Does technical analysis actually work?

The answer

Some of it works and most of it does not, and the parts that work are not the parts that get taught. Trend and momentum are documented across many markets and many decades; named chart patterns and candlestick shapes have very weak support and much of that support disappears when the tests are corrected for the number of rules that were tried.

Why this costs you money

There are 2 ways to lose money on this question, and most people find one of them.

The first way is believing. You learn 30 chart patterns. You get good at spotting them on completed charts, which is easy, and you conclude you can spot them in advance, which is a different skill. Then you trade them. Your win rate is around 45%, which feels close enough to work, so you keep going. Costs and spreads take a slice of every trade. After 200 trades you are down and you cannot point to a single decision that was obviously wrong. That is the shape of the loss: not a disaster, an erosion.

The second way is dismissing. You read that professors proved charts are worthless, so you ignore price entirely. You buy a good company and hold it while it falls 70%, because your method contains no mechanism that can ever say stop. You were right about the business and it did not matter.

Both errors come from treating "does technical analysis work" as a single question with a yes or no answer. It is at least 5 separate questions and they have different answers. This article separates them.

How it works

The claim the evidence is aimed at

The Efficient Market Hypothesis is the idea that prices already reflect available information, so you cannot beat the market using information that is already reflected. Eugene Fama set out 3 versions of it in a 1970 review paper in the Journal of Finance.

Technical analysis lives or dies on the weak form. That is the whole reason this argument exists. If prices reflect all past price information, then a chart contains nothing about the future and every method in this cluster is decoration.

The related idea is the random walk: the claim that price changes are independent of each other, so what happened yesterday tells you nothing about today. Fama's 1965 study in the Journal of Business is the standard early reference. Burton Malkiel's A Random Walk Down Wall Street (1973) took the argument to the public.

One thing to understand before reading any test

There is a problem at the centre of all of this, and Fama named it himself. It is called the joint hypothesis problem.

To test whether a market is efficient, you need a model of what returns should be. If a strategy earns more than the model says it should, you can conclude the market is inefficient, or you can conclude your model of expected returns is wrong. You cannot separate the 2 from the data.

This is why the argument has run for 60 years without resolving. It is not that one side refuses to look at the evidence. It is that "the market is efficient" is not, on its own, a testable statement.

The 5 questions, and their separate answers

Here is the honest scorecard. Each row is a different claim, and lumping them together is why this subject is such a mess.

Claim 1: Price changes are perfectly independent, so past prices contain nothing. — Not supported. Returns show measurable dependence at some horizons in many markets. This is settled. It is also much weaker than it sounds, because a small statistical dependence is not the same as a profit.

Claim 2: Cross-sectional momentum exists — stocks that outperformed over the past 3 to 12 months tend to keep outperforming over the next few months. — Well documented. Narasimhan Jegadeesh and Sheridan Titman published the main result in the Journal of Finance in 1993, using US stocks from 1965 to 1989. They published a follow-up in 2001 confirming it held in the 1990s, which matters because that period was outside their original sample. The effect has since been found in many countries and in other asset classes. It is one of the most replicated results in finance.

Claim 3: Time-series trend following works — an instrument above its own recent average tends to keep going. — Documented, mainly in futures. Tobias Moskowitz, Yao Hua Ooi and Lasse Pedersen published "Time Series Momentum" in the Journal of Financial Economics in 2012, testing 58 liquid futures and forward contracts across equities, bonds, currencies and commodities. They found positive excess returns from a simple rule based on the past 12 months. Long-run studies extending trend following back to the 19th century have been published by practitioners at AQR. Note who publishes this research. It is often produced by firms that sell trend-following products. That does not make it wrong. It does mean you should look for independent replication, and independent replication does exist.

Claim 4: Named chart patterns predict direction — head and shoulders, flags, triangles, candlestick patterns. — Weakly supported, and the strongest evidence claims less than traders think. The most careful study is covered in the cases below. The general finding is that some patterns carry a small amount of statistical information, and that this is not the same as a strategy you can trade after costs.

Claim 5: A retail trader, using discretionary chart reading, can make money after costs. — The direct evidence is strongly against it. This is the only claim that matters to the reader, and it is the one nobody tests when they are selling a course. The evidence is in the "In India" and "In the United States" sections below, and it is not close.

Why so many published studies find that technical analysis works

Cheol-Ho Park and Scott Irwin surveyed the modern literature in the Journal of Economic Surveys in 2007. They reviewed 95 studies and found that a majority reported positive results for technical trading rules. Their conclusion was not that technical analysis works. It was that most of those studies suffered from specific, correctable problems.

There are 4 problems and every reader should be able to name them.

1. Data mining, also called data snooping. If you test 8,000 rules on the same price history, some will look excellent purely by chance. The standard significance test assumes you tested 1 rule. Campbell Harvey, Yan Liu and Heqing Zhu made this point forcefully in the Review of Financial Studies in 2016. They argued that given the number of strategies that have been tested in finance, the usual statistical hurdle is far too low, and proposed a much stricter one.

2. Ex post rule selection. A researcher looks at the data, notices what worked, and then reports the rule as though it had been chosen in advance. This is not fraud. It is what happens naturally when you explore a dataset before writing down your hypothesis.

3. Transaction costs and slippage. Many published rules trade frequently. A rule that earns 0.06% per trade before costs earns nothing after them. Studies that ignore the spread are not describing anything a person could do.

4. Survivorship and publication bias. Failed rules are not published. Failed funds close and disappear from the databases. The record you can read is the record that survived, and that is a biased sample by construction.

The eye is the problem

Here is the deepest issue, and it has nothing to do with markets.

In 1959 Harry Roberts published a short paper in the Journal of Finance in which he generated price charts from random numbers. The charts contained trends, support levels, reversals and, famously, a shape that looked like a head and shoulders top. They were noise. They looked exactly like the charts in the textbooks.

Human beings are built to find structure in randomness. This is a well documented feature of perception, not a personal failing. The consequence for a chart reader is severe: you cannot use your own sense of "that looks like a real pattern" as evidence, because that sense fires just as strongly on noise.

This is also why the counting exercise in this cluster matters more than anything else in it. Counting is the only thing that separates a pattern that is real from a pattern your eye assembled.

One honest complication is worth adding, because it cuts the other way. Thomas Gilovich, Robert Vallone and Amos Tversky published a famous 1985 paper arguing that the "hot hand" in basketball was an illusion. In 2018 Joshua Miller and Adam Sanjurjo showed that the original analysis contained a subtle statistical bias, and that a hot-hand effect probably does exist. The lesson is not that streaks are real after all. It is that "that is just randomness" is also a claim, and it also has to be tested.

What it tells you, and what it does not

The evidence tells you that the direction of past price movement carries information about future price movement, in many markets, over horizons of weeks to months. That is a real finding and it is the foundation under trend following, momentum funds and most systematic managed futures.

It does not tell you that this information is large. Momentum's edge is a modest average across many positions over long periods. It is not a signal about your one stock this week.

It does not tell you the information is stable. Momentum has severe crash periods. Kent Daniel and Tobias Moskowitz documented these in the Journal of Financial Economics in 2016. The worst episodes come after a market crash, when the previous losers rebound violently and a momentum portfolio is short exactly those names.

And it does not tell you the information survives at retail scale. The documented results come from portfolios of dozens or hundreds of positions, rebalanced mechanically, with institutional costs. A person trading 3 stocks on discretion, paying retail costs, is not running that strategy in any recognisable form.

The decision rule

Before you accept any claim about a chart method, ask 3 questions in this order.

1. Was the rule written down before the data was looked at? If not, the result tells you what happened, not what will happen.

2. How many rules were tried? A single rule with a strong result is interesting. The best of 8,000 rules is a lottery winner.

3. Does it survive costs and an out-of-sample period? Out-of-sample means data the rule was not built on, usually a later period.

If a method cannot be stated precisely enough to fail those 3 questions, it is not a method. It is a description.

That last line is the one to keep. A rule that can be adjusted after the fact to explain any outcome is unfalsifiable, and an unfalsifiable rule cannot be evidence for anything, including itself.

Try this now

This is the most useful 10 minutes in this cluster. It measures whether you can read a chart forward, which is a completely different skill from reading one backward.

1. Open a daily chart of a liquid index or a large stock you follow. Set the range to about 3 years.
2. Pick a date at least 6 months into the chart. Cover everything to the right of that date with a piece of paper or your hand. On a phone, scroll so the date sits at the right edge.
3. Write down 2 things: up or down over the next 20 trading sessions, and your confidence on a scale of 1 to 5.
4. Uncover. Mark it right or wrong.
5. Repeat 20 times, on 20 different dates. Do not skip a date because it looks unclear. Skipping the unclear ones is the exact bias this test exists to measure.
6. Count. Then count a second number: how many of the 20 periods were up, regardless of what you said.

What you should see. Most people score between 10 and 13 out of 20 and feel they did reasonably well. Now look at your second number. Over most 3-year windows in a rising market, roughly 11 to 13 of the 20 periods were up anyway. So a person who wrote "up" 20 times without looking at the chart would have scored about the same as you did.

That comparison is the whole point. The bar is not 10 out of 20. The bar is whatever "always up" scored. If you did not beat it, your chart reading added nothing on this sample, and you now know that from your own data rather than from an argument.

Then check your confidence numbers. Compare your accuracy on the trials you marked 5 against the trials you marked 1 or 2. If those 2 rates are the same, your confidence carries no information, and confidence that carries no information is the most expensive thing a trader can own.

Three real cases

1. Alfred Cowles versus Brown, Goetzmann and Kumar — the same record, 2 answers, 65 years apart In 1933 Alfred Cowles published "Can Stock Market Forecasters Forecast?" in Econometrica. Among other things he examined the market calls made by William Peter Hamilton in his Wall Street Journal editorials, which applied Dow Theory. Cowles concluded that Hamilton's record did not beat simply holding the market. The paper is one of the founding documents of the sceptical case, and its 3-word summary of the whole question is often quoted.

In 1998 Stephen Brown, William Goetzmann and Alok Kumar re-examined the same editorials in the Journal of Finance. Using modern methods, they found that Hamilton's calls did add value on a risk-adjusted basis, largely because his strategy was out of the market during volatile periods. Same person, same editorials, opposite conclusion. The difference was the adjustment for risk.

That is the joint hypothesis problem in a single example, and it is why you should be suspicious of anybody who tells you this question was settled.

2. Brock, Lakonishok and LeBaron, then the correction — what data mining does to a good result In 1992 William Brock, Josef Lakonishok and Blake LeBaron published a study in the Journal of Finance testing 26 simple technical rules, mostly moving-average crossovers and range breakouts, on the Dow Jones Industrial Average from 1897 to

1. They found the rules had genuine predictive power. It was the most

influential pro-technical result of its era.

In 1999 Ryan Sullivan, Allan Timmermann and Halbert White published a follow-up in the same journal. They expanded the universe to a very large set of rules and applied a statistical method that corrects for the number of rules tested. The original result survived in the original period. But the best rules did not perform significantly in the years after 1986, once the correction was applied.

Pierre Bajgrowicz and Olivier Scaillet extended this in the Journal of Financial Economics in 2012 using a false-discovery-rate approach and found no evidence of profitable technical rules in the recent period after transaction costs.

Read the sequence carefully. Nobody cheated. The 1992 result was honest work. It simply did not survive being asked how many other rules were in the drawer.

3. Lo, Mamaysky and Wang, 2000 — the strongest pro-chart evidence, and exactly what it claims Andrew Lo, Harry Mamaysky and Jiang Wang published "Foundations of Technical Analysis" in the Journal of Finance in 2000. Their contribution was methodological. Chart patterns had always been defined by eye, which made them untestable. The authors used a smoothing technique called kernel regression to define 10 classic patterns mathematically, including head and shoulders and broadening tops, and then searched for them automatically in US stock data from 1962 to 1996.

They found that several patterns did provide incremental information. The distribution of returns after a pattern differed from the unconditional distribution by a statistically significant amount.

Now read what they did not claim. They did not claim the patterns were profitable. They were explicit that statistical significance in conditional return distributions is not the same as a trading strategy, and that costs and implementation were outside the scope of the paper.

This is the single most cited study in defence of chart patterns, and its actual finding is: the shapes are not pure noise, and that is a much smaller statement than the one usually made in its name.

The question that resolves it

A novice asks: does this pattern work?

An expert asks: compared with what, over what sample, and how many other patterns did I consider before choosing this one?

Every real disagreement about technical analysis dissolves into those 3 sub questions. "Compared with what" is the benchmark. "Over what sample" is out-of-sample testing. "How many did I consider" is the multiple-testing correction. A person who asks all 3 will never be badly fooled by a backtest, and a person who asks none of them will be fooled by every one.

What would make this wrong

The conclusion of this article is that trend and momentum have real support and that named patterns have weak support. Here is what would overturn each half.

The trend half would be wrong if momentum and time-series trend produced no excess return in data after the periods they were discovered in. This is a live question, not a settled one. Published anomalies have a documented tendency to weaken after publication, and there have been long stretches — including much of the period after 2009 — where trend-following returns were poor. If the next 20 years look like that, the honest conclusion changes.

The pattern half would be wrong if somebody defined patterns precisely, in advance, tested them on markets and periods not used to develop them, corrected for the number of rules tried, subtracted realistic costs, and still found tradable returns. That is a fair test and nothing prevents it. Parts of it have been done. The full version, with a positive result, has not been published as far as I am aware.

The honest limits of the sceptical case matter too. Most academic tests use mechanical rules on daily data for large US stocks. A practitioner uses several inputs together, on instruments and timeframes the studies do not cover, with position sizing that the studies do not model. It is fair to say the studies do not test what practitioners actually do.

But that defence has a trap in it, and you should see the trap. If every negative result can be dismissed with "that is not how real technicians work", then the method has been placed beyond testing. At that point it is a belief. A serious technician accepts the burden: state the rule, test it, and let it fail.

Carry this

• The question is not 1 question. Trend has support. Named patterns have very little.
• The best pro-chart study found information, not profit. That distinction is the whole subject.
• Ask 3 things of any result: written in advance, how many rules tried, does it survive costs and a later period.
• Your eye finds patterns in random data. It always has. Count instead.

Knowledge check

Related

• What technical analysis is, and what it is for
• The history of technical analysis and Dow Theory
• Reading candlesticks and basic patterns
• Algorithmic trading, and who you are actually trading against
• ← Technical analysis: foundations