Nonzero Lab Research · Evidence · 2026-08-19 · 6 min read

Do Moving Averages Work? Thirty Years of Evidence

Moving averages looked strong in early tests, faded after data-mining corrections and costs, yet still inform allocation. The answer depends on use.

1992: the optimistic opening

Brock (1992) is the modern starting point for testing technical rules. On a long sample of the Dow Jones Industrial Average from 1897 to 1986, they tested moving-average and trading-range-break rules and found that returns following buy signals were higher than returns following sell signals, and that the difference could not be explained by several common stochastic models for stock prices.³

The weight of that result comes from the sample length and the testing discipline: not a screenshot from a bull market, but nearly ninety years of daily data. It made the strong claim "simple rules contain zero information" hard to defend. If your prior was that moving averages are pure superstition, 1992 was a counterexample you had to take seriously.

But be precise about what it did and did not say. It said: in this one market, over this one century, returns after rule signals showed a statistically detectable asymmetry. It did not say any rule stayed profitable after costs, and it did not say the asymmetry survived in other markets or other decades.

1999: the data-mining correction

Sullivan (1999) asked the next question: if you test enough rules, some will look significant by luck — how much of the significance survives that correction? Testing roughly eight thousand technical rules with a data-snooping adjustment, they found a two-part answer. The best rule remained significant in the original sample. In the out-of-sample decade of 1987–1996, it was no longer significant.⁴

That was the first watershed. "Best in history" and "usable going forward" are different claims, separated by a data-mining discount. Every rule selected because it looked good on one stretch of history owes that discount.

The logic of the discount is worth keeping. Among eight thousand rules, a few will always land on the noise of that particular era by luck. They get selected precisely because they landed on it — which means the selected rule is, by construction, the one that got luckiest. Using the winner of a beauty contest to predict next year's winner is the everyday version of the same mistake.

2012: false discoveries and transaction costs

Bajgrowicz (2012) computed the discount more carefully. After controlling for multiple testing with false-discovery-rate methods, rule performance showed no persistence; once transaction costs were included, most of the remaining excess return was consumed.¹

The two corrections compound. The data-mining adjustment removes the rules that were merely lucky; the cost adjustment removes the rules whose small real edge cannot survive friction. What remains after both is close to nothing — and "close to nothing" is the honest starting point for any new rule you meet.

By this point the picture was: the in-sample edge of simple rules, after data-mining corrections and costs, is mostly gone. It is also why "I backtested it and the curve looked great" carries almost no information — the costs your backtest didn't charge, the market will charge for you.

A backtest that skips costs is not a rough estimate of a strategy; it is an estimate of a different, friendlier world. Spread, slippage, and the simple fact of crossing the bid-ask spread on every entry and exit compound across exactly the high-frequency trading a signal rule demands. The flatter the edge, the faster the costs eat it.

What if you ask a different question?

The story does not end in pessimism, because later work asked different questions.

Zhu (2009) stopped asking "can moving-average signals predict returns" and asked instead "does using moving averages for timing inside an asset-allocation framework beat a fixed allocation?" The answer: when return predictability exists, strategies that incorporate moving averages can outperform fixed proportions.⁵

Neely (2014) went more macro: do technical indicators forecast the equity risk premium? They found the predictive power of technical indicators to be on par with macroeconomic variables, and complementary to them — each side carried information the other lacked.²

At a different level of use, the answer to "does it work" changes. As a trading signal, the evidence is weak. As an input to allocation and risk-premium views, evidence remains.

Why would the two levels differ? Because the questions differ. At the signal level you ask "up or down tomorrow" — the hardest forecast there is. At the allocation level you ask "how much risk exposure makes sense now" — a question about position size, with far more room for error. The same moving average can be a noise generator for the first question and useful information for the second.

For the overall ledger, Park (2007) surveyed the modern literature: of 95 studies, 56 reported positive results, 20 negative, and 19 mixed.⁶ That distribution is itself the conclusion. This was never a question with a yes-or-no answer.

What does this mean for a trader?

For someone making decisions in front of a chart, thirty years of evidence offer not an answer but a sharper question: in which market, in which period, used in which way, does a rule retain information out of sample?

There is no universal answer, because all three variables drift with the environment. The only reliable method is out-of-sample testing in the markets you actually trade — and a blind test is the cheapest out-of-sample test there is. You do not know what comes next, so every call you make is genuinely out of sample.

A practical reading of the whole literature then looks like this. Do not adopt a rule because a study blessed it; the blessing is sample-specific. Do not dismiss a rule because a study buried it; the burial is sample-specific too. Treat every rule as a hypothesis about your market, your timeframe, your costs — and test it where the outcome cannot leak into the decision.

This connects directly to sample-size discipline: even if a rule has a real edge, a few dozen attempts cannot detect it (see How Many Trials Do You Need?). A second common misreading is treating an indicator's default parameters as laws of nature — RSI is the canonical case (see RSI Overbought and Oversold: Fifty Years of Misreading).