OOS Sharpe at 87% of IS: Walk-Forward Results That Hold Up

The Walk-Forward Framework

S31 implements the full walk-forward validation that S24 configures. While S24 handles the fold structure and optimization, S31 executes the actual strategy in each OOS window and collects performance metrics. The separation exists because validation logic and execution logic should live in different modules.

Each OOS window runs the complete V7 pipeline: L1 signal generation, L2 gating, L3 exit management, and all S-series risk controls. Nothing is simplified or approximated. The OOS walk is a full simulation that mirrors production execution as closely as possible.

The 87% Transfer Rate

OOS Sharpe ratio at 87% of IS is a strong result. For context, academic literature considers 70% transfer acceptable and 80%+ good. The 87% suggests that V7's edge is robust and not heavily overfit to specific training data characteristics.

The consistency across folds matters more than the average. Individual fold transfer rates range from 78% to 94%. The lowest fold (78%) covers a period including a major correlation breakdown event that the model had not seen in training. The highest fold (94%) covers a trending period that closely matched training conditions.

Walk-Forward as Living Validation

Walk-forward validation is not a one-time exercise. Every time the model is considered for retraining, S31 runs fresh walk-forward tests on the updated model. The transfer rate must remain above 75% for the retrained model to be accepted. This creates a validation gate that prevents deploying models that test well in-sample but fail to generalize. Combined with PBO (S21) and Monte Carlo validation (S41), walk-forward provides the middle layer of a three-tier validation framework. PBO tests for structural overfitting. Walk-forward tests for temporal generalization. Monte Carlo tests for tail risk survival.