The Head and Shoulders Pattern: What It Is and What Testing Found

A head and shoulders is a sequence of three price peaks in which the middle one — the head — stands above the two on either side of it, the shoulders. Chartists read it as a sign that a trend has finished, and take the signal when price crosses the line drawn under the two troughs that sit between those peaks.

That is what the pattern is. Whether it forecasts anything is a separate question, and it happens to be one of the few chart patterns with real published evidence behind it, because researchers at the Federal Reserve Bank of New York wrote it into code twice and tested what it produced. This post covers the definition, the conventions attached to it, how the pattern was made objective enough to test, and what the tests found, including where they found nothing.

What is a head and shoulders pattern?

A head and shoulders is a formation of three consecutive peaks where the second is higher than both the first and the third. Osler, in a 1998 Federal Reserve Bank of New York staff report, treats it as a reversal signal: the version that forms after prices have been rising is a top, and the mirrored version at the end of a decline, built out of troughs rather than peaks, is a bottom.

That bottom is the same shape turned upside down. Lo, Mamaysky and Wang, in the NBER working paper discussed below, label the mirror image the inverted head and shoulders and define it as the identical specification with the highs and the lows swapping roles.

The convention is old. Osler dates belief in the pattern to before 1930 and cites R. W. Shabacker's Stock Market Theory and Practice of that year, the spelling used in the paper's own reference list. Eight technical manuals went into building the paper's algorithm, and Osler notes how closely they line up with each other on the pattern's details, which the paper finds remarkable precisely because so little rigorous work on the pattern existed. That is worth holding on to before going further. Consistency across textbooks tells you the convention is stable and widely taught. It tells you nothing about whether the shape carries information.

What is the neckline, and when is a head and shoulders confirmed?

The neckline is the line drawn from the trough to the left of the head to the trough on its right, extended forward in time. Osler's reading of the manuals is that nothing counts until price breaks through that extended line somewhere after the third peak. That break is what makes the pattern confirmed, and it is also the moment the trader is told to act: for a top, by going short.

Two further conventions travel with it. The first is the price objective: the manuals measure the vertical distance from the neckline up to the head and project it from the crossing, treating that distance as the smallest move the pattern should be expected to deliver rather than a ceiling. The second is the bounce: the manuals allow that price may drift back toward the neckline before the objective is reached and then continue away from it, which is why the convention does not read an immediate retrace as failure.

Turning that into something testable forced further choices. Osler recorded entry at the closing price on the day the close crossed the neckline, held the position until price stopped moving in the predicted direction, and applied a one percent stop loss. Those are the researcher's decisions about how to obey a vague instruction, not part of the pattern itself.

How do you define a head and shoulders objectively?

By writing rules strict enough for a computer to apply without a human eye. Lo, Mamaysky and Wang, in their NBER working paper on technical analysis, count subjectivity among the main obstacles to academic study of the field, because whether a shape is present in a chart is "often in the eyes of the beholder". Both research groups answered that the same way: fix what counts as a peak, then fix how symmetric the result has to be.

Osler's algorithm starts by throwing away small moves. Every turning point has to be followed by a reversal of some minimum size — the cutoff — before it is admitted as a peak or a trough, which leaves the series as a zig-zag. Set the cutoff at five percent and a local high is only recorded as a peak after price has fallen five percent below it. The scan is repeated with cutoffs running from 1.5 to 6.0 times the standard deviation of that stock's daily returns, so patterns of different sizes are caught. Beyond the three-peak rule, the formation has to follow an advance, and it has to be roughly symmetric: the number of days from left shoulder to head must fall within a factor of 2.5 of the number from head to right shoulder, with comparable constraints on the heights of the shoulders and troughs.

Lo, Mamaysky and Wang instead smooth the price series using kernel regression, then require five consecutive extrema — maximum, minimum, highest maximum, minimum, maximum — with the two outer maxima within 1.5 percent of their average and the two minima between them within 1.5 percent of theirs. Their inverted version is the same specification beginning on a minimum.

Does the head and shoulders pattern actually work?

In the published test on US equities, it did not. Osler ran the algorithm across 100 firms drawn at random from the CRSP database, from the set with price data covering 2 July 1962 to 31 December 1993, found roughly 27 confirmed patterns per firm, and measured what a trader following the manuals would have earned. The average position returned -0.24 percent, held for about 10 business days.

The comparison matters more than that number on its own. For each stock Osler built 12,500 artificial price series by resampling that stock's own past returns at random, which leaves the return distribution intact but destroys any dependence between one day and the next, and ran the same pattern-finding and trade-management code over them. Those meaningless patterns averaged -0.03 percent per position. The real ones did slightly worse, and the difference was not statistically significant, with a marginal significance of 0.12 on the paper's first test. The paper concludes that head and shoulders trading was not profitable, and that unprofitability is one of the two grounds on which it classifies the traders following the pattern as noise traders, the other being the volume they generate.

Currencies gave a different answer. The earlier of the two staff reports, by Osler and Chang, ran the same kind of exercise on the floating-rate era in foreign exchange: daily rates for major currencies against the dollar, March 1973 through June 1994, with the trading profits judged against what 10,000 bootstrapped series produced under a random-walk null. Both later papers report the outcome of that work. Osler's 1998 paper says the pattern was profitable for the mark and the yen, and Lo, Mamaysky and Wang put it as two of the six currencies in the sample, the yen and the Deutsche mark, reaching statistically significant profits. Osler still declines to call currency use of the pattern rational, because other technical strategies dominated it.

What did Lo, Mamaysky and Wang find about the head and shoulders?

That conditioning on a head and shoulders shifted the distribution of subsequent returns, which is not the same thing as making money. Lo, Mamaysky and Wang applied their smoothed-extrema definitions to US stocks from 1962 to 1996 and, for each of ten patterns, compared the one-day return beginning three days after a pattern ended against the unconditional distribution of daily returns.

For the NYSE and AMEX sample, a Kolmogorov-Smirnov test separated five of the ten patterns from the unconditional case, and the head and shoulders was among them. The inverted version was not, at a p-value of 0.104. In the Nasdaq sample, all ten patterns cleared the five percent level. The authors' own reading of that is the careful one and deserves repeating: some patterns add information the unconditional distribution does not already contain, and they are explicit that this on its own establishes nothing about whether a trader could turn it into profit above the market's.

Frequency counts from the same paper are a useful check on whether these shapes are unusual at all. The NYSE/AMEX sample contained 1,611 head and shoulders patterns; a simulated geometric Brownian motion series, calibrated to match each stock's mean and standard deviation, contained 577. More than chance would produce, then, but a random price series still drew hundreds of them.

Has the head and shoulders pattern been tested on futures?

Not in any of the research cited here. The 1998 paper sampled US equities, the 1995 paper sampled dollar exchange rates, and Lo, Mamaysky and Wang sampled US stocks listed on NYSE, AMEX and Nasdaq. None of the three tested an exchange-traded futures contract, and a result drawn from one market is evidence about that market rather than about futures.

This matters because both algorithms work, in their base form, from nothing but a series of closing prices, so the pattern can be looked for in any price series at all — an index future, a crude oil contract, a five-minute bar as readily as a daily one, on any of the platforms we support. Being drawable everywhere is not the same as being tested everywhere. Every result above also came from daily bars, so none of it says anything about what the shape does inside a single session, which is the timeframe most people asking about this pattern are actually looking at.

What does a head and shoulders mean on a funded futures account?

Mostly that the holding period in the research does not fit the account. Osler's average position lasted about 10 business days of daily bars, and a funded account with a flat-by-the-close rule cannot carry anything like that. At Lumen, every position closes by 4:45 PM ET and nothing is held overnight or over a weekend, so a trade meant to run for two weeks has nowhere to live. Using the pattern on that account means using it on intraday bars, in a market and on a timeframe none of these studies examined.

The second constraint is the drawdown. A convention that expects price to drift back toward the neckline before resuming assumes you can sit through an open loss, and on a funded account the amount you can sit through is a fixed number rather than a judgement call: the drawdown line is monitored in real time through the session and does not care what shape is on the chart.

None of that is an argument against reading charts. It is an argument for knowing which claims about a pattern have been tested and which are inherited from a textbook, in the same way the published research on individual day traders is worth knowing before assuming results follow effort. Our own risk disclosure says most people who attempt this will fail.

FAQ

Is an inverse head and shoulders the same pattern?

It is the same geometry mirrored. Osler's paper calls the version forming at the end of a downtrend a bottom, where troughs take the role peaks play in a top, and Lo, Mamaysky and Wang define the inverted pattern as five extrema starting on a minimum instead of a maximum. The evidence is not symmetric, though. In the NYSE/AMEX sample of that paper, the head and shoulders separated statistically from the unconditional return distribution and the inverted version did not, at a p-value of 0.104.

How is the head and shoulders price target calculated?

By measuring the vertical distance from the neckline to the head and projecting that same distance from the point where price crosses the neckline. The technical manuals Osler surveyed present that distance as a floor on the move rather than a forecast of where price stops. Osler's test did not exit on it: positions were held until price stopped moving in the predicted direction, with the objective used only to size the allowance for a bounce back toward the neckline, so the profit figures in the paper are not a measure of how often the target was hit.

How often does a head and shoulders appear on a chart?

Less often than it feels. Osler's algorithm found about 27 confirmed patterns per firm across 31 and a half years of daily data, slightly fewer than one a year for each stock. Lo, Mamaysky and Wang counted 1,611 in their NYSE/AMEX sample, against 577 in a simulated random series built to match the same statistical properties, a useful reminder that a random walk draws this shape too.

Does heavy volume on the neckline break confirm the pattern?

It confirms that people are trading it. Osler measured unusual volume around neckline crossings and found that the crossing day and the two days after it together accounted for about one fifth of a day's volume, roughly half of it on the crossing day itself, while a control window centred 60 days after the head showed no significant unusual volume at all. Lo, Mamaysky and Wang separately found that conditioning on volume trend added little information for most patterns.