Dormant awakening (sustained quiet + sudden spike)

Thesis (rejected unconditional; CONFIRMED only with day-2 confirmation): A stock with 60 days of sustained dormancy (annualised σ ≤ 0.30, range ≤ 15 % of midpoint) that suddenly closes up ≥ R on ≥ V × the dormant-window mean volume is institutional accumulation breaking surface — forward drift should follow.

Distinctness from existing: - coil_breakout requires a prior thrust (+50 %/30d). Here we require the OPPOSITE — sustained quiet. - volatility_squeeze (null) tested band-tightness alone, no trigger-day return / volume condition. - burst_continuation has no dormancy precondition. - accumulation_validate is a 20-day vol z-score above the 200DMA, not a single-day spike inside a quiet base.

Code: research/dormant_awakening.py Report: research/dormant_awakening_report.json (n=2,309 base events)

Controls are dormancy-matched: random days from the same name that themselves passed the 60-day dormancy gate but did NOT trigger. This is apples-to-apples and harder to beat than random-day controls.

Cell sweep (cumulative — events at the threshold or stricter):

Strictness does not monotonically improve the signal — the loosest base cohort sits at the top of the long-horizon edges. That's inconsistent with "this is a real cohort-specific edge" and more consistent with "we are picking up noise from a moderately positive prior."

Strata (base = X >= 5 % & Y >= 3x):

Two findings worth keeping:

1. Same day-2-confirmation pattern that defines burst. Among 2,309 dormant-awakening events, 216 (~9 %) have a T+1 follow- through (>= +3 % on >= 1.5x dormant-window vol). That sub-cohort drifts ~+6 %/20d, ~+6 %/40d, ~+6 %/60d at hit-rate 0.79. The mirror cohort (no follow-through, n=2,093) is FLAT across every horizon. As with burst_continuation, the signal is entirely in the names that confirm on the next day — the single-day trigger is noise on its own.

2. OOS 2020+ is negative. The base cohort flips sign post-2020 (20d -0.44 %, 40d -2.23 %, 60d -1.23 % at n=659). Pre-2020 the edge was modestly positive (+1.2 % avg 20d). Likely decay as dormant names became more algo-traded; the day-2-confirmation sub-cohort may have survived (n is too small in OOS to split further with confidence), but the unconditional signal has not.

Verdict: Reject the unconditional dormant-awakening cohort. The day-2-confirmation sub-cohort is a noisier version of the burst-chain already in the scanner (burst's +9 %/20d edge on n=720 vs this one's +6 %/20d on n=216), so productizing it adds n=216 borderline-similar events on top of n=720 cleaner ones — not worth the cohort surface.

Meta-finding: Day-2 follow-through is becoming THE selector. It carries the signal in: - burst_continuation: +9.2 % / +13.3 % 20d/40d edge (n=720). - dormant_awakening: +6.0 % / +6.0 % 20d/40d edge (n=216).

Both cohorts have flat-to-null forward drift in their "no second day" mirror sets. If we want a single meta-cohort, "any name that closed up materially on heavy volume AND followed through on day 2" looks like the productizable abstraction — distinct from the prior-day conditioning (dormancy / burst / coil base).