Version Log
All models evaluated on held-out storm years with strict temporal splits.
| Version | AUC-ROC | Key Changes | Training Data | Status |
|---|---|---|---|---|
| v5.0 | 0.814 | Initial XGBoost baseline | ERA5 2000–2022 | RETIRED |
| v5.1 | 0.882 | HRRR integration, extended features | ERA5+HRRR 2000–2023 | RETIRED |
| v5.2 | 0.882 | Isotonic calibration, feature pruning | ERA5+HRRR 2000–2023 | RETIRED |
| v8.0 | 0.919 | Spatial clustering, synoptic forcing index, Vortex v2 grid ensemble | ERA5+HRRR+SPC 2000–2025 | OPERATIONAL |
| v9.0 | — | HRRR-native 0.5° grid, full CONUS, no SPC polygon dependency | HRRR 2014–present | IN DEV |
Forecast Pipeline
Five deterministic stages — versioned, logged, and scored against ground truth every run.
Ingest
HRRR output, ERA5 reanalysis, and SPC reports pulled and aligned to a common grid.
Compute
CAPE, SRH, bulk shear, LCL, STP, and spatial clustering features per grid cell.
Infer
XGBoost ensemble generates raw tornado probability for every metro and grid cell.
Calibrate
Isotonic regression aligns output to observed base rates. 15% means 15%.
Verify
Rolling AUC and Brier scored against NWS storm reports. Retraining triggered if performance slips.
v9.0 — In Development
The next generation: full CONUS independence, grid-native resolution, no SPC anchoring.
v9.0 In Development
v9.0 is a ground-up rebuild trained natively on HRRR model output at 0.5° resolution — no SPC polygon boundaries, no region-based aggregation. Full CONUS coverage. Target AUC exceeds 0.93. It runs parallel to v8.0 and will only replace it after independent validation confirms superior performance across all storm types.