Scored from … onwards · best_match + ECMWF blend (matches production) ·
GFS, ECMWF, and 24h-persistence shown as independent baselines
PHFN-7 forecasts remain unchanged. This view tests whether versioned,
location-derived surface features explain residual error on the same
observation sample used for Best Match.
Advanced location-balanced and confidence diagnostics
Observation-Conditional Extremes
RMSE and Bias grouped by hour of day (UTC) across all lead bins.
Reveals systematic daytime/night-time model errors.
Categorical verification of cloud cover using sky condition classes (SKC · FEW · SCT · BKN · OVC).
HSS and Gerrity Score replace RMSE here — continuous metrics are invalid for discrete METAR observations.
Manual Pipeline Triggers
Enter your VALIDATION_ADMIN_TOKEN and click a step. Each step may take 30–90 s.