HazardPulse
Every model is versioned and tracked.
We never run an unregistered model. Every version is pinned, every promotion is justified, and every retirement is recorded.
Our models
HazardPulse uses four prediction models, each specialized for a different type of hazard:
Earthquake model
Uses patterns in past earthquake activity to estimate future risk. Global AUC 0.907, temporal AUC 0.799.
Hurricane model
Uses ocean temperature, wind patterns, and satellite data to predict rapid intensification. 94% accuracy at ranking risk levels.
Tornado model (day-ahead)
Uses weather parameters to predict tornado environments 12-24 hours ahead. 64% accuracy (this is a hard problem).
Tornado model (storm-level)
Scores individual thunderstorms using radar rotation and atmospheric physics. 89% accuracy at ranking risk. Accumulating live verification data.
Models are only promoted after they beat the previous version on holdout data. Models that fail quality checks are rejected.
Active models
| Model | Hazard | Status | Promoted | AUC | Brier | Architecture |
|---|---|---|---|---|---|---|
| hp-etas-v3.2.1 | Earthquake | active | 2026-02-28 | 0.907 / 0.799 | 0.182 | GBM + Bayesian cal. |
| hp-hurricane-ri-v8.1 | Hurricane RI | active | 2026-03-13 | 0.938 | 0.034 | 50-bag logistic + histogram GBT (d3+d4) + Platt cal. |
| hp-tornado-meso-v1.4.2 | Tornado (day-ahead) | active | 2026-03-05 | 0.644 | 0.231 | Random forest + Platt cal. |
| hp-tornado-coherence-v1 | Tornado (storm-object) | research | 2026-03-15 | 0.894 | -- | Coherence Field Theory + Platt cal. (41 features) |
Promotion / retirement history
| Date | Action | Model | Reason |
|---|---|---|---|
| 2026-03-05 | promote | hp-tornado-meso-v1.4.2 | +0.02 AUC, improved sharpness on 2025 spring holdout |
| 2026-03-05 | retire | hp-tornado-meso-v1.4.1 | Superseded by hp-tornado-meso-v1.4.2 |
| 2026-03-13 | promote | hp-hurricane-ri-v8.1 | Advisory-time eval: AUC 0.938 avg (0.910-0.957), BSS +0.29, +0.009 delta vs OFCL across 5 holdouts. 65 features, real NOAA SST, Platt calibration. |
| 2026-03-13 | retire | hp-hurricane-ri-v3.2 | Superseded by hp-hurricane-ri-v8.1 - v3.2 used IBTrACS best-track eval (AUC 0.84); v8.1 uses honest advisory-time protocol |
| 2026-03-01 | promote | hp-hurricane-ri-v3.2 | +0.04 AUC, -0.015 Brier vs hp-hurricane-ri-v3.1 on holdout |
| 2026-03-01 | retire | hp-hurricane-ri-v3.1 | Superseded by hp-hurricane-ri-v3.2 |
| 2026-02-28 | promote | hp-etas-v3.2.1 | +0.03 AUC, -0.011 Brier vs hp-etas-v3.1.0 on holdout |
| 2026-02-28 | retire | hp-etas-v3.1.0 | Superseded by hp-etas-v3.2.1 |
| 2026-02-15 | reject | eq-v3.2.0 | Failed calibration floor G4 on validation set (Brier 0.261) |
| 2026-02-10 | reject | to-v1.4.0 | Regression in false alarm rate (+8%) on CONUS spring holdout |
Retired models
| Model | Hazard | Retired | Final AUC | Final Brier | Forecasts issued |
|---|---|---|---|---|---|
| hp-etas-v3.1.0 | Earthquake | 2026-02-28 | 0.78 | 0.153 | 2,841 |
| hp-hurricane-ri-v3.2 | Hurricane RI | 2026-03-13 | 0.84 | 0.168 | 1,891 |
| hp-hurricane-ri-v3.1 | Hurricane RI | 2026-03-01 | 0.80 | 0.183 | 1,247 |
| hp-tornado-meso-v1.4.1 | Tornado | 2026-03-05 | 0.75 | 0.124 | 3,102 |
Promotion policy
Requirements for promotion
- Statistically significant improvement on holdout set (AUC or Brier)
- No regression in calibration or false alarm rate
- Passes all gate checks (G0-G12) on validation corpus
- Training data cutoff documented and frozen
- Feature set and architecture documented in registry
Rejection criteria
- Brier score above calibration floor threshold (0.25)
- False alarm rate regression exceeding 5%
- Calibration error exceeding 8% in any bin
- Inability to produce replay-verified deterministic outputs
- Missing or incomplete feature documentation