# Why snow forecasts disagree, and how to read the spread

> Why weather models give one mountain different snow totals, how grid size and ensembles explain it, and how to read the spread yourself.

Source: https://skibudsapp.com/guides/why-snow-forecasts-disagree
Updated: 2026-10-17

Snow forecasts disagree because no weather model sees the mountain you ski, and each one blurs it differently. NOAA's global ensemble divides the world into squares about 25 kilometers across, each close to 20 times the area of [Whistler Blackcomb](https://skibudsapp.com/resorts/whistler-blackcomb), the largest ski area in North America, so a whole resort averages into one gentle bump. Its sharper HRRR model uses 3-kilometer squares but looks no more than two days ahead. On top of that, every model starts from a slightly different picture of today's atmosphere and turns water into snow its own way. So the useful question is how far apart the forecasts sit, how many days out you are, and whether the gap is closing, rather than which number is right.

## Even a tight ensemble is right less often than it looks

NOAA runs its Global Ensemble Forecast System as 31 versions of the same forecast, out to 16 days, each started from a slightly different version of the current atmosphere. Where the runs bunch together, small errors in the starting picture aren't changing the answer much. Where they fan out, the honest forecast is a range.

NOAA's own [GEFS page](https://www.emc.ncep.noaa.gov/emc/pages/numerical_forecast_systems/gefs.php) shows how much that bunching is worth. It walks through one run from November 4, 2020. Where the members agreed more closely than they had on 9 of every 10 forecasts over the previous 30 days, forecasts with that much agreement had landed in the same range as what actually happened only 61 to 84% of the time. NOAA's explanation is one sentence: "the underlying forecast model is imperfect, the initial conditions are imprecise, and the atmosphere behaves chaotically."

That check measured the upper-air pressure pattern about 18,000 feet up, five days ahead, over one 30-day stretch in late 2020. It says nothing directly about snowfall at a single resort, which depends on that pattern and on everything between it and the ground. Agreement is worth having. It isn't a promise.

## Every member of the ensemble sees the same blurred mountain

All 31 runs share one model, so they share its blind spots, and for skiers the biggest is grid size. A model divides the world into boxes and gives each box one elevation, one temperature and one snowfall amount. A box about 25 kilometers on a side covers roughly 600 square kilometers; Whistler Blackcomb covers 33. Inside that box, a whole resort and the valleys around it average into one gentle bump. That bump is lower than the real summit, so the model can put rain where the upper lifts get snow, or spread snow evenly across a range when the windward side should get most of it; [orographic lift](https://skibudsapp.com/guides/orographic-lift-and-snow) explains why that matters.

Finer models see more terrain, at a cost. NOAA describes the [High-Resolution Rapid Refresh](https://rapidrefresh.noaa.gov/hrrr/) as a "3-km resolution, hourly updated" model, with forecasts extended to 48 hours every six hours. It resolves ridges and valleys a global model flattens, and it stops before the weekend you're trying to plan on a Tuesday.

Smaller differences pile on top. A storm arriving six hours later can move most of its snow from Friday into Saturday, so two forecasts can agree on the storm and disagree on any single day. A degree or two of temperature moves the snow level by hundreds of feet. Models turn the same inch of water into 10 inches of snow or 18, which [snow-to-liquid ratio explained](https://skibudsapp.com/guides/snow-to-liquid-ratio) covers. And some sources report a calendar day, others the next 24 hours from now, others overnight, and those totals can't be compared directly.

## Measure the spread yourself

You don't need an app to see how far apart the forecasts are. Collect each source's snow total for the same day over the same window, a calendar day against a calendar day and never against the next 24 hours. Sort them. With five sources, set aside the highest and the lowest; the three left span roughly the middle half of the range, and the bottom of that span is the number to plan on. Then watch two things as the day gets closer: whether the span narrows, and whether it moves. For US mountains, the National Weather Service publishes [probabilistic snowfall forecasts](https://www.weather.gov/prob-snow) with a most likely, minimum and maximum amount for winter storms, a ready-made range.

The rebuilt SkiBuds app, coming soon, will do this arithmetic for you. For a US resort, it will line up forecasts from the National Weather Service, MET Norway and WeatherAPI for the same local date and window, and give each day a 0 to 100 confidence score built from lead time and the middle-half spread of the sources' snowfall, temperature and wind. With its optional ensemble source switched on, it will also count how many ensemble members forecast at least your powder-alert threshold. Both numbers describe agreement, not the chance that a particular amount will fall.

## One Saturday at Alta, read from Wednesday to Friday

Here is a **hypothetical** week at [Alta](https://skibudsapp.com/resorts/alta). It's Wednesday, and five sources give these snow totals for Saturday: 3, 5, 6, 9 and 14 inches. Set aside the 3 and the 14, and the middle half runs from 5 to 9 inches.

Saturday is still three days out, and forecast skill fades with every day of lead time, so read Wednesday's numbers as permission to pencil the day in and nothing more. The plan worth making works at 5 inches and gets better at 9.

On Thursday the same sources say 6, 7, 8, 10 and 12. The middle half has moved up to 7 to 10, and the low outlier is gone. A forecast that rises and tightens from one day to the next is more convincing than a single big number, because new data is pushing the models the same way.

On Friday, Saturday is tomorrow, and the sources read 7, 8, 9, 9 and 11. The middle half is 8 to 9. That's the moment to commit, so confirm the carpool and the meeting spot. If Friday had instead shown the middle half sliding back to 3 to 6, the better move would have been to keep it an ordinary ski day and skip the dawn start.

The rule underneath that week is short enough to say in the car. **Pencil in a day from the week-out outlook, plan for the bottom of the middle half, and commit when the day becomes tomorrow and the range has stopped moving.** [How to read a ski resort snow forecast](https://skibudsapp.com/guides/how-to-read-snow-forecasts) covers the elevation and timing checks that go with it.

## Sometimes the high number is the right one

The method has a blind spot, and Colorado's March 2003 storm shows where it is. In a [study of that storm](https://www.weather.gov/media/gjt/Research/Exceptional_Mesoscale_2003.pdf), meteorologists describe the main run of NOAA's global model repeatedly forecasting a more northern track than other models, and "a far less significant potential storm" for the area that was eventually hit hardest. About a third of the same model's ensemble members had the storm's upper low in the right place more than three days ahead, and other centers' models, the European one among them, kept the storm farther south. Within NOAA's model, the minority was right: the [National Weather Service in Boulder](https://www.weather.gov/media/bou/March03Snowstorm.pdf) recorded 31.8 inches in Denver, its biggest storm since 1913, and 87.5 inches in the foothills. The authors offer three possible reasons NOAA's main run missed, and one of them is this page's blurred mountain: the terrain it resolved was "considerably less detailed" than that of the Eta, NOAA's regional model, which had kept the storm on its southern track. The other two were how the runs were started and differences in the models' inner workings, and the paper doesn't say which mattered.

Setting aside the highest total treats it as noise, and in 2003 the big numbers weren't noise. A third of one model's runs and several other models pointed to the big storm, which makes a second scenario with real support. So before you set the top total aside, check whether it stands alone or has company, and if the company grows from one day to the next, keep the big storm in the plan as a real possibility.

There's a quieter version of the same problem: five apps aren't always five opinions. For North American mountains, [MET Norway's forecasts](https://docs.api.met.no/doc/locationforecast/datamodel) come from the European Centre's high-resolution model, so if another of your sources draws on that model too, a tight spread can be one forecast counted twice.

Once snow starts falling, observations beat every model, and [snow depth vs snowfall vs SWE](https://skibudsapp.com/guides/snow-depth-vs-snowfall) explains how to read them. The last step of that hypothetical week happens at 6 a.m. Saturday, when Alta's morning report replaces all five forecasts. If it reads 8 inches or more, you got the day you planned for. If it reads 4, the sources were wrong together, which a tight spread can never rule out, and the time to switch to the backup plan is before you start the car.