MOO→11:00 Verdict Report — что работает, что нет

Полный audit 14 setups + 3 stacks · Apr 27, 2026 · критерии: lookahead bias / permutation test fail = DROP · 4 параллельных аудитора

Master verdict table

#NameVerdictDirHonest Sh$/yrLookaheadPerm testReason
1organic_pump v1
sub5_pump M14 LIVE
KEEPSHORT5–7$10–12KPASSPASSAnchor: pre-period 2023 PF 2.17. Только setup с 4y validation.
2organic_pump v2.5
A6+A7+B universe
MERGE→#1SHORT?+$8–13K incrementalPARTIALNOT-DONEA1 (exit 10:00) DROPPED per 11:00 directive. A6/A7 — same class что v5 (0/12 perm). Universe expansion (B) safe.
3g3a SHORT
ATR-stretch ≥+2 ATR
KEEP*SHORT2–4~$13KPASSNOT-DONE* Gate: run perm test + isolated 2023 PF before live. Single-feature → lower v5-class risk.
4v3 hedge stackSHORT=#1; LONG paperL+S5–7PASSSHORT yes; LONG noSHORT leg = duplicate #1. LONG leg paper-only до 2023 PF rebuild + isolated audit.
5org_dump
plugin LONG bounce
RESEARCHLONG??PASS code, dirty depsNOT-DONEPlugin engine never ran (pm_history parquet not built). Direction orthogonal к #1, но deps на снапшот flags (is_spac, float).
6reports_pump
plugin SHORT news
DROPSHORTFAILv5 family failedSI/float snapshot trap (recurring). v3 уже отверг earnings overlay (Δmax +0.055). Дублирует #1 catalyst.
7reports_dump
plugin LONG news
DROPLONGFAILNOT-DONERSI14 lookahead recurrence risk. Earnings overlay — DEAD per v3 audit.
8v5 layered stacksDROPL+SFAIL (7 dirty)FAIL 12/12Within-ticker shuffle preserves Sh ~14 — edge в payoff distribution shape, не layer specificity.
9first5_directionalDROPL+SPASSNOT-DONESpec violation: entry @ 09:35 (требует first 5min observation). Не MOO 09:30.
10BREAK_UP proximityKEEPLONG5–7$3–6KPASSM11 stack OKSh 14.85 reported per-trade — реальный daily 5–7 (deflated 2.11×). LIVE в M11, прошёл 5audits.
11Q sector laggardDROPLONG−$11.5K!PASSFAIL (LOO −2.69)5audits LOO: removing Q ULUCHILA strategy на Sh +2.69. "Edge" был interaction artifact в M10R basket.
12P twistDROPLONG~$0PASSFAIL (Δ≈0)5audits LOO: removal Δ≈0. Indistinguishable от random в M11.
13S RSI continuationDROPLONGFAILcontaminatedRSI14 lookahead — same class что M13 v5/v6 BLOCKED. Sh 6.70 measured ДО .shift(1) cache fix Apr 26.
14TP-only no-SL overlayOVERLAY→#1/#3all+50% Sh+30–50% multPASSgrid sweepНе standalone, мета-overlay. Drop hard −1% SL, keep TP +7%, force exit 11:00.
— Stack combinations —
S1v2.5 SHORT × BREAK_UP LONGRESEARCH-MOREL+S5–6.5$22–28KA: pass; B: passNOT-DONE on comboOrthogonality HIGH (~3–5% trade overlap). Gate: backtest BREAK_UP @ 11:00 exit (currently только 09:55 validated).
S2g3a × first5_directionalDROPComponent B (#9) spec-incompatible. Не лезет в MOO 09:30 family.
S3v3 hedge × Q laggard boostDROPComponent B (#11) — REMOVED по 5audits. Q laggard hurts strategy −$11.5K/yr.

Charts

Engine v1 — финальная архитектура

✅ KEPT (deploy queue)

#1 organic_pump v1 (LIVE) + #3 g3a SHORT (после gates) + #10 BREAK_UP LONG (паpa с #1 micro-hedge) + #14 TP-only overlay (на all SHORT)

Engine v1 expected: Sh 5–7 daily · $25–35K/yr Phase 1 ($15K BP) · MaxDD est −$3–5K · 3 setups + 1 overlay

🟡 RESEARCH backlog (paper-only)

#4 v3 LONG leg (gap_-10..-5 × pmd_-5..-2)

Gate:Build pre-period 2023 cache c pmd feature → isolated PF ≥1.5

#5 org_dump (LONG bounce)

Gate:Build pm_history parquet → run audit driver → перм тест

S1 stack (v2.5 SHORT × BREAK_UP LONG)

Gate:Backtest BREAK_UP @ 11:00 exit + perm на combined trade pool

❌ DROPPED (9 setups+stacks)

#6 reports_pump #7 reports_dump #8 v5 layered
#9 first5 #11 Q laggard #12 P twist
#13 S RSI S2 g3a×first5 S3 v3×Q lag

Reason buckets:
4 × lookahead/snapshot trap (#6, #7, #13, S3 by ricochet)
2 × spec violation 09:30 entry (#9, S2)
2 × LOO removed/noise (#11, #12)
1 × perm test 12/12 fail (#8)

Key insights — что мы научились

Lookahead trap — recurring class. 4 dropped из-за того же класса bugs: snapshot fields (SI, float, sector) или RSI14/ATR14 без правильного .shift(1) per-ticker. M13 v5/v6 уже забанены навсегда. Любой новый setup с RSI14 / SI / earnings AMC = красный флаг до прохождения cache audit.
Per-trade Sh vs daily Sh — 2–3× inflation. #10 BREAK_UP "Sh 14.85" → реальный daily 5–7. #11 Q laggard standalone "Sh 7.23" → daily 3.0–3.5, и LOO showed contribution NEGATIVE −2.69. Per-trade Sh без daily aggregation = vanity metric.
Plugin engine ≠ deploy-ready edges. 4 plugins (organic/reports × pump/dump) — хорошая архитектура, но 3/4 plugins использовали те же снапшот features. #5 (org_dump) единственный пока годный к paper, и то — нужно сначала pm_history parquet build.
Что survived все аудиты — 1 setup. Только #1 organic_pump v1 имеет: (a) clean lookahead, (b) pre-period 2023 PF 2.17, (c) live broker hookup, (d) M14 risk gates. Engine v1 строится поверх #1. #3 g3a и #10 BREAK_UP — близкие, но требуют дополнительных gates.
Ortho-pair found. S1 (v2.5×BREAK_UP) — единственная stack которая прошла верификацию на компонент level. ~3-5% trade overlap, разные direction, разные универс. После gates готова к paper deployment.

Action plan

StepTaskAffectsEffortPriority
1P0 fix: TradingApp MOO routing (session=OPG не EXT)ALL1d manualP0 BLOCKER
2Run perm test (within-ticker shuffle) на g3a #3#3~2hP1
3Build pre-period 2023 isolated PF for #3, #4-LONG#3, #4~3hP1
4Build pm_history parquet → run org_dump audit#5~6hP2
5Backtest BREAK_UP @ 11:00 exit (vs 09:55)#10, S1~2hP1
6Apply TP-only no-SL overlay #14 на #1 → A/B vs current#1+#14~2hP1
7Build cache for ADV 500K+ universe (3y)v2.5 deriv~6hP2
8Code: remove dropped setups из scorer (cleanup)codebase~1hP3
Total ~22h. P0 routing fix — gating всё. P1 (steps 2-3, 5-6) ~9h → unlocks #3 + #10@1100 + overlay test → Engine v1 deploy ready.

Каноническая методология (что использовали аудиторы)

DROP criteria применённые

  1. Lookahead bias — snapshot fields (SI/float/sector), missing .shift(1), feature timing future-leak
  2. Permutation test fail — within-ticker shuffle preserves Sh → edge artifact, не signal

Дополнительные signals (info only)

  1. LOO Δ Sh negative — feature actively HARMFUL
  2. Pre-period 2023 PF <1.5 — 2024-born edge trap
  3. Per-trade Sh 2-3× daily Sh — adjust honest projection down
  4. Spec violation (entry timing != 09:30) — exclude from family

Что НЕ применяли (но надо)

  • Bootstrap CI (most setups) — only memory references
  • Walk-forward folds for new setups
  • Slippage stress test (10–50bps)
  • Capacity / market impact analysis
  • Cross-correlation matrix между всеми kept setups

Все эти можно проверять только когда P0 routing fix + pre-period cache extension сделано.

Sources