BTC vol clusters, but the half-life will not sit still

Verdict: WATCH (descriptive) · Asset/TF: BTCUSDT 1h (primary), 15m and 4h · Sample: chronological 60/20/20

This is the one effect that kept surviving every directional falsification: absolute returns cluster. The question was whether that persistence has a stable horizon you can size on.

What happened

Q1 — |r| ACF / half-life. Persistence is real and AC1-stable. 1h AC1 = 0.290 / 0.271 / 0.254 (discovery / validation / OOS); shuffle ≈ 0.002. Half-life is not stable: lag-to-half is 29h vs 5h (83% relative miss, fails the 25% gate). Discovery has a fat tail (ACF still ~0.15 at 48h); validation decays fast. Same fail on 15m (37 vs 11 bars) and 4h (exp 14.3 vs 3.9 bars).

Q2 — sign after |r| control. The prereg “direction-null |AC1(sign)| < 0.03” was not confirmed. 1h val AC1(sign) = −0.068; partial after residualizing on |r| is still −0.068. Small, significant, economically weak (~46.6% same-sign next). Same family as the thin vol-shock / streak / loc fades. Not an edge.

Q3 — exploratory risk. High rolling |r| does mark higher subsequent vol (~2×). It does not mark worse Sharpe (1h val top Sharpe 3.00 vs bottom 0.19 — a bull-regime artifact). One-bar trend-follow gated by tercile is dead; 5 bps RT annihilates it.

Trading

No directional claim. No stable sizing horizon. Do not treat “vol is high, therefore fade/size-down by N hours” as a calibrated rule on this sample.

Final Verdict

WATCH

Cannot pin a stable half-life for risk sizing, so not promoted. Not a clean NO_EDGE either: AC1 clustering is the surviving effect across this research log. Revisit with a regime-conditional decay, not a single exponential.