Low turnover: the cost problem is solved, the signal problem replaces it
Full surface in results/turnover_sweep_2026-09-09.txt. 12 configurations per
universe (lookback 60/120/250 × hold 21/63/126/252), costs applied through
costs.sensitivity() rather than by re-running, position-age rule disabled so
it cannot force closes inside a 252-session hold.
Costs stop mattering. Turnover falls from 22–30x to 0.6–3.6x. In small caps the spread drag collapses from 43 points at the weekly baseline to under 2 at an annual hold. The 14.27bps toll that killed the fast version is no longer the binding constraint.
But the surface is a random field. Excess over equal-weight ranges from −228 to +225 points with a standard deviation of ~129, and the largest swing from a single parameter step is 302 points (large cap, hold 126, lookback 120→250). A real effect produces a gradient. This does not.
Main effects are the only honest read, and they are weak:
| large | small | |
|---|---|---|
| lookback 60d | −18.8 | −159.3 |
| lookback 120d | −45.2 | −42.2 |
| lookback 250d | +109.8 | +0.8 |
| hold 21d | +151.4 | −16.2 |
| hold 252d | −97.8 | −97.7 |
Long formation with a short hold — the canonical 12-1 momentum specification — is the only direction that shows up in both universes, and it is a prior from the literature rather than a post-hoc pick, which is the one thing that makes it worth noting at all.
The standout cell is what selection looks like. look=250/hold=21 in large caps returned +225% over equal-weight, or +4.82%/yr. But it was chosen from 12 cells whose mean is +15.2 and whose sd is 129 — under a pure-noise null the expected maximum of 12 draws sits ~210 points above the mean. The observed 225 is almost exactly what noise produces when you sort it. And at 9.2 years, even an information ratio of 0.5 only reaches t=1.52.
Slowing down fixed the cost problem and revealed that there was never enough statistical power to see the signal underneath it.
Pre-registered test of the standout cell
look=250 / hold=21, single configuration, no sweep. Year-by-year excess over
equal-weight:
| large cap | small cap | |
|---|---|---|
| years positive | 5 / 9 | 4 / 9 |
| mean excess | +5.5%/yr | +3.9%/yr |
| t | 1.07 | 1.32 |
The large-cap +225% comes almost entirely from three years — 2020 (+26.3%), 2024 (+19.9%), 2026 (+24.9%). Six of nine years are flat or negative. That is the shape of a few lucky stretches, not an edge.
Both t-statistics are what a sweep's best cell looks like when you test it honestly. Confirmed noise.