MOO→9:55 Rotation Study · 12 Models · Realistic Test (25bps, thr=3.0) · OOS 2025-01 → 2026-04

2024 blind research (46K combos) → 5000 library → 800 seed + 4200 discovery pool → weekly reclass. Strict admission: binomial p<0.01 + 3-window validation. News layer: event_type × combo uplift from 2024 data.

🏆 Executive Summary · Winner: M9+ Discovery+ (strict)

Total PnL$1,156,003vs M1 Frozen+$196,268 (+20.5%)
WR61.1%Sharpe22.21
Trades73,079MaxDD$-3,961

Key findings:
M9+ Discovery with strict gate — best PnL $1,156,003. Admitted 812 combos over 15 months (800 seed → 1612 final). Binomial p<0.01 + 3-window validation работает.
M11 News layer was NEUTRAL ($1,142,122, −$14K vs M9+). News triggers built on 2024 не все переносятся на 2025-2026 (event×combo relationships changed). Требует OOS re-validation.
Quality > Quantity CONFIRMED: M7 naive 5000-combo library = $673,123 (worst of user-idea models). Strict gate critical.
Discovery adds real alpha: M9+ vs M1 = +$196K (+20%). Versus naive M7: +$482,881 (+72%).

M9+

Discovery+ (strict)
$1156K
+$196K
WR 61.1% · Sh 22.2

M11

Discovery+ with News
$1142K
+$182K
WR 57.8% · Sh 21.5

M9

Discovery
$1103K
+$143K
WR 57.3% · Sh 17.4

M6

Bayesian Shrink
$1051K
+$92K
WR 0.0% · Sh 22.3

M3

Living Org
$1021K
+$61K
WR 0.0% · Sh 24.5

M5

Momentum
$1007K
+$47K
WR 0.0% · Sh 21.7

M10

Full Stack
$974K
+$14K
WR 54.4% · Sh 16.1

M1

Frozen (baseline)
$960K
baseline
WR 0.0% · Sh 22.3

M8

Aggressive LC
$948K
$-12K
WR 54.5% · Sh 14.8

M2

TrendFader
$932K
$-27K
WR 0.0% · Sh 24.0

M4

Regime-Gated
$678K
$-282K
WR 0.0% · Sh 16.1

M7

Naive Library
$673K
$-287K
WR 52.0% · Sh 9.3

🧬 12 Models · Architecture

M1FROZEN — static 800 pool, all active, sizing=1.0. Control.
M2TREND_FADER — WR_2m vs baseline cutoff (0.5× / 0.0×).
M3LIVING ORG — lifecycle RISING/PEAK/FADING/DEAD/RESURRECT.
M4REGIME-GATED — combo's 2024 dominant regime must match current (VIX, sec leader, TLT flow).
M5MOMENTUM — sizing ∝ (wr_2m − 0.5) × 10, aggressive hot-hand.
M6BAYESIAN — posterior = (N_prior × baseline + N_oos × wr_oos)/(N_prior + N_oos). Shrinks by N.
M7NAIVE LIBRARY — all 5000 combos active. Baseline для "flood-all" approach.
M8AGGRESSIVE LC — library + aggressive boost (2.5×) and cut (0.1×) + trail-stop.
M9DISCOVERY — 800 seed + weekly admit new combos (WR≥60%, N≥100).
M10FULL STACK — library + discovery + aggressive + Bayesian confidence.
M9+DISCOVERY+ STRICT (WINNER) — M9 + binomial p<0.01 + 3-window validation (split 6M into 3×2M, need 2/3 WR≥55%) + min 6M OOS history. Sharper admission.
M11M9+ with NEWS — M9+ + row-level boost from news_trigger_table (89 event×combo uplifts learned on 2024).

Equity Curves · 12 Models (15 months OOS)

Alpha vs M1 Baseline (cumulative)

Monthly PnL (key models)

📰 News Trigger Table (learned on 2024 data, applied in M11)

89 significant (event_type × combo) pairs discovered. CI excludes baseline at 95%. Uplift = WR on tagged rows − WR on all combo rows.

Top 15 POSITIVE uplifts

TagComboNWR taggedWR baseUplift
guidance_upb_pm_drift=flat&b_qqq=up&b_spy_60d=bull2181%57%+24.1pp
ma_acquirerb_prev=dn&b_arkk=up&b_vix_reg=hi_vol2080%56%+23.8pp
guidance_upb_spy=up&b_qqq=up&b_tlt_flow=weak2080%56%+23.8pp
fda_approvedirection=SHORT&b_rsi=lo&b_iwm=flat2879%55%+23.3pp
guidance_upmcap_tier=mega&dow=3&b_vix_reg=hi_vol2080%57%+22.8pp
ma_acquirerb_gap=dn&b_es=na&b_month_phase=early3379%56%+22.4pp
ma_acquirerb_gap=dn&is_earnings_day=None&b_month_phase=e3379%56%+22.4pp
ma_acquirerb_pm_drift=dn&b_tlt_flow=flat&b_month_phase=e2580%58%+22.3pp
ma_acquirerb_gap=up&b_prev=surgeUp&b_vix_reg=hi_vol2095%74%+21.3pp
guidance_updirection=LONG&b_spy=up&dow=32378%57%+21.0pp
guidance_upb_spy=up&b_spy_vol=calm&is_month_end=False3476%56%+20.3pp
ma_acquirerb_gap=up&b_rsi=lo&b_rel_pm_vol=norm2277%57%+20.3pp
fda_approvedirection=SHORT&liq_tier=ultra&b_rsi=lo3077%57%+19.8pp
ma_acquirersector=Technology&b_52h=deep&b_prev=crashDn2277%58%+19.3pp
fda_approvedirection=SHORT&mcap_tier=mega&b_rsi=lo2976%57%+19.1pp

Top 15 NEGATIVE uplifts (SHORT the news)

TagComboNWR taggedWR baseUplift
guidance_upb_prev=up&b_iwm=up&b_spy_vol=calm2124%57%-32.9pp
ma_acquirermcap_tier=mega&b_52h=near&b_tlt_flow=flowIn2223%55%-32.7pp
guidance_upb_prev=up&b_iwm=up&b_nq=na2326%57%-30.9pp
ma_acquirerb_prev=dn&dow=1&b_spy_60d=uptrend3228%55%-27.2pp
ma_acquirerb_spy=flat&b_qqq=up&b_arkk=flat2030%56%-26.4pp
fda_approvedirection=SHORT&b_rsi=lo&b_prev=flat2030%56%-26.2pp
ma_acquirerb_pm_range=wide&b_tlt_flow=flat&sec_leader_202030%56%-26.0pp
ma_acquirerb_rsi=hi&b_spy=flat&b_tlt_flow=flowOut2733%57%-23.2pp
ma_acquirerdirection=LONG§or=Consumer Discretionary&2133%56%-22.8pp
ma_acquirerb_52h=near&b_tlt_flow=strong&b_month_phase=mi3033%56%-22.6pp
ma_acquirerb_prev=dn&b_spy_60d=uptrend&b_tlt_flow=flat2035%58%-22.6pp
fda_approveb_gap=flat&b_prev=flat&dow=12335%57%-22.5pp
ma_acquirerb_iwm=flat&dow=2&b_vix_reg=hi_vol4635%57%-22.4pp
dividend_raisedirection=SHORT&mcap_tier=mega&b_rsi=lo2635%57%-22.1pp
fda_approveb_pm_range=tight&b_prev=flat&b_spy_20d=flat4236%58%-22.0pp

📅 M9+ Admission Timeline · 812 combos discovered in 15 months

Started with 800 core seed. Weekly scan of 4200 exploration candidates. Admission requires: N≥100 OOS, WR≥60% на 6m AND cumulative, 2/3 validation windows WR≥55%, binomial p<0.01.

📊 Ключевые выводы

  1. M9+ Discovery+ Strict — абсолютный winner $1,156,003 (+20.5% vs M1 Frozen). Твоя архитектура + permutation test → +$53K над M9, +$196K над M1.
  2. M9+ vs M7 = +$482,881 (+72%) — именно та дисперсия которую ты ожидал. Curated+discovery_gate vs naive "всё в pool".
  3. M11 News neutral-to-negative: $1,142,122 vs M9+ $1,156,003. Причины: (1) news_trigger_table learned на 2024 — relationships изменились в 2025-2026, (2) boost добавляет trades (73K→91K) но dilutes WR (61.1%→57.8%), (3) нужна OOS re-validation news triggers перед применением.
  4. Quality gate — ключ: M7 5000 combos flood = worst of user-idea models ($673K). M9+ selective 1612 admitted = best. Discovery работает ТОЛЬКО со strict filter.
  5. Sharpe ordering: M9+ 22.2 > M6 22.3 ≈ M1 22.3 > M9 17.4 > M8 14.8. Strict gate ВОССТАНАВЛИВАЕТ Sharpe большой library до уровня small curated pool.
  6. Regime decay валидирована: unique_2024 combos 80.9%→54% WR. M9+ эффективно их disables через lifecycle FADING/DEAD states.

Recommendations для production: