SPY 0DTE — Risk Profile

Loss-attribution catalog. Aggregates the research artifacts in data/reference/loss_attributions/ into a queryable view of "which features distinguish loss days for each strategy family." Two halves: latest patterns (most recent attribution per family, top features by Z-score) and promoted patterns (features that recurred across ≥2 attributions at |Z|≥1.5 — the regime markers worth treating as decision inputs). Currently the recommender's swap rules only fire on promoted patterns; latest patterns are exploratory.

Loading…

Promoted regime patterns

Features with |Z| ≥ 1.5 in ≥ 2 attribution runs for the same strategy family. Stable across runs = regime markers worth treating as decision inputs. Empty when no patterns have crossed the promotion threshold yet.

Strategy family	Feature	Direction	Mean Z	Z range	Runs
Loading…

Latest patterns per strategy family

Most recent loss-attribution artifact per family. Top features ranked by |Z-score| of the difference between mean-on-common- loss-days and mean-all-days (in std units of the all-day distribution). Color: red = elevated on loss days; blue = depressed.

Loading…

Counter-strategy verdict (session 15)

From the loss-attribution v2 with the full 12-counter pool. On the 27 IC common-loss days observed in the 500-day window:

Counter strategy	Loss-day WR%	All-day WR%	Lift	Total $ on loss days	Verdict
afternoon_long_strangle	88.9%	22.6%	+66pp	+$3,996	🏆 Best — validated counter
afternoon_long_call	74.1%	27.0%	+47pp	+$3,020	Strong #2 — Tier A validated
morning_long_call	51.9%	48.2%	+4pp	+$358	Mild positive
morning_long_put_slight_up	29.6%	40.4%	-11pp	+$44	Flat
morning_ic_strong_up	66.7%	49.0%	+18pp	-$245	Smaller loss
morning_jade_lizard	59.3%	43.8%	+16pp	-$293	Smaller loss
morning_scs_heavy_up	55.6%	55.8%	−0pp	-$101	Flat
morning_sps_heavy_down	77.8%	64.6%	+13pp	-$117	Mild positive
afternoon_sps	74.1%	86.4%	-12pp	-$936	Loss
afternoon_scs	37.0%	87.4%	-50pp	-$1,617	Catastrophic
afternoon_ib	0.0%	59.8%	-60pp	-$4,011	❌ Same-regime victim
afternoon_jade_lizard	11.1%	77.4%	-66pp	-$3,996	❌ Hypothesis falsified

Implication for the live recommender: when today's features match the IC-loss-regime catalog at ≥medium severity, swap IB → LongCall in the validated 3-way ensemble (IB + STRANGLE + MorningLP). LongCall is a Tier A coverage-gap structure validated as IC counter via this analysis — captures upside continuation that kills IC.

How this feeds decisions

Catalog flow per D13:

Tier 1 — each loss-attribution run writes a JSON artifact to data/reference/loss_attributions/.
Tier 2 — scripts/build_loss_patterns_index.py aggregates the artifacts into frontend/data/loss_patterns.json (the file rendered on this page). Promotion rule: feature with |Z|≥1.5 in ≥2 attributions for the same family.
Tier 3 — src/briefing/regime_risk.py matches today's features (from MarketState) against the catalog; emits severity-coded warnings. src/briefing/checkpoint_runner.py runs regime_conditional_ensemble() on each checkpoint and embeds today's recommendation in the session JSON.
Tier 4 — methodology promotion: stable patterns become candidate negative filters or parameter rules, walk-forward validated, then promoted to validated findings in PROJECT_STATE.md / CLAUDE.md.

Engine-bias caveat. Z-scores are computed from feature distributions, so they are bias-resistant. Absolute P&L numbers (e.g., the counter-strategy table above) carry the documented BSM-IV-undershoot bias; relative rankings, correlations, and Z-scores are engine-honest.