Define the concept before you test it
Mean Reversion Range is a backtesting idea where you look for situations in which price appears to be “too far” from a central tendency and might later move back toward it. A “range” is the measurable band that defines “too far” and “back in.” Examples of measurable definitions include a band around a moving average, a fixed-width channel, or a band based on dispersion.
Responsible backtesting starts by writing down the exact rule that turns market data into three states: (1) when the price is outside the band, (2) what “target” means inside the band, and (3) what counts as an exit. Without precise definitions, two people can run “the same” test and get incompatible results.
Specify inputs, assumptions, and costs explicitly
Backtests often fail because they assume away important frictions. To make results independently verifiable, list every assumption used in your calculation.
Common assumption categories:
- Data resolution and timestamping: Are you using tick data, minute bars, or daily bars? Are you using close-to-close, open-to-open, or intrabar estimates?
- Price used for decisions vs. execution: If the rule triggers at the end of a bar, you cannot also assume you executed at that same bar’s best price. You must define an execution time.
- Costs (kostensoorten): Include at least transaction costs relevant to your environment, such as spread, commissions, and fees. If you cannot obtain them, run sensitivity ranges instead of using a single optimistic number.
- Slippage and liquidity limits: If your band is tight, small execution differences can dominate outcomes. Model slippage as a parameter, not as an afterthought.
When you change any assumption, rerun the backtest and record how results move. This helps separate stable mechanics from variable market or provider conditions.
Control bias with disciplined testing (aannames and controlebron mindset)
A responsible backtest treats performance as a hypothesis, not a conclusion. Bias can be introduced by how you choose parameters, how you clean data, or how you select the time periods used to build the rule.
Practical bias controls:
- Parameter tuning separation: If you adjust band width or central tendency length, do it on a training period only. Do not reuse the same period to claim final performance.
- Walk-forward validation: Refit or re-choose parameters only at defined points in time, then test on the next segment. This reduces look-ahead bias.
- Multiple regimes: Test across different volatility and trend conditions. Mean reversion behavior can vary with regime; a rule that works in one regime can fail in another.
In this context, a “controlebron” is the validation segment you did not use for selection. If the rule only performs where it was tuned, the result is likely overfit.
Evaluate using out-of-sample checks and failure modes
Historical relationships do not establish future results. Therefore, the key question is not only “what was the return,” but “does the behavior persist when conditions change?”
Use out-of-sample checks:
- Holdout period: Reserve a segment never used for parameter selection.
- Rolling windows: Evaluate performance in successive windows to see whether it degrades.
- Robustness to dispersion changes: If your band is based on dispersion, try alternative dispersion measures and verify whether conclusions remain consistent.
Also report at least one material limitation or failure mode. Examples include:
- Regime shift: If price stops oscillating around the central tendency, “too far” may keep moving further.
- Band mismatch: If your band width is too narrow, executions can be dominated by noise and costs (kostensoorten).
- Bar-resolution artifacts: With coarse data, you may miss that the rule would have exited earlier using finer timestamps.
What to document so others can reproduce your reasoning
To verify the relevant facts independently, produce a concise “test specification” that includes:
- The exact band rule (definition of outside/inside, central tendency, and range width).
- The decision and execution timing rules.
- The cost model with parameter ranges for uncertain items.
- The dataset scope (dates, instrument universe, and data frequency).
- The splitting method for training vs. out-of-sample.
- Summary metrics and how they were computed.
Finally, treat results as evidence about the mechanics under your specific assumptions—not as a guarantee. Outcomes vary with market conditions, costs, execution, and other variable factors that a backtest may not fully capture.