v3 pump-fade — research cockpit

SHORT-only premarket→MOO 09:30. Lookahead-clean, 1-min. OOS 2026-03-31…05-22 (37d, 3 483 arms) · 6m 2025-11-24…05-22 (123d, 6 304 arms). Updated 2026-05-31. v3 = research scorer (not live; live=v2.5)

Bottom line — what to ship

branchdecisionevidence (OOS / 6m)
Entry timingR0 — short immediately at signalSHIPSh 14.85/10.77; waiting hurts every window/grade (p=0)
Tail-stopA-grade only, −8% buy-stop; A+ noneSHIPA Sh +2.7/+0.9; A+ stop hurts mean p<.001
Entry windowDown-weight 08:15; favour 08:00 & 09:10SHIP08:15 Sh ~1.2 vs 08:00/09:10 ~13
QQQ hedgeDrop it (or keep light for DD)SHIPunhedged +0.6..1.2 Sh; v3 is beta-free
ExitHold to MOO auctionSHIPcovering early strictly worse, monotonic
vwap_ext_revDiagnostic only — NOT a ruleWATCHreplacement Sh is non-causal; edge on A/micro
Scaling/avg-upNoAVOIDvext-add dilutes Sharpe 13.3/9.6 < R0
Raise vol-gateNo — keep 50KAVOID100K→300K loses PnL+Sharpe, no edge gain
EXTREME flagCan removeAVOIDdormant lottery 0.1% trades, 0% PnL, fat tail
ext_at_S gateNoAVOIDentry-time stretch cannot time wait-vs-now

0 · Equity & daily PnL (OOS · R0 · $500×grade)

1 · Entry timing — short immediately (R0)

ruleWR%mean%Sharpe
R0 immediate (=v3) SHIP62.7+0.42014.85
R1 first red57.0+0.24610.74
R2 two red55.9+0.2088.88
R3 lower-high57.5+0.26211.67
R4 peak (oracle) ORACLE80.4+1.42519.75
Waiting for a reversal (R1/R2/R3) loses on every window & grade (sign-flip p=0). The prior "reversal wins" was a 5-min-bar lookahead. R4 = oracle ceiling.

2 · Tail-stop — A-grade only, −8%

blanket stopWR%mean%Sharpeworst%MaxDD$
NONE62.6+0.41714.85-54.1-637
-2%57.6+0.31614.18-12.1-629
-3%60.1+0.34614.23-12.1-651
-5%61.6+0.37514.85-12.1-637
-8%62.2+0.38613.85-12.1-637
-10%62.3+0.38413.91-12.1-637
-20%62.6+0.39014.83-20.4-637
Blanket stop ≈ Sharpe-neutral (confirms H6). The edge is grade-conditional:
A gradeSh OOSSh 6mmean OOSmean 6mworst%
no-stop8.195.34+0.305+0.270-54.1
-5%11.196.40+0.370+0.306-12.1
-8%10.926.23+0.372+0.306-12.1
A+ gradeSh OOSSh 6m
no-stop13.7610.52
-5%13.5710.66
-8%12.4810.03
A grade −8%: Sharpe +0.9 (6m) … +2.7 (OOS), mean unchanged, worst −54%→−12%. A+: stop hurts mean (p<.001 both) → no stop. Variance/tail fix, not a mean edge. Helper: v3_tailstop.py

3 · vwap_ext_rev — diagnostic, not a rule

where R0 bleedsR0 OOSvext OOSedgeR0 6mvext 6medge
A grade-1.84+0.66+2.49-1.26+0.81+2.07
A+ grade-0.03+0.82+0.85-0.13+0.67+0.79
micro<300M-0.37+1.90+2.27-0.37+1.90+2.27
small-1.05+1.21+2.27-1.00+1.21+2.21
large>10B+0.10+0.83+0.73-0.03+0.66+0.69
"Replacement" Sharpe (16.8/11.6) is non-causal (needs to know ex-ante which names stretch). Executable forms don't beat R0: pure-wait 9.4/6.3, add/scale 13.3/9.6. Use vwap_ext_rev only to locate the bleed (A-grade + micro/small, robust both periods) → motivates the A-tail-stop. Causal next: stop-then-reenter-on-reject.

4 · Siblings

VOL-GATE — don't raise (keep 50K)
floorN OOSkept%Sh OOSN 6mSh 6m
50K3483100%14.85630410.77
75K288883%13.82522310.20
100K252572%13.1645549.37
150K203958%12.5336538.35
200K172650%12.1730737.65
300K130337%9.6422866.35
ENTRY-WINDOW — 08:00 & 09:10 money; 08:15 dead
windowN OOSmean OOSSh OOSN 6mmean 6mSh 6m
08:00 BEST807+0.79012.951619+0.84411.31
08:15 DEAD509+0.0871.18868+0.1291.69
08:30531+0.3027.191094+0.1693.54
08:45594+0.3436.64847+0.3064.52
09:00411+0.3146.03682+0.2704.54
09:10 BEST299+0.55913.78565+0.4008.41
09:20332+0.3276.47629+0.1292.15
QQQ HEDGE — drop it (hurts Sharpe)
Sh OOSDD OOSSh 6mDD 6m
hedged QQQ 1:114.85-63710.77-637
unhedged (raw)16.02-64011.41-921

5 · Exit-side — hold to MOO

exitSh OOSSh 6m
09:155.285.99
09:205.416.34
09:2510.298.51
09:2912.369.51
MOO 09:30 BEST14.8510.77
Covering before the open is strictly worse on both periods (monotonic). The fade fully realizes at the 09:30 auction.

Open / next

① stop-then-reenter on vwap_ext_rev reject (causal way to monetize the diagnostic) · ② combine ship-items (R0 + drop-hedge + skip-08:15 + A-tailstop) into one config & re-measure · ③ port/ship gate when v3 goes live (currently research-only).
Generated by build_v3_dashboard_2026_05_31.py · all numbers lookahead-clean, hedged unless noted