Glossary
Walk-Forward Optimization
Out-of-sample testing protocol that rolls optimization and validation forward in time. The institutional standard against single-period backtest overfitting.
Walk-forward optimization (WFO) is a rolling out-of-sample validation protocol. Parameters are optimized on an in-sample window, then tested untouched on the immediately following out-of-sample window. The window then advances and the process repeats.
Mechanics
- Split history into N segments.
- In-sample (IS): optimize parameters on segments 1..k.
- Out-of-sample (OOS): apply those parameters unchanged to segment k+1, record performance.
- Roll forward by m segments and repeat.
- Stitch all OOS segments into a continuous walk-forward equity curve.
Anchored vs rolling
- Anchored (expanding) WFO — IS window grows; OOS window advances. Preferred when long history is informative (slow regimes).
- Rolling WFO — IS window slides forward at fixed length. Preferred when older data is structurally different from current regime.
Why WFO beats single backtests
A single optimization run picks the parameter set that fits the entire history post hoc. WFO forces parameters to be selected on data the strategy didn't yet "know about." A WFO equity curve compounded across OOS-only segments is closer to what live performance looks like.
Where WFO fails
WFO is not immune to overfitting. If you re-run WFO with different IS/OOS ratios, search grids, or scoring metrics until a clean equity curve appears, you've overfit at a meta-level. See backtest overfitting.
WFO also can't detect regime change that lasts longer than an IS window — the optimizer adapts to the new regime as it appears in IS, masking the underlying break.