MOO-955 EXIT 11:00 — Full Re-Cross-Check

Apr 26, 2026 · Сверка всех 12 факторов на baseline 9362 trades 2023-2026 (deployed M11_APR25 за вчера)

🚨 КРИТИЧЕСКАЯ НАХОДКА — M11 DEPLOYED OVER-RESTRICTIVE

Re-cross-check показал что deployed stack B+C1M+D4+F+noRsi30+noS12+noAVOID СОКРАЩАЕТ capacity на 91% (N 9362→845) ради marginal Sh trade-off, при этом BACKTEST Sh ниже базового (10.33 vs 14.17 для LIGHT-варианта).

Hard data: M11 2026 Sh = 3.50 ($2,065/yr), LIGHT 2026 Sh = 11.55 ($17,769/yr). LIGHT_LONG_only 2026 Sh = 20.70.

Recommendation: Откатить избыточные фильтры C+D, оставить только F (no_macro) + noRsi30 LONG forbid + noBREAK_DN_LONG. Это и есть suboptimization из Apr 25 sprint — стек был fit-on-train для inflated Sh не daily-aggregated.

📊 TL;DR — Joint Sharpe at EXIT 11:00 (daily-aggregated)

M11 Deployed (yesterday)

Sh 10.33

N=845 · $22.5K/yr · 2026 Sh 3.50 ⚠️

LIGHT (only no_macro)

Sh 14.17

N=7926 · $160K/yr · 2026 Sh 11.55 ✓

LIGHT_LONG_only

Sh 13.45

N=4096 · $117K/yr · 2026 Sh 20.70 🏆

RAW Baseline

Sh 13.92

N=9362 · $191K/yr · 2026 Sh 11.91

1. Per-Factor Effects (ALL_3y, sorted by Sharpe)

Factor	N	WR%	Mean%	Sharpe (daily ann)	$/yr@$5K	Verdict
F (no macro day)	7755	69.9	+1.218	14.14	$157,418	KEEP
F skip FOMC/PPI only	8675	70.0	+1.249	13.96	$180,531	KEEP (lighter)
noRsi<30 LONG	9320	69.6	+1.224	13.94	$190,160	KEEP
BASELINE (no filter)	9362	69.6	+1.227	13.92	$191,417	—
B drop BREAK_DN_LONG	9326	69.6	+1.224	13.87	$190,270	KEEP (free)
E no_earnings (drop)	9091	69.7	+1.213	13.74	$183,792	KEEP
LONG only	4716	77.7	+1.792	13.35	$140,872	STRONG
D drop sector_worst (rank≤10)	5500	67.8	+1.225	12.54	$112,259	KEEP (gentle)
B BREAK_UP LONG	148	82.4	+1.575	12.13	$3,884	SCORE BOOST
BREAK_UP_B (all)	331	82.5	+1.463	11.43	$8,071	BOOST OK
C dvol≥1M	3339	66.0	+1.257	10.41	$69,966	DROP — −Sh, −2/3 N
D sector_top4	2256	70.4	+1.392	10.40	$52,334	DROP — −Sh
D sector_top3	1707	70.7	+1.430	10.13	$40,692	DROP
P twist top25 LONG	993	74.4	+1.800	8.91	$29,782	SCORE BOOST
SHORT_S1 only	2423	55.5	+0.378	6.50	$15,256	marginal
SHORT_S12 only	99	56.6	+0.501	3.52	$826	DROP (deployed)
B BREAK_DN SHORT	174	58.6	+0.529	3.44	$1,535	weak

2. Per-Year Stability — M11 vs LIGHT vs LIGHT_LONG

Variant	2023	2024	2025	2026	Last_6m	ALL_3y
RAW baseline (no filter)	Sh 16.83 N=2562	Sh 12.99 N=1608	Sh 13.18 N=3598	Sh 11.91 N=1594	Sh 12.88 N=2745	Sh 13.92 N=9362
LIGHT (no macro)	Sh 16.83 N=2562	Sh 13.26 N=1393	Sh 13.18 N=2751	Sh 11.55 N=1220	Sh 13.29 N=2043	Sh 14.17 N=7926
🏆 LIGHT_LONG_only	Sh 13.68 N=2165	Sh 16.72 N=679	Sh 11.65 N=1008	Sh 20.70 N=244	Sh 13.22 N=507	Sh 13.45 N=4096
LIGHT_SHORT_only	Sh 13.25 N=397	Sh 8.40 N=714	Sh 9.59 N=1743	Sh 8.27 N=976	Sh 9.76 N=1536	Sh 9.82 N=3830
⚠️ M11 DEPLOYED (B+C1M+D4+F+...)	N=0	Sh 9.97 N=169	Sh 12.54 N=531	Sh 3.50 N=183	Sh 6.60 N=304	Sh 10.33 N=845

📉 M11 vs LIGHT 2026 разрыв: M11 в 2026 даёт Sh 3.50 и $2,065/yr против LIGHT'а Sh 11.55 и $17,769/yr — 8.6× меньше прибыли при 3.3× худшем Sharpe. Фильтры C/D уменьшили N в 6.7× а Sh упал в 3.3×. Это over-fit.

3. Pairwise Interaction Matrix (Sh)

Sharpe joint = mask1 ∩ mask2. Diagonal = solo factor. Большая часть пар: Sh ≈ solo (фильтры orthogonal, можно складывать без cost). Но C1M/D4 везде режут Sh.

	B	C1M	D4	E	F	Q	P	noRsi30	noS12	noAVOID
B	13.87	10.38	10.28	13.69	14.10	9.85	10.78	13.89	13.83	13.87
C1M	10.38	10.41	10.47	10.24	10.43	4.51	7.32	10.45	10.06	10.41
D4	10.28	10.47	10.40	10.21	10.54	6.51	8.95	10.38	10.48	10.40
E	13.69	10.24	10.21	13.74	14.01	9.73	10.71	13.76	13.70	13.74
F	14.10	10.43	10.54	14.01	14.14	9.69	10.80	14.17	14.09	14.14
Q	9.85	4.51	6.51	9.73	9.69	9.85	9.85	9.85	9.83	9.85
P	10.78	7.32	8.95	10.71	10.80	9.85	10.83	10.86	10.79	10.83
noRsi30	13.89	10.45	10.38	13.76	14.17	9.85	10.86	13.94	13.90	13.94
noS12	13.83	10.06	10.48	13.70	14.09	9.83	10.79	13.90	13.88	13.88
noAVOID	13.87	10.41	10.40	13.74	14.14	9.85	10.83	13.94	13.88	13.92

📌 Pattern: F + noRsi30 ячейка 14.17 = глобальный максимум. C1M/D4 строки и столбцы везде <11 — они системно режут Sh при любых партнёрах.

4. Greedy Stack Search (start = baseline, add factor that maximizes Sh)

Step	Add	Selected	N	Sh	$/yr	Note
0	BASELINE	—	9362	13.92	$191,417	start
1	F	F	7755	14.14	$157,418	+0.22 Sh
2	noRsi30	F+noRsi30	7724	14.17	$156,443	+0.03 Sh ⭐ OPTIMAL
3	noAVOID	F+noRsi30+noAVOID	7724	14.17	$156,443	0 effect (no rows)
4	B	+B	7697	14.12	$155,514	−0.05 Sh
5	noS12	+noS12	7619	14.07	$154,896	−0.05 Sh
6	E	+E	7431	13.94	$150,312	−0.13 Sh
7	P	+P	4189	10.66	$58,860	−3.28 Sh ⚠️
8	Q	+Q	3592	9.58	$40,163	−1.08 Sh
9	D4	+D4	880	6.65	$10,601	−2.93 Sh ⚠️ capacity collapse
10	C1M	+C1M (full M11+P+Q)	294	5.62	$2,929	−1.03 Sh

✓ Greedy optimum: Step 2 — F + noRsi30 — Sh 14.17, N=7724, $156K/yr at $5K basket. Каждый последующий фильтр уменьшает Sharpe. P/Q/D4/C1M катастрофически режут capacity.

5. M11 Deployed — Compare to Optimal

Stack	N	WR%	Mean%	Sh	$/yr	2026 Sh	Verdict
OPTIMAL: F+noRsi30 (LIGHT)	7724	69.7	+1.214	14.17	$156,443	11.55	🏆 keep simple
LIGHT_LONG_only	4096	77.4	+1.717	13.45	$117,203	20.70	🏆 strengthening 2026
M11_APR25 deployed	845	72.5	+1.598	10.33	$22,505	3.50	⚠️ over-restricted
M11 + P boost	563	67.1	—	8.46	$10,910	—	worse
M11 + Q boost	295	59.0	—	5.62	$2,924	—	much worse

6. Hedge Comparison (NO_HEDGE vs SPY vs QQQ) at EXIT 11:00

Hedge	N	WR%	Sh (daily)	$/yr@$5K	Note
NO_HEDGE	6217	65.7	12.24	($568M raw)	highest gross
QQQ_1to1 (deployed)	6217	64.7	12.10	($538M raw)	good — risk reduce
SPY_1to1 (was)	6217	64.6	12.15	($533M raw)	—

📌 QQQ vs SPY разница ≈ 0.05 Sh = noise. Smart Apr 25 рекомендация QQQ остаётся корректной (не вредит). NO_HEDGE даёт +0.14 Sh но больше MaxDD — trade-off оставлен.

7. Recommendation — Откат M11 → LIGHT

🎯 Action items для Apr 27:

strategies/moo_955/m10r_rules.py — закомментировать apply_apr25_layer в части C+D (dvol/sector), оставить только B+F+noRsi30+noAVOID hard-gates.
strategies/moo_955/config.json — выключить apr25.boosts.dvol_filter и apr25.boosts.sector_filter (или поднять threshold до 0).
Оставить: F (skip FOMC/PPI), noRsi30 LONG forbid, noBREAK_DN_LONG, BREAK_UP boosts (S+/S/A/B), Q/P как score boosts (не gates).
QQQ hedge оставить — не вредит.
Размер basket: вернуть к ~30-40 позициям/день вместо ~5 (capacity).

⚠️ Риски downscope M11:

Apr 25 sprint research отчитал Sh 7.28 для full stack — это per-trade aggregation (raw mean/std). Daily-aggregated показывает ~10-14 на различных stack — методология different. Стратегические рекомендации Apr 25 (BREAK_UP S+/S, Q laggard, P twist) валидны как score boosts, не как hard gates.
2026 sample малый (N=183 для M11) — Sh 3.50 может быть noise. Но LIGHT N=1220 даёт Sh 11.55 = надёжный сигнал.
Удаление dvol_filter откроет penny-stock спам. Mitigations: keep universe filter ADV ≥ 2M (уже есть в universe_adv_2m.csv).

8. Summary — что осталось рабочего после re-cross-check

Factor	Solo Sh	Joint Sh delta	2026 stable?	Final verdict
F (skip FOMC/PPI)	14.14	+0.22	✓ Sh 11.55	HARD GATE
noRsi30 LONG forbid	13.94	+0.03	✓	HARD GATE
noBREAK_DN_LONG (B partial)	13.87	~0	✓	HARD GATE (free)
BREAK_UP S+/S/A/B	11-16	—	✓	SCORE BOOST
noSHORT_S12	3.52 solo	−0.05	—	DROP S12 (pos free)
Q laggard LONG	9.85	−1.08	?	SCORE BOOST only
P twist LONG	8.91	−3.28 ⚠️	?	SCORE BOOST only
E earnings overlay	13.74	−0.13	?	OK as gate (no_earnings)
C dvol≥1M	10.41	−4 везде	?	DROP — over-restrictive
D sector_top4	10.40	−2.93	?	DROP — over-restrictive
SHORT_S12	3.52	−0.05	?	DROP (free)
H short float	N=0 (LA disabled)	—	—	SKIP (LA snapshot trap)

Generated: 2026-04-26 · Source: revalidation_apr25/baseline_3y_1100.parquet + 11 factor parquets · Daily-aggregated Sharpe (×√252) · Source script: full_recheck_1100_apr26.py
Compare to memory: research_moo955_FINAL_SYNTHESIS_apr25.md, project_moo955_M11_DEPLOYED_apr26.md