How we predict snow

1. Start from the best weather models

Every forecast blends four global weather models hour by hour: ECMWF (European), GFS (US), ICON (German) and GEM (Canadian). Where a high-resolution regional model covers the resort (HRRR in the US, HRDPS in Canada, AROME in France and the Alps, ICON-D2 in central Europe, MSM in Japan), we add it at double weight, because it sees mountains far better than a global grid does.

How much the global models disagree about the next three days of summit snowfall sets our confidence. That spread gives the low, likely and high storm totals and the chance of a powder day.

2. Bring the weather down to each run

A weather model sees a mountain as a few grid boxes. Skiers ski runs. So we map every run from OpenStreetMap and measure its top, bottom, steepness and the direction it faces. Resorts without mapped runs get 12 virtual slopes: three elevation bands times four directions.

Then, at the top, middle and bottom of each run, hour by hour:

• Temperature follows a lapse rate fitted to the model's freezing level, so the summit can be snowing while the base rains.
• Rain or snow splits between −0.2 °C and 2.2 °C. Snow density follows the Kuchera snow-to-liquid ratio: cold storms give light powder, warm ones give heavy cement.
• More snow higher up: about 4% more precipitation for every 100 m of elevation.
• Wind loads snow onto sheltered slopes and scours exposed ones, more so above treeline.
• Sun: the real sun position on each slope drives melting, overnight crusts and spring corn.
• Groomers and skiers: grooming at 4 a.m. on groomed runs, skier traffic from 9 to 4 depending on difficulty.

The result is classified into a surface you would recognize (deep powder, corduroy, corn, crust, ice) and scored 0 to 100. Runs with too little snow show as predicted closed.

3. Look further ahead: the season outlook

Beyond 10 days, weather forecasts stop being useful, so for later dates we forecast the odds, not the weather. Three layers, each checked against the others:

• What is normal. The last 10 seasons of reanalysis (ERA5) at each summit, week by week. We add what has already fallen this season and the 7-day forecast. Comparing a resort with its own history cancels most of the bias in reanalysis data.
• What the seasonal model expects. ECMWF's seasonal forecast (SEAS5, issued on the 5th of every month) predicts each month's snowfall against its own climate. A ratio of 1.2 means 20% snowier than normal. It already includes El Niño, La Niña, ocean heat and the rest of what the model knows. We use it for every resort, October to April.
• What El Niño has done before. For each resort and month we compare snowfall with that winter's December-to-February Oceanic Niño Index (NOAA) across the 44 winters from 1981-82 to 2024-25, and ask what the fit says for a strong El Niño (+2.2, like 1997-98; the 2015-16 record reached +2.5, and NOAA forecasts a very strong event this winter). We keep that estimate only where leave-one-winter-out testing shows it predicts better than assuming a normal month.

The forecast is the normal times a weighted ratio: the average of the two estimates' departures from normal (the El Niño fit counts as normal where it has no skill), scaled by a calibration factor, and capped at half to one and a half times normal.

How we weight them. Open-Meteo archives ECMWF's seasonal forecasts only from September 2025, so winter 2025-26 is the one winter we can score both estimates on, for every resort-month from October to April (against what ERA5 says fell, and out-of-sample for the El Niño fit). In the Rockies (287 resort-months, 41 resorts):

• ECMWF alone did slightly worse than assuming a normal month (skill -0.01) and called the right side of normal 57% of the time. It overshot: the scale that would have worked best was 0.45.
• The El Niño fit alone was level with a normal month, and mostly stayed quiet, as designed.
• The average of the two did best: skill +0.03, 60% right, and a best scale of 0.88.
• Outside the Rockies (US and Canada, 70 resort-months so far), ECMWF alone scored -0.12 and needed a scale of 0.12, so it deserves far less weight there.

The calibration factor for each region is the measured scale of the average, held between 0 and 1 and pulled halfway to a skeptical 0.5, because one winter is thin evidence (and it was a weak La Niña-neutral winter, when seasonal models have the least to go on). Today that is 0.69 for the Rockies, 0.36 for the rest of the US and Canada, and 0.5 where we have not scored yet. The Snow page uses the same weights, so its seasonal outlook is calmer than ECMWF's raw numbers: Heavenly's raw +60% for January becomes about +10%.

What the El Niño testing shows. In the Wasatch, University of Utah atmospheric scientists told us El Niño has little practical predictive skill for snowfall there and that reanalysis precipitation does not resemble reality, so Utah resorts get no El Niño adjustment and are marked low confidence. In the Rockies El Niño is a weak, patchy signal. Only 13% of resort-months beat a plain normal in 44-winter testing, mostly December and March, and where they do, El Niño explains about 5% of the month-to-month swing (up to 17% at best). So for most months the seasonal model carries the forecast, and we say so instead of forcing an El Niño story. The two methods point the same way in about three-quarters of the resort-months where both show a real signal.

Where it runs. The '26-27 Snow Report uses it for the Rockies today. Heavenly, the Northeast, Lutsen and the rest of the US, the Alps, and Japan follow as the history is processed and tested region by region; the Andes and the rest of the Southern Hemisphere wait for the May 2027 seasonal run, because El Niño is expected to fade before their 2027 winter. These outlooks are odds, not storm forecasts.

4. Check ourselves every day

How we intend to improve

1. Verify outside the US. SNOTEL only covers the western US. Next we are adding Canadian and European station networks, so the skill-weighted blend works as well at Whistler and Zermatt as it does at Snowbird.
2. Correct each resort's bias. Some resorts are consistently over- or under-forecast: lake effect, rain shadows, narrow valleys that models smooth out. Once a resort has enough verified seasons, we will correct its snowfall forecast for that known bias.
3. Weight models by lead time and storm type. Today skill is measured on days 1 and 2. The model that is best tomorrow is not always best five days out, or in every kind of storm. We plan to weight models by lead day and by storm type.
4. Score the surface, not just the snowfall. Snowfall is easy to measure; corduroy and crust are not. We want to check our surface and open/closed predictions against resort grooming and lift reports, and against skier feedback.
5. Account for settling. Snow-depth gain undercounts new snow because fresh snow compacts. Modeling that settling will make our station scores fairer and our bias corrections more accurate.
6. Score more winters. We scored the seasonal outlook on one winter because that is all that is archived. Every month we will save ECMWF's forecast as issued and score it once the month is over, so each region's calibration firms up winter by winter. We will also test other signals that move snowfall (the Pacific Decadal Oscillation, the Arctic Oscillation) the same way: used only where they beat normal in testing.

Forecasts are guidance, not guarantees, and never a substitute for avalanche bulletins or ski patrol. Spotted a forecast that missed? Tell us. Misses are how the model gets better.

Open the snow forecast