October 12, 2026 7 min read The Chartix Editorial Team

“Does technical analysis work” is the question every trader asks and almost nobody answers properly, because as phrased it has no answer. Technical analysis is not one thing. It is a few hundred pattern definitions, applied across every market and timeframe, tested (when tested at all) with different rules for entering, holding and exiting. Most of the argument online is two people testing different things and each calling the other a fool.

Does technical analysis work? The published candlestick tests disagree by market and period: large US stocks over 1992 to 2002 showed no value and Thai stocks over 2006 to 2016 little, while S&P 500 stocks over 1992 to 1996 and Taiwanese stocks over 2002 to 2008 showed positive results. Market, period and definition decide the answer.

What follows is a plain summary of the main studies, described only from what their published abstracts and text actually state, then the practitioner numbers everyone quotes, then an honest reading of why the results disagree and what that leaves you with.

What “does it work” has to mean before it can be tested

A testable version of the question names four things: a pattern with a written definition, a market, a period, and a rule for what you do after the pattern appears and when you stop. Change any one and you have a different experiment. Two of the studies below use different exit conventions for the same kind of pattern, one exiting on a fixed date and one at an average price over the holding period, and the Thai study explicitly tests both because the choice matters.

That is the first thing to hold on to. When someone tells you candlestick patterns are proven or debunked, ask which pattern, which market, which years and which exit. If they cannot say, they are reporting a feeling.

The main published tests

Marshall, Young and Rose (2006), published in the Journal of Banking and Finance and also available as Marshall’s 2005 Massey University thesis of the same title, tested candlestick trading strategies on Dow Jones Industrial Average component stocks over 1992 to 2002, with the short holding periods candlestick analysis traditionally uses (ten days or fewer). They used a bootstrap method that generates random open, high, low and close series to compare against. Their conclusion, as the thesis abstract states it, is that candlestick technical analysis does not have value and there is no evidence that a trader following it would outperform the market. The authors also note their test is out of sample in a strong sense: candlesticks were developed on Japanese rice data centuries before this US stock data existed.

Caginalp and Laurent (1998), in Applied Mathematical Finance, tested three-day candlestick patterns on all S&P 500 stocks over 1992 to 1996. Their abstract reports out-of-sample statistical significance and a profit of almost 1% over a two-day holding period, and describes the result as, to their knowledge, the first scientific test to provide strong evidence in favour of any trading rule on a large unrestricted scale. Note the overlap with the study above: a partly shared universe and a shared stretch of years, with opposite conclusions. Different patterns, different exit rule, different test.

Lo, Mamaysky and Wang (2000), in the Journal of Finance, is the study most often cited as vindicating technical analysis, and it is worth being exact about what it tested. It is not a candlestick study. The authors built an automatic pattern recogniser using kernel regression, applied it to a large number of US stocks over 1962 to 1996, and compared the distribution of daily returns after chart patterns such as head and shoulders or double bottoms against the unconditional distribution. Their finding is that several patterns provide incremental information and may have some practical value. “Incremental information” is a statement that returns look different after the pattern. It is not a statement that a trading rule built on the pattern made money after costs, and the abstract makes no such claim.

Lu, Shiu and Liu (2012), in the Review of Financial Economics, tested two-day candlestick reversal patterns on the component stocks of the Taiwan Top 50 Tracker Fund from 29 October 2002 to 31 December 2008, buying on bullish patterns and holding until a bearish pattern appeared. They examined three bullish and three bearish reversal patterns and report that the three bullish patterns were profitable in that market, with an out-of-sample test and a bootstrap as robustness checks.

Tharavanij, Siraprapasiri and Rajchamaha (2017), in SAGE Open, tested bullish and bearish candlestick reversal patterns on the SET50 stocks of the Stock Exchange of Thailand from 3 July 2006 to 30 June 2016, over holding periods of one, three, five and ten days, using both of the exit conventions from the studies above and with and without filtering by common indicators. They found little use for the patterns: the mean returns of most were not statistically different from zero, the few with significant returns carried high standard deviations, a binomial test found the patterns could not reliably predict direction, and the indicator filters generally did not improve either profitability or accuracy.

Five careful studies: two negative on candlesticks, two positive on candlesticks in different markets with different rules, and one positive on chart patterns but silent on profit. That is the state of the evidence, and any page that quotes you only one of them has chosen for you.

The practitioner numbers, and why they do not travel

The per-pattern percentages that circulate online mostly trace to Thomas Bulkowski’s thepatternsite.com, and they are a different kind of evidence. They are one analyst’s counts over his own database, using his own definitions and his own measuring rules, and they are useful precisely because those rules are written down.

Take the hammer. Bulkowski’s identification guideline is a single candle in a downward price trend with a lower shadow at least two or three times the height of the body and little or no upper shadow. In the traditional reading it needs a decline before it, it is confirmed by the next candle closing above its high, and the read is invalidated by a close below its low.

Hammer, one candle, after a decline. Drawn to the real proportions the formation rules describe. Full hammer guide

His page reports that the hammer is supposed to act as a bullish reversal and that in his testing it does so 60% of the time, ranking 65th for performance out of 103 candlestick patterns, with the statistics described only as based on hundreds of perfect trades and no period stated on the page. Read that carefully. It is a count of how often price went up afterwards, by his definition of “up”, in his sample. It says nothing about how far, nothing about after costs, and nothing about a 5-minute chart of a currency pair.

His chart-pattern statistics carry the same warning in their definitions. His glossary measures average rise from the opening price the day after a breakout to the “ultimate high”, which is the highest peak before price falls at least 20%, and his “bull market” figures exclude two specific bear periods. Those are reasonable choices. They are also choices, and a figure produced by them cannot be lifted into a different market, a different timeframe, or a different holding rule and remain true.

Why the evidence is mixed, and why that is not a cover-up

Four things account for most of the disagreement, and none of them is a conspiracy.

Definitions vary. Bulkowski’s hammer wants a lower shadow of two or three times the body; other definitions accept twice; a screener may accept anything with a long lower wick. Each produces a different set of instances and a different result.

Exits vary. A pattern followed by a fixed ten-day hold is a different bet from the same pattern held until the opposite pattern appears. The Thai study tested two exit conventions side by side for exactly this reason.

Data snooping. Test enough patterns over enough windows and some will look significant by chance. Marshall’s out-of-sample framing and Lu, Shiu and Liu’s bootstrap are both attempts to defend against this, and the Lo, Mamaysky and Wang design of comparing conditional against unconditional return distributions is another.

Predictive is not profitable. Returns after a pattern can be measurably different from normal without that difference surviving spreads, slippage and the cost of the times it is wrong. The distinction is why Lo, Mamaysky and Wang can be quoted by both sides.

The position worth taking

A candlestick pattern is a description of what buyers and sellers just did. That is all it ever was, and described that way it is unambiguously true: a long lower wick really does record a price that was sold down and bought back. Whether that fact predicts tomorrow is exactly the question the studies above disagree on, so treat it as evidence rather than a verdict.

What the pattern does reliably give you, regardless of which study you believe, is a discipline. It has a context requirement, so you check what came before. It has an invalidation, so you know the price at which the read is wrong. And it asks for confirmation, which is the step that separates a shape from an observation; the confirmation candle post covers why skipping it is where most of the damage happens. The reversal patterns post takes each pattern through that logic in turn. A pattern read that way is not a forecast. It is a hypothesis with a built-in way to be proven wrong, which is more than most trading ideas can say.

This is also the only honest way to build a tool around it, which is why Chartix names the context requirement and the invalidation price with every read rather than a probability. If you want to see what that looks like on a real screenshot, the AI chart analysis page shows it. For the patterns themselves, the bullish patterns post lists the seven worth knowing, and the candlestick pattern library draws every pattern with the two things that make it checkable.

Sources

Every figure above was checked against these before publishing.