Bollinger Bands strategy backtest results

The short answer

Answers here: Does the Bollinger Bands strategy work? · What is its win rate? · 2 or 2.5 standard deviations? · Long only, or both ways?

What the script does

Bollinger Mean-Reversion Strategy, from the strategy library.

The numbers, run by run

Net and holding the symbol are after costs over the same period. Held-out is the verdict on the last third of the history judged alone: survived, decayed, did not survive, or too few trades to judge.

Year by year

20, 2.0 on SPY, long and short, the same rule long only, and SPY itself, each calendar year after costs. A partial first or last year counts the sessions it has.

Every trade

The 138 trades of 20, 2.0 on SPY, long and short, oldest first: side, entry and exit days, fill prices after slippage, and the return after both commissions. The other runs' trades open below.

The script

Pine v6, exactly as it ran for 20, 2.0 on SPY, long and short. The variants change only the inputs listed in the table.

//@version=6
strategy("Bollinger Mean-Reversion Strategy", overlay=true, default_qty_type=strategy.percent_of_equity, default_qty_value=100, initial_capital=100000, commission_type=strategy.commission.percent, commission_value=0.05, slippage=1, margin_long=0, margin_short=0)

src      = input.source(close, "Source", group="Setup")
len      = input.int(20, "Length", minval=2, group="Setup")
mult     = input.float(2, "StdDev multiplier", minval=0.5, step=0.1, group="Setup")
exitAt   = input.string("Basis", "Exit at", options=["Basis", "Opposite band"], group="Setup")
useLong  = input.bool(true, "Trade longs", group="Direction")
useShort = input.bool(true, "Trade shorts", group="Direction")
atrLen   = input.int(14, "ATR length", minval=1, group="Risk")
stopMult = input.float(2, "Stop (ATR multiples)", minval=0, step=0.5, group="Risk")

[basis, upper, lower] = ta.bb(src, len, mult)
atr = ta.atr(atrLen)

longSig   = ta.crossover(close, lower) and barstate.isconfirmed
shortSig  = ta.crossunder(close, upper) and barstate.isconfirmed
// both exits computed every bar, the setting picks one (ta.* never sits inside a branch)
xBasisUp  = ta.crossover(close, basis)
xBasisDn  = ta.crossunder(close, basis)
xUpperDn  = ta.crossunder(close, upper)
xLowerUp  = ta.crossover(close, lower)
exitLong  = (exitAt == "Basis" ? xBasisUp : xUpperDn) and barstate.isconfirmed
exitShort = (exitAt == "Basis" ? xBasisDn : xLowerUp) and barstate.isconfirmed

pU = plot(upper, "Upper band", color=color.new(#E5484D, 20), linewidth=1)
pL = plot(lower, "Lower band", color=color.new(#16A56A, 20), linewidth=1)
plot(basis, "Basis", color=color.new(#A3A39E, 0), linewidth=1)
fill(pU, pL, color=color.new(#A3A39E, 94), title="Band fill")

if longSig and useLong
    strategy.entry("Long", strategy.long)
    strategy.exit("Long stop", "Long", stop=stopMult > 0 ? close - stopMult * atr : na)
if shortSig and useShort
    strategy.entry("Short", strategy.short)
    strategy.exit("Short stop", "Short", stop=stopMult > 0 ? close + stopMult * atr : na)
if exitLong
    strategy.close("Long")
if exitShort
    strategy.close("Short")

How it was tested

Alpaca daily data since 2016, and 5-minute data from the last 600 sessions, 0.05% commission and 0.025% slippage on every fill, the last third of the history held out, textbook settings and nothing tuned. The engine, fills and what it does not model are written out once, in data and methodology.

Questions