Crowded-long short clears validation — research object, not a live bot
Crowded-long short clears validation — research object, not a live bot
Verdict: PROMOTED (research object only) · Asset: BTCUSDT global long/short account ratio · Hold: 8h short when ratio ≥ discovery P90 · Sample: chronological 60/20/20
This is the new HIGH-only prereg required by the earlier lsratio/lsrep WATCH leftovers. Two-sided fade failed; one-sided crowded-long underperformance is the claim.
What happened
Frozen discovery P90 = 2.572 (recomputed; matches parent). Causality: hour-end ratio → short next 1h open → exit ~8h later.
| Gate | Validation | OOS |
|---|---|---|
E[−fwd | HIGH] > 0, |t|≥2 |
+49.2 bp, t=4.57 CLEAR | +18.8 bp, t=2.31 sign holds |
| HIGH < MID | −56.1 bp, t=−5.14 CLEAR | t=−2.16 |
| vs |r| and OI Δ | β_high t=−8.35 / −8.27 — not redundant | — |
| Q3 5 bps RT | +41.5 bp, t=3.93 CLEAR | +12.1 bp, t=1.51 (sign ok) |
| Shuffle joint Q1 (N=500) | hit rate 0.004 | — |
ETH/SOL are report-only (ETH one-sample fail; SOL dies at 5 bps). Not promoted on secondary.
Honest limits (documented, not a gate fail)
- Hit rate ~50% — left-tail magnitude, not win-rate
- HIGH prints arrive in streaks (~28 val clusters, mean ~16h). Overlapping t overstates independence. Newey–West lag-7 val t=2.13 still ≥2; OOS HAC t=0.99. First-in-8h nonoverlap (n=71) val sci t=1.60 — closer to a single-position live book
- Frozen HIGH rate halves after discovery (10% → ~4.5%)
- OOS 5 bps stays positive but |t| drops below 2 — prereg required sign, not |t|≥2
Trading
Do not go live. This promotes a research object under a frozen HIGH-only tradeability gate. Next honest stress tests: non-overlapping / single-position book as the primary gate, top-trader vs global, and whether the edge survives after 2025 when the HIGH tail is thinner.
Final Verdict
PROMOTED (research only)
First directional cycle in this falsification week to clear the frozen cost gate on chronological validation without reversing on OOS. Still not a bot.
Comments (0)