Backtest evidence

Is Bollinger period 20 optimal, and is there one bar count that works on every timeframe?

Bollinger Bands Period: Is 20 Really Best? 3,024 Settings Tested

Finding

There was no universal bar count. Averaging seven breakout deviations across 2024 and 2025, period 50 led the 15-minute chart, period 20 the hourly chart and period 5 the four-hour chart — different counts that represent similar lookbacks of 12.5 to 20 hours. On hourly bars period 20 had the highest two-year mean at +1,562.7 pips with all seven deviations profitable in both years, but its seven-deviation mean was −211.0 pips in the first half of 2025 and only two deviations were profitable in both halves. Lengthening the period widened the mean half-width at 1σ from 12.79 pips at period 5 to 88.68 at period 200 and cut annual breakout trades from 934 to 46, yet the share of closes inside the bands stayed between 44.9% and 50.3% rather than rising with the period. Period 5 at 2σ is impossible in this implementation and recorded no trades in all 12 timeframe-and-window cells.

Key results

Leading period by two-year mean
50 on M15, 20 on H1, 5 on H4 — 12.5 to 20 hours of elapsed time
Period 20 on hourly bars
all seven deviations profitable in both years, +1,562.7 pip two-year mean, but −211.0 pips in the first half of 2025
Period 5 at 2σ
arithmetically impossible; no trades in all 12 cells

Scope

  • USDJPY. Calendar 2025, calendar 2024, and the two halves of 2025
  • 15-minute, hourly and four-hour bars; 24,903 M15 bars in 2025, 6,226 H1 and 1,610 H4
  • 12 periods (5 to 200) x 7 deviations (0.5σ to 2σ) x 3 trading rules = 252 settings, on 3 timeframes and 4 windows: 3,024 runs
  • 0.3 pip spread, 0.1 lots, zero slippage

Method

Measured with Formiq's production backtester across three trading rules: long fade, short fade and close breakout. Fills use the signal bar's close. An opposite-side reversal may occur on the exit bar, while same-side re-entry waits for a later bar. No stop or target in the main grid.

Limitations

  • One pair and two calendar years.
  • The best period changed by timeframe and cannot be assumed to transfer to another pair or to daily bars.
  • Selecting from 12 periods and seven deviations creates multiple-comparison uplift. The opposite year is a check, not untouched future data.
  • Period 5 and 2σ is impossible under this population-standard-deviation implementation. A sample deviation, or a calculation that excludes the current close, changes the boundary.
  • Touches use highs and lows; breakouts use closes.
  • An open position is closed at the final bar of each window, so the two half-year results do not add exactly to the calendar-year result.