear. If a method finds patterns in data known to contain none, finding one in real data is not evidence.

The collective version

The most uncomfortable part, and it is in the source:

Collective data snooping is potentially even more dangerous because it is not easily recognized by each individual researcher.

Each person examining charts believes they are observing a market. In aggregate, millions of people examining the same price history, keeping what appears to work and discarding what does not, constitute an enormous undirected search. The shapes that survive that process are selected for memorability and apparent success on the historical record — which is exactly what the search optimises for, whether or not anything predictive exists.

So the popularity of a pattern is not evidence for it. Popularity is what you would observe either way.

What this does not establish

In fairness, and because the rest of this course insists on it:

None of the above shows that chart patterns contain no information. It shows that the usual evidence for them is of a kind that cannot distinguish a real effect from a search artefact. Those are different claims, and the stronger one is not available from this argument.

Some patterns have been tested with pattern-recognition algorithms, which is the honest attempt to make the shapes machine-identifiable. The research literature notes that academic work has historically been limited to mathematically expressible techniques, with visual patterns tested only more recently by such algorithms. Chapter 9 reports what the testable work found.

Working the problem

Designing a genuine test. What must be fixed in advance:

1. A mechanical definition of the pattern. Precise enough that a program identifies instances without human input — how many touches, what proportions, over how many bars, with what tolerance. This is the hardest step, and the difficulty is itself informative: writing the definition forces every judgement the chart reader makes silently into an explicit parameter, and there are more of them than anyone expects.

2. The universe. Which instruments, which exchange, which period. Chosen before looking, or you have selected the sample that works.

3. The interval. Chapter 2's free parameter.

4. The outcome and its horizon. What counts as the pattern "working" — a move of what size, measured over how many periods, from what reference price?

5. The comparison. This is the one most often missing. What happened after randomly chosen dates in the same instruments over the same period? Without that baseline, "price rose 60% of the time after the pattern" is meaningless if price rose 60% of the time generally.

6. Costs. The literature's standard is that a rule is profitable only if risk-adjusted profits exceed transaction costs. An effect smaller than the spread is not an effect anyone can use.

7. Out-of-sample data. Hold back a period, or replicate on data generated after the rule was fixed. This is the step that separates a finding from a fit.

Which is hardest: step 1. And it is hardest for a revealing reason. When you try to write down what a pattern is, you discover that practitioners do not agree, that the tolerances are elastic, and that elasticity is what allows the pattern to be found whenever it worked and disclaimed whenever it did not. The difficulty of formalising it is not an obstacle to testing the claim — it is a substantial part of the answer.

The point

Chart patterns enter the vocabulary by searching historical charts for shapes that preceded a move, which is the textbook form of in-sample data snooping, so the evidence usually offered for them cannot distinguish a real effect from a search artefact. They have no published error rate because failures get reclassified as not-genuine instances, their identification requires judgement that two readers will dispute, and the human eye reliably finds such shapes in series known to be random. Popularity is no evidence, since collective searching selects for memorability on the historical record. None of this proves patterns are empty — it shows the usual case for them is untestable, and formalising a pattern well enough to test is itself the hard part.