MOO→11:00 Engine v1 — FINAL re-research verdict REJECTED

2026-04-27 · 3 параллельных агента · joint backtest + permutation tests + new hypothesis mining · real numbers

Layers C5+C6+C7+C8, заявленные как survived candidates после первого аудита, на полном joint backtest и permutation тесте показали: не альфа, а capacity reduction. Все 4 фильтра NET NEGATIVE на daily Sharpe (Engine v1 7.30 < baseline 8.16), permutation 0/5 PASS. Тот же класс провала что v5 layered stacks (Apr 26). Recommendation: deploy чистый organic_pump v1 baseline (gap≥10 × rvol≥3 SHORT MOO→11:00, TP+7%/SL-1%).

📊 Phase A — Joint backtest (real numbers)

Data: events_v2.parquet (234,798 PM events) + cache_intraday_3y_pre/ для 2023 OOS · Sized for 1% stop = $150 max-loss/trade

Baseline (#1 alone)

Sh 8.16

N=3,200 · $175.5K/yr · DD -$2.9K · WR 24.9%

Engine v1 (4 layers)

Sh 7.30

N=1,326 · $97.4K/yr · DD -$3.3K · WR 26.3%

N reduction

−58.5%

$/yr cut −44%, DD slightly worse

2023 OOS validated

Sh 4.43

Direction survives but $/yr -76% vs raw 2023

Per-layer ablation (drop one — Sh delta)

Layer dropped	N	Sh	Δ vs full Engine v1 (Sh 7.30)	Verdict
C5 (price < $10)	1,999	7.50	+0.20	HURTS Sh
C6 (pm_dvol ≥ $5M)	1,844	7.72	+0.42	WORST — biggest hit
C7 (skip vol_mult 5–10×)	1,373	7.43	+0.12	marginal hurt
C8 (exclude ETFs)	1,369	7.34	+0.04	~neutral

Yearly breakdown (Engine v1)

Year	N	Sh	PnL	WR	Note
2024	325	5.61	$37,269	24.9%	normal
2025	632	6.54	$81,771	24.2%	normal
2026 YTD	369	12.54	$73,134	31.2%	REGIME OUTLIER — 75% of $/yr

Phase A verdict: Engine v1 — NOT additive alpha. Filters cut N by 58.5%, cut $/yr by 44%, slightly INCREASE DD. Each layer individually hurts Sharpe (drop-one ablation: ALL 4 deltas positive when removed). Higher WR (26.3% vs 24.9%) is real but doesn't compensate for capacity loss. 2026 регимная аномалия — без неё цифры ещё хуже.

🎲 Phase B — Permutation tests (within-ticker shuffle)

Data: walker_trades.parquet (1,770 SHORT trades) · 1,000 iter · seed=42 · same methodology that killed v5 stacks (0/12 PASS Apr 26)

Layer	N+/N	Real Sh	Null Mean	Null P95	p-value	Verdict
C5 pm_open<$10	1202/1770	+0.169	+0.167	+0.176	0.340	FAIL
C6 pm_dvol≥$5M	1258/1770	+0.202	+0.199	+0.213	0.395	FAIL
C7 skip rvol 5–10×	107/1770	+0.049	+0.072	+0.184	0.345	FAIL (directional only)
C8 exclude ETFs	1572/1770	+0.183	+0.183	+0.183	1.000	DEGENERATE
COMBINED 4-layer	741/1770	+0.222	+0.219	+0.240	0.428	FAIL

Phase B verdict — 0/5 PASS:

C5/C6/C8: Real Sh sits at null mean. Layers add NO selection edge over a random within-ticker mask of equal size. Marginal Sh lift = pure selection-frequency artifact, not date-specific edge.
C7: "Noise zone" 5–10× IS directionally weaker (+0.049 vs null +0.072), но p=0.345 — не значимо. Need bigger N for power.
C8 degenerate: ETF flag inseparable per-ticker (each ticker is wholly ETF or wholly non-ETF) — within-ticker shuffle is a no-op. Untestable by this method.
Combined p=0.428 — joint stack indistinguishable from random selection.

Конечное: edge сидит в asymmetric payoff distribution shape (gap≥10 SHORT pump-fade), не в filter specificity. Same conclusion as v5 layered audit Apr 26.

🧠 Phase C — TOP 5 new hypotheses queue (orthogonal to C5–C8)

Mined from research_30_hypotheses_apr25.md + 4 other memory files · pre-screened for lookahead

#	Hypothesis	Source	LA	Est. effect	Data path
1	H1 Auction money pressure (rolling-60d)	30hyp H1 + 3aud A3 PASS	CLEAN*	+0.3–0.7 Sh	`imbalance_2y_full.parquet`
2	H6 Mega-cap leader divergence SHORT-reframe	30hyp H6	CLEAN	+0.5–1.0 Sh	`cache_intraday_v3_2y/`
3	H18 HY credit spread regime (HYG/LQD)	30hyp H18	CLEAN	+0.4–0.8 Sh conditional	`cache_index_1m/HYG.parquet`
4	H27 Unfilled gap-fill magnet (prior 10d)	30hyp H27	CLEAN	+0.3–0.5 Sh	`enrichment_data/calc_*`
5	H30 ATR-normalized sizing (.shift(1))	30hyp H30 + Stream D	CLEAN*	DD −20–30% → indirect Sh +0.3–0.5	`backfill_atr_rsi_5y.py`

Blacklist — DO NOT test

B1 H21 Analyst upgrade 24h — 0/3 audit PASS, snapshot LA, N=33–61
B2 Any Short Float variant — TIER RED 3× snapshot trap
B3 H4 Halt + EOQ — DIRTY both (same-day future-leak + same-quarter rank-includes-future)

🎯 Final recommendation

DEPLOY: чистый organic_pump v1 baseline
gap_pct ≥ +10% × rvol_pm ≥ 3 × SHORT MOO 09:30 → TP +7% / SL −1% / hard exit 11:00 ET
Backtest: Sh 8.16 daily, $175K/yr, MaxDD −$2.9K, WR 24.9% (2024–2026, N=3,200)
2023 OOS confirmed: Sh 4.55, $43.5K/yr, WR 23.5% (N=439)

DROP from Engine v1:

C5 (price < $10) — perm FAIL p=0.34, drop-one Δ+0.20 Sh
C6 (pm_dvol ≥ $5M) — perm FAIL p=0.40, drop-one Δ+0.42 Sh — biggest hit
C7 (skip 5–10× zone) — perm FAIL p=0.35, drop-one Δ+0.12 Sh
C8 (no ETFs) — degenerate test, drop-one ~neutral; keep as soft tilt only

Treat C5–C8 as capacity controls / soft tilts, NOT alpha layers. They reduce trade count without proven edge.

NEXT SPRINT: test TOP 5 hypotheses (Phase C list) on raw baseline, with mandatory permutation gate before adoption. Min N≥200/fold, daily Sh + bootstrap-by-dates, slippage stress 30 bps, |corr|<0.10 vs base.

📐 Phase 1 deployment spec ($15K BP)

Parameter	Value	Notes
BP slice	$15,000	1% max-loss = $150/trade
Entry	MOO 09:30 ET	strict auction routing required (TradingApp routing fix P0 — see memory)
Filter (base)	gap_pct ≥ +10 × rvol_pm ≥ 3	NO C5/C6/C7 stack
Direction	SHORT	asymmetric pump-fade
TP / SL	+7% / −1%	R:R 7:1, breakeven WR 12.5%
Exit	hard 11:00 ET	force-close all open positions
Concurrent	max 2 positions	per memory M14 deploy
Skip days	FOMC, SPY 5d≥+3%, earnings ±1d	per Tier 1 unified spec
Expected	Sh ~8 daily, $/yr ~$15–25K (scaled to $15K BP)	2023 OOS supports lower bound