Live Market Overview
End-of-day regime + term structure
Last close: 2026-08-03
· Calibration: 2026-07-23
· Regime: 2026-08-03
S&P 500
7,600.50
+1.48%
vs prev close
2026-08-03
VIX 30-DAY IMPLIED
15.86
🟢 NORMAL range
VVIX VOL-OF-VOL
90.81
Normal VVIX
Term Structure
CONTANGO
Front=13.28 → Back=21.2
Slope=+7.92 pts
Current Regime — rule-based label
R1 · SHORT GAMMA
Deterministic rule (rv₂₀ vs VIX · VVIX gate) — the label the backtest
trades on. As of 2026-08-03.
ML classifier · research only
R1 · 76.4%
63.4% out-of-sample, it loses to the 90% “predict-yesterday” baseline, so it is not used for trading (shown for research only).
R0
Low VIX / backwardation → buy straddles
R1
Normal contango → sell premium
R2
High VVIX → vomma trades
PDV Model Forecast vs ATM Implied
CURRENT VIX (30D IMPLIED)
15.86%
PDV FORECAST (REALISED)
8.37%
IMPLIED − REALISED SPREAD
+7.49pp
Moderate vol risk premium → favours short vol strategies
0.6th
percentile
2015–2025 mean: +3.92pp
2015–2025 mean: +3.92pp
ⓘ Spread at 0.6th percentile
— vol appears cheap relative to 10-year history.
Historical mean: +3.92pp.
Current: +7.49pp.
Implied–Realised spread — last 60 trading days
σ₁ (5d EWMA): 8.71%
σ₂ (60d EWMA): 14.32%
Model: OLS + GARCH(1,1)
Updated: 2026-08-03
VIX Term Structure
C8 Classifier — Full-Period Statistics (2015 – 2025)
R0 LONG GAMMA
12.7%
350 / 2,757 days
Buy straddles
R1 SHORT GAMMA
42.8%
1,180 / 2,757 days
Sell premium
R2 VOMMA ACTIVE
44.5%
1,227 / 2,757 days
Vomma trades
Overall Accuracy: 63.4%
Persistence baseline: 90.0%
Features: 5 (vvix excluded — C16)
Top Feature: fear_premium (39%)
Model: XGBoost 3-class
Status: RESEARCH ONLY (C17)
C16: vvix removed from training — the R2 label is defined as VVIX>threshold, so including
it inflated accuracy to 86.2% by letting the model recover its own labelling rule; 63.4% is
the honest out-of-sample number. C17: 63.4% loses to the
no-skill persistence baseline ("predict yesterday's regime" scores 90.0% on 2020+,
and yesterday's label is fully observable). The classifier is demoted from the trading loop —
the backtest uses the lagged deterministic rule labels directly. Kept here as a documented
negative result: regimes are too sticky for daily ML classification to add value.
PDV Model — Guyon-Lekeufack (2023)
σ̂(t) =
0.354
× σ₁ +
0.241
× σ₂ −
1.496
× lev +
0.0346
σ₁ half-life: 5 days (EWMA)
σ₂ half-life: 60 days (EWMA)
Walk-fwd R² (linear): 0.31
Walk-fwd R² (kernel): 0.23
GARCH persistence: 0.979
GARCH β: 0.890
Active model: PDVLinear3F
· PDVLinear4F requires VIX column in stored features.
COVID Stress Test — 2020-03-16
PDV predicted 92.4% ann vol.
Actual realised: 202.6%.
Error: −110pp — structural limitation of backward-looking models.