Chart patterns
The well-known formations are visually compelling but hard to define precisely. Where they've been rigorously tested, the evidence is mostly weak.
Head-and-shoulders, double top, triangles, flags: these patterns fill entire books and always look impressive in hindsight.
The problem is precision. When exactly is a shoulder a shoulder? How crooked can the neckline be? Ten people will draw ten different patterns on the same chart.
That's exactly where verification fails. Whatever can't be precisely defined can't be tested, and whatever can't be tested can't be called proven.
That doesn't mean it's all nonsense. Some patterns describe plausible processes, like orders getting worked through during a sideways phase. But that's a description, not a forecast, and missing the difference costs money.
Attempts to automate pattern recognition require a formal definition, typically via smoothing the price series, identifying local extremes, and setting conditions on their relative position. Studies using such methods find weak statistical signals for individual formations that mostly don't survive once trading costs and multiple-testing issues are accounted for.
The multiple-testing problem is central here. Search across many patterns, time frames, and markets, and individually significant results turn up by chance alone. Without correcting for the number of attempts, say through appropriate statistical methods, a single positive finding isn't interpretable.
There's a cognitive component on top of that. People reliably see patterns in random series, an effect shown in experiments with simulated price series: viewers identify the same formations there as in real charts. A pattern's recognizability is therefore no proof of its informational value.
Summary
- Whatever can't be precisely defined can't be tested.
- People see the same patterns in random price series as in real ones.
- A single positive result out of many attempts means nothing.
Did you get it?
Why are chart patterns hard to verify?
Because they can barely be defined precisely, and recognizing them stays subjective.
What is the multiple-testing problem?
Search across many combinations, and significant results turn up by chance alone.
What do experiments with simulated price series show?
That viewers recognize the same formations there as in real charts.
Related
- Moving averagesStage 3
- OverfittingStage 4
- Evaluating sourcesStage 3