Technical Analysis · Chapter 9 of 12
Candlestick patterns: which ones survive a test
The most taught and least evidenced part of the discipline — and a worked method for finding out which parts are real.
Candlestick patterns are where most people start and where the evidence is thinnest. That is not a reason to ignore them; it is a reason to be precise. Some have a plausible mechanism and some are decoration, and the difference is testable because the definitions can be made entirely mechanical.
Making the definitions codeable
Every pattern below reduces to comparisons of open, high, low, close and the body-to-range ratio from chapter 2. Once written this way they stop being a matter of opinion:
| Pattern | Mechanical definition | Claimed meaning |
|---|---|---|
| Bullish engulfing | Candle 1 bearish and NORMAL (body ≥ 80% of range); candle 2 bullish and NORMAL; candle 2's body fully contains candle 1's body | Sellers overwhelmed in one period |
| Bearish engulfing | The mirror | Buyers overwhelmed |
| Morning star | Candle 1 bearish NORMAL; candle 2 NEUTRAL (body < 80% of range) and gapped or small; candle 3 bullish NORMAL closing above candle 1's midpoint | Selling exhausted, then reversed |
| Evening star | The mirror | Buying exhausted |
| Bullish harami | Candle 1 bearish NORMAL (the mother); candle 2 bullish NEUTRAL, body fully inside candle 1's body | Momentum has stalled |
| Hammer | Lower wick ≥ 2× body; upper wick minimal; occurring after a decline | Price was rejected at the low |
| Shooting star | The mirror after an advance | Price was rejected at the high |
| Doji | Body < 5% of range | Equilibrium. Not directional on its own |
Note how much of the work the context clause does. ‘After a decline’ needs its own mechanical definition — how many bars, how big a decline — and different choices there change the results far more than the candle definition does. Most published pattern statistics do not specify it at all, which alone makes them unreproducible.
A worked test you can run
Take bullish engulfing on the NIFTY 500, ten years of daily data:
- Find every occurrence mechanically. Do not filter by whether it looks good.
- Fix the outcome in advance: the return over the next 5 days, net of a realistic round-trip cost.
- Build the control. This is the critical step, and the naive control is wrong. Since a bullish engulfing occurs after a decline, comparing it to random days measures short-horizon mean reversion, not the pattern. The correct control is a similar-sized decline without the pattern.
- Report expectancy in R, and split by year.
- Then vary the context definition — three bars down, five bars down, a 1-ATR decline — and see whether the result survives. If it only works with one definition, it does not work.
When this is done carefully, the typical finding is that single-candle and two-candle patterns have small effects that mostly disappear against a matched control and after costs. Some show a modest edge in specific contexts — usually at a well-defined level, in a specific volatility regime, or in the direction of a higher-timeframe trend. That is a genuinely informative result, and it points at what patterns are actually good for.
What candlestick patterns are actually good for
Not as standalone entries. As a timing trigger inside a decision you have already made for other reasons.
The structure that survives testing looks like: the higher timeframe says up; price has pulled back into a level defined mechanically; and now a bullish engulfing gives you a specific bar to act on, with a specific low to put a stop under.
The candle is not providing the edge. It is providing a non-arbitrary entry bar and a non-arbitrary stop reference, which is a real and underrated contribution — because the alternative is entering at a price you chose by feel, and the stop placement is where a large part of a strategy's expectancy actually lives.
Being able to say that — ‘the pattern is not the edge, it is the trigger and the stop reference’ — is the single most credible thing you can say about candlesticks.
The larger chart patterns, briefly and honestly
Head and shoulders, double tops, flags, pennants, wedges, cup and handle. The problem with all of them is the same and it is fatal to testing: identification is subjective. How asymmetric may the shoulders be? How many bars? How much overlap? Different automated implementations of ‘head and shoulders’ find substantially different sets of instances on the same data.
The consequence is not that they are meaningless — a head and shoulders is a recognisable description of a trend making a lower high after a failed extension, which is just market structure by another name. It is that the structural description is testable and the pattern name is not. If you find yourself relying on a chart pattern, express what it means structurally, and test that instead.
Check yourself