Does Trend Following Work? What Time-series Momentum Research Says

Trend following is one of the few trading approaches backed by academic research across many markets and many decades. Academics call it time-series momentum. But what the research measures is a portfolio spread across dozens of markets at once, not the single chart a retail trader has open. This article explains what the research found, what the results actually look like, and then simulates an SMA 50/200 rule to show that it wins only when the market really trends.

Time-series vs cross-sectional momentum

“Momentum” in research comes in two kinds. The names are similar but the ideas are different.

	Time-series momentum	Cross-sectional momentum
Question	Is this market up or down against its own past?	Which markets are stronger than the others?
Method	Up over the lookback, hold Buy. Down, hold Sell	Rank them, buy the strong group, sell the weak group
Overall direction	Can be all Buy or all Sell	Usually paired so longs and shorts offset
Most cited study	Moskowitz, Ooi & Pedersen (2012)	Jegadeesh & Titman (1993), in stocks

Moving-average crossovers and channel breakouts like the Donchian Channel are all time-series momentum, because they only look at the market’s own price. A currency strength meter that ranks which currency is strongest is closer to the cross-sectional kind.

What the research found

Moskowitz, Ooi & Pedersen (2012), “Time series momentum”, published in the Journal of Financial Economics, tested 58 futures and forward markets covering equity indices, currencies, commodities and bonds. The core rule looks at roughly the past 12 months of return: if positive, hold Buy; if negative, hold Sell. Each market’s position is sized so all markets carry similar volatility. The main findings:

• A market’s past return helped predict the direction of the same market over the next month, by a small amount, and this showed up across asset classes
• The effect lasted about a year, after which part of it reversed
• The diversified portfolio did well when stock markets made extreme moves, both up and down
• The same study used COT report data and found speculators tended to hold positions in the direction of momentum

Later, Hurst, Ooi & Pedersen wrote “A Century of Evidence on Trend-Following Investing”, testing the same kind of rule over more than a hundred years, with a consistent main conclusion. Another study worth knowing is Fung & Hsieh (2001), which explained that trend-following fund returns behave like holding an option that pays when price travels far in either direction.

Before you apply any of this: these studies used low-cost futures across dozens of markets, and several authors of the first two worked with AQR, a firm that runs this kind of fund. Research results are historical and do not tell you future results.

The shape of results you have to accept

Trend following has a personality most people can’t stand:

• Many small losses. With no direction, the system enters, gets stopped out, re-enters, gets stopped out again
• A few big wins. Almost all the profit comes from a handful of long trends, so the distribution of results is positively skewed, with a long tail on the profit side
• Long quiet stretches. For several years after the 2008 crisis, many funds of this type were flat or weak for a long time
• Crisis alpha is a tendency, not a promise. If markets fall slowly and for long, the system ends up short with them. If they crash fast and bounce straight back, no trend forms in time

Trend following doesn’t make money by being right often. It makes money by letting the right trades run far and keeping the wrong ones small. Cut off the big trades and only the small losses are left.

Long losing streaks are normal for this method. Read more in Losing streaks are normal.

Try it: an SMA 50/200 rule on two simulated series

To see the mechanism clearly, we built two simulated price series of 3,000 bars each, both starting at 100 and both with the same volatility of 1% per bar.

• Trending series: each bar carries a 0.06% tilt in one direction, and each bar has a 1 in 250 chance of the direction flipping, so trends last about 250 bars on average
• Mean-reverting series: on every bar, price is pulled back toward 100 by 1% of its distance from it

The rule: at each bar’s close, if the SMA 50 is above the SMA 200, hold Buy; if below, hold Sell; flip when the lines cross. It uses only closes already known, with no look-ahead. Each trade costs 0.1% of price (in and out). Results are in % of price, not compounded, with no leverage.

One trade from the trending series, step by step:

1. Bar 2,299: SMA 50 = 35.5945, still below SMA 200 = 35.6880. The system is holding Sell
2. Bar 2,300: SMA 50 = 35.7050 moves above SMA 200 = 35.7047. Close the Sell and open a Buy at the close, 39.05
3. Price climbs to a high of 58.22 on bar 2,619. At that point the trade is up 58.22 ÷ 39.05 − 1 = 49.1%
4. Bar 2,661: SMA 50 = 53.96 drops below SMA 200 = 54.02. Close the Buy at 51.87
5. Result = 51.87 ÷ 39.05 − 1 = 32.8%, minus the 0.1% cost leaves +32.7%, held for 361 bars

Even the best trade gave back about 16 percentage points before exiting, because a moving average always lags price. That is the price of staying in a trend for a long time.

Over the whole trending series, the system made 16 trades and won 7 (43.8%). The average winner was +15.3%, the average loser −6.36%, for a total of +49.8%. The three best trades (+29.7%, +29.3%, +32.7%) add up to 91.7%, more than the total, so the other 13 trades together lost money. On the mean-reverting series it made 26 trades, won only 4, and ended at −88.0%.

Notice that even on the trending series, the cumulative result went 1,497 bars without a new high. On a daily chart that’s about 6 years. Its deepest fall from a peak was 42.4%, and the rule never changed.

One series can be lucky or unlucky, so we generated 500 of each kind and ran the same rule:

	Trending	Mean-reverting
Series ending positive	78.0%	0.6%
Median total	+53.5%	−90.3%
Win rate per trade	39.5%	22.1%
Profit factor	1.78	0.20
Longest stretch without a new high (median)	1,252 bars	—

Even the trending series ended negative 22% of the time, and the worst 10% were below −32.7%. On the mean-reverting series the same rule lost almost every time, because when the SMA 50 rises above the SMA 200, price is already above the middle, which is exactly where it gets pulled back. All these numbers come from simulated prices, not a real market.

Trade results are positively skewed

Pooling every trade from the 500 trending series gives 8,227 trades.

The median trade lost −2.49%. The average winner was +18.58% and the average loser −6.80%. The skewness is 3.57, clearly positive, and the best 10% of trades made 145% of the net profit, so the other 90% together lost money.

An interesting detail: on the mean-reverting series, trade results were still positively skewed (0.51), because the rule’s shape is to cut a loser when the lines cross back and let a winner run. But the Profit factor was only 0.20. The skewed shape comes from the rule. The profit can only come from a market that actually trends.

Why it has to be diversified across many markets

One market is a bet that this market will trend while you hold it. Research avoids that problem by holding dozens of markets. In a year when gold goes nowhere, bonds or oil may run instead.

We extended the simulation: 200 portfolios, each with 20 trending markets of the same kind but unrelated to each other, equal money in each.

	Single market (4,000 runs)	20-market portfolio (200 portfolios)
Ending positive	78.9%	100%
Median total	+48.1%	+50.3%
Deepest fall from peak (median)	52.8%	6.4%

The average result is about the same, but the depth of the fall shrinks to roughly one eighth. With half the trend strength (0.03% per bar), a single market ended positive only 58.1% of the time, while the 20-market portfolio ended positive 79.0% of the time.

Real markets aren’t independent. If the average correlation between markets is ρ, the ratio of return to volatility improves by about √(N ÷ (1 + (N − 1) × ρ)) times:

• 20 markets, ρ = 0: √20 = 4.47 times
• 20 markets, ρ = 0.2: √(20 ÷ 4.8) = 2.04 times
• 20 markets, ρ = 0.5: √(20 ÷ 10.5) = 1.38 times

Pairs that share the USD are often highly correlated, so trading five USD pairs is not diversifying across five markets. Read more in Correlation and portfolio risk.

Costs and leverage

Cost per trade. A slow rule like SMA 50/200 trades only about 16 times per 3,000 bars, so costs don’t eat much. In the simulation, raising the cost from 0.1% to 0.5% per trade lowered the median total from +53.5% to +47.8%. But shrink the same rule down to a small timeframe and the trade count multiplies while trend distance shrinks relative to the spread. See Spread cost on small timeframes.

Overnight financing. Trend following holds positions for a long time. In this simulation a position was open for 2,800 of the 3,000 bars. On a daily chart (about 250 bars a year) that’s 11.2 years. Assume a net swap cost of 2% a year of contract value, and you pay 11.2 × 2 = 22.4%, nearly half the +53.5% median result. Real swap rates differ by asset and broker, and can be positive in some directions. Always check your broker’s contract specification.

Leverage. A diversified portfolio falls much less deeply, which is why funds can size up. A single market that falls about 50% of contract value without any leverage will, with high leverage, empty the account before the next trend arrives. Size positions with 1% risk per trade and understand leverage and stop out.

What we saw in our own backtests on gold

Our results on gold (XAUUSD) match the picture from the research: a single market depends on whether that year trends.

• Riders win, faders lose in trend years. One trend-following system on gold M1, 0.10 lot per trade, 10,000-dollar balance, lost 2,764 dollars in 2024, made 7,404 in 2025, and made 17,836 from January to September 2026. A morning counter-trend system made 1,230 in 2024 and lost 728 in 2025 (2024–2025 used 1-minute price data, 2026 used real ticks). A separate stepped counter-trend system made 11,791 in 2026 but emptied the account in 2024. Details in Ride the trend or fade it
• The regime wall. A pullback system on gold M1, tried at 7 parameter values, was positive in all 7 cells in 2026 but lost in 25 of 28 cells across 2022–2025. The real variable was the year, not the parameter. See Market regime change
• Far targets beat near ones. On gold M5 from 5 January to 24 July 2026, real ticks, 0.02 lot, 3,000-dollar balance, with the same entries, an exit that locked in profit early won 78.5% but lost 493 dollars, while a far target at Fibonacci 161.8 won only 38.9% but made 3,185 dollars. That is the same positively skewed shape as above. But the same far target lost money in 2022 and 2023. See Near or far Take profit

All of these are backtest results on a single market. Past results do not predict future results, and they are not returns you should expect.

How to use this

• Judge a trend system by winners against losers, not by win rate. Winning 40% can be normal. See Win rate vs break-even
• Don’t cut the big trades. A trailing exit pulled too tight kills the positive skew. Compare approaches in Trailing stop methods compared
• Test many years and many regimes. A trend system has to survive sideways years; a counter-trend system has to survive strong trend years
• Set expectations for quiet stretches. Before starting, find the longest stretch without a new high in your test, and ask yourself honestly whether you could sit through it

Research says diversified trend following has historical evidence behind it. It does not say your one chart will trend this year.

Common mistakes

• Applying research results to one market. The academic results come from dozens of volatility-sized markets. One market swings far more
• Quitting during a quiet stretch, then coming back after the big trend has passed, which collects the losses and misses the big trade
• Making the averages faster to exit sooner. The win rate rises, but so do trade count and costs, and it cuts off the profit tail that is the real source of money
• Picking the best parameter from the past. Try many values and pick the prettiest, and the result usually looks better than it is. See Multiple testing and data snooping
• Forgetting swap. Holding a CFD for months, overnight financing can eat a large share of the profit