Burst + volume continuation (Nanexa-style chain)

Thesis (PARTIALLY CONFIRMED — second-day variant): A single trading day with a large close-to-close gain (>= 10 / 15 / 20 %) on markedly elevated volume (>= 2x or 3x trailing 20d-average) is the first leg of a multi-day mark-up. Inspired by Nanexa.ST around Dec 2025 (not in our cached universe — validated on the US universe where the same setup is testable).

Distinct from: - explosive_moves.py (which looked at predictors of +30 %/week moves, not the trigger day itself). - accumulation_validate.py (slow 20-day volume z-score, not a single-day burst). - gap_patterns.py (open-vs-prior-close gap, not close-to-close).

Code: research/burst_continuation.py Report: research/burst_continuation_report.json (n=8,080 base events)

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

Strictness in return (raising X) lifts the 40-day edge from +1.4 % to +3.6 %; strictness in volume (raising Y) does the opposite. So the burst-day return carries the signal, not the volume multiple per se — once vol is at 2x avg, going to 3x removes information.

Strata — where the signal actually lives (base = X>=10 % & Y>=2x):

Three readings:

1. The base single-day burst is too weak to trade. +0.79 %/20d and +1.39 %/40d at hit-rate 0.54 is not strong enough to support position sizing after frictions. Raising X to 20 % helps (+3.58 %/40d at n=1,284) but n drops.

2. Stage-2 hurts, drawdown helps. Opposite of the coil-breakout finding. These bursts are reversion events, not trend continuation: they fire most usefully on names already drawn down (dd <= -15 %): +1.69 %/20d and +3.25 %/40d. On already- extended (dd > -5 %) names the edge is flat. The Stage-2 split confirms this — non-Stage-2 names (which is where deep drawdowns live) carry the edge.

3. The Nanexa-style chain is the real signal. Filtering to the 720 events where T+1 also closes >= +5 % on >= 1.5x volume produces +9.2 %/20d and +13.3 %/40d at hit-rate 0.70. The mirror cohort (no second burst, n=7,360) is flat — every drop of the base signal lives in the names that confirm with a second burst.

   Important entry-timing caveat: the second-burst flag is only known after T+1's close. So this is not a same-day setup — it's a day-2 confirmation setup. The +9 % / +13 % drifts are measured from close[T] forward, which means about half the move (+5 %/+1d) is the second-burst day itself; the remaining +4-8 % is genuine post-T+1 drift. Practical entry is at close[T+1] and the realisable drift is the post-T+1 component (see "second-burst (post-T+1)" cohort below).

4. Earnings coverage is partial. Many large-cap earnings.parquet files in the current checkout are empty, so the on_earnings split is unreliable for those names; with the names that DO have data, the on-vs-off split is small. Re-run after a full data refresh to confirm.

Verdict: Add the second-burst T+1 confirmation cohort as a short-term scanner candidate. The single-day burst alone is too thin. The two-day confirmation is among the strongest setups surfaced in this file — comparable to the death-cross + deep-drawdown long signal at similar n. Match the Nanexa-style chain the user described, validated quantitatively.

Recommended scanner cohort spec: - Trigger day T: close-to-close >= +10 % AND vol >= 2x 20d-avg. - Confirmation day T+1: close-to-close >= +5 % AND vol >= 1.5x 20d-avg. - Entry: close of T+1. - Holding: 20-40 days. - Sizing: small — the per-trade std is wide and the hit-rate is 0.70, not 0.85 — there will be obvious losers in the cohort.