Winter weather
Snow forecast models: which to trust, and when
Last updated: 14 November 2026
Every winter the same thing happens: one model run shows 40 cm of snow, the next shows rain, and social media fills with maps of both. Snow is the hardest everyday forecast there is. This guide explains why, which models are worth checking at each stage of an approaching storm, and how to read them together.
Why snow is so hard to forecast
A snowfall forecast is three forecasts multiplied together, and an error in any one of them changes the answer.
First, how much precipitation falls. Precipitation is already the least reliable quantity in any weather model, because it depends on small-scale processes the grid cannot fully resolve.
Second, whether it falls as snow. Near freezing, a difference of one degree through a shallow layer of the atmosphere decides between snow, sleet, freezing rain and plain rain. A storm track that shifts 50 km can move that line across a whole city.
Third, how much snow that precipitation makes. The familiar rule that 1 cm of water gives 10 cm of snow is only an average. Wet snow near freezing can be closer to 5 to 1. Cold, dry powder can exceed 20 to 1. The same amount of water can produce a slushy coating or a deep, fluffy cover.
Why model snow maps differ so much
Many of the snow maps shared online are not direct model output. They are calculated afterwards from the model’s precipitation and temperature, and different websites use different methods. A map that assumes a fixed 10 to 1 ratio and counts sleet as snow will show far more than one that estimates snow density and melting.
So two maps “from the GFS” can disagree with each other before you even compare the GFS with another model. When you look at a snow map, check what it actually shows: snowfall, snow depth, or a simple conversion.
Five to ten days out: is there a storm at all?
At this range, ignore amounts completely. The only question worth asking is whether the global models agree that a storm system will exist in roughly the right place.
Check ECMWF, the GFS and, out to a week, ICON and the UK Met Office model. If one run of one model shows a major snowstorm and the rest do not, it is a possibility, not a forecast. If several independent models show a low in a similar position on consecutive runs, the signal is real, even though the details will change.
Ensembles help here. If most ensemble members of ECMWF and the GFS produce the storm, confidence in the event is growing.
Three to five days out: track and temperature
Now the question becomes where the low will track and where the cold air will be. This is where the global models are at their most useful, and where comparing them matters most.
Look at the pressure pattern and the temperature fields, not snow totals. A cluster of models taking the low along the same path means the rain–snow line is becoming predictable. A wide spread in track means the forecast for your location could still go either way, and anyone quoting precise amounts is guessing.
ECMWF is the best single guide at this range on average, but its lead is not large enough to overrule several other models that agree with each other.
One to two days out: amounts and bands
Inside 48 hours, high-resolution regional models take over. Models with grid spacing of 1 to 3 km — the HRRR in the United States, ICON-D2 in central Europe, AROME in France, the UK Met Office’s UKV — can represent the narrow bands of heavy snow that produce the biggest totals, as well as lake-effect snow, upslope snow on mountains and cold air trapped in valleys.
Global models at 9 to 13 km will still have the overall storm right, but they smooth out those features. For final amounts, the regional models and your national weather service’s forecast, which blends all of this with local experience, are the best sources.
Which model is best for snow?
There is no consistent winner. Verification studies of winter storms show the best-performing model changing from one event to the next. What does hold up is that a blend of several models is more reliable than any single one, and that agreement between independent models is the best available measure of confidence.
Elevation is a further complication. A global model sees a smoothed version of the terrain, so a town in a valley and a ski resort 1,000 metres above it can share one grid cell. In mountains, finer models and local knowledge matter far more than the choice between two global models.
How to check a winter storm on MeteoModels
MeteoModels overlays the global models for precipitation, temperature, wind and pressure. For an approaching winter storm, a useful routine is:
- Switch to pressure and overlay ECMWF, GFS and ICON to compare the track of the low.
- Switch to temperature and look at where the models place the freezing line.
- Switch to precipitation and turn on the certainty mask: bright areas are where the models agree on where it will fall.
- Tap your location for the multi-model consensus and the spread between models.
- Repeat with each new run. A forecast that holds steady from run to run is more trustworthy than one that jumps.
The short version
A week out, ask whether the models agree a storm exists. Three to five days out, compare track and temperature across the global models. Inside two days, use high-resolution regional models and your national weather service for amounts. At every stage, trust agreement over any single dramatic run.
Frequently asked questions
Which weather model is most accurate for snow?
No single model is consistently best. ECMWF is the strongest on average for the storm track several days ahead, and high-resolution regional models are best for amounts within one to two days. A blend of models is more reliable than any one.
Why do snow forecasts change so much?
Snow totals depend on precipitation amount, a temperature profile close to freezing and the snow-to-liquid ratio. A small change in storm track or temperature shifts all three, so totals can swing widely between model runs.
How far ahead can snow be forecast accurately?
The existence of a storm can often be seen five to seven days ahead. Reliable snowfall amounts for a specific place are usually only possible one to two days ahead.
What is the snow-to-liquid ratio?
The depth of snow produced by a given depth of water. 10 to 1 is a common average, but it ranges from about 5 to 1 for wet snow to more than 20 to 1 for cold, dry powder.