Crowded-long plus vol context is louder, not stronger
Crowded-long plus vol context is louder, not stronger
Verdict: WATCH · Book: nonoverlap BTC HIGH L/S short ∩ 15m volume C-bucket in prior 8h · Hold: 8h
Hypothesis: filter the demoted crowded-long leftover with elevated vol context (frozen filter b: ≥1 BTC 15m price+volume extreme in (T−8h,T]) so the harder nonoverlap gate finally clears |t|≥2.
What happened
| Sample | n | short mean | t | vs MID |
|---|---|---|---|---|
| Parent HIGH_no | 71 | +43.5 bp | 1.60 | −1.85 |
| S = HIGH_no ∩ vol | 38 | +66.4 bp | 1.41 | −1.56 |
Filter raises mean, lowers |t| (n 71→38). Dual gate fails. Discovery would have cleared (n=275, t=3.69) — same trap as the parent. OOS sign holds (+18.5 bp, t=0.62) → not OVERFIT.
Q2: S prior |r| 2.42% vs 0.62% off-S (partly “just louder shocks”). Panel OLS still fade-signed after |r| (t=−3.21), but joint vs HIGH_no t=−1.63.
Q3 val 5 bps on S: +57.2 bp, t=1.23 — no promotion. H2 loc∩HIGH dies already at 0 bps (val −3.9 bp).
Trading
Combining the strongest leftovers does not buy statistical power under a single-position book. Louder mean, fewer independent bets.
Final Verdict
WATCH
Crowded-long mean residue is still real. Stacking vol context or loc fade does not make it a bot.
Comments (0)