Walk-forward analysis for mean reversion strategies: special considerations
Mean reversion breaks when liquidity, spreads, and fat tails change. Use walk-forward windows that respect microstructure, watch turnover and capacity, and validate fills separately from signal logic.
Mean reversion strategies often look excellent in smooth backtests and fragile in live trading because their edge is concentrated at the hardest part of execution: the bid-ask boundary.
Walk-forward analysis must reflect that reality, not hide it.
Special consideration 1: costs dominate the edge
Small per-trade edge multiplied by high turnover means your WFA metric should prioritize net performance after fees and slippage stress, not gross signal accuracy.
Special consideration 2: window length interacts with half-life
Mean reversion signals can decay intraday or over weeks depending on the model. If your OOS window is too short, you measure noise. If too long, you blend regimes and obscure failure modes.
Use multiple window scales and compare stability, not a single "best" configuration.
Special consideration 3: liquidity regime shifts invalidate old fills
A pair that mean-reverts at tight spreads can become untradeable when spreads widen.
Add a liquidity filter that is fit only on IS data, then evaluated OOS without peeking.
Special consideration 4: fat tails break naive stop logic
Stop distances derived from normal-world intuition will be wrong. Validate tail behavior explicitly.
Special consideration 5: inventory and funding constraints (perps)
If you trade perpetuals, mean reversion often interacts with funding and basis dynamics. Your WFA must include funding in the PnL path for the same periods you trade live, or you will misclassify winners and losers.
Special consideration 6: capacity and queue position
Mean reversion often competes for the same liquidity. If your model assumes top-of-book fills but live trading walks the book, your WFA is optimistic.
Add at least one stressed fill model for OOS evaluation.
Special consideration 7: define what "failure" means before you look
Mean reversion strategies can look bad in trends while still being valid if your risk policy is regime-aware.
Predefine:
- max acceptable time underwater
- max acceptable drawdown for the strategy class
- minimum trades per window for any conclusion