Bollinger Bands: Mean Reversion or Breakout? 315 Settings Tested

Bollinger Bands are a moving average with a band above and below it, set at some multiple of how much price has been varying. Most price action stays inside that envelope. Two common trading rules are fade the band, which treats a touch as a reversal, and follow the break, which treats a close outside as continuation.

Which one is right depends on who you ask. So we measured both, on the same pair, over the same period, paying the same costs.

For 2025 (1 January to 31 December) on USD/JPY, seven periods crossed with five deviations, plus four choices of where the fade takes profit, gives 140 fade settings, run separately for the long and the short side, and 35 breakout settings: 315 in all. Each was run on three timeframes, and the whole grid was run again on 2024.

The answer first. On the hourly chart, the fade averaged a 60.4% win rate and came through both years in only 25 of 140 settings; the breakout averaged 40.4% and came through in 23 of 35. Selling the upper band did worse still: zero settings out of 140 were profitable in both years on either the 15-minute or the hourly chart. Ranked by win rate, the order reverses.

fade the lower band (140)fade the upper band (140)follow the break (34)one-hour chart, 2025

What Bollinger Bands measure

The middle line is a simple moving average (period). The bands sit above and below it at a multiple of the standard deviation of closes over that same period (deviation).

The usual claim is that 1σ holds about 68% of the action and 2σ about 95%. That is what a normal distribution would do. Counting the actual bars gives something else: here are the 6,226 hourly USD/JPY closes of 2025 against a 20-period band:

Every band holds less than the textbook says. At 2σ, roughly one bar in eight closes outside. A fade built on "price rarely goes there" is going to meet price out there more often than the assumption allows.

The bands themselves are a ruler that stretches with the market: wide when it is moving, narrow when it is not. The same 2σ is a different number of pips on different days.

The shared parameters are period and deviation. The fade also has an exit deviation.

The counts bear it out. Averaged over the fade-long settings on the hourly chart in 2025, a period of 10 traded 319 times a year and a period of 50 traded 53. By deviation it runs 213 trades at 1σ down to 77 at 3σ. Raising period or deviation reduces trade count and therefore reduces the total spread paid.

How to put them on a chart

Bollinger Bands ship with every platform.

Drawing them is the same everywhere. The decision that matters is whether the rule trades a touch or a close outside the band. This article tests a band-touch fade and a close breakout separately.

How this was tested

Entering this table reproduces these results.

The long fade and short fade are reported separately because a band touch is directional: buying the lower band and selling the upper one produce entirely different results.

Fade the band or trade the break?

The fade won more often per trade; the breakout kept more settings profitable in both years.

(Mean win rate is over the 2025 settings; the settings that never took a trade are left out of that column.)

The figure at the top plots the hourly row of that table, one dot per setting: win rate across, net pips up. The red and purple dots (fading) gather to the bottom right; the blue ones (following) to the top left. The settings that are right more often are the ones below zero.

Selling the upper band produced no setting at all that was profitable in both years on the 15-minute or hourly chart. USD/JPY rose 1,632 pips through 2024, so repeatedly selling upper-band touches kept the short fade positioned against that rise.

Why the higher win rate loses

The two sides of the trade are different sizes.

fade longfade shortbreakoutleft = mean loss, right = mean win, in pips, 2025

Averaged over every 2025 setting:

A fade wins a little when the touch snaps back and loses a lot when price simply keeps going: the exit waits for the far band, so a trade that was wrong stays wrong for a while. The loss runs 1.7 to 1.9 times the win. A 60% hit rate does not survive winners worth 60% of the losers.

The breakout is the mirror image: wrong more often than not, but the trades that work are the ones that run.

The relationship mostly holds on the other timeframes. The fade-long wins 20.8 and loses 40.6 pips on M15, wins 84.9 and loses 140.9 on H4. The breakout wins 66.5 against 42.5 on M15, and 286.1 against 136.4 on H4.

There is one exception. On the four-hour chart the short fade wins 99.3 pips against a 94.5-pip loss: the only row where the winner is the larger side. It still came through both years in 1 setting out of 140: a favourable size ratio did not cover what 2024's rise took out of it.

The textbook 20 and 2σ, used both ways

fade (solid = 2025)breakout (solid = 2025)period 20, 2σ; the fade takes profit at 1σ

The only rows profitable in both years are the 15-minute and hourly breakouts. Faded, the same two numbers lost on M15 in both years; on H1 they made a small profit in 2025 and lost in 2024.

The four-hour chart changes sign between years on both sides. Its breakout took 41 trades in 2025 at a 19.51% win rate for −2,126.7 pips; that is not a sample worth drawing a conclusion from.

How to choose period and deviation

The only fade settings that survived were the slow, rarely touched ones. Period and deviation are varied one at a time on the hourly chart and averaged over the remaining settings. The last column counts settings profitable in both 2024 and 2025.

Breakout, by period

Periods 20 and 25 went 5 for 5. The textbook period really is near the top on the breakout side.

Breakout, by deviation

A 1σ deviation went 7 for 7. Narrower bands break more often, which gives the setting a sample to stand on. Past 2.5σ a break becomes rare and the year starts deciding the result.

Fade, by deviation

For the fade, wider deviations reduced the average loss. Every deviation still averaged negative in 2025, and a 1σ fade was profitable in both years in only one of its 28 settings.

Period runs the same way. On the hourly chart the fade at period 50 came through both years 7 times out of 20; at period 20, 0 times out of 20. What survives on the fade side is the setting that almost never fires.

Does last year's pick hold up?

The best fade of 2024 finished below the 2025 median; the best breakout stayed above it.

The fade setting lost 268.0 pips in 2025, below that year's −251.7-pip median. The breakout setting made +1,488.1 pips, above the +788.6-pip median.

The direction does not run both ways. Period 10 at 1.5σ made +1,838.4 pips over 363 trades in 2025 and lost 156.5 pips in 2024. Even for the breakout, one year's annual net did not select a setting that worked in the other year.

Filters, stops and targets

The ADX filter turns the fade positive in 2025 and deepens its loss in 2024. Added to the breakout it came in under the unfiltered version in both years. Screening a trend-following rule for trend strength did not help it.

Fixed stops and targets are executable levels measured from the filled entry. A 60-pip target therefore closes at exactly +60 pips when reached; the engine does not subtract a second exit-side adjustment from that configured level.

Only the 24-bar time exit beat the opposite-band exit in both years, and only on the breakout (+1,177.5 → +1,413.8 and +1,434.6 → +1,536.2). Tightening the stop makes the breakout worse in a straight line because the rule depends on the few trades that continue far enough to offset its losses, and a close stop removes those trades.

Nothing helped the fade in both years.

How much does spread take?

About 40% of the fade's profit.

At a zero spread the fade makes 121.1 pips; at 0.3 pips it keeps 71.0. Costs remove about 40% of the pre-cost profit: the 50.1-pip difference is exactly 167 trades × 0.3 pips.

The breakout still keeps 953.1 pips at a 2.0-pip spread. It loses 264.0 pips going from 0 to 2.0, which is 132 × 2.0. The same relationship held without error in QQE and moving average crossovers.

Choosing the fade because it wins more often is not something this test supports. The fine sweep from 0.5σ through 4σ takes the separate question of which deviation works best. The 125-bar BBW squeeze test uses the distance between those bands as a volatility condition. Williams %R also produced high win rates without correspondingly high annual net results. The RSI fade and level-cross test found mean win rates of 65.3% for the fade and 23.7% for the cross. Fibonacci retracements compare buying a level with trading its break across 401 depths. MA disparity settings sets the same two lines by a percentage of price instead of a standard deviation, and judges them on closes rather than wicks. Building the rules without writing code is how to repeat any of this on your own pair.

Notes

• One pair, USD/JPY, and two years. Nothing here guarantees the same pattern elsewhere
• The fade holds until the far band with no stop, so a trade that goes wrong shows its full loss. Stops are measured in their own section
• The breakout reads closes. A bar whose wick pierced a band but whose close came back inside is not a signal
• Entries and exits fill at bar closes. Real fills differ
• The spread is a flat 0.3 pips throughout. Real spreads move with the session and around releases
• Four-hour settings trade between 4 and 153 times a year on the fade side, and one breakout setting took no trades at all. Win rates and profit factors at those counts are wide numbers before anything else is said