What Is a Moving Average Crossover Strategy, and Does It Work?

A moving average crossover strategy buys when a short average of recent prices passes above a longer one and sells when it passes below. Whether it works has been tested: in Park and Irwin's study of twelve US futures markets, the returns twelve trend-following systems earned over 1978 to 1984, the crossover among them, had turned negative over 1985 to 2003. Park and Irwin trace the earliest analysis of moving averages they located to the 1930s, and name this family the one practitioners reach for most among trend-following methods, citing two surveys of foreign exchange dealers.

The tests are the reason this post exists. Park and Irwin's 2005 AgMAS research report ran the rule on twenty-six years of US futures prices, charged commissions on every round turn, and re-chose its parameters every year from data that ended before the year being traded.

What is a moving average crossover strategy?

A moving average crossover strategy compares two averages of the same price series over different lookback lengths and holds a position based on which one is higher. In the version Park and Irwin specified for their 2005 futures study, both averages are taken over closing prices; the position goes long at the next day's open when the short average sits above the long one, and short at the next day's open when it sits below.

Two details of that specification matter more than they look. First, it reverses and never stands aside: there is no flat state, so each signal closes one position and opens the opposite one. Second, the decision uses the averages as of the close and executes at the following open, which is why Park and Irwin can cite Neftci for the rule being statistically well defined — nothing in the signal requires knowing a price that has not printed yet.

That is the whole mechanism. Everything else — which lengths, simple or exponential, whether to add a band — is a parameter choice on top of those three lines.

Are the golden cross and death cross defined terms?

Not in the place a US futures trader would look them up. The CFTC's public glossary runs to several hundred entries on a single A-to-Z page, and as of 24 September 2026 that page contains no entry for "golden cross", "death cross", "crossover", "whipsaw", or "moving average" — the phrases do not appear on it at all.

What the glossary does carry is the vocabulary underneath. Its entry for "Trend" is one sentence describing the general direction prices have been heading, up or down, with no method attached to it. It has entries for "Charting", "Chartist" and "Counter-Trend Trading". It has a heading for "Trendline" with no definition printed under it. And its "Technical Analysis" entry attaches a condition to the whole approach, which in the CFTC's words "can work consistently only if the theory that price movements are a random walk is incorrect."

The studies cited here do not use the nicknames either; they name rules by their two lengths. Brock, Lakonishok and LeBaron's 1992 study of the Dow tested five pairs: a 1-day short average against long averages of 50, 150 and 200 days, a 5-day short against 150, and a 2-day short against 200. Each pair was run with a 1% band and without one. To look up whether "the golden cross" has been tested, search the pair of numbers.

Does a moving average crossover work in futures markets?

On Park and Irwin's twelve-market test it worked in the earlier period and not in the later one. Their 2005 AgMAS report ran twelve trend-following systems — the dual moving average crossover among them — across twelve US futures markets on daily data from 1975 to 2003. The markets were corn, soybeans, sugar and cocoa; silver and copper; the British pound, the Deutsche mark and US T-bills; plus lumber, live cattle and pork bellies.

The result splits cleanly in two, and the split is the finding. In the earlier window, an equally weighted portfolio of all twelve systems earned statistically significant returns in four of the markets: 24.48% a year in sugar, 21.65% in silver, 7.64% in the Deutsche mark and 2.37% in T-bills.

Over the full 1978–2003 stretch the aggregate came to −3.14%, or −1.67% under the lower-cost assumptions. Park and Irwin also regressed annual returns on a time trend: the coefficient was negative in ten of the twelve markets and significantly so in six, and it was significantly negative for all twelve systems. For the crossover specifically, the estimated decay was 0.58 percentage points a year with a t-statistic of −4.16. Statistical significance throughout was assessed with a stationary bootstrap of 1,000 resamples rather than a plain t-test.

Has the crossover been tested on stock index futures?

Yes. Taylor (2000) ran moving average pairs on S&P 500 and FT 100 index futures. It was not the twelve-market study above that did it: those twelve markets include no stock index contract at all, so nothing in those results carries across to the ES or the NQ. What follows comes from Park and Irwin's companion review of 92 modern studies, and only as that review describes the underlying papers; neither original was read for this post.

The review's account of Sullivan, Timmermann and White (1999) is the sharpest. They searched roughly 8,000 rules drawn from five system families, moving averages among them, over a century of the Dow and over S&P 500 index futures from 1984 to 1996. On the futures sample the best rule returned a mean 9.4% a year at a nominal p-value of 0.04, which taken on its own would count as significant. Correct that p-value for the size of the search and it rises to 0.90, which is why the review reads the return as a product of the search rather than a finding.

Taylor (2000) is the other entry that reaches index futures. The review lists its data series as including the S&P 500 index and index futures and the FT 100 index and index futures, tested on daily data with short averages of 1, 2 and 5 days against long averages of 50, 100, 150 and 200, with and without a 1% band. On the result line the review prints, neither the FTSE 100 nor the S&P 500 index produced anything significant; nothing is reported separately for the futures series.

Be careful with one number here. The same 9.4% appears elsewhere in the review attached to a 50-day variable moving average with a 1% band, applied to the Dow over 1897–1996, where its Reality Check p-value was zero — the opposite conclusion, on a different market over a hundred-year sample rather than a thirteen-year one.

Which moving average lengths should you use?

The sources cited here point in two directions. Park and Irwin's review describes Brock, Lakonishok and LeBaron picking five combinations it calls popular, which is one way to choose; their own futures study chose systematically instead, and had to re-choose every year. That crossover grid ran eight short lengths from 2 to 25 days against thirteen long lengths from 5 to 65 in steps of five — 104 combinations per market, per year.

Their procedure was to simulate the previous three years, take the combination with the best mean net return over those three years, and trade only that one for the following year. The 1993 parameters came from 1990–1992; at the end of 1993 the process repeated on 1991–1993. That is the honest version of "optimising", and it is the one that makes the results out-of-sample.

What the resulting table shows is how unstable the answer was. For corn alone, the winning pair was 15 and 40 days going into 1978, 10 and 25 going into 1995, and 5 and 55 going into 2003. A trader who settled on any one of those and left it would have been carrying a setting three years of data no longer supported. Two other assumptions hold the model's returns to what the rule itself earned: one contract per transaction, and no pyramiding or reinvestment of profits. How many contracts a funded account can actually carry is a separate question, worked through in our post on futures position sizing.

Does it matter whether you use a simple or an exponential average?

It matters mechanically, and the difference is not only about weighting. A simple average weights every price in the window equally; an exponential average leans on recent prices, through a recursion that never fully discards anything.

Sierra Chart's studies reference, whose exponential-average entry carries a last-modified date of April 2026, documents the calculation as a smoothing multiplier of 2/(n+1) applied to the current bar's input, with the remainder carried from the previous bar's value. Because that previous value carries the one before it, the platform's page states that the recursion reaches the chart's first bar. The consequence the page flags follows from that: changing the Days To Load setting changes the exponential average printed at a given chart column.

The same vendor's combined Moving Averages study draws three averages, and each one takes its own input, length and type from a list of seven. Since Park and Irwin specified simple averages of closing prices, a backtest that silently used a different type is not testing the rule they tested — a reminder that a chart line is a platform setting, the same way VWAP is a platform setting rather than a fixed market fact.

How much do transaction costs matter to a crossover system?

Enough to decide the answer. Park and Irwin's base case charged $100 per contract per round turn across the whole sample, counting commission and the spread together in that single figure — against published bid-ask spread estimates for those twelve markets running from $3 to $25 per contract. Their second scenario cut the commission component to $50 over 1985–1994 and to $25 from 1995 onward, which put total costs at $100, then $75, then $50 across the three stretches.

Cutting those costs did not rescue the later period: the aggregate across all twelve markets and all twelve systems moved from −5.82% to −3.80% and stayed statistically insignificant. The report also notes a cost the spread does not capture at all — market impact — and puts clearing, exchange and floor brokerage fees at roughly $2 per contract.

Why costs matter this much here is structural. A reversing, always-in-the-market system trades on every crossing, including the ones that reverse again days later. The moving-average section of Park and Irwin's futures report names the problem directly: moving averages do badly in congested markets and are prone to whipsawing, and that goes especially for a system that keeps a trader in the market with no criterion for standing aside — which is what a reversing crossover is. Their review gives the remedy as a percentage band around the average, so a crossing only counts once it clears by a set margin, which leaves fewer trades to pay for. The trade-off the review records on the other side: in Brock, Lakonishok and LeBaron's variable-length moving average results, the 1% band increased the measured buy-minus-sell spread in every case — and those tests took no account of transaction costs at all.

What a crossover system runs into on a funded account

Every study described above runs on daily bars, and a reversing daily system carries its position across nights and weekends by construction. That collides with the rule this account runs on: our trading hours rule requires every position to be flat by 4:45 PM ET, anything still open is closed automatically, and the result counts toward your balance and your drawdown.

So a daily crossover cannot be traded here as the studies specified it. Compressing the same rule onto intraday bars is a different test, and none of the results above speak to it: every price study described here — Park and Irwin's, Sullivan, Timmermann and White's, Taylor's, Brock, Lakonishok and LeBaron's — ran on daily data. Treating a five-minute crossover as though it inherits a daily study's findings is a step the evidence cited here does not support.

Two further constraints change the arithmetic even if you do compress it intraday. The drawdown line is monitored live, not at the close, so an always-in-the-market system that reverses into a losing swing can end the account mid-session. And the contract limits cap what you can hold at once across all positions, which means a reversing system cannot scale its way out of a run of whipsaws. Whatever the rule, the base rate does not change: our risk disclosure puts it plainly, that most people who attempt this will fail.

FAQ

Is a golden cross a reliable buy signal?

Nothing in the sources cited here supports calling it reliable, and none of them uses the term. The CFTC glossary has no entry for it. Sullivan, Timmermann and White searched about 8,000 rules, moving averages among them; on S&P 500 index futures over 1984 to 1996 the best rule in that search returned 9.4% a year at a nominal p-value of 0.04, and 0.90 once the size of the search was accounted for. That is the closest thing to an answer these sources hold, and it points away from reliability.

What is a whipsaw in a moving average crossover?

The moving-average section of Park and Irwin's futures report treats whipsawing as the problem moving averages run into in congested markets, and says it applies particularly to a system with no flat state to fall back on. Their review ties whipsaws to periods when the short and long averages move closely together, and gives the remedy as a percentage band around the average, so a crossing only counts once it clears by a set margin — which leaves fewer trades, and so less to pay in costs. The CFTC glossary, checked on 24 September 2026, carries no entry for the word.

Do these studies mean technical analysis does not work?

They do not say that. Park and Irwin's review counted 92 modern studies and found 58 positive, 24 negative and 10 mixed, so most of the studies it counted reported profitability. Their criticism is aimed at the testing procedures, chiefly data snooping. Their own futures test is where they address that directly: returns that were statistically significant over 1978 to 1984 across twelve markets and twelve systems had turned negative over 1985 to 2003.

Do professional traders actually use moving averages?

The two surveys Park and Irwin's review summarises both cover foreign exchange dealers rather than futures traders, and both say yes. In a 1988 survey of chief foreign exchange dealers in London, 64% reported using moving averages or another trend-following system, and around 90% reported using some technical analysis at the shortest horizons, which that survey bounded as intraday out to a week. A 1995 survey of Hong Kong dealers put moving averages and trend-following at the top of the technical methods for usefulness, and found daily the most popular data frequency.