Research

Does technical analysis work? What research shows

Results depend on the rule, market, sample and costs


Can a chart pattern provide a tradable edge? The studies discussed here test specific rules, markets and periods. Some report effects; others find no robust return after costs. None of these results settles the question for every chart method and every market.


Part one. What the science says

Technical analysis includes different rules: candlestick patterns, moving averages, breakouts and other ways of describing prices. Evidence about one rule does not automatically apply to the others. It matters what was tested, on which data, and after which costs.

One review is Cheol-Ho Park and Scott Irwin's "The Profitability of Technical Analysis: A Review", an AgMAS Project report from the University of Illinois. It brings together empirical tests of technical trading systems across stocks, currencies and futures.

The review must be read within the markets, rules and historical periods it covers. It is not evidence that every technical method stopped working after a particular year. Important issues when interpreting reported profitability are data snooping, choosing rules after seeing the result, and transaction costs.

The first of these deserves an explanation, because it is the main trap of the whole industry. If you try a thousand rules on the same history, the best of them will look brilliant — simply by the laws of probability, even if all thousand are empty. That is how most "working" strategies are born: not because someone found a regularity, but because someone kept searching long enough.

Candlestick patterns did no better. Piyapas Tharavanij and co-authors tested bullish and bearish reversal candlestick patterns on the Stock Exchange of Thailand — "Profitability of Candlestick Charting Patterns in the Stock Exchange of Thailand", in the journal SAGE Open. The conclusion comes in two parts. On their own: for most patterns the average return is statistically indistinguishable from zero, and for the few where a difference was found, a return of 0.07–0.84% sits next to a standard deviation of four to seven percent — that is, the signal drowned in the noise around it. And the second part, more painful: filtering by the stochastic oscillator, RSI or the money flow index "in general does not increase either profitability or prediction accuracy". The popular recipe "pattern plus indicator confirmation" does not pass the test. This is a result for the patterns and market tested, not a test of every possible indicator confirmation.

A more narrowly defined example is Mathias Mesfin's 2026 arXiv preprint. The figures below refer to version 3, revised on 15 September 2026. Its test concerns particular intraday rules on MNQ futures; it is not a test of every method mentioned in the introduction.

In 2026 Mathias Mesfin published "Structural Limits of OHLCV-Based Intraday Momentum Signals in MNQ Futures: A Systematic Falsification Study". He took the Nasdaq 100 future, 947 trading days of five-minute data for 2021–2025, and tested fourteen families of intraday momentum signals. Each had to pass five conditions at once: a t-statistic of at least 2.0 on out-of-sample data, at least thirty trades in every test window, a positive return after realistic costs, the same direction in 2023, 2024 and 2025 separately, and a permutation test with a significance level below 0.001.

None passed.1

Eleven of the fourteen families had gross returns of 0.07–1.50 points, below the study's 2.0-point friction floor. In this sample, a positive average before costs did not provide an edge that could be captured after costs. The sign of a gross average alone does not establish robustness or transferability.

Three more families cleared the costs but fell on robustness: the long breakout of the morning range gave a t-statistic of 0.88, the reversal based on a volatility-and-volume classifier gave 1.26, and the short gap continuation earned a solid 16.53 points gross with a t-statistic of 1.46 and fell apart on the requirement to work in each of the three years separately. A familiar story: a strategy that "worked" usually worked in one particular year.

And what the author did at the end deserves separate respect. He ran two signals through the same grinder that he knew in advance had an edge — and they passed: t-statistics of 3.11 and 4.30. This is called a positive control, and it answers the main objection to any negative study: "maybe your method just can't see anything". It can. It did not see precisely what was being tested.


Research

ResearchChoosing the best result afterwards distorts the estimate.. List all tested variants, including failures. ['Many variants on one history', 'Selection', 'Published results']Many variants on one history↓Selection↓Published results
  • Choosing the best result afterwards distorts the estimate.
  • List all tested variants, including failures.
Teaching diagrams with hypothetical numbers. They explain concepts, not the next price move.

Part two. Positive findings and their conditions

If it all ended here, the article would be pointless. But the same research has a second half too — about what does pass the test.

The examples below document effects under particular assumptions and risks. They are not an exhaustive list of what can work, nor a promise that the same result will continue.

Before costs → After costs

Before costs → After costsA: Before costs +1.5; B: Costs −2; C: After costs −0.5. Hypothetical units. Illustrative calculation, not a strategy result. [1.5, -2, -0.5]1.5A-2B-0.5C
  • A: Before costs +1.5; B: Costs −2; C: After costs −0.5
  • Hypothetical units
  • Illustrative calculation, not a strategy result.
Teaching diagrams with hypothetical numbers. They explain concepts, not the next price move.

Trend following

A long-horizon study provides another perspective. A group of researchers from Capital Fund Management — Lempérière, Deremble, Seager, Potters and Bouchaud — collected data spanning two centuries: "Two Centuries of Trend Following". Spot prices of commodities and indices since 1800, government bond yields since 1918, futures since 1960. The conclusion: the ten-year return of trend following "has never been negative in two centuries".

But this coin has a heavy reverse side, and the authors do not hide it. The typical length of a drawdown, they write, equals one divided by the square of the Sharpe ratio — in years. For a strategy with a Sharpe of 0.7 that means two-year drawdowns as the norm. For 0.5 — four years. Think about it: a correct strategy that has worked for two hundred years can give you nothing but losses and explanations for four years in a row. No amount of discipline replaces understanding this number in advance. This is an approximation in the approach discussed by the authors, not a timetable for future drawdowns of every strategy.

There is a third side too, the most unpleasant one. Geetesh Bhardwaj, Gary Gorton and Geert Rouwenhorst studied how much of this premium reaches the investor — a paper with a telling title, "Fooling Some of the People All of the Time: The Inefficient Performance and Persistence of Commodity Trading Advisors". Managers of futures funds earned 5.4% of gross excess return over 1994–2007, while the investor received 85 basis points — eight tenths of a percent over Treasury bills, statistically indistinguishable from zero. The premium was there. It was taken away in fees.

The volatility premium

Another study examines the variance risk premium. Peter Carr and Liuren Wu in "Variance Risk Premiums" (Review of Financial Studies) showed that the variance risk premium on the S&P 500 index is persistently negative: investors accept a negative expected return just to insure themselves against rising volatility. The other side of the same trade — selling volatility — has historically delivered very high returns per unit of risk.

And the main thing: it is clear why this does not disappear under the pressure of arbitrageurs. It is not a pricing error that can be corrected. It is payment for risk. The option seller gets money for taking on the risk of a crash that hedgers buy their way out of. The risk is real — and so is the payment. Anyone who was selling volatility in February 2018 or in March 2020 found out exactly what they had been paid for all the years before.

Carry

In crypto, one carry mechanism involves funding payments on perpetual contracts. A record of payments shows what was transferred in the past. It does not by itself establish the future net return of a position.

Funding payments help link a perpetual contract to its reference price. Their timing, formula and sign depend on the contract and venue. With positive funding, long positions generally pay short positions; with negative funding, the direction reverses. A spot-long/perpetual-short position can reduce directional exposure, but funding can change and basis, fees, collateral and liquidation risks remain. "Designing Funding Rates for Perpetual Futures in Cryptocurrency Markets" studies the design of funding rates and replicating portfolios; it does not guarantee the return of this trade.

But the premium here depends on the regime more than is usually thought. When the market is overheated with longs, the funding rate shoots up; when capital comes into this trade, it compresses toward the yield of ordinary dollar instruments — and then the whole point of the construction disappears: the same money could have been held in dollars with no exchange risk, no leverage and no risk of the short leg being liquidated. Carry is not a money-printing machine but payment for being willing to hold a position that, at the critical moment, there may be no one to hand over to.


Part three. Risks and limits of interpretation

The examples differ in their mechanisms, test conditions and risks.

A documented return must be considered together with the risk and conditions under which it was measured. Recognizing a shape is not enough to establish an advantage.

Selling volatility, you take on the risk of a crash that others pay to avoid. Following the trend, you accept the risk of long reversals and four-year drawdowns that most people cannot psychologically endure. In carry — the risk that the rate flips at the worst possible moment and the short leg gets wiped out by a margin call.

To evaluate a chart rule, define it before examining the test period and check its result on new data after costs. An explanation of why it might work is not a substitute for that test.

The second point follows from this.

Interpretation limit · Outcome

Interpretation limit · Outcome4 × 1 − 8 = −4. A positive mean does not exclude large losses.. Illustrative calculation, not a strategy result. [1, 1, 1, 1, -8]11121314-85
  • 4 × 1 − 8 = −4
  • A positive mean does not exclude large losses.
  • Illustrative calculation, not a strategy result.
Teaching diagrams with hypothetical numbers. They explain concepts, not the next price move.

Why discoveries disappear

Some reported effects weaken outside the original sample or after publication. How much they weaken is an empirical question, not a rule that every discovery must disappear.

David McLean and Jeffrey Pontiff study 97 published stock-return predictors in "Does Academic Research Destroy Stock Return Predictability?", Journal of Finance. They report returns 26% lower out of sample before publication and 58% lower after publication, in both cases relative to the original in-sample return. The difference is 32 percentage points between these reductions, not a further 32% decline from the out-of-sample return. On a hypothetical original alpha of 5%, a 58% relative decline leaves 5% × (1 − 0.58) = 2.1%. This illustrates the stated comparison; it is not a return forecast for a particular strategy.

Antoine Falck, Adam Rej and David Thesmar also examine decay in "Why and How Systematic Strategies Decay". The reported relationships between publication year and decay describe the strategies and period they study. They do not establish a fixed lifetime, measured in months, for every newly published rule.

Popularity on YouTube neither proves that a method works nor proves that its edge has disappeared. A specific claim still needs a defined rule, independent data and realistic costs.


Part four. The numbers they don't mention in courses

Since we have taken on uncomfortable facts, here are these too. They are not about methods but about people.

Fernando Chague, Rodrigo De-Losso and Bruno Giovannetti obtained complete regulatory data on retail accounts in Brazilian futures — "Day Trading for a Living?". Of the 19,646 people who started trading, 1,551 made it to three hundred or more trading sessions — seven point nine percent. Of these persistent ones, 97% lost money. 1.1% earned more than the Brazilian minimum wage. More than a bank teller's salary — half a percent. Eight people out of one and a half thousand.

The Brazilian study reports no evidence of learning by day traders over the periods observed. That finding concerns the studied traders, market and period. It does not show that practice can never improve any skill; it does warn against assuming that time spent trading is evidence of improving net performance.

Taiwan, a different era and a different market, the same result. Brad Barber, Yi-Tsung Lee, Yu-Jane Liu and Terrance Odean went through fifteen years of complete data for the entire market, 1992–2006: "The Cross-Section of Speculator Skill: Evidence from Day Trading", Journal of Financial Markets. Their wording: "less than one percent of the day trader population is able to predictably and reliably earn positive abnormal returns net of fees". In absolute numbers — about four thousand people out of roughly four hundred and fifty thousand trading in a year.

Note: these four thousand exist. Skill in speculation is not a myth; it is measurable and repeatable. It is just that its distribution looks not like "learn and you can", but like the distribution of skill in professional sports. The only difference is that in the second case nobody sells courses promising you a spot in the NBA starting lineup in three months.


Independent observations

Independent observationsA mechanism explanation does not replace independent testing.. Validation must include costs and new data. ['Mechanism explanation', 'Rules fixed + New data', 'Costs + Control comparisons', 'Limits of the inference']Mechanism explanation↓Rules fixed + New data↓Costs + Control comparisons↓Limits of the inference
  • A mechanism explanation does not replace independent testing.
  • Validation must include costs and new data.
Teaching diagrams with hypothetical numbers. They explain concepts, not the next price move.

What to do with all this

We do not think chart markup is useless. It is useful — but not in the role in which it is usually sold.

An unfilled gap on a chart is a fact. A level from which the price reversed twice is a fact. An extreme that the price went through is a fact. These are landmarks: where to put a stop, where to set a target, where the price has already been. A map of the terrain.

What a map does not do is tell you where the car will go.

The difference between these two things is the whole difference between a tool and a promise. A tool saves time and structures the view. A promise sells better, but is not kept.

For a concrete reading task, start with market structure: it separates observed pivots from interpretation. The guide to correction confirmation and invalidation shows why a plausible count is not certainty. When examining similar historical episodes, read what their statistics do and do not show before treating a past frequency as a forecast.

And we publish the numbers from this article because we believe the user has the right to know what exactly they are holding in their hands.


Sources

  1. Park C.-H., Irwin S. The Profitability of Technical Analysis: A Review. AgMAS Project Research Report 2004-04, University of Illinois. ageconsearch.umn.edu/record/37487 — a consolidated review of empirical tests of technical trading systems.
  2. Tharavanij P., Siraprapasiri V., Rajchamaha K. Profitability of Candlestick Charting Patterns in the Stock Exchange of Thailand. SAGE Open, 2017. doi:10.1177/2158244017736799 — candlestick reversal patterns with and without indicator filters.
  3. Mesfin M. Structural Limits of OHLCV-Based Intraday Momentum Signals in MNQ Futures: A Systematic Falsification Study. arXiv:2605.04004, 2026. arxiv.org/abs/2605.04004 — fourteen signal families, 947 days of five-minute data, none passed the five criteria.
  4. Lempérière Y., Deremble C., Seager P., Potters M., Bouchaud J.-P. Two Centuries of Trend Following. arXiv:1404.3274, 2014. arxiv.org/abs/1404.3274 — trend following on data since 1800; drawdown length as 1/S².
  5. Bhardwaj G., Gorton G., Rouwenhorst K. G. Fooling Some of the People All of the Time: The Inefficient Performance and Persistence of Commodity Trading Advisors. NBER Working Paper 14424. nber.org/papers/w14424 — 5.4% before fees versus 85 basis points after, 1994–2007.
  6. Carr P., Wu L. Variance Risk Premiums. Review of Financial Studies, 22(3), 2009. ideas.repec.org — a persistently negative variance risk premium on the S&P 500.
  7. Designing Funding Rates for Perpetual Futures in Cryptocurrency Markets. arXiv:2506.08573. arxiv.org/abs/2506.08573 — how the funding rate of perpetual contracts is built.
  8. McLean R. D., Pontiff J. Does Academic Research Destroy Stock Return Predictability? Journal of Finance. PDF — 97 predictors; both reported reductions use the original in-sample return as their reference.
  9. Falck A., Rej A., Thesmar D. Why and How Systematic Strategies Decay. Capital Fund Management, 2021. cfm.com — the year of publication explains 30% of the variation in decay, plus 5 percentage points for each year.
  10. Chague F., De-Losso R., Giovannetti B. Day Trading for a Living? SSRN 3423101 — results for the Brazilian day-trader sample and observed period, including the absence of evidence of learning.
  11. Barber B., Lee Y.-T., Liu Y.-J., Odean T. The Cross-Section of Speculator Skill: Evidence from Day Trading. Journal of Financial Markets, 18, 2014. escholarship.org — Taiwan 1992–2006, less than 1% of day traders predictably profitable.

  1. "None passed all five" — none passed all five (from the paper's abstract). ↩