Lasso (L1)
Feature-importance + sparse weights. Used for the multi-asset rolling-window pipeline (see /backtests). Typical R²~0.18-0.25 on 1-quarter forward.
Random Forest
Non-linear, robust to outliers. Used for crisis-risk classifiers + regime-aware bagging. Permutation feature-importance for interpretability.
Logistic curve
For bounded metrics (fentanyl share, vaccination rate, recidivism rate). Saturation-aware fitting.
Regime-aware modeling
High-vol vs normal split with per-regime coefficients. Covers the Fed-policy + crisis-event heteroskedasticity.
Sentiment fusion
Weighted blend of keyword tonality + (planned) NewsAPI.ai / Permutable / EventRegistry. Used in /sentiment.
GTAP-style trade
Trade-elasticity + Phillips-curve pass-through. The "trade pass-through" pattern in the US-Impact engine.
Hormuz oil model
P_t = P_0 + k·R_t + M_t. Region-risk × event-multiplier. Calibrated against 1990-2024 Brent shocks.
Doomsday composite
10-stream weighted sum (9 base + US-Impact + Domestic-Chaos + trust-inverted). Documented at /doomsday.
Retraining cadence
- Lasso / RF — weekly walk-forward refit on a rolling 5-year window.
- Logistic / sigmoid — monthly.
- Doomsday weights — quarterly review against ground-truth events.
- Sentiment keyword sets — quarterly (or after a regime-change event).