Chart Patterns and What They Are Not

The short answer

A chart pattern is a name for a shape that price has already drawn. "Head and shoulders", "flag", "double bottom" — each one is a label applied to a segment of history that has finished printing.

The shape is real. The name is a convenience. What neither of them contains is a probability about the next bar, and most of the statistics you will see attached to patterns were produced in a way that guarantees they look better than the pattern is.

What a pattern actually encodes

Strip a pattern of its name and you are left with a small amount of information about how price moved.

A double bottom says: price fell to an area, rallied, returned to roughly the same area, and rallied again. That is a factual statement about two reactions at one zone, which is the same observation as a support level, drawn with extra steps.

A flag says: a fast directional move was followed by a shallow, orderly move in the other direction on smaller ranges. Again factual — and again, a description of something that has finished.

This is the honest content of every pattern: a compressed summary of past OHLC data. It is genuinely useful compression. "Consolidation after an impulse near prior resistance" is a lot to hold in your head; "bull flag at resistance" is not. Compression is not prediction, though, and the slide from one to the other happens quietly, usually inside a single sentence.

Why the names feel predictive

Three things do the work, and none of them is evidence.

The names are outcome-shaped. A "reversal pattern" is called that because it is defined partly by reversing. When the shape appears and price continues instead, the shape is usually renamed — it becomes a continuation pattern, or a failed one, or it stops being discussed. The vocabulary absorbs every outcome, which is precisely what makes it feel reliable.

Humans complete shapes automatically. Visual pattern completion is fast, involuntary and indifferent to whether the pattern means anything. You will see a triangle in a random series. The seeing is not a judgement, so it does not feel like one, and it arrives with a confidence that was never earned.

Recall is selected. You remember the head and shoulders that preceded a large decline. You do not remember the eleven that did not, because you never labelled them. This is confirmation bias operating on your own chart history, and it is covered in more detail in the cognitive biases lesson.

The base-rate problem

The question a pattern claim needs to answer is not "how often did this pattern work?" It is: out of every occurrence of this shape, how often was the outcome different from what would have happened anyway?

That question requires three numbers that pattern literature almost never supplies.

1. A mechanical definition. Before you can count occurrences, the pattern must be specified so precisely that a script could find it without human judgement — exact tolerances on the height, symmetry, duration and retracement. Most published definitions are prose, and prose definitions get applied more loosely to charts that already look promising.
2. Every occurrence, not the clean ones. A complete count must include the messy, ambiguous, half-formed instances. Those are the majority, and they are the ones a human eye skips.
3. The base rate. If the instrument rose over the test window more often than it fell, a bullish pattern will "work" often, with or without the pattern. Without comparison to the unconditional outcome, a success rate is unreadable.

Add the usual problems of any historical test — sample size, the specific period chosen, the exit rule assumed, fees and slippage excluded — and you can see why the same pattern is reported with quite different numbers by different authors. When you cannot verify a figure of this kind, the correct move is not to use a smaller one. It is to carry no figure at all.

The general failure here is overfitting: finding a rule that describes one history extremely well and mistaking that description for a property of markets. The backtesting lesson covers how to test in a way that would let a pattern fail.

Context, not trigger

None of this makes patterns useless. It makes them one kind of input rather than a reason to act.

Used as a trigger, a pattern says "the shape is complete, enter now". This requires the shape to carry predictive content, which is exactly what has not been established, and it puts your entry at the most crowded and most obvious point on the chart.

Used as context, a pattern says something much weaker and much more defensible: price has been behaving in a structured way around this area, and that structure has an edge beyond which it is no longer true. You are not buying the shape. You are using the shape to locate the level that disproves your idea.

That second use survives the criticisms above, because it does not depend on the pattern predicting anything. It depends only on the pattern being a description of where recent activity sat — which it is.

The numbers below are invented for teaching. They describe no real instrument and are not a trade idea.

Suppose price has held a zone around 100 twice, pulled back to 104 after the second reaction, and is consolidating. The pattern reader sees a double bottom and a pullback. The useful translation is: buyers appeared near 100 on two occasions, so a decisive trade below that zone — say a close under 98 — says they are no longer appearing. That gives an entry region, an invalidation at 98, and a distance you can hand to the risk-reward calculator along with whatever target the surrounding structure justifies.

Notice what the pattern did there. It did not supply a probability. It supplied a level. The plan would be identical if you had never learned the word "double bottom".

Working rules

• Describe before you name. Say what price did in plain words, then attach the label if you still want it. If the plain description sounds unremarkable, the label was doing the persuading.
• Define completion in advance. Write down the exact condition that makes the pattern present. If you have to soften the definition to include the chart in front of you, the pattern is not there.
• Never state a pattern success rate you have not computed yourself, on data you selected before knowing the outcomes.
• Keep a record of the patterns you named and did not trade. Without that column, your memory of how patterns perform is being assembled from survivors.
• Ask what the pattern adds beyond the levels it contains. Often the answer is nothing, and that is fine — the levels were the useful part.

A pattern is a way of talking about a chart. Talk with it. Do not let it place the order.