Here is a situation that frustrates everyone who checks snow forecasts: one site says 35 cm for the same resort and the same week where another says 130 cm. Both claim to be “using the models”. So what is actually going on?
The short version: snowfall is not something the model hands you. It is derived from three things — how much precipitation falls, what fraction of it falls as snow rather than rain, and how light or dense that snow is. Two of those three depend heavily on elevation. The first one is where the models genuinely disagree.
It starts with precipitation, and models disagree a lot
Take Saas-Fee, a high, dry valley in the Swiss Alps. On a recent run, the four global models we pull forecast anywhere from 32 mm to 139 mm of water over seven days. Same resort, same week, a four-fold spread. The highest-resolution model (ICON, 7 km) said 32 mm. The coarser ones — GFS, GEM, ECMWF — said 91 to 139 mm.
Why the gap? Global models are coarse, and mountains are small. A 13 km grid cell cannot see the rain shadow behind a ridge, so it spreads moisture where a real valley stays dry. The finer the model, the better it resolves that. ICON, the finest of the four, was the closest to the truth.
Then you convert water to snow
Once you have precipitation, you work out how much of it is snow. That depends on the freezing level: snow reaches roughly 300 metres below it. At the summit of a high resort nearly everything falls as snow; at the base a lot of the same storm can fall as rain.
Then there is the snow ratio — how many centimetres of snow you get per millimetre of water. Wet snow near freezing might be 9:1; cold, dry snow can be 15:1 or more. Get the ratio wrong on a big precipitation number and the error multiplies.
What we changed
We used to average the models equally, which let the wet, coarse models pull the number up. Two changes fixed most of the overprediction:
- We now weight the models by resolution. ICON (7 km) carries three times the weight of GFS or GEM, and ECMWF carries twice.
- We cap the snow ratio at 12:1 instead of 18:1, which was too generous for most storms.
- We scale the raw precipitation down — global models systematically over-forecast rain and snow over mountains, so we take a conservative fraction of it.
On that Saas-Fee example, the headline summit number came down from about 128 cm to 92 cm. Still a big forecast — three of the four models genuinely saw a wet week — but no longer inflated by the worst offenders.
How to protect yourself
Look at the range, not just the headline. We show the spread between the wettest and driest model on every resort page. A wide range means low confidence, whatever the headline says. And always check the elevation that matters to you: the summit number is the most flattering one.