Foundations of Technical Analysis: Making Chart Patterns Testable
Do chart patterns retain predictive information under a reproducible test?
Define chart patterns mechanically, use the same detection rule across samples, and compare conditional return distributions after detected patterns with a prespecified benchmark. The cited evidence concerns distributional differences, so net profitability still requires execution costs and prospective testing.
The study checks conditional distributions, not guaranteed net trading profits.
Evidence map
| Aspect | Finding |
|---|---|
| What it is | The attempt to convert the subjective vocabulary of chart reading into definitions a machine can apply, so that the claims can be tested at all. |
| Key result / formula | The method is the contribution. A price series is smoothed with kernel regression, which removes noise while preserving the shape of local extrema, and patterns are then defined by explicit conditions on the sequence of maxima and minima of the smoothed curve: a head and shoulders is a specified ordering of five extrema within tolerance bands, and so on for the other patterns. |
| Why it matters for backtesting | The transferable part is the discipline of the method more than the finding. |
What it is
Lo, Mamaysky and Wang define a set of classic patterns algorithmically and measure whether their occurrence carries information.
Key result / formula
Applying these definitions mechanically to a large sample of United States stocks over several decades, they compare the distribution of returns conditional on a pattern's occurrence with the unconditional distribution, using a goodness-of-fit test on the two distributions. Several patterns show statistically distinguishable conditional distributions, which they interpret as evidence that the patterns carry some incremental information. They are explicit about the limits: statistical distinguishability is not profitability, the effect is not shown to survive costs, and the smoothing parameter is a choice that affects what counts as a pattern.
Why it matters for backtesting
A pattern that cannot be defined without human judgement cannot be tested, and a user who claims a pattern works should be asked for the definition first; if it cannot be written as a rule Stochastly can evaluate, the claim is outside what any backtest can settle. The paper also models the right comparison: conditional against unconditional distribution rather than conditional mean against zero, which detects effects on shape as well as on location. Two cautions for an agent citing it: the smoothing parameter is a free choice that should be fixed in advance, and distinguishable is not tradable, so the result supports testing patterns on one's own data before relying on them.
Source
Lo, Mamaysky & Wang, "Foundations of Technical Analysis: Computational Algorithms, Statistical Inference, and Empirical Implementation", Journal of Finance 55(4), 2000, 1705-1765. Primary source