Science

The science behind our forecasts and projections

How boreas2 works, how we test it, and the data behind it.

About us

We are a team of AI/ML engineers and climate researchers who met through our time at Stanford and Cambridge. We are also skiers. Facing a strong El Niño winter, we wanted to know where and when to ski more than two weeks out, and we couldn’t find a tool that told us honestly. So we built one. We are based in Jackson Hole, Wyoming.

What is boreas2?

boreas2 is our own time-series weather forecasting and prediction model. We incorporate AI and physics-based weather models with decades of station records, over millions of data points, to spot snow signals for over 5,000 ski areas across the whole season, so that you can plan.

What makes boreas2 special?

Illustrative values. See individual ski area reports for actual top, middle, and base forecasts. Click to replay.
Mountain-scale modeling, developed for skiers.

Global weather models typically divide the world into squares roughly 10 to 25 km across. A whole ski area, valley floor and summit together, can fall inside one square and get one averaged answer. That’s why a global forecast can show rain at a resort where the upper lifts got snow. boreas2 learned from years of past forecasts set beside what mountain weather stations actually measured. boreas2 focuses on what matters on a mountain: how much new snow, how much rain, and how cold it will be, day by day, at the base, middle, and top of each resort.

Climate signal detection beyond 15 days.

We have built boreas2 to detect periods throughout the season when the weather is more likely to be colder, warmer, or wetter than usual. These periods happen because the jet stream (the fast river of air, 10 km up, that steers storms) sometimes settles into a shape and stays there for days or weeks. Slow-moving parts of the climate can push the jet stream the same way for weeks.

How good are boreas2’s short-term forecasts?

We score ourselves against the world’s best forecasts, and we show you how we do, including where we fall short. Below are some results:

boreas2 (our estimate)OpenSnow’s published results

Average missLower is better
boreas2peaksignal1.33 °C
PEAKS 2.0OpenSnow1.64 °C
NWS forecastUS National Weather Service1.81 °C
WeatherNext 3Google DeepMind1.91 °C
ICONGerman Weather Service (DWD)1.99 °C
IFS HRESECMWF2.02 °C
GFSNOAA2.13 °C
GDPSEnvironment and Climate Change Canada2.17 °C
Big missesLower is better
boreas2peaksignal30.5 %
PEAKS 2.0OpenSnow37.9 %
NWS forecastUS National Weather Service42.5 %
WeatherNext 3Google DeepMind44.3 %
ICONGerman Weather Service (DWD)44.7 %
IFS HRESECMWF46.5 %
GFSNOAA47.7 %
GDPSEnvironment and Climate Change Canada49.5 %
Most often bestHigher is better
boreas2peaksignal55.4 %
PEAKS 2.0OpenSnow43.1 %
NWS forecastUS National Weather Service38.2 %
WeatherNext 3Google DeepMind36.1 %
ICONGerman Weather Service (DWD)35.2 %
GFSNOAA33.3 %
IFS HRESECMWF33.0 %
GDPSEnvironment and Climate Change Canada31.9 %

boreas2 misses the least of any forecast in these charts, for temperature, wind, and precipitation (rain and snow), and it also has the fewest big misses.

How we score. OpenSnow published a test of its PEAKS 2.0 model against six leading forecasts. We score boreas2 under the same rules. We can’t rerun their test ourselves, so our row is an estimate. Note these results are only for US weather stations.

How good are boreas2’s seasonal forecasts, beyond 15 days?

We know of no other forecast that projects this far ahead for ski resorts, so there is no leaderboard to put ours on. Instead we test them against past winters. The further ahead we look, the harder it is to predict conditions accurately. So, instead of providing false precision, boreas2 detects signals for periods that are more likely than usual to be colder, warmer, or wetter, and we show what “usual” has meant for each ski area in past winters.

A period is a run of days, usually 5 to 14, that’s more likely than usual to be colder, warmer, or wetter for those dates. A lean is a longer, gentler tilt over three to four weeks.

Every signal on the site carries its own record. Here is the summary over 20 past winters (2004–05 to 2023–24), re-forecast with the same model as if each one were live, so that you know how much to trust the long-range signals.

15–46 days ahead
Snowy or wet periodabout 8 in 10 came true
Cold periodabout 7 in 10 came true
Warm periodabout 7 in 10 came true
Wet leanabout 7 in 10 came true
Cold or warm leanabout 6 in 10 came true
47–100 days ahead
Warm periodabout 8 in 10 came true
Cold periodabout 8 in 10 came true
Cold leanabout 7 in 10 came true
Wet leanabout 6 in 10 came true

Signal strength. Each signal is tagged Strong, Medium, or Weak, depending on how much more often calls like it came true in past winters than that weather happens anyway. A strong signal came true far more often than usual. A weak signal came true only slightly more often than usual. We still show it, but plan around it lightly.

Data sources

Our forecasts and projections come from our own model, boreas2, built on data published by the organizations below. We thank them for making it available. None of them endorses this site.

Forecasts and climate

© ECMWF, CC BY 4.0, modified by us; ECMWF accepts no liability for the data. Contains modified Copernicus Climate Change Service information 2026; neither the European Commission nor ECMWF is responsible for its use. WeatherNext © 2024–2026 Google LLC, under its terms of use; experimental data, not validated for real-world use. Also NOAA and the National Weather Service, NASA NEX-GDDP-CMIP6 (CMIP6 via WCRP), CHELSA (Karger et al. 2017), WorldClim 2.1 (Fick and Hijmans 2017), Daymet (ORNL DAAC), and CONUS404 (USGS). Climate signals: Bureau of Meteorology (the MJO index) and NOAA Climate Prediction Center (El Niño and La Niña).

Weather and snow observations

USDA NRCS SNOTEL, California DWR (CDEC), NOAA NCEI; Environment and Climate Change Canada, under the Open Government Licence – Canada and – British Columbia; Météo-France (updated 9 October 2026); MeteoSwiss and WSL-SLF; GeoSphere Austria; Deutscher Wetterdienst; Finnish Meteorological Institute (CC BY 4.0); MET Norway; Japan Meteorological Agency (出典:気象庁ホームページ), processed by Archean Labs, Inc.

Maps and terrain

© OpenStreetMap contributors, OpenSkiMap, © OpenMapTiles. Elevation: Copernicus WorldDEM-90 © DLR e.V. 2010–2014 and © Airbus Defence and Space GmbH 2014–2018, provided under COPERNICUS by the European Union and ESA; all rights reserved. Imagery: NASA GIBS.

Resort facts

Some from Wikipedia (CC BY-SA 4.0), linked on each resort page.