Monolith Research
The NAS pass was a test of memory. A strategy can look composed over the full sample and still fail the moment the sample is cut into time. Six rolling folds were used: one year of training, 120 days of testing, then a 120-day step forward. Costs were stressed. The result is less flattering than a full-sample chart, and therefore more useful.
The best risk-adjusted V3 candidate, recorded as optuna_v3_full_risk01, produced an average walk-forward Sharpe of 1.216, summed test return of 47.32%, worst fold return of -3.19%, and worst drawdown of 13.26% across 360 trades. It did not have the largest headline return in the table. It had the cleaner balance of return, fold stability, and drawdown.
The older notebook baseline still had teeth: under stressed costs it summed to 66.57% across the same walk-forward structure. But the penalty was a rougher risk shape. The job of this pass was not to flatter the old model; it was to find what could survive more mechanical handling.
For now the NAS logic remains a research branch with a dry-run boundary. The public record can say this much: the edge is not assumed from the full sample. It is made to answer fold by fold.