Drawdown-Risk Lens (20-day, transparent weighted z-score composite)

A probabilistic drawdown-risk gauge over the credit / inflation / rate LEVELS the full-history study flagged for the platform’s multi-week drawdown lane, plus a validated positioning input — dealer net gamma. Renders as the “Drawdown-Risk Lens · 20d” block in the regime card with a 0–100 score, a low / elevated / high band, and the per-leg contributions. The 15-year re-audit (U1, 2026-07-13) upgraded it to a validated cross-regime drawdown-severity gauge — but the edge is between-regime severity, never within-episode timing (see the caveats).

This is not a directional predictor, not a timing trigger, and not the validated logistic from the research. It is a transparent weighted z-score composite, chosen deliberately for explainability and graceful degradation over fitted accuracy.

What it measures

Each of the nine macro levels is standardized against its trailing baseline (default ~2 trading years), oriented by economic risk-direction, and averaged into a single number. The dealer-gamma leg adds a positioning read on top. Higher = more stress + more fragile dealer positioning relative to the recent norm; lower = calmer.

The dealer-gamma positioning leg

An important positioning input in the platform’s drawdown research is dealer net-gamma positioning (a “dealer-gamma analytics” feed). When option dealers sit at historically low (more-negative) net gamma, they are short gamma — forced to sell into a falling market and buy into a rising one, which amplifies drawdowns. The leg is net_gex_rank: the 252-day trailing percentile rank of the aggregate dealer net-gamma exposure. A low rank (historically negative gamma) reads as elevated drawdown risk; a high rank (positive gamma, stabilizing dealers) reads as calmer.

The discovery-era read (rank-vs-drawdown IC ≈ +0.25 non-overlapping, on two episodes) did not survive the 15-year re-validation. The U2 re-audit (2026-07-13) halved the non-OL rank-IC to ≈ +0.10 to +0.11 at 15y power (boot-CI lower bound ≤ 0), and each macro leg (nfci / hy_oas / gz_spread / ofr_fsi) individually out-ICs it — so on rank-IC the gamma leg is now mid-pack, not the strongest. What did transfer (and to the 2011–2020 out-of-sample span) is the quintile tail-lift: ~5–6× Q20/Q80 separation on the −5%-drawdown days, monotone. So the leg’s value is tail-capture + positioning-class diversification + freshness, not day-ordering rank-IC. Its weight was re-sized down from 2× to 1.5× accordingly (the residual premium pays for tail-lift, not IC; 1.0 is defensible on strict IC parity).

The rank needs a trailing-history warm-up; on a cold start (under ~6 months of gamma history) the leg abstains rather than report a misleading mid-rank.

Why this shape, not a fitted model

The full-history study (.research/SYNTHESIS.md) validated the multi-week drawdown signal as a parsimonious L2 logistic over seven of these macro features (the composite later added gz_spread + ofr_fsi_funding). That model is sklearn-based, and the platform’s prod image deliberately ships no machine-learning runtime — scikit-learn and its peers live in requirements-dev.txt and the no-Dockerfile-reference lint (verify-docs.sh:27) blocks them from ever being COPYed into a production container.

The deployable shape is therefore the transparent composite. Trade-offs, on the record:

• Transparent + auditable. Every contribution is a z-score the operator can recompute by hand. The card surfaces the per-leg contributions so it never moves without showing why.
• Weighted, not fitted. The macro levels are equal-weight (1.0 each) with four config-listed exceptions — the funding-stress leg carries 1.5 (highest forward-drawdown IC of any tested component); ebp + gz_spread carry 0.5 each (the GZ spread is predicted-default plus EBP by construction, so weighting both at 1.0 double-counted the EBP information; the split keeps the pair’s combined credit-premium mass at one leg, backtested within noise or better on all horizons); and the Treasury-curve leg carries 0.0 — it is oriented for a 12–18-month recession lead, mis-horizoned against the 10–20-day target, and was the only leg with a wrong-signed standalone relationship to forward drawdown (it flipped the whole composite negative in the 2022 bear); zero-weighting it improved the drawdown read on every horizon and both validation episodes, so it is kept visible in the breakdown at weight 0. The dealer-gamma leg carries a 1.5× over-weight (re-sized down from 2× by the U2 15y re-validation). This is a coarse, hand-set weighting — it encodes “gamma is a diversifying, tail-lift-strong positioning input” without smuggling in an opaque single-bear fit.
• Degrades gracefully. A missing feature is just dropped from the weighted average; the gamma leg abstains until it has enough history; below four available legs the gauge suppresses itself (status: insufficient_data) rather than report off a thin set.

The nine macro inputs + risk-orientation

The orientation table lives in app/signals/drawdown_risk.py:_FEATURE_ORIENTATION and is documented inline with the rationale. Summary:

A +1 feature contributes +1·z to the mean (positive z = up = more risk). A −1 feature contributes −1·z (so a deeply inverted curve, where z is very negative, still pushes risk up).

The dealer-gamma leg’s risk-z is derived from the percentile rank instead of a level z-score: oriented_z = (0.5 − rank) × √12, so a rank of 0 (most-negative gamma) maps to a strongly positive risk-z and a rank of 1 (most-positive gamma) to a strongly negative one. The √12 scaling puts a uniform [0,1] rank on a unit-variance footing comparable to the level z-scores — so the leg’s weight, not an accidental scale mismatch, is the lever.

How the gauge is computed

For each available macro feature with a non-zero baseline standard deviation:

1. z = (level − trailing_mean) / trailing_std
2. oriented_z = z × orientation, weight = feature_weights.get(leg, 1.0) (most legs 1.0; funding-stress 1.5; ebp and gz_spread 0.5 each per the double-count split above; t10y3m_spread 0.0 per the wrong-signed / mis-horizoned note above — computed and shown, contributes 0)

For the dealer-gamma leg (when enough history exists):

• oriented_z = (0.5 − net_gex_rank) × √12, weight = 1.5 (re-sized from 2.0 by U2, 2026-07-13)

Then:

3. mean_z = Σ(oriented_z × weight) / Σ(weight) over available legs
4. score = 100 / (1 + exp(−mean_z)) — a logistic squash with steepness 1.0, so mean_z = 0 maps to 50, +1σ to ~73, +2σ to ~88, −1σ to ~27.
5. Band: low under +0.5σ, elevated between +0.5σ and +1.25σ, high at or above +1.25σ.

Each leg’s response carries a contribution field — its signed share of mean_z — so the card shows how much each input moved the gauge. If fewer than four legs are populated the gauge returns status: insufficient_data and the dashboard suppresses the block.

Caveats — read these before you act on it

• Cross-regime validated composite, but per-leg trust is uneven (U1, 15y). The 15-year re-audit validated the composite as a cross-regime drawdown-severity gauge — the HIGH band fired in every 2011–2026 bear with 0 calm-year false fires, and it beats VIX-only at the IC level (partial IC ≈ +0.14, survives BH-FDR). But leg-level trust is uneven: only 4/10 legs (nfci, ofr_fsi_funding, hy_oas, gz_spread) survive per-leg BH-FDR at 15y, the hy_oas survival is partly synthetic-substrate inflation (FLAG-U1a, re-graded forward on real data), and the breakevens are regime-fragile — wrong-signed pre-2021 (FLAG-U1b). Treat the composite band as trustworthy; treat any single leg’s contribution with the per-leg caveat in mind.
• Between-regime severity, NOT within-episode timing (U1 bound + U4). The edge is between-regime discrimination. Inside a named drawdown episode the composite reads CONTRARIAN (7/8 negative within-episode IC) — a HIGH print mid-episode is a severity/state read, never a step-aside-before-the-drop timer. And as a thresholded top-decile tail-capture flag (rather than an IC-level read) the composite does NOT beat VIX-only under causal expanding-window thresholds (U4, 2026-07-14) — VIX-only dominates every causal comparison. The beats-VIX result is an IC-level claim only; the gauge is coincident, not leading.
• The gamma leg’s rank-IC is mid-pack at 15y power. The non-overlapping out-of-sample sample behind the discovery-era gamma weight was ~57–116 observations (overlapping 20-day windows autocorrelate, inflating the nominal count); at full 15y power the rank-IC halved to ≈ +0.10/+0.11 (U2). The 1.5× weight (re-sized from 2× by U2) is sized for the leg’s quintile tail-lift + positioning-class diversification, not its day-ordering rank-IC. Treat the leg as a diversifying tail-capture tilt, not the strongest input.
• Hand-set weights, not fitted. The fitted logistic moved the AUC; this composite does not. Treat the gauge as situational awareness, not a quantitative probability.
• The breakeven sign is regime-specific. 2022 was an inflation-driven drawdown; elevated breakevens read as risk here. A deflationary bear would invert that. The orientation is the load-bearing assumption of the gauge — flag the gauge as suspect if the regime story changes.
• It does not contribute to the health score. This is a separate risk-side read; the 19-component health score is unchanged.

Where it lives in the code

• Pure helper: app/signals/drawdown_risk.py (no DB, no network, no ML). Takes the macro levels + baselines and the pre-computed net_gex_rank.
• Route: GET /api/v1/signals/drawdown-risk (computes trailing baselines + current levels from daily_signals, computes the 252d gamma rank from the gex history, calls the helper). Schema-versioned by DRAWDOWN_RISK_SCHEMA_VERSION.
• UI: rendered in the regime card on / directly below the regime line, above the historical base-rate block.
• Tests: tests/test_drawdown_risk.py pins macro orientation, banding, the dealer-gamma leg’s orientation + over-weight + abstain-on-thin-history, per-leg contributions, the insufficient-data fallback, and a non-mocked route test that drives the gamma leg from seeded gex history.