Strategy Optimization

Principle

Optimization that maximizes in-sample performance is curve-fitting. The goal is robust generalization, not peak historical return.

Walk-forward

1. Split history into rolling windows: train (in-sample) → test (out-of-sample) → roll forward.
2. Optimize parameters only on each train window; record performance on the following test window.
3. Aggregate out-of-sample results — that is the honest estimate.

Anti-overfitting

• Parameter parsimony — fewer knobs = more robust. Reject params with no economic rationale.
• Plateau, not peak — prefer parameter regions where neighbors also perform well; a lone spike is noise.
• Out-of-sample gate — if OOS Sharpe < ~50% of IS Sharpe, the edge is likely fitted.
• Deflated Sharpe / multiple-testing correction when scanning many combos.
• Cross-asset / cross-regime check — does the same logic hold on a correlated pair or a different period?

Workflow

1. Define hypothesis (why should this edge exist?).
2. Coarse grid → identify plateau.
3. Walk-forward validate the plateau center.
4. Paper/forward-test before any capital.

Pairs with [[backtesting-discipline]] (bias-free measurement) and [[risk-management]] (sizing of the result).