Crowded-long short does not survive a single-position book

Verdict: WATCH · Asset: BTCUSDT global L/S ≥ P90, 8h short · Sample: chronological 60/20/20, non-overlapping entries

Stress test of the HIGH-only PROMOTED research object. Parent allowed overlapping 8h holds (~28 val clusters). Frozen rule: greedy first-available HIGH, 8h cooldown, per split. P90 2.572453 not recut.

What happened

Val (decides): HIGH_no n=71 (from 450 overlapping). E[−fwd|HIGH_no]=+43.5 bp, t=1.60 — mean still > 0, |t|<2. Welch vs MID t=−1.85 (fade sign yes, |t|<2). Dual Q1 FAIL.

Q3 val 5 bps: +35.9 bp, t=1.34 FAIL. Not scored NOT_TRADEABLE because Q1 failed first.

OOS: n=74, +16.3 bp, t=0.76 — sign holds, not OVERFIT.

Mean almost unchanged vs parent (~49 → ~44 bp). The |t| drop 4.57 → 1.60 is the independence correction. Cluster-first (n=28, val t=2.05) and HAC (t=2.48) were not switched in after seeing them; cluster-first OOS reverses.

Trading

A live single-position 8h book looks like this sample, not like 450 overlapping hourly prints. Do not treat the parent PROMOTED as a tradable occupancy. Keep it as a WATCH marker: the mean is still there, the independent bets are not.

Final Verdict

WATCH

Demote the overlapping HIGH-only object under single-position 8h occupancy. Research leftover, not a bot.