Does Technical Analysis Work? A Testable Question

The short answer

Some of it is measurable and some of it is not defined precisely enough to test. We publish which is which.

The question has to be narrowed before it can be answered

"Does technical analysis work" is not one question. It bundles a claim about pattern recognition, a claim about market efficiency, and a claim about your own execution, and the three have different answers.

The testable version is narrower: does this specific rule, applied mechanically to this data, produce outcomes different from an arbitrary entry on the same bars? That question has an answer, and the answer can be checked.

What has to be true before a test means anything

1. The rule is written as an executable condition, so it fires the same way every time and someone else can run it.
2. The comparison is against a benchmark on identical data, because instruments drift and drift is not edge.
3. The result carries its sample size and an interval.
4. If many rules were tested at once, the significance is corrected for that. Testing 107 formations at p<0.05 produces roughly five false positives by construction.

Most published technical-analysis research fails at the first or the fourth. A rule described in prose cannot be re-run, and a study reporting only the formations that worked has selected on the outcome.

Where this site currently stands

All 125 formations are written as executable rules with published thresholds, and each carries the conditions that invalidate its own detection. The Python engine and the browser mirror have been run against each other across three timeframes: 81,301 detections, 0 differing.

The measured outcomes are not published at the moment. The run that produced them was computed on a superseded engine with the still-forming bar included, and it was withdrawn rather than published with a caveat. Hit rates return when the re-run completes.

What is worth doing in the meantime

The parts of trading that do not depend on the answer are the parts that matter most anyway: position sizing, expectancy, and the drawdown arithmetic. None requires a pattern to work, and all three are decided by you rather than by the market.

Start with how trading expectancy works and how drawdown affects recovery.

What the academic literature actually found

The published research is less one-sided than either camp suggests, and the disagreements are mostly about method rather than about results.

1. Early studies found little. Work through the 1970s and 1980s largely reported that simple rules did not beat buy-and-hold after costs — but tested few rules, on few markets, with the costs of that era.
2. Later work found some effects and then corrected them away. Several rules that looked significant across large rule families stopped being significant once the family was corrected for.
3. Where effects survive, they are small and decay. Published anomalies tend to weaken after publication, which is consistent with them being either arbitraged or overfitted to begin with.

None of that settles the question for any specific rule. It does establish the standard a claim has to meet: a benchmark, a correction across everything tested, and an out-of-sample period.

Efficiency is a matter of degree, not a yes or no

The efficient markets argument is often stated as "prices already reflect all information, so charts cannot help". The stronger and more useful version is narrower: any effect large enough to trade profitably attracts capital until it is no longer large enough.

That predicts exactly what the data shows — small effects that persist, large ones that do not — and it explains why an edge measured on a decade of history is not evidence it will be there next year.

What would change the answer

The useful position is not "it works" or "it does not" but a statement about what evidence would move you. For a specific rule, that is a short list.

1. A benchmark-relative effect that survives correction across everything tested, not just the rule being reported.
2. Replication on markets not used to develop it — an effect present on one instrument and absent on forty is a property of that instrument.
3. Survival out of sample, on a period held back and tested once.
4. A size that clears costs. An effect of a fifth of a percentage point is real and untradeable at the same time.

Written down in advance, that list is falsifiable. Written afterwards, it becomes whatever the result happened to satisfy.

Do the arithmetic

Put your own numbers in rather than trusting the example.

Open the tools

Read next

How to Read Candlestick Charts How to Calculate Position Size How Much Should You Risk Per Trade? What Is a Good Risk/Reward Ratio? All guides →