Model accuracy · method
How we grade weather model snow forecasts
4070 compares historical model snowfall estimates with observer reports and NOAA's gridded snowfall analysis, using model runs selected before each measurement day, and ranks models by their average miss for each region and mountain range.
Method updated October 14, 2026.
The measurements and estimates
Town-area rankings use snow observers: CoCoRaHS volunteers, National Weather Service cooperative observers and official weather stations. Mountain and regional rankings use NOAA's National Snowfall Analysis, a gridded estimate built from observations and model guidance. Mountain rankings are checked against NRCS SNOTEL snow-depth gain, which is affected by settling and is not a fresh-snow measurement. Resort snow reports are not used; agreement with the analysis is not independent resort validation.
Each place's storm total adds its complete reporting windows. Missing reports remain unavailable. The public map and headline totals show new-snow observer reports; area rankings can use a different set of locations and a gridded estimate. Each result names its reference data.
What Day 1, Day 3 and Day 5 mean
Day 1 uses the newest model initialization within the 24 hours before each measurement day starts. Day 3 uses the 48-72 hour bracket before that day, and Day 5 the 96-120 hour bracket. Daily estimates are added over the storm, so later days can use runs initialized after the storm began. These retrospective archive comparisons do not establish when a forecast was available to users.
A model needs a complete estimate for every measurement window in the storm to enter a comparison. These pages show lead-day brackets 1, 3 and 5; members see available leads from 1 to 15. A missing comparison can reflect incomplete data as well as a model's forecast horizon.
How model snowfall is calculated
Town and area estimates use published snow depth where available. Snow-water fields are converted at 10:1; GFS and GEFS estimate snow water from precipitation and a snow-category field. Models without a snow field use precipitation at 10:1. Resort and pass comparisons use height-adjusted historical estimates; ensemble estimates from averaged inputs can differ from Ski Forecast, which converts each member before averaging.
Ensemble means average a model's members and appear as separate entries. Historical archive reconstructions can differ from the forecast customers saw at the time. Only customer-visible models are ranked.
Ranking and the clear-winner rule
Area models are ranked by their average absolute difference from the reference snowfall, using the same comparison locations at that lead. Models need forecasts at 90% of eligible locations; only locations covered by every retained model are compared. A smaller average miss ranks higher.
A clear winner needs at least five comparison locations, a margin over the runner-up of at least the larger of half an inch or 10% of the mean reference snowfall, and a paired 90% uncertainty interval above zero. Mountain and regional rankings also require the same top model in the observation cross-check. Otherwise the result is too close to call. Areas averaging under an inch do not receive a snow-event winner.
Limits
Snow varies across mountain elevations and exposures. Snow-depth gain is affected by settling, and the snowfall analysis uses model guidance, so neither is an independent resort new-snow measurement. A matching winner does not remove these limits.
The published storm sample does not evaluate all dry days or all false alarms. Nearby locations and models can be correlated. One storm does not establish general forecast skill, and preliminary results can change when reports are completed.
Questions about the method
How far ahead are forecasts graded?
Day 1 uses the newest model initialization within the 24 hours before each measurement day starts. Day 3 uses the 48-72 hour bracket before that day, and Day 5 the 96-120 hour bracket. Daily estimates are added over the storm, so later days can use runs initialized after the storm began. These retrospective archive comparisons do not establish when a forecast was available to users. Members see available lead days from 1 to 15.
Which snowfall number is used for each model?
Town and area estimates use published snow depth where available. Snow-water fields are converted at 10:1; GFS and GEFS estimate snow water from precipitation and a snow-category field. Models without a snow field use precipitation at 10:1. Resort and pass comparisons use height-adjusted historical estimates; ensemble estimates from averaged inputs can differ from Ski Forecast, which converts each member before averaging.
What does "too close to call" mean?
It means the smallest average miss did not pass the separation test, or the required second comparison was missing or disagreed. A clear winner needs at least five comparison locations, a margin over the runner-up of at least the larger of half an inch or 10% of the mean snowfall, and a paired 90% uncertainty interval above zero. Mountains and regions also need the same top model in the observation cross-check. This describes the selected storm and reference data, not general forecast superiority.
Why are mountains checked against both an analysis and sensors?
Snow sensors measure depth gain, which is affected by settling and is not fresh snowfall. The snowfall analysis fills gaps between gauges using model guidance. A mountain range has a clear winner only when both comparisons name the same top model; agreement does not remove their measurement limits.
Where do the measurements and estimates come from?
Town-area rankings use snow observers: CoCoRaHS volunteers, National Weather Service cooperative observers and official weather stations. Mountain and regional rankings use NOAA's National Snowfall Analysis, a gridded estimate built from observations and model guidance. Mountain rankings are checked against NRCS SNOTEL snow-depth gain, which is affected by settling and is not a fresh-snow measurement. Resort snow reports are not used; agreement with the analysis is not independent resort validation.
Measurements and estimates
- CoCoRaHS. Volunteer snow observers
- NWS Cooperative Observer Program. Daily snowfall from cooperative observers
- NRCS SNOTEL. Mountain snow-depth gain; not a fresh-snow measurement
- NOHRSC National Snowfall Analysis. NOAA's gridded snowfall estimate, including model guidance
Licences and required credits are on the data credits page.