What Is a Double Top or Double Bottom Pattern?

A double top is a price run that stalls at one level, falls away, comes back, and stalls at roughly that level again. A double bottom is the same event upside down: two declines ending at about the same price with a rally in between.

Whether the shapes carry information is a separate question, and unusually for a chart pattern there is published work answering part of it — work whose results for the top and for the bottom did not come out the same. What follows is the definition, the evidence, and what it means for sizing a position on a funded account.

What is a double top and a double bottom?

A double top is two peaks that reach approximately the same price with a dip between them; a double bottom is two troughs that reach approximately the same price with a rally between them. Both are charting conventions: the CFTC's glossary, which does define a good deal of the surrounding vocabulary, carries no entry for either.

The vocabulary underneath them is defined officially. Under Resistance, the CFTC glossary describes a price area where fresh selling shows up and damps a rise; under Support, the mirror of that — an area where fresh buying arrives and a decline stops. A double top is therefore a claim about resistance: price reached an area, sellers met it, price came back, and the same thing happened. A double bottom makes the matching claim about support.

The glossary also defines technical analysis itself, and attaches a condition worth carrying through the rest of this post. Its Technical Analysis entry says the method works from how price has moved, how quickly, and how trading volume and open interest have shifted, and that it leaves the underlying fundamentals out — and it adds that the approach can work consistently only if the random-walk view of price movement is incorrect. That is a testable statement, and the work below is essentially that test.

Searching the glossary returns nothing for "double top", "double bottom" or "neckline". It does carry a Head and Shoulders entry, and even there the wording hedges — that shape is described as generally considered predictive of a price reversal, rather than as one that is. Our post on the head and shoulders pattern goes through it. The double top has no such entry at all.

How is a double top defined precisely enough to test?

By fixing two numbers: how close the two peaks have to be, and how far apart in time. Lo, Mamaysky and Wang, in their NBER working paper on technical analysis, set out formal definitions for ten classical patterns so a computer could find them; their Definition 5 covers the double top and bottom.

Their rule requires the two tops to sit within 1.5 percent of their average, and requires the second to occur at least 22 trading days — about a month — after the first. The month-apart requirement is credited in the paper to Edwards and Magee's Technical Analysis of Stock Trends, the 1966 fifth edition, the same book the authors name as an example source for all five pairs of patterns they formalise. So the tolerance is the researchers' choice and the separation is the chartists'.

Getting there needs one more step. Raw prices are too noisy for "peak" to mean anything, so the paper smooths each price history with a kernel regression and takes the local maxima and minima of that curve, in overlapping windows of 38 trading days. Peaks are features of the smoothed line, not of individual closes.

Two things follow. First, "roughly the same price" is a band, not a match: each peak may sit up to 1.5 percent away from the midpoint between them — 75 cents either side on a $50 stock, and since the tolerance is a percentage, 75 points either side of a 5,000-point level. Second, the definition stops at the two peaks. It contains no entry rule, no confirmation trigger and no price objective, so nothing that follows tests the trade most people mean when they say they trade double tops.

How often does a double top appear on a chart?

Often enough to be the most common of the ten in the NYSE/AMEX sample. Across NYSE and AMEX stocks from 1962 to 1996, the algorithm found 2,076 double tops and 2,075 double bottoms — more than the 1,611 head and shoulders and 1,654 inverted head and shoulders in the same data.

The sample matters before those counts mean anything. Prices came from CRSP, split into seven five-year subperiods running from 1962; within each one, fifty stocks were drawn at random, ten from each fifth of the market-capitalisation range. A separate Nasdaq sample built the same way produced 1,208 double tops and 1,147 double bottoms.

The comparison that matters is not between patterns but against noise. The authors ran the identical detector over simulated geometric Brownian motion — a random walk calibrated to match each stock's own mean and standard deviation — and it produced only 535 double tops and 574 double bottoms on the NYSE/AMEX side. Real prices threw up roughly three to four times as many as the random walk did.

Be careful what that gap shows. It shows the shape is not merely an artefact of randomness; broadening tops and bottoms went the other way in the same table, with the simulation producing more of them than the data did. It does not show the shape predicts anything. Frequency is not forecasting power, and the authors treat their simulation as one realisation rather than a general result.

Does the double top pattern actually work?

The one study behind this post tested both halves of the pair and got different answers for each, in one of its two samples. Lo, Mamaysky and Wang tested whether returns following a detected pattern are distributed differently from returns generally, using the one-day return beginning three days after the pattern ends. That is a test of information content, not of a trading strategy.

On the NYSE/AMEX sample, seven of the ten shapes cleared their goodness-of-fit test, meaning the returns that followed them looked statistically unlike returns in general. The double top produced the largest test statistic of the ten there, Q = 50.97. The double bottom was one of three that failed, at a p-value of 16.6 percent. A second test, the Kolmogorov-Smirnov, separated five of the ten from the unconditional case; the double top was the last of those five to clear, at p = 0.021, and the double bottom was not among them, at p = 0.215. That figure is read from the paper's Table VII; the paper's own text on page 22 instead gives 0.393 for the double bottom, which Table VII assigns to the triangle top. The table is the source cited here, because the five values below 0.05 in that row are exactly the five patterns the text names as significant.

The Nasdaq sample behaved differently: every one of the ten patterns was significant at the 5 percent level, despite fewer detections and therefore a weaker test. The authors read this as a difference in how informative the patterns are between the two markets, not as a blanket endorsement.

What the authors themselves draw from it is narrow. Some of the patterns carry information the unconditional distribution does not already hold, most clearly on Nasdaq; the authors say in the same breath that this need not mean a trader can convert it into profits above the market's, and that shapes inherited from the charting tradition — they name head and shoulders and rectangles — may not be the best ones for the job. Adding a volume-trend condition did not rescue either double: in the NYSE/AMEX sample neither the double top nor the double bottom reached significance under rising or falling volume.

Why would price stop at a level it has already reached?

Because of where other people's orders are sitting. The clearest evidence among the sources behind this post comes from currency markets rather than futures: Osler, in a Federal Reserve Bank of New York staff report, examined 9,667 conditional orders placed at one large dealing bank between 1 September 1999 and 11 April 2000, across three dollar pairs.

The requested execution rates were not spread evenly. Around 8.7 percent of orders sat at rates ending in 00, with weaker clusters at 50 and at other rates ending in zero or five; an Anderson-Darling test rejected uniformity for all three pairs at better than the 0.01 percent level. The two order types also clustered differently: take-profit orders were executed exactly at 00 in 9.3 percent of cases against 4.4 percent for stop-loss orders, a difference the author tested by bootstrap and found significant at 1.1 percent.

The asymmetry does the work. When price rises into a level ending in 00, roughly 10.5 percent of take-profit sell orders are triggered against only 2.8 percent of stop-loss buys, and the amounts behind each order type are similar on average. Selling dominates there, which is a mechanical reason for a rise to stall. Just past the level the imbalance flips: 7.4 percent of stop-loss buys sat at rates ending 90 to 99 against 14.4 percent at rates ending 01 to 10 — a reason for a move to accelerate once through.

Note the boundaries, because they are wide. One bank's order book, three currency pairs, seven months, and the levels studied are round numbers — not prior highs, and not double tops, which the paper never mentions. It supports the general idea that a level can hold because orders sit at it. It is not evidence about a double top on an index future.

Double top vs head and shoulders: what is the difference?

A double top has two peaks at approximately the same height; a head and shoulders has three, with the middle one standing above the other two. Both were formalised in the same study, and it found different things about each.

The last row is the one to take away. In each of these two pairs, the top version separated from the unconditional return distribution and the bottom version did not — same data, same algorithm. That direction is not a rule, though: among the broadening patterns in the same table, the bottom cleared the threshold and the top did not.

What does a double top mean on a funded futures account?

Practically, it means a level to place a stop against and a number of failed attempts your drawdown can absorb — arithmetic, not analysis. The research above says nothing about whether the next double top resolves downward, so the part you control is the size of the attempt.

Work it from the account rather than the chart. A $50K Classic account here is simulated, and carries a $2,000 maximum drawdown with a four-contract limit, or forty micros; every size and its drawdown sits on the pricing page. Risk $250 on each attempt and eight losses in a row end the account. Risk $100 and you get twenty. Neither number says anything about whether the pattern works; they say how many times you may be wrong before the question stops being yours to ask.

The drawdown here is end-of-day trailing, which changes how a winning trade feels. The line follows your highest closing balance, never retreats, is monitored live during the session, and stops moving once your profit equals the drawdown amount — how the drawdown moves sets out the mechanics with a worked example. A pattern that runs in your favour and gives most of it back before the close has cost you nothing in the drawdown, but has not bought you room either.

One more thing worth being blunt about: a shape that produced a p-value of 0.215 in one sample is no reason to increase size, and neither is one at 0.021. Our position sizing post works through setting risk per trade first, and our risk disclosure says plainly what the outcomes look like for most people who attempt this.

FAQ

Is a double bottom just a double top turned upside down?

Geometrically yes — the formal definition in Lo, Mamaysky and Wang's NBER paper mirrors the top, swapping maxima for minima. Statistically the two behaved differently in their NYSE/AMEX sample. The double top's conditional returns differed significantly from unconditional returns under a Kolmogorov-Smirnov test, at p = 0.021; the double bottom came in at p = 0.215, short of the 5 percent threshold. Their goodness-of-fit test agreed, failing the double bottom at 16.6 percent. On their Nasdaq sample, all ten patterns were significant at the 5 percent level, both doubles included.

How far apart do the two peaks of a double top have to be?

In the definition used for testing, at least 22 trading days — roughly a month. Lo, Mamaysky and Wang attribute that requirement to Edwards and Magee's Technical Analysis of Stock Trends. They also require the two tops to fall within 1.5 percent of their average price — a percentage, so on a 5,000-point level it works out to about 75 points either side of the midpoint. Their test ran on daily stock prices; a shape marked up on a five-minute chart is not the thing that definition was built to find.

Has the double top been tested on futures markets?

No published test of the double top on futures was located for this post. The main study covered US stocks from CRSP, sampled from NYSE, AMEX and Nasdaq between 1962 and 1996. The nearest adjacent evidence, Osler's New York Fed staff report on order clustering, comes from an FX dealing bank and concerns round numbers, not chart patterns. Applying either to an equity index contract extends it past what the data covers.

Does volume confirm a double top?

The one study behind this post found no help there for the double top. Lo, Mamaysky and Wang re-ran their Kolmogorov-Smirnov test with each pattern paired to a rising or falling volume trend. In the NYSE/AMEX sample, statistical significance mostly fell away rather than improved; the exceptions they pick out are the triangle bottom with declining volume and the broadening top with rising volume. Neither the double top nor the double bottom cleared the 5 percent threshold under either volume condition there. The CFTC glossary's Technical Analysis entry lists volume among the approach's inputs, but listing an input is not evidence it works.