🎯 Déjà-vu Regime Classifier — k=5 KMeans

Iter 2 of déjà-vu strategy · 2026-05-07 · OOS 130 days (2026-02 → 2026-04)
Reuses 49-feature fingerprint, switches selection from cosine top-K to discrete cluster membership · advisory only · raw PnL no slippage

Headline: regime classifier beats analog matcher

moo_to_close aggregate Sharpe

+2.12

vs analog matcher -0.33 (was FAIL)

Sign-flip p-value

0.022

passes Apr 26 gate (< 0.05)

Selectivity ratio

5×

401 signals vs 1991 in analog matcher

5 регимов — таксономия и интерпретация

R0 · n=73

Yields-up rally

dgs2+0.9 nfci+0.9 dgs10+0.8 arkk_5d+0.8 qqq_5d+0.8 spy_5d+0.8

Risk-on, ставки растут, growth тянет

R1 · n=21

Panic risk-off

vxx+2.2 vix+1.9 spy-1.8 xly-1.7 qqq-1.7 xlf-1.7 xlk-1.6

VIX spike, всё красное

R2 · n=89

Growth→energy rotation

arkk_1d-0.6 xle_minus_spy+0.5 arkk_minus_qqq-0.5 qqq-0.5 xly-0.5 xlk-0.5

Tech-fade при сильном XLE

R3 · n=1

Extreme melt-up

spy+8.8 qqq+8.4 xlk+7.9 xlc+7.7 (single outlier)

Один экстремальный день — бытовой outlier

R4 · n=97

Rates-easing

dgs2-0.8 nfci-0.7 dgs10-0.7 t10y2y+0.6 hyoas-0.6 vix-0.5 gld_5d+0.4

Ставки падают, кривая steepens, gold up

Per-regime OOS performance

Regime	Days fit	OOS signals	Mean PnL	Win-rate	Verdict
R0 Yields-up rally	73	152	+0.575%	60%	winner
R1 Panic risk-off	21	40	-0.298%	32%	model fails (mean-reversion bet, trend extends)
R2 Growth→energy rotation	89	71	-0.462%	36%	model fails (same dynamic)
R3 Extreme melt-up	1	108	+0.194%	56%	singleton, OK
R4 Rates-easing	97	30	+0.556%	70%	winner

Post-hoc filters (NOT for deploy — data-snooped)

Filter	Window	N	Sharpe	p-value	Mean	WR
All clusters (honest)	moo_to_close	223	+2.12	0.022	+0.28%	—
Drop R2	moo_to_close	176	+3.71	0.0005	—	—
Drop R1+R2	moo_to_close	156	+4.27	0.0005	+0.61%	62.8%
Only R0+R4	moo_to_close	98	+5.21	0.001	+0.85%	63.3%
Drop R1+R2	moo_to_1100	134	+2.47	0.036	+0.23%	56%

⚠ Эти фильтры идентифицированы пост-факт на OOS-выборке (HARKing). Sh 5.21 нельзя claim'ить как edge — требуется forward validation на отдельные 60 дней до deploy.

Сравнение: Analog matcher vs Regime classifier

Метрика	Analog matcher (May 06)	Regime classifier (May 07)
moo_to_1100 aggregate Sh	0.48	+1.17
moo_to_close aggregate Sh	-0.33 ❌	+2.12 ✅
moo_to_close aggregate p	0.76	0.022
moo_to_close STRONG-tier Sh (p)	2.96 (0.023)	-2.02 (small N=32)
moo_to_close best tier	STRONG (Sh 2.96)	WEAK (Sh 4.30 p=0.002)
Total signals	1991	401 (5× selectivity)
Interpretability	low	high (named regimes)

Caveats

OOS 130 days too short. Sh 2.12 needs ≥6mo for robustness; HHI 0.71 fails because only 2 quarters covered (структурно gate inapplicable, не реальный fail).
Post-hoc cluster filters не deploy-ready. R0+R4 only Sh 5.21 — это data-snooping; нужна forward validation.
Counter-intuitive WEAK-tier лидерство: regime-conditional pool 40-90 days vs analog top-K=10. CI bounds разные интерпретации; deploy-логика дашборда без изменений.
R1/R2 систематически проигрывают. Causal hypothesis: model bets on mean-reversion, но в panic/rotation режимах тренд extends. Требуется regime-aware sizing.
Schtask still NOT installed. Same gating as analog matcher — никто не deployed.
Slippage не моделируется (per May 05 directive).

Next steps

Extend OOS to 6mo — нужен backfill futures_premarket_12m.parquet или drop futures features (drop ~7 фич, остаётся 42).
Forward-validate R1+R2 block — paper-trade with cluster filter 60 days, проверить asymmetry persistence.
Sensitivity sweep — k ∈ {4,5,6,7}, fit_pool ∈ {90,120,180}; убедиться Sh ≥ 2 не зависит от seed/k.
Add regime tab to broker UI — показать today's cluster + regime label + per-regime signal table рядом с analog matcher.
Если PASS на 6mo OOS: deploy паралельно с analog matcher — два сигнала ансамблем.

Generated 2026-05-07 · advisory only · raw PnL no slippage · 42/42 unit tests pass · NO production deploy