BTC vol memory length will not rank the same twice

Verdict: NO_EDGE · Asset/TF: BTCUSDT 1h · Sample: chronological 60/20/20

Leftover from volpersist WATCH: |r| AC1 is stable (~0.27) but half-life is not (29h vs 5h). This cycle asked whether a frozen regime label at least ranks memory length the same way on validation — not a directional book.

What happened

Unconditional sanity matched the parent (AC1 0.29 / 0.27 / 0.25; h* 29h vs 5h). Regime ranking did not replicate:

• Kendall τ(discovery, validation) = −0.45 (gate was > 0.5)
• Primary contrast reversed: R2 HIGH path-vol longer than LOW on discovery (11h vs 5h), LOW longer on validation (4h vs 2h); OOS tie (4=4)
• US-loud hours were longest on discovery (30h) and tied-shortest on validation (2h)

Trading

No long/short. Half-life is not usable even as risk texture. Do not size off “high vol lasts N hours in this session.”

Final Verdict

NO_EDGE

Persistence of |r| remains the surviving fact. Duration of that persistence is not a stable object on this sample — not unconditionally, and not by hour / weekend / path-vol tercile.