Model Skill
AUC-ROC across held-out storm years. Strict temporal split — held-out years never touch training data.
AUC-ROC by Version
Higher is better. 0.5 = coin flip. 1.0 = perfect discrimination.
EVALUATION: STRICT TEMPORAL SPLIT · HELD-OUT YEARS NEVER TOUCH TRAINING
Calibration · Reliability Diagram
When the model says 15%, tornadoes should occur ~15% of the time.
ISOTONIC-CALIBRATED · v8.0 · BRIER-SCORED VS NWS LSRs
Feature Importance
What the model actually weighs when computing tornado probability — v8.0 XGBoost gain scores.
Top Features · v8.0
Normalized feature gain. Higher = more influential in final probability output.
SPATIAL CLUSTERING OF SIGNIFICANT TORNADO REPORTS (150KM RADIUS) DOMINATES — TORNADOES TRAVEL IN PACKS
Verification Standards
How we measure performance — and what triggers automatic retraining.
0.919
AUC-ROC · Held-out storm days
The model correctly ranks a tornado day above a non-tornado day 91.9% of the time across years it has never seen.
EF1+ / 25MI
Verification target
A hit is any EF1+ tornado within 25 miles of the forecast point. No soft definitions, no EF0 inflation.
90-DAY
Retraining trigger
If rolling AUC drops below 0.87 over any 90-day window, the model automatically enters retraining. No manual intervention required.