What are Bollinger Bands, and what exactly are you testing?
Bollinger Bands are a volatility indicator built from two components: a middle line and upper/lower bands. The middle line is typically a moving average of price over a chosen lookback period. The upper and lower bands are set at a distance from the middle line based on the standard deviation of price over the same lookback window.
A responsible backtest does not test “the market.” It tests a specific, fully stated set of rules applied to historical data. That means you must specify (a) the price series used (for example, close vs. typical price), (b) the lookback length and any deviation multiplier, and (c) the decision rule that converts band behavior into an action—if you are evaluating an action at all. Without explicit rules and inputs, “backtesting” becomes an interpretation rather than a reproducible experiment.
Mechanism: make your assumptions explicit before you run any numbers
Start by writing down the mechanics of your calculation and evaluation pipeline.
- Indicator calculation inputs
- Window length: the number of periods used for the moving average and standard deviation.
- Standard deviation definition: whether you use the population or sample form is a choice; state it.
- Deviation multiplier: the factor that sets the band width.
- Price data: which field and timeframe the bands are computed from.
-
Backtest timing rules Indicator values depend on historical candles. A common failure mode is accidentally using information from the same bar you would claim to trade at. You need a clear rule for when the band value becomes “known” and when the action is evaluated. Specify whether actions are evaluated on the next bar’s open, close, or another timestamp.
-
Event handling and data hygiene Define how you handle missing bars, corporate-action-like discontinuities (if applicable to your data), and any interpolation. If your dataset has gaps, the band calculation and any rule triggers can change.
Costs and data realism: separate model fit from tradability
Even without giving any trading recommendations, a responsible backtest must account for costs and constraints, because they can dominate results.
- Costs as a cost model: include at least a simple estimate for transaction costs such as spread and commissions. If you ignore costs, you often bias outcomes upward.
- Execution realism: define whether you assume mid-price, bid/ask, or a conservative fill model. If your rules rely on band touches, the fill assumption matters.
- Slippage: in real markets, execution can deviate from the idealized price. Decide whether to include a fixed slippage assumption or a conservative cap.
Also separate indicator behavior from performance evaluation. Bollinger Bands can describe volatility, but your backtest is evaluating your complete pipeline: calculation, rule triggering, execution, and cost accounting.
Bias controls: how to avoid overfitting and look-ahead errors
Backtests often fail because they are “too clever” about the past. You can reduce that risk with bias controls.
-
Fix parameters before you test If you try many lookback lengths and deviation multipliers until results look good, you are selecting what happened to fit history. A responsible approach is to define parameter candidates in advance or reduce the degrees of freedom.
-
Out-of-sample validation Use at least two phases: one to set choices (even if only for rule clarity), and another to evaluate performance. A rolling or walk-forward scheme helps because it tests sensitivity to different market regimes.
-
Repeat across multiple non-overlapping periods Volatility indicators can behave differently across quiet and volatile periods. Evaluate the same rules across multiple time windows to reduce the chance that you are fitting one specific environment.
Evidence or example: what to measure without assuming future success
A backtest should produce metrics that reflect the full definition of your evaluation.
- Define your outcome: what does a “result” mean (for example, the return of a rule-defined action, or a forecast-like error if you are predicting a statistic).
- Use consistent measurement: window size, entry/exit timing, and cost inclusion must match your stated assumptions.
- Check stability: compare results across subperiods rather than relying on a single headline number.
If you only compute how often the price “touches” the bands, you are measuring indicator interaction, not necessarily an actionable edge.