🪞 Déjà-vu thin-slice — Walk-Forward Report

Period: 2025-07-01 → 2026-04-24 · 12-month horizon · advisory only (Apr 29 directive) · raw PnL no slippage (May 05 directive)
Spec docs/superpowers/specs/2026-05-06-dejavu-strategy-design.md · Plan docs/superpowers/plans/2026-05-06-dejavu-strategy.md

Aggregate verdict

moo_to_1100 (all tiers)

Sharpe 0.48

N=878 · mean +0.039% · sign-flip p=0.174 · CI [-0.89, 1.95] · HHI 0.32

FAIL

moo_to_close (all tiers)

Sharpe -0.33

N=1113 · mean -0.035% · sign-flip p=0.759 · CI [-1.72, 1.22] · HHI 0.31

FAIL

Tier-stratified — где edge живёт

Window	Tier	N	Sharpe (raw)	Win-rate	Mean PnL	Verdict
moo_to_close	STRONG	115	+2.96	62.6%	+0.46%	PASS (sign-flip p=0.023)
moo_to_1100	STRONG	55	+1.35	56.4%	+0.114%	UNDERPOWERED (p=0.27, N=55)
moo_to_1100	MEDIUM	198	+0.90	50.0%	+0.080%	borderline
moo_to_1100	WEAK	625	+0.25	49.8%	+0.019%	noise
moo_to_close	MEDIUM	331	-0.82	45.3%	-0.092%	net-negative
moo_to_close	WEAK	667	-0.98	49.2%	-0.093%	net-negative

Главный вывод: aggregate FAIL потому что WEAK+MEDIUM нетто-отрицательны и тянут вниз сильный STRONG-tier. moo_to_close STRONG-only сигналы проходят все 4 ворот Apr 26 (Sharpe ≥ 1.0, sign-flip p < 0.05, block-bootstrap CI lo > 0, HHI < 0.4).

Top / Bottom тикеры по cumulative signed PnL

Top 10 — где работает

Ticker	N	Σ PnL %	Mean %	WR %
XBI	69	+36.05	+0.52	66.7
SLB	47	+25.20	+0.54	61.7
XOM	43	+24.44	+0.57	60.5
MPC	51	+19.06	+0.37	56.9
BKR	48	+16.49	+0.34	54.2
TSLA	35	+16.08	+0.46	57.1
HAL	39	+15.96	+0.41	64.1
XLE	40	+9.57	+0.24	62.5
PSX	35	+9.23	+0.26	54.3
KRE	50	+8.23	+0.16	58.0

Кластер: oil-services + refiners + biotech

Bottom 10 — где сигнал ломается

Ticker	N	Σ PnL %	Mean %
GDX	62	-22.68	-0.37
PLTR	57	-21.17	-0.37
NEM	55	-17.86	-0.32
RTX	56	-15.75	-0.28
AEM	31	-13.08	-0.42
LHX	26	-12.35	-0.47
JETS	39	-11.88	-0.30
USO	48	-10.26	-0.21
GLD	45	-8.30	-0.18
COIN	48	-8.03	-0.17

Кластер: gold-miners + defense + consumer-growth

Sample run — 2025-12-15 (smoke test)

Top-K analog dates

Date	Cosine sim
2025-09-03	0.521
2025-10-06	0.504
2025-01-28	0.476
2025-10-15	0.475
2025-08-06	0.460
2025-11-05	0.456
2025-11-21	0.451
2025-09-18	0.449
2025-10-08	0.446
2025-08-04	0.444

STRONG signals (CI lower bound > 0.5%)

Ticker	Window	Dir	Mean	CI
AAL	moo_to_close	LONG	+1.72%	[+0.75, +2.83]
ARKK	moo_to_close	LONG	+1.43%	[+0.88, +1.98]
GDX	moo_to_close	LONG	+1.18%	[+0.61, +1.80]
XBI	moo_to_close	LONG	+1.14%	[+0.53, +1.71]

Architecture & performance

Walk-forward run time

197 s

218 days × 50 tickers × 2 windows = 1991 signals

Fingerprint speedup

3700×

45 s → 12 ms via pre-pivoted closes lookup

Unit tests

35 / 35

incl. 96-second no-lookahead invariant test

Pipeline (06:00 → 07:00 ET)

fingerprint.py — 49 фич: cross-asset 1d/5d (10) + sector ETF 1d (12) + spreads (4) + FRED (8) + futures pre-market (7)
analog_match.py — z-score + cosine + top-K=10 + ±5d exclusion (autocorrelation guard)
bootstrap.py — 1000 percentile resamples, deterministic seed
behavior.py — CI lo > 0 → LONG / hi < 0 → SHORT, tier по distance from 0
ttn_filter.py — drop cells с corporate news (guidance/M&A/FDA/buyback и т.д.)
walkforward.py — sharpe + sign-flip + block-bootstrap + quarter HHI
runner.py — orchestration + Telegram STRONG-tier alert

Decision & iteration

Schtask NOT installed. Aggregate verdict FAIL — нельзя deploy дашборд с WEAK/MEDIUM tier'ами.

Reduced-surface viable: moo_to_close STRONG-only сигналы (Sh 2.96, p=0.023) — броkep tab уже подсвечивает STRONG отдельным badge'ем, дисциплина пользователя = торговать только STRONG.

Iter 2 candidates (max 3 итерации до DEAD per Apr 26 framework):

Drop high-collinearity features в fingerprint → top-20 PCs
Tighten K from 10 → 5 (более селективный пул)
Mahalanobis вместо cosine (использует 12m covariance matrix)
Stricter STRONG threshold: CI lo > 0.7% вместо 0.5%

Iter 3 candidates:

Switch trigger от latent-regime к event-class (FOMC days, tariff-headline days)
Extend horizon до 5y для richer pool

Generated 2026-05-06 · advisory only · raw PnL no slippage · NO production deploy