Bollinger Band Settings: Which Sigma Works Best?

The Bollinger Band dialog has a box called “Deviation” or “Standard Deviations.” Two is the familiar default, but whether 1σ, 2σ or 3σ is best depends first on what you do at the band.

We tested USD/JPY from 0.5σ through 4σ in 0.25 steps, crossed with seven periods from 10 to 50. A lower-band fade, an upper-band fade and a close-based breakout were run on 15-minute, hourly and four-hour bars across 2024, 2025, and the two halves of 2025: 3,780 backtests in all.

The short answer is that 1σ is the most defensible starting point for an hourly breakout. From 0.5σ through 1.25σ, all seven periods were profitable in both 2024 and 2025. Within that group, 1σ ranked first of the 15 deviations in 2024 and fourth in 2025. The 2025 winner, 1.75σ, fell to twelfth in 2024.

That does not make 1σ universally optimal. It means that, among these candidates, 1σ made the best compromise between profit, survival across periods and transfer to another year. The fade did not produce a repeatable “best sigma.”

0.5 to 1.25σ1.5 to 4σseven periods; USD/JPY hourly breakout

What Bollinger Bands sigma changes

Let SMA(n) be the simple moving average of the last n closes, SD(n) their standard deviation and k the deviation setting. The bands are:

Upper = SMA(n) + k × SD(n)

Lower = SMA(n) − k × SD(n)

The 2 in the settings dialog does not mean 2%. It multiplies the rolling standard deviation by two. Raising sigma pushes both bands away from the mean, producing fewer touches and fewer close-based breaks. Raising period also tends to reduce frequency because the centre line reacts more slowly.

The often-quoted 68%, 95% and 99.7% figures for 1σ, 2σ and 3σ describe a normal distribution of independent observations. A market has serial dependence, trends and jumps, so those percentages are not market signal frequencies.

closes finishing insideperiod 20; 6,226 USD/JPY H1 bars in 2025

On 6,226 hourly USD/JPY bars in 2025, using period 20, 45.0% of closes finished inside 1σ, 87.1% inside 2σ and 99.1% inside 3σ. That leaves 12.9% outside 2σ. Merely reaching 2σ does not establish that price is abnormal or due to reverse.

The settings dialog

For the conventional 20-period, 2σ display in MT4 or MT5, the inputs are below. TradingView and Formiq expose the corresponding period and deviation fields.

Formiq's backtester adds a trigger choice.

The same 20 and 2σ can give opposite results when a touch is faded and a close outside is followed. The trading direction has to be fixed before sigma can be compared.

How this was measured

The fade's opposite band stays fixed at 1σ so the experiment changes the entry sigma only. Moving the entry and exit together would not reveal which one caused the result.

The hourly sigma results

Each row below averages its seven periods. “Positive both” counts the periods that made money in both 2024 and 2025.

All seven periods survived both years from 0.5σ through 1.25σ. One sigma ranked first in 2024 and fourth in 2025, the most even placement in that robust group. Two sigma was also viable, with six of seven periods positive in both years, but its result depended more on period.

The apparent return to four survivors from 3.5σ upward needs the trade column beside it. Those settings averaged 6.3, 4.1 and 2.3 trades in 2025. A few successful outliers are not the same as a stable setting.

The whole 105-setting breakout grid changed with timeframe.

Hourly bars retained 67 settings in both years. Four-hour bars went from 84 winners in 2024 to 32 in 2025, which is why the answer here is explicitly an hourly one.

Period 20 side by side

Fixing period at 20 makes the frequency change easier to see.

At period 20 alone, 1.75σ beats 1σ in both years. Across all seven periods, however, 1.75σ falls to twelfth in 2024. The best sigma after fixing one period is not the same question as the sigma that survives changing the period.

No stable sigma for the fade

The 105 hourly settings point in a different direction when the band is faded.

Widening the entry often made the fade-long loss smaller. Averaged over seven periods, 1σ made −833.0 pips in 2025 and −112.5 in 2024. At 2.75σ the figures were +150.5 and +1.7; at 3σ, −2.8 and +46.4.

But the best 2025 fade-long sigma, 2.75σ, was not the 2024 winner; that was 3σ. Their year-to-year sigma rank correlation was only +0.182. The short fade retained just one of 105 settings in both years. A wider band may simply trade less and therefore lose less; it did not establish a repeatable mean-reversion optimum.

Out-of-sample transfer

Choosing sigma by 2025 alone selects 1.75σ for the breakout, with a seven-period mean of +1,850.2 pips. In 2024 it made +590.2 and ranked twelfth of 15. Choosing on 2024 selects 1σ at +1,568.8; it made +1,081.9 and ranked fourth in 2025.

Optimizing period and sigma together is less stable still. The 2025 winner, period 10 and 1.75σ, made +2,753.0 pips but ranked 63rd in 2024 at +686.0. The 2024 winner, period 25 and 2.25σ, made +2,793.5 but ranked 68th in 2025 at +231.3.

The rank correlation across all 105 settings was +0.040 between years and +0.104 between the two halves of 2025. Profit can remain positive while the fine ordering fails to transfer. That is why the conclusion favors a neighborhood of surviving settings over the single best row.

Can win rate choose sigma?

No. Averaged over all 105 hourly settings in 2025, the fade long won 61.6%, the fade short 60.5% and the breakout 41.0%.

The fade is right more often but loses more when wrong. The breakout's average winner is about twice its average loser. Looking only for the sigma with the highest hit rate discards that payoff shape.

Filters, exits and costs

We added common filters to the period 20 and 1σ breakout used as the robust baseline.

None of the ADX thresholds beat the baseline in both years. London and New York hours improved it slightly in both, from +1,052.5 to +1,073.5 in 2025 and +2,071.7 to +2,154.3 in 2024, but that session result needs more pairs and years before it becomes a setting recommendation.

Fixed stops and targets improved 2025 at the cost of 2024. None beat the plain rule in both years.

With spread varied on the 2025 period 20 and 1σ breakout, net profit moved from +1,135.6 pips at zero to +1,052.5 at 0.3, +858.6 at 1, +581.6 at 2 and +304.6 at 3 pips. Its simple break-even spread was about 4.10 pips. One sigma trades often, so low cost is part of the setting.

What the test supports

1. For an hourly breakout, 1σ is the most defensible baseline. Every period from 10 to 50 survived both years at 0.5σ through 1.25σ, and 1σ ranked first in 2024 and fourth in 2025
2. The single-year winner is not the best setting. The 2025 winner, 1.75σ, ranked twelfth in 2024; the year-to-year sigma rank correlation was +0.068
3. Two sigma is viable, not uniquely correct. Six of seven hourly breakout periods were profitable in both years, including period 20
4. The fade has no defensible best sigma here. Only 25 of 105 long settings and one short setting survived both years; widening mainly reduced frequency
5. Results above 3.5σ need their sample size attached. Two to six trades a year cannot separate a setting from a few large moves

Which Bollinger Bands period works best? compares lookbacks from 5 through 200 and tests whether the best bar count changes with timeframe. The broader comparison between fading and following the bands is in 315 Bollinger Band backtests. RSI periods and 30/70 levels also show why a higher fade win rate need not produce more profit. MA disparity settings put the band at a percentage of price rather than a standard deviation, while backtesting without code shows how to rebuild the condition on your own pair.

Limits of this test

• One pair, USD/JPY, and two independent calendar years. Nothing here guarantees the same order on other pairs or regimes
• USD/JPY rose 1,632 pips in 2024 and finished 56 pips lower in 2025. That regime difference is the point of splitting years and also a limit on generalization
• Picking a winner from a period-and-sigma grid creates a multiple-comparison uplift. The other year checks it, but is not an untouched future sample
• Three sigma and wider settings have small samples. A 4σ mean of 2.3 trades cannot be compared with 260.4 trades at 1σ at equal confidence
• Touch rules use highs and lows; breakouts use closes. Counting wick breaks as signals would produce a different test
• Fills are at bar closes and spread is fixed. Live fills and changing spreads are not reproduced
• The fade exit stays at the opposite 1σ band to isolate entry sigma. A study that optimizes the exit as well answers a different question