How SnowTally works

About the accuracy figures: they come from an after-the-fact test on archived NBM v4.x forecasts, winters 2023-24 to 2025-26 (November–April), with the published scores from the held-out 2025-26 winter. They are not a live track record.

Resort pages will show the chance of 3 and 6 inches of new snow in 24-hour periods (12:00 to 12:00 UTC), days 1–7, for each resort, plus a beta 12-inch figure.

Thresholds

Pages say "at least 6 inches" (6″ or more). NBM's own thresholds are formally "greater than" a value in metres (for example more than 0.1524 m); at the precision snowfall is measured, we treat the two as the same. Because each threshold is calibrated separately, a higher threshold could occasionally come out above a lower one (196 of 131,546 held-out forecasts, all above 96%); we cap each higher threshold at the next lower one, so a page never shows a bigger chance of 6 inches than of 3 inches, and this left the published skill scores unchanged at the precision shown.

Forecast input

NOAA's National Blend of Models (NBM) v5.0, extended runs at 01, 07, 13 and 19 UTC. We take the 24-hour snowfall exceedance chances at the nearest 2.5 km grid cell. NBM v5.0 has no 3-inch threshold, so the 3-inch chance is interpolated between the 2- and 4-inch chances (linear in log-odds against log of the threshold); in our test this beat reading it off NBM's percentiles.

Calibration

Raw chances are recalibrated per threshold with a logistic model whose inputs are the raw chance, lead day, the difference between the target elevation and the grid cell's mean terrain height (USGS 3DEP), and both relative to NBM's median snow level. It is one model with the same coefficients for every station, not fitted per station or per resort. Because one model serves every station, it can be well calibrated on average and still be off at a single station. In the held-out 2025-26 test, calibrated 6-inch chances of 40–60% (average 49%) came true 58% of the time across all stations (1,016 forecasts). At the 35 stations with at least 10 such forecasts, the share ranged from 9% to 100%, and at half of them it was between 54% and 69%. Mid-range chances were therefore somewhat too low on average, and at a single station they are less certain than the pooled scores suggest. Calibrated headline values are computed for the elevation of each resort's reference station, where they can be checked. Upper and lower mountain values use the same model at the highest and lowest lift points and are labelled as estimates, not verified at summit level. Grid cells nearest the top lifts sit a median 844 ft below them across the 27 resorts in our list.

Accuracy: an after-the-fact test

All accuracy figures come from an after-the-fact test on archived NBM v4.x forecasts (v4.1–v4.3, 01 UTC run), November to April, winters 2023-24 to 2025-26. New snow at SNOTEL stations is derived from hourly snow depth, which reads low when snow settles. Because the calibration learns from these observations, calibrated chances inherit some of that low bias, most of all for 12 inches. GHCN-Daily stations report new snow measured by observers. This is not a live record.

The published skill comes from a validation model fitted on 2023-24 and 2024-25 and tested on the held-out 2025-26 winter, pooled over all 119 truth stations (not only the stations used on resort pages), at each station's own grid cell, lead days 1–7.

For 12 inches the improvement over raw NBM is not statistically certain: the 95% interval of the change in Brier score runs from -0.0018 to +0.0011 and includes zero. That is why 12-inch figures are shown as beta.

The test used the 01 UTC run only; pages may show a later run, which was not part of the test. 'November to April' refers to issue dates: scored days in the held-out winter run from Nov 2 to May 7 (up to 187 days). Skill falls with lead time: across all stations, the calibrated 6-inch skill score was +0.44 on day 1 and +0.11 on day 7.

Skill score here is the Brier skill score (BSS) against how often the station itself saw that much snow in the two training winters (2023-24 and 2024-25): 0 means no better than that frequency, 1 means perfect, and below 0 means worse. 'Calibrated BSS' is the score of SnowTally's calibrated chances, 'raw' that of NBM as issued.

The calibration we have prepared for the site (version v1) is the same model refitted on all three winters. Because every winter was used to fit it, it has no separate held-out score; the scores above are for the validation model. NBM v4.x is also not v5.0, which the live site will use, and v1 lowers many v5.0 chances strongly. In early November we will check v1 against our own logged v5.0 forecasts. Only then will we decide whether resort pages show calibrated chances or NBM's own chances; this page will say which, and why. We plan to start live verification this winter and will publish the results here.

Reading the reliability chart

Each resort page will show a chart from the held-out 2025-26 winter. The x-axis is the forecast chance, the y-axis how often that much snow was then observed at the reference station. Points on the dotted diagonal mean the chances were well calibrated; points below it mean over-forecasting. Blue circles with a solid line are calibrated, orange squares with a dashed line are raw NBM. Each day is counted once per lead day 1–7, so one day appears in up to seven forecasts. Bins with fewer than 10 forecasts are left out of the chart; every bin is listed in the chart-data table under it. The chart is drawn by the site build from the same held-out predictions that give the skill scores (calibration fitted on 2023-24 and 2024-25 only).

12 inches

Shown as beta. The 12-inch figure goes through the same calibration as 3 and 6 inches, but twelve-inch days are rare at the reference stations (in the test winter, between 0 and 7 days per reference station), too few to verify it at a single station, and across all stations the improvement over raw NBM is not statistically certain (table above). Treat it as a rough indication: it was tested, but the test could not show that calibration improves on raw NBM.

How pages are made

On resort pages, all numbers and sentences are generated by code from the data. There is no AI-written description of any resort. Explanatory pages such as this one are drafted with AI assistance and reviewed before publishing. SnowTally is ad-funded and has no affiliate links.

Calibration version prepared: v1 (frozen), fitted on all three archived winters. Its use on resort pages will be decided after the NBM v5.0 check in early November. Skill figures on this page come from the validation fit on 2023-24 and 2024-25, scored on 2025-26.