MOC→19:00 Phase-Aware (Apr 27)

Замена binary skip-flags на graduated size_mult per (SPY phase × today event). 4 phases × 3 events = 12 combos.

Headline metrics

Phase × Event recommendations (size multiplier)

Phase \ Event	NORMAL	PANIC (SPY ≤ −1.5%)	GREED (SPY ≥ +1.5%)
BULL_CALM	0.75× (C-only, +0.12%)	0.25× (A∪B, marginal)	0.75× (A∪B, +0.09%)
BULL_VOL	1.50× (A∪B, +0.38%)	— (no data)	0.00× (skip)
BEAR_CALM	0.00× (skip)	0.00× (skip)	0.00× (skip)
BEAR_VOL	1.50× (C-only, +1.38%!)	0.25× (C-only small)	0.00× (skip)

Edge column meaning: A∪B = (intraday_fade ≤ −7% OR pm_high ≥ +10%); C-only = today_dropped ≤ −7% bounce play. Strategy picks edge-flavor per combo based on which has stronger mean.

Phase distribution (3 yrs, 496 days)

Phase	Days	%	Note
BULL_CALM	328	66.1%	SPY 20d ≥ 0 + 5d realvol ≤ 20 — workhorse
BEAR_CALM	105	21.2%	SPY 20d < 0 + low vol — slow grind down, edges DEAD
BEAR_VOL	49	9.9%	SPY 20d < 0 + 5d realvol > 20 — capitulation phase, BEST mean +1.38%
BULL_VOL	14	2.8%	Vol-up regime in uptrend — small N but +0.38%

Comparison: phase-aware vs v7 hard-skip

Metric	Phase-aware (15)	v7 hard-skip (14)	Δ
Trades / yr	550	835	−285 (−34%)
Basket / day	3.6	5.4	−1.8
Trading days used	296 / 496	303 / 496	−7
Net % per trade	+0.184% (weighted)	+0.155%	+0.029%
Sharpe daily	1.92	2.48	−0.56
$/yr @ $10K base	+$10,115	+$12,912	−$2,797
MaxDD @ $10K base	−$1,905	−$3,635	+$1,730 (better)
$/MaxDD ratio	5.31×	3.55×	+1.76

Verdict

Phase-aware vs hard-skip — trade-off ясен:

Hard-skip v7 даёт выше абсолютный $/yr ($12.9K vs $10.1K) и выше Sharpe (2.48 vs 1.92).
Phase-aware даёт лучше DD контроль (−$1.9K vs −$3.6K) и выше $/MaxDD ratio (5.31× vs 3.55×) — capital-efficient.
Mean per trade higher (0.184% vs 0.155%) — quality over quantity.
5 из 11 (phase,event) комбинаций всё равно получили size_mult=0 — но это narrow combos (BEAR_CALM × all events, BULL/BEAR_VOL × GREED), не blanket "skip день если ANY flag".

Рекомендация: phase-aware подходит если приоритет — стабильность DD и risk-adj returns. Hard-skip v7 — если приоритет абсолютная прибыль. Можно гибрид: использовать phase-aware multipliers (1.5/0.75/0.25) для "good" combos, но без zero-out — оставить минимум 0.25× даже на BEAR_CALM × NORMAL чтобы capture редкие положительные дни.

Edge-by-edge per phase (top combos by N×mean)

Combo	Edge	N (3yr)	Mean %	WR %	Mult
BULL_CALM × NORMAL	C-only	682	+0.123	55.0	0.75×
BEAR_VOL × NORMAL	C-only	53	+1.379	62.3	1.50×
BULL_VOL × NORMAL	C-only	11	+0.637	63.6	1.50×
BEAR_VOL × PANIC	C-only	424	−0.009	49.3	0.25×
BEAR_CALM × NORMAL (A∪B)	A∪B	566	−0.088	42.4	0.00×
BEAR_CALM × PANIC (A∪B)	A∪B	140	−0.238	52.9	0.00×

⚠️ Surprise: BEAR_VOL × NORMAL — лучший combo (+1.38% mean!) — capitulation bounce setup. Только N=53 за 3 года, но эти дни приносят непропорционально много. BEAR_CALM × NORMAL (566 trades, mean −0.09%) — основной "noise eater", killing all his trades fixes Sharpe.

Files

research_results/moc_1900_apr27/scripts/15_market_phases.py — phase classifier + size_mult logic
SPY regime source: enrichment_data/spy_10yr_daily.parquet (1255 rows)
Trade pool: 89,057 trades 2024-04 → 2026-04