Trend Following vs Mean Reversion: 387 Backtests Compared
Trend following or mean reversion: which style works better? We labelled every strategy version we have backtested by the kind of rule it is, before looking at any result by label, and compared what each style did before costs, after costs, against random entries and under our six gates.
Guide · 387 versions comparedNeither style won after costs. Breakouts came closest to an edge.
Tests: 54 · Versions: 387 · Positive after costs: 46% trend following, 43% mean reversion · New backtests: none
Short answer: in our tests neither won after costs: 46% of trend-following versions and 43% of mean-reversion versions were positive after costs. Trend following, and breakouts above all, beat random entries more often (22% against 9%), but not by more than chance allows in our sample.
00The short answer
- After costs, the two styles finished level in our tests. 46% of trend-following versions and 43% of mean-reversion versions were positive after costs, and only 3 versions passed all six of our gates, all of them trend following.
- Trend following was positive before costs more often (77% against 69%) and beat random entries more often (22% against 9%), but neither difference is larger than shuffling the labels produces by chance in a sample of 12 mean-reversion tests.
- Breakouts did best of all the kinds: 54% positive after costs and 34% above the random control, against 13% for other trend rules and 9% for mean reversion.
- Mean reversion traded half as often and still gave more to costs: a median 86 trades a year against 164, yet costs took 149% of its profit before costs against 80%. Its edge per trade was smaller.
01How we labelled the strategies
Every version got one label from its rule, decided by how the entry picks a direction relative to the move before the signal. The label file was written before any result was joined to it (the labels, the specification).
- Trend: enters with a trend measured by an average, an indicator or a recent move: moving-average crosses, MACD, Supertrend, Parabolic SAR, Heikin Ashi, ADX, Ichimoku, TDI, pullbacks bought in the trend’s direction, intraday momentum.
- Breakout: enters when price leaves a range or channel: Donchian and Turtle, opening ranges, the initial balance, noise-area bands, inside bars, the TTM Squeeze, the London range.
- Mean reversion: enters against the preceding move: RSI(2) dips, Bollinger fades, RSI 30/70, WaveTrend, divergences, gap fades, fading the VWAP cross.
- Trend following = trend + breakout, as the term is usually used (Turtle trading is the classic trend-following system). Patterns (engulfing, head and shoulders, the H1 Open Trap) and a time-of-day rule are shown but kept out of the main comparison. The gold tests keep the labels of our gold strategies guide.
Each test weighs the same: a rate is the mean over tests (a test with several styles counts once in each), so our 96-version FX study does not outvote a 3-version test. 24 versions with a debatable label are flagged; section 06 removes them.
02Trend following vs mean reversion
| Share of versions | Trend following 42 tests, 284 versions | Mean reversion 12 tests, 88 versions | Pattern and time 4 tests, 15 versions |
|---|---|---|---|
| Positive before costs | 77% | 69% | 81% |
| Lost before costs | 23% | 31% | 19% |
| Lost to costs | 31% | 26% | 46% |
| Positive after costs | 46% | 43% | 35% |
| Beat random entries (top 5% of the control) | 22% | 9% | 44% |
| 4 or more of the 6 gates | 8% | 9% | 13% |
| Passed all six gates | 3 | 0 | 0 |
| Median share of the profit before costs taken by costs | 80% | 149% | 267% |
| Median trades a year | 164 | 86 | 275 |
The gaps that look large, 8 points before costs and 13 points against random entries, are within what chance gives: shuffling the style labels among the tests 10,000 times produced a gap at least as large 46% and 18% of the time. After costs the gap is 3 points. With only 12 mean-reversion tests, our sample cannot separate the two styles.
03Trend, breakout and mean reversion
| Share of versions | Trend 25 tests, 192 versions | Breakout 18 tests, 92 versions | Mean reversion 12 tests, 88 versions |
|---|---|---|---|
| Positive before costs | 78% | 75% | 69% |
| Lost before costs | 22% | 25% | 31% |
| Lost to costs | 38% | 21% | 26% |
| Positive after costs | 40% | 54% | 43% |
| Beat random entries (top 5% of the control) | 13% | 34% | 9% |
| 4 or more of the 6 gates | 7% | 10% | 9% |
| Passed all six gates | 1 | 2 | 0 |
| Median share of the profit before costs taken by costs | 85% | 75% | 149% |
| Median trades a year | 167 | 161 | 86 |
Breakout rules stand out on two counts: the most versions positive after costs and the most above the random control, and two of the three versions that passed all six gates (the New York noise area breakout on NQ). The third, the long-only 1-hour golden cross on gold, is a trend rule; it was later changed to Inconclusive after a data check.
04Long-only rules and rising markets
Most of our markets rose over the test years, which helps every long-only rule. We expected trend following’s lead before costs to come from its long-only versions. It did not:
| Positive before costs | Trend following | Mean reversion |
|---|---|---|
| Long-only versions | 83% | 82% |
| Long-and-short versions | 72% | 60% |
Long-only versions of both styles were positive before costs equally often; the gap is in the versions that also go short (shuffle: 34%). After costs, long-and-short versions were positive 38% of the time for trend following and 33% for mean reversion.
05By market
Only index futures (ES and NQ) have enough tests of both styles to compare (19 trend following, 5 mean reversion): 45% against 50% positive after costs, 14% against 15% above the random control. On gold, forex and Bitcoin we ran only two mean-reversion tests each, against 5, 8 and 11 trend-following ones, too few to compare. Costs by market are in how much trading costs ate across our backtests.
06Without the debatable labels
Four groups of versions could be argued either way (the Coppock curve, Stochastic + RSI + MACD, the Fibonacci scalping rule and one opening-range variant). Without these 24 versions: trend following 48% and mean reversion 47% positive after costs; 22% and 10% above the random control. The picture does not change.
07Every test and its label
* Includes versions with a debatable label (section 06).
08What this does not say
- It is not every strategy. These are popular retail rules we chose to test; 42 trend-following tests against 12 mean-reversion tests.
- The styles were tested on different markets and timeframes. 5 of the 12 mean-reversion tests are on ES futures, and trend following dominates gold, forex and Bitcoin.
- Positive is not proven. Only 3 of 387 versions passed all six gates.
- A label is a simplification. Many real systems mix the two (a pullback in a trend buys weakness in an uptrend); our rule for the label is in the specification.
09So which style should you trade?
- The style matters less than the edge. In our tests the share of rules that survived costs was about the same for both.
- Check the result against random entries. That is where breakouts stood out; being positive after costs alone did not separate the styles.
- Mean reversion lives on small wins: put your round-trip cost next to its average trade before anything else.
- Long-only results on rising markets flatter every style; look at the versions that also go short.
- Decide the rules before the test and count every variant; see how to backtest a trading strategy without fooling yourself. Whether a rule also works on another market: our cross-market scoreboard.
FAQCommon questions
Is trend following better than mean reversion?
Not clearly, in our 387 backtested versions: positive after costs 46% against 43%, and the three versions that passed all our gates were trend following. The differences before costs and against random entries were within what chance gives with 12 mean-reversion tests.
Does trend following work?
Some trend-following rules were positive after costs (46% of versions), but only 22% beat the same trades at random times and three passed all six of our gates. Breakout rules did best: 34% beat random entries.
Does mean reversion work?
In our tests 43% of mean-reversion versions were positive after costs and 9% beat random entries; none passed all six gates. Costs took a median 149% of their profit before costs, because the average trade is small.
Is a breakout strategy trend following?
Usually yes: a breakout enters in the direction price leaves a range, and the classic trend-following system, Turtle trading, is a breakout rule. We show breakouts separately because they did better than other trend rules in our tests.
Which style has lower trading costs?
Mean reversion traded less often in our tests (a median 86 trades a year against 164), but costs took a larger share of its profit before costs, because each trade aims for a small move. How much costs ate across our backtests.
10Check our arithmetic
fm_spec.md · fm_label.js · fm_labels.json
fm_run.js · fm_ref.js · fm_logs.js · fm_rows.json · fm_summary.json · check results
11Corrections
None since publication.