Every number on this site is one of three things, and the page says which: a model's own output, a quantity derived here, or an observation.
An aerialway way is the one object in OpenStreetMap whose two ends are, by construction, the bottom and the top of ground somebody rides. Reading the terrain under all 4,183 lower-48 lifts and taking each area's lowest and highest lift terminal gives, against published resort specifications: Jackson Hole −62 ft, Big Sky −54, Vail +52, Breckenridge +34, Park City +13, Steamboat −41, Beaver Creek −31, Telluride −89. Median absolute error about fifty feet on a 27 m DEM. The residual is mostly one-signed and explainable: a resort quotes its peak, this measures its highest terminal.
The alternative — a box around a map pin — ranks a day lodge, an avalanche training area and Tuckerman Ravine among the biggest mountains in the country, and gives Colorado 49 ski areas where there are about 34.
Each resort is cut into 500-foot bands over the lift-served footprint: the convex hull of the uphill lift lines, buffered 400 m. Each band carries its area, mean slope, mean aspect and the PRISM winter normal at that height. A band's aspect is the direction of the mean unit vector, not the mean of the angles — 350° and 10° average to 0°, not to 180°.
Measured on 432 paired western SNOTEL stations (<25 km apart, >300 m apart in height) over DJF 2025–26 — 37,582 station-pair-days:
| Mean | Median | 10th | 90th | Inverted | Within ±1 of 6.5 | |
|---|---|---|---|---|---|---|
| Daily mean T | 3.33 | 4.31 | −4.29 | 9.37 | 22.5% | 18.2% |
| Daily min T | 1.53 | 2.60 | −9.03 | 10.07 | 36.3% | 12.4% |
| Daily max T | 6.00 | 6.67 | −0.95 | 12.03 | 11.8% | 17.7% |
K/km; "inverted" means the higher station was warmer. 6.5 K/km is right about one winter day in six and costs 7.7 °F at the summit of a 3,000-foot mountain before the model contributes any error of its own. Overnight the mountain is upside down more than a third of the time, and on those days the assumption misses the summit by a median 16.6 °F.
A better constant does not help. The regional observed mean beats 6.5 by only 7%, because the daily spread is enormous (σ = 5.87 K/km), and surface predictors explain r² = 0.07. So the profile cannot come from a constant or from a surface station: it has to be read out of the model's own resolved vertical structure, which is the only thing that knows an inversion is on.
Caveats. SNOTEL sites sit in mountain clearings that pool cold air, so this likely over-represents inversions relative to an exposed slope; the pairs are western only; and the method attributes the whole difference to height when some of it is horizontal.
Precipitation uses PRISM 800 m monthly normals the way the NWS's own Mountain Mapper QPE does — as the shape of terrain precipitation a model cannot resolve: R(x) = PRISM(x) / <PRISM> over the model's own footprint. Dimensionless, about 1 over a footprint, so it conserves the model's total while redistributing it the way four decades of gauges say the terrain does.
The footprint must be the model's effective resolution, not its grid spacing. A mesoscale model does not resolve a feature one cell wide — effective resolution is roughly 6–7Δx. Over a 3 km box HRRR's ratio comes out 1.00–1.03 and the correction appears to do nothing; over the ~20 km HRRR actually resolves it spans 0.80–1.30. A ski mountain is about 5 km wide — entirely inside the unresolved band.
What R does not do. It is climatological, so it encodes the average storm's wind direction. A storm from an unusual quarter loads the other side of the range, and R will confidently enhance a slope that is that day in the lee. It is blended with a flow-dependent upslope term from the model's own low-level wind, and it is published per band rather than hidden inside a number.
HRRR 3 km / 48 h, the convection-allowing workhorse and the only model in the network's archive carrying the full winter suite; it resolves a lake-effect band, which matters more here than anywhere. RRFS 3 km / 18 h, HRRR's replacement, carried alongside rather than instead, because the two disagree about band placement more than about totals. NBM 2.5 km / 10 days, the calibrated blend. ECMWF 0.25° / 10 days, the synoptic check.
The trap. asnow in HRRR and RRFS is GRIB "Total snowfall" in metres of snow depth — the model's own snow-to-liquid ratio has already been applied. ECMWF's sf is "Snowfall" in metres of water equivalent. They are different physical quantities, both called snowfall and both delivered in metres, and plotting them on one axis understates ECMWF by roughly the snow-to-liquid ratio — a factor of ten to twenty. Everything here is carried as both, and which one a number is is never implicit.
The winter suite is only written on the 6-hourly synoptic cycles, though these models run hourly, so a 15Z RRFS has temperature and no snowfall. Accumulations in this archive are run totals; difference them for a rate, never sum them.
Aspect, not exposure. Bands carry slope and aspect from the DEM. Wind loading, scouring and where the cornice builds are not modelled, and a 13 m DEM cannot see a gully that holds snow. A band is not a trail: the forecast resolves 500 vertical feet of a mountain, not a named run. Footprint acreage is a lower bound on marketed acreage. Eastern verification is thin: SNOTEL is a western network, and east of the Mississippi the fallback is NOHRSC, COOP and ASOS, all coarser and none at summit height. Snow-to-liquid ratio is the model's, and it is the largest single error term in turning liquid into inches; it is carried through here, not improved.