Does technical analysis work?

🎯 In 30 seconds: The evidence is mixed, and the details matter. Price momentum over roughly 3 to 12 months is one of the best-documented effects in finance. Simple trend and moving average rules showed real signals in older data, but much of the edge faded in later years or after accounting for how many rules were tried. Chart patterns carry a little information but have rarely produced reliable profits after costs. The biggest danger is not that technical analysis is useless; it is that testing many ideas on the same data always finds a winner by luck. This chapter shows how to test an idea honestly.

What "works" should mean

Before looking at evidence, be clear about the claim. A trading method "works" if it does all of these:

Makes money after all costs: commissions, spreads, slippage and taxes. See What trading really costs.
Beats a simple alternative, usually buying and holding an index fund, for the same risk.
Holds up on data it was not designed on (out-of-sample).
Is unlikely to be luck, given how many trades it made and how many ideas were tried.

There is a weaker, more defensible claim too: charts help with risk management. Knowing where the trend changes, where support sits and how volatile a stock is helps you set stops and size positions, even if none of it predicts price. That use does not require any forecasting edge.

Why many economists were skeptical

The efficient market hypothesis, in its "weak form", says past prices are already reflected in today's price, so they cannot be used to predict future returns in a way that beats the market after costs. If a chart pattern reliably predicted a rise, traders would buy early, and the pattern would stop working.

The counter-arguments are behavioral and mechanical. People react slowly to news, chase trends, and place orders at the same obvious levels. Those habits could leave patterns in prices. The research below is, in large part, a decades-long test of which side is right.

The evidence, topic by topic

Idea	What research found	Strength
Cross-sectional momentum (buy recent winners, sell recent losers)	Strategies ranking U.S. stocks on past returns earned significant positive returns over 3 to 12 month holding periods, not explained by systematic risk. Part of the gain faded over the following two years.	Strong, widely replicated
Time-series momentum (trend following)	In 58 futures markets, past 1 to 12 month returns tended to continue, then partly reverse. A study going back to 1880 found trend following profitable over the following 110 years as a whole.	Strong for diversified baskets of markets
Short-term reversal	Individual U.S. stock returns showed a significant tendency to reverse from one month to the next.	Documented; costly to trade
Moving average and range breakout rules	Showed real signals on the Dow from 1897 to 1986. After testing nearly 8,000 rule variations, the best rule still looked real in that period, but was not statistically significant from 1987 to 1996.	Mixed; weaker in later data
Chart patterns	Computer-detected patterns in U.S. stocks (1962 to 1996) gave "incremental information". Head and shoulders was unprofitable on average in U.S. stocks, and in currencies was beaten by simpler rules.	Weak
Support and resistance	Levels published by currency firms helped predict intraday trend interruptions, with results varying by firm and currency.	Some evidence, mostly in currencies
Short-term retail trading	In Taiwan, the vast majority of day traders lost money; about 15% earned abnormal returns after fees in an average year.	Strong evidence of poor average outcomes

A survey of 95 modern studies on technical trading rules summed up the field well. It also explains why the positive count cannot be taken at face value:

Most modern studies found positive results for technical rules, at least through the early 1990s. The same survey warned that most had problems in their testing procedures, including data snooping, choosing rules after the fact, and difficulty measuring risk and trading costs.

Note that the strongest evidence (momentum) comes from ranking many stocks or many markets and holding diversified portfolios for months. That is quite different from reading one stock's daily chart. The momentum chapters, Momentum trading and Trend following, cover that style.

Data mining: why backtests lie

A backtest applies a trading rule to past data to see how it would have done. It is the only way to test an idea before risking money, and it is very easy to fool yourself with.

Data mining (or data snooping) is what happens when you try many variations and keep the best. Even on data with no pattern at all, one variation will look good by chance. The more you try, the better the best one looks, and the less it means.

To show this, we simulated it. We generated random price histories with zero expected return, so by construction no rule can have a real edge. On each one we tested 93 moving average crossover rules on the first five years, picked the best, and then ran that rule on the next five years.

One of 60 simulations, chosen because it is close to the average. Across all 60, the best rule averaged +4.9% a year in-sample and minus 1.2% a year out-of-sample, after a 0.1% cost per trade. Only 20 of the 60 best rules made money out-of-sample. The in-sample "edge" was pure luck.

Researchers have put numbers on this. One paper calculated that with five years of daily data, a researcher who tries more than about 45 independent strategy variations should expect to find one with an impressive in-sample risk-adjusted return (a Sharpe ratio of 1) even when the true expected out-of-sample result is zero.

Other ways a backtest misleads

Look-ahead bias: using information that was not available at the time, such as trading at a day's close using a signal calculated from that same close, or using revised data.
Survivorship bias: testing only on stocks that exist today. The companies that went bankrupt or were delisted are missing, which makes buying dips look safer than it was.
Unrealistic costs and fills: assuming you buy at the exact breakout price, ignoring spreads and slippage, or ignoring that a stop can gap.
Too few trades: 15 trades can look brilliant by luck.
Changing markets: a rule that worked in a 1990s market with high costs and slow information may not work today, especially once it is widely known.

Common mistake: tuning until it looks perfect

Changing a 50-day average to 47 days because it made the backtest better is fitting noise. So is adding "but not on Mondays" to remove a few bad trades. If a small change in a setting wrecks the result, the original result was luck. A real effect should work roughly as well at 40, 50 and 60 days.

How to test an idea honestly

Write the rule before you look. Entry, exit, stop, position size, which stocks, which dates. Ambiguous rules get interpreted kindly after the fact.
Split the data. Develop on one period (in-sample) and keep a later period untouched (out-of-sample). Look at the out-of-sample result once.
Count your variations. Write down every version you tried. The more you tried, the more skeptical you should be of the best one.
Include realistic costs. Commissions, the bid-ask spread and some slippage on every trade.
Compare with buying and holding the same asset, and look at drawdowns, not just returns.
Check robustness. Nearby settings and other similar stocks or markets should give similar results.
Walk forward. Re-test repeatedly: develop on years 1 to 5, test on year 6, then develop on 2 to 6, test on 7, and so on.
Paper trade it. Real-time results with no hindsight are the final test. See Paper trading.

Worked example: is this backtest real?

Hypothetical backtest of a breakout rule on one stock: 40 trades, 45% winners, average win $300, average loss $180.

Expectancy before costs: 0.45 x $300 minus 0.55 x $180	$36 per trade
Costs per round trip (spread and slippage)	$8
Expectancy after costs	$28 per trade
Total over 40 trades	$1,120
Typical spread of individual trade results (standard deviation)	about $250
Standard error of the average: $250 / square root of 40	about $39.50
Average divided by its standard error: $28 / $39.50	about 0.7

A ratio of 0.7 is well within the range that luck produces. As a rough guide, statisticians look for around 2 before taking a result seriously, and more if many versions were tried. At this win rate and payoff, you would need several hundred trades before a $28 edge could be told apart from zero. Expectancy is explained in Risk/reward and expectancy.

This is not a reason to give up. It is a reason to size small, keep testing, and treat early results as noise in both directions.

Practical tip: keep a "graveyard" file

Every time you test an idea and drop it, write it down with the result. When you finally find something that works, the graveyard tells you how many tries it took, which tells you how much of the result might be luck. It also stops you re-testing the same failed idea under a new name.

So, should you use technical analysis?

A balanced answer:

Use it for structure and risk: trend direction, levels for stops and targets, volatility for sizing. This is the most defensible use.
Treat signals as hypotheses: any pattern or indicator rule is an idea to test, not a known edge.
Respect the base rate: most short-term traders in large studies lose money. Assume you will be average until your own records, over many trades, say otherwise.
Keep a core: for most young people, a long-term index fund core with a small trading satellite is the sensible structure. See A plan for a young person.

Try it

Pick one simple rule from Moving averages or Breakouts and gaps and write it down exactly.
In the strategy backtester, run it on the first half of the available history. Record trades, win rate, average win and loss, drawdown and return compared with buying and holding.
Try at most five variations. Write each one down.
Run only your chosen version on the second half. Compare. Then run it on two other stocks or an index.
Compute the average trade divided by its standard error, as in the worked example. If it is below 2, keep it on paper in the paper portfolio for at least 30 more trades.

Check yourself

Which chart-based effect does the chapter call one of the best documented in finance? Answer: Price momentum over roughly 3 to 12 months.
What is data mining (data snooping)? Answer: Trying many variations of a rule and keeping the best. Even on data with no pattern, one variation will look good by chance.
What is survivorship bias in a backtest? Answer: Testing only on stocks that exist today, so the companies that went bankrupt or were delisted are missing.
A backtest shows 45% winners, an average win of $300 and an average loss of $180, with $8 of costs per trade. What is expectancy after costs? Answer: $28 per trade ($36 before costs: 0.45 x $300 minus 0.55 x $180).

🔬 Does technical analysis work? What the research says