Study 2 · Volatility Dead-Zone Sizing Overlay

Pre-registered (docs/PREREGISTRATION_STUDY2.md) · 1 lever · 2 trials · conservative costs · 2026-06-12 23:33 UTC

✅ IMPROVEMENT (candidate for forward test)

YES — 'halfsize' improves IS MAR (2.706->2.926) within guardrails and does not collapse OOS (3.108->3.158). Candidate for a forward test. NOTE: OOS is confirmatory-on-burned-data, not virgin OOS.

The hypothesis

Study 1 found mid-vol-band entries (ATR percentile-rank 0.33–0.66) were the only statistically-significant dead zone (OOS mean-R −0.10 vs +0.54). This overlay half-sizes those entries. V5 entries/exits are otherwise untouched — baseline reproduces bit-exact (gate: passed, 86/1.647 on 15m IS BTC).

In-sample (10 assets, halving-split) — selection

config	trades	win%	PF	MDD%	MAR	t
V5 baseline	600	62.67%	1.825	-11.44%	2.706	5.11
skip mid-vol	573	61.61%	1.794	-9.55%	2.879	4.89
halfsize mid-vol ← chosen	601	62.56%	1.883	-10.08%	2.926	5.34

Both modes improved MAR within the MDD guardrail; halfsize won on MAR (2.926 vs 2.879) and uniquely raised PF (1.825→1.883). Selection used IS only.

Out-of-sample (10 assets) — confirmatory

config	trades	win%	PF	MDD%	MAR	t
V5 baseline	315	61.27%	1.825	-8.07%	3.108	3.79
halfsize mid-vol	315	61.27%	1.888	-7.97%	3.158	3.95

⚠ Honest caveat: the dead-zone band was identified using OOS data in Study 1, so this OOS is confirmatory-on-burned-data, NOT virgin out-of-sample. It shows the overlay doesn't break OOS — it cannot, by itself, prove the edge.

Bottom line

The overlay improves every OOS metric — PF 1.825→1.888, MDD −8.07%→−7.97%, MAR 3.108→3.158, t 3.79→3.95 — by half-weighting a population of near-zero-expectancy trades. The magnitude is small (+1.6% OOS MAR). It is the first lever across both studies to clear its pre-registered bar, but given the burned-data caveat and modest size, the correct next step is a forward (paper/live-shadow) test, not an immediate change to the live bot.