What Technical Analysis Claims (and Doesn’t)

Three claims under one name

Technical analysis is usually argued about as if it were one idea, and that is why the argument never ends. It is at least three separate claims, and they stand on very different ground. The first is that **trends persist**: an asset that has risen over the past several months is somewhat more likely to keep rising than to reverse, and the same for falls. This is the best-supported claim in the field. Moskowitz, Ooi and Pedersen found “time-series momentum” in 58 futures and forward markets — equity indexes, bonds, currencies and commodities — and Hurst, Ooi and Pedersen extended trend-following evidence back to the 1880s, with positive average returns in every decade and particularly good results in the worst equity drawdowns. The effect is modest, noisy and has long dry spells, but it is real enough that a whole industry (managed futures) runs on it. The second is that **prices react at levels**: old highs and lows, round numbers and moving averages. Here the evidence is partial, but there is a mechanism. Orders genuinely cluster at those prices — stop-losses just beyond them, take-profit orders at round numbers, unfilled bids where a stock last turned. Osler documented exactly this in currency markets: stop orders bunch just past round numbers and take-profit orders at them, which produces the reversals and accelerations chartists describe. A level is a memory of orders, not a line with powers. The third is that **named shapes predict on their own** — that a head-and-shoulders or a hammer carries a forecast wherever it appears. This is the weakest claim. Large samples put most patterns only modestly better than chance, with wide failure rates, and the edge that exists usually comes from the context (a level, a trend, volume) rather than from the shape. • Trends persist — best evidenced, modest, with long droughts (time-series momentum). • Levels matter — a mechanism in clustered orders, partly evidenced. • Shapes predict alone — weakest; context carries most of whatever edge exists. The rest of this subject is built on that ranking: trend and volume first, levels second, named patterns last and always qualified by location.

Dow’s tenets, and what survived them

The modern field descends from Charles Dow, whose editorials in The Wall Street Journal around 1900 were later codified as Dow theory. Some of his ideas aged well, some did not, and sorting them is a good first exercise in reading the field critically. Dow argued that the market moves in trends at several scales at once — a primary trend lasting a year or more, secondary reactions against it, and minor daily noise. That survives almost intact: it is the basis of multi-timeframe analysis (T11) and of every trend filter. He argued that averages should confirm each other (his industrials and transports), which survives in a looser form as the idea that breadth and related markets should corroborate a move (T12, T14). He argued that volume should confirm the trend, which survives as relative volume (T4). What did not survive is the idea that a trend can be identified with certainty while it is happening. Dow theory signals arrive late by design, and the rules for when a primary trend has “changed” are vague enough that two practitioners routinely disagree on the same chart. The lesson the field took from that is the one this subject insists on: write the rule down precisely enough that a computer could follow it, or it cannot be tested. Dow theory, audited — Trends at three scales: Survives — multi-timeframe analysis, trend filters · Averages must confirm each other: Survives loosely — breadth and intermarket confirmation · Volume confirms the trend: Survives — relative volume on breakouts · A trend change can be called precisely in real time: Does not survive — signals are late and rules are vague ←

Random walks, and why charts might still carry information

The case against technical analysis is the efficient-market argument in its weakest form: if past prices predicted future prices, traders would trade on it until the prediction disappeared. In its strict form — prices as a pure random walk — that argument has been rejected empirically more than once. Lo and MacKinlay found weekly U.S. stock returns did not behave like a random walk in their 1988 study, and the trend evidence above is another rejection. But rejecting a pure random walk is not the same as proving chart patterns work, and this is where most of the confusion lives. Small departures from randomness can exist and still be too small to profit from after the spread, commissions and slippage. Brock, Lakonishok and LeBaron reported that simple moving-average rules on the Dow beat buy-and-hold in data from 1897 to 1986 — and a decade later Sullivan, Timmermann and White showed how much of that kind of result can come from searching many rules on the same data, with the best rule’s performance weakening out of sample. Why would any information survive at all? Three reasons recur in the research. Investors under-react to news and then herd, which stretches moves into trends. Institutions trade slowly to limit their market impact, which makes their buying or selling persist for days or weeks. And orders cluster at levels, as above. Each is a mechanism you can name — which is the standard every chart claim in this subject is held to. • Strict random walk: rejected in several studies. • Profitable after costs: a much higher bar, and the one that matters. • Mechanisms that could leave information in prices: under-reaction then herding, slow institutional execution, clustered orders. A backtest that searched many rules and kept the best one has already used up the data it was tested on. That is data snooping, and T17 is about how it fools people.

How to test any chart claim

Every lesson after this one hands you claims: a moving average filters noise, a divergence warns of a turn, a breakout on high volume is more reliable. The habit to build now is a short checklist that turns any of them into something you can test instead of believe. First, state it as a rule a computer could follow: exactly which bar, which threshold, which exit. Second, name the mechanism — why would this work, and for whom is it a loss? Third, compare it with a base rate: how often would a random entry in the same market, held the same time, have done as well? Fourth, include the costs — the spread, commissions and slippage of every entry and exit. Fifth, test it on data that played no part in designing it. Applied to the three claims, the checklist produces the ranking above. Trend rules pass it in many markets with modest results. Level-based rules pass it sometimes, mostly when combined with trend and volume. Shape-only rules mostly fail at the third or fourth step: they beat a random entry by less than the cost of trading. One claim through the checklist — Claim: “A breakout above a 20-day high starts a trend” · Rule: Buy at the close above the prior 20-day high; exit at the close below the 10-day low · Mechanism: Stops cluster above the old high; trend persistence · Base rate: Random entries in the same market, held as long · Costs: Spread + commission + slippage on every round trip · Out of sample: Years the rule was not designed on ←

What the tests found: some signal, much noise, a lot of mining

The academic record on technical analysis is neither the dismissal it is often given nor the endorsement its sellers claim. Brock, Lakonishok and LeBaron tested simple moving-average and range-breakout rules on the Dow Jones Industrial Average from 1897 to 1986 and found that returns after buy signals were higher than after sell signals — by more than standard models of returns could explain. That paper made technical rules respectable for a while. Then came the correction for data mining. Sullivan, Timmermann and White asked how good the best of those rules would look if you counted all the rules a researcher could have tried — about eight thousand of them — and tested the winner on the decade that followed. The best rule held up inside the original sample and did not in the later one. Lo, Mamaysky and Wang took a different route, teaching a computer to detect classic chart patterns in U.S. stocks from 1962 to 1996; they found that some patterns did shift the distribution of later returns a little — information, but modest. The strongest evidence is for the plainest rule. Moskowitz, Ooi and Pedersen showed that a futures contract’s own return over the past twelve months predicted its next month’s return across 58 markets — equity indexes, bonds, currencies and commodities — a pattern they called time-series momentum, which is the statistical core of trend-following (T13 and T20). The honest summary: sharp, simple trend rules applied across many markets have the best support; precise patterns have a little; and most of what is sold as technical analysis has never been tested at all. • Brock, Lakonishok & LeBaron (1992): moving-average and breakout rules on a century of the Dow looked informative. • Sullivan, Timmermann & White (1999): corrected for about 8,000 rules tried, the best did not survive the next decade. • Lo, Mamaysky & Wang (2000): computer-detected patterns carried modest information. • Moskowitz, Ooi & Pedersen (2012): trend in 58 futures markets — the strongest evidence. The evidence, strongest to weakest — Trend rules across many markets: Robust across 58 futures markets ← · Simple moving-average rules on one index: Informative in-sample; fragile after correcting for mining · Classic chart patterns: Modest information when defined precisely · Most retail chart lore: Never tested — no evidence either way

Why edges fade once they are found

Even a real pattern has a half-life. McLean and Pontiff studied 97 published predictors of stock returns and found their returns were about 26% lower outside the original research sample and about 58% lower after the research was published. Part of the in-sample result was luck that the out-of-sample period did not repeat; part of the remaining edge was traded away once the publication told everyone where it was. A pattern you read about is, on average, a smaller pattern than the one in the paper. Two forces work in opposite directions on chart levels. Some levels are **self-fulfilling**: round numbers, a prior high, a widely watched 200-day average can matter because many traders place orders there, and studies of order flow find orders cluster at round numbers. Others are **self-defeating**: once enough people act on a signal, prices move before the signal fires, and the pattern disappears into the anticipation. Neither force is visible on a chart, and both change over time. For a learner the practical rule follows directly. Treat every chart claim as a hypothesis with an expiry date: write it so it can be tested, test it on data it was not built from (T17), expect what survives to be smaller than it looked, and keep checking it after you start using it. • Published predictors: about 26% weaker out of sample and 58% weaker after publication (McLean & Pontiff). • Self-fulfilling levels: round numbers and watched averages where orders cluster. • Self-defeating signals: anticipation trades the pattern away. • Treat a chart claim as a hypothesis with an expiry date. What happens to a published edge — Inside the original sample: The full reported return · Outside the sample, before publication: About a quarter smaller · After publication: More than half smaller ←

Sources

• Time-series momentum across 58 futures marketsMoskowitz, Ooi & Pedersen, “Time Series Momentum” (Journal of Financial Economics, 2012)
• A century of trend-following evidenceHurst, Ooi & Pedersen, “A Century of Evidence on Trend-Following Investing” (Journal of Portfolio Management, 2017)
• Patterns tested with modern statisticsLo, Mamaysky & Wang, “Foundations of Technical Analysis” (Journal of Finance, 2000)
• Moving-average rules and data snoopingBrock, Lakonishok & LeBaron (Journal of Finance, 1992); Sullivan, Timmermann & White (Journal of Finance, 1999)
• Stop-loss clustering at round numbersOsler, “Currency Orders and Exchange Rate Dynamics” (Journal of Finance, 2003)