Walk-forward on 112 names
Full output in results/. Primary configuration was pre-registered as
rolling/return before running, precisely so the anchored variant could not be
cherry-picked afterwards.
| rolling (primary) | anchored (check) | |
|---|---|---|
| walk-forward selection | +100.2% | +283.3% |
| average candidate | +114.5% | +117.6% |
| always mom-k20-20d | +287.0% | +298.0% |
| always mom-k20-60d | +282.3% | +303.5% |
| buy-and-hold SPY | +185.4% | +185.4% |
| perfect hindsight | +605.0% | +638.9% |
| selection hit rate | 36% | 71% |
| rank correlation | +0.11 | +0.39 |
| churn | 54% | 15% |
Primary result: selection failed again, harder. At 36% it did worse than a coin flip and worse than simply averaging every candidate. Committing to one momentum variant beat adaptive selection by 187 points. This is the second independent reproduction, now on a universe five times larger.
In 0 of 14 folds did selection pick the actual best candidate. In the last
three folds it had rotated into trend-monthly — which had won folds 8, 9 and
10 — exactly as momentum won the next three by +13.9%, +12.1% and +24.8%. Plain
performance chasing.
Anchored disagrees, and there is a mechanism. Training on all history rather than a rolling two years produces a more stable ranking (churn 15% vs 54%) and never chases into trend-monthly. That is a plausible hypothesis, not a result: p=0.09 on its own, ~0.17 across two configurations, and adopting it because it scored better is the cherry-pick this section exists to prevent.
The robust finding is about momentum, not about selecting. A momentum variant was the best candidate in 10 of 14 out-of-sample folds (p=0.004 against a 2-in-6 null), in both configurations, at both a 20-day and a 60-day lookback. Two different parameterisations landing within 5 points of each other (+287% / +282%) is weak evidence the effect is not a fitted artifact.
Caveat that walk-forward does not address: momentum turns over ~29x NAV/year and
ExecutionModel charges no spread on limit orders. At that turnover a 1-2bps
effective spread is roughly 0.5-0.9%/yr compounding against these numbers.