Head to head
Error by lead time
mean absolute error, every hour of the runTime of day
bias by local solar hour, first 24 hours of each run · above zero = too warm / too moistStation by station
last 30 days, leads 1–48 hRain and flight category
equitable threat score by lead · 0 = no skill, 1 = perfect · frequency bias beside itIs the ensemble honest?
six RRFS members, runs from 9 SeptemberAcross the whole map
every 3 km cell against MRMS radar and rain and the URMA analysisRadar and rain
fractions skill score by lead · above the dashed line is useful skillSurface, everywhere
against the URMA 2.5 km analysis, land onlyBias map
day-1 bias on 30 km blocks, last 30 daysHow this is scored
Every 00, 06, 12 and 18Z run of the RRFS since NOAA began publishing it on 12 August 2026, and every HRRR run at the same hours, is read at 2,200 US ASOS airports: the nearest 3 km grid cell (the two models share one grid), no interpolation and no height correction, the same for both. Truth is the airport's own hourly report (NCEI's GHCNh plus aviationweather.gov, as weather.city's Station Interrogator parses them).
Comparisons are on a common sample: an airport-hour counts only if the observation, the RRFS and HRRR all have it. HRRR runs 48 hours at these cycles, so later leads are RRFS alone. The ensemble mean joins from 9 September, when NOAA began publishing the members; its numbers compare runs from then on only.
Wind and gust are in mph. An ASOS report carries a gust only when there is one, so an hour without one is scored against its sustained wind — which is why both models' gusts run high. Rain events are hourly totals at or above 0.01 and 0.10 in. Flight category comes from visibility (capped at 10 miles, as reported) and ceiling: MVFR or worse is a ceiling at or below 3,000 ft or visibility at or below 5 miles; IFR or worse is below 1,000 ft or 3 miles.
Across the whole map, both models are read on their shared 3 km grid and compared cell by cell with NOAA's gridded truth: MRMS composite reflectivity (the scan nearest the top of the hour, as the strongest echo in each 3 km cell) and MRMS gauge-corrected hourly rain (the average over the cell), and the URMA 2.5 km analysis for temperature, dew point, wind and gust, over land only. Runs at 00, 06, 12 and 18Z, at hours 1, 2, 3 and every third hour to 48. Radar and rain use the fractions skill score, which gives credit for putting a storm in the right place give or take the neighbourhood shown (3 to 81 km); 1 is perfect, and above 0.5 plus half the observed frequency counts as useful. Source: fresh.gridscores, /api/v1/gridscores.
From the RRFS going operational at 12Z on 6 October, these scores continue unchanged, so the parallel and operational periods can be compared directly. Numbers are recomputed as each run's last hour is observed; the raw sums are kept, so any window adds up exactly. Source: fresh.scores, /api/v1/scores.