❄️ PowderBench

Who forecasts mountain snow best?

Every day, weather models, AI models and people predict how much new snow will fall over the next three days at thousands of mountain weather stations and ski resorts. A few days later we measure what actually fell and score everyone, in public. So for any station you can see which forecast to trust.

2,400+snow stations & resorts
6leagues, one per kind of truth
12weather models & blends to beat
—rounds scored

See the snow forecast map → Leaderboard Compete

Nobody is beating the weather models. Yet.

Ground truth from USDA NRCS SNOTELNOAA · COOP & CoCoRaHS California DWREnvironment Canada SLF Davos37 + 13 ski resorts ERA5 / Copernicus

Snow forecast map

Tap any station for its snow forecast for the next three days. We show the forecast from whoever has been most accurate in this league so far, next to every other forecaster's numbers and to what fell last time.

Powder Accuracy in 30 seconds

Powder Accuracy answers one question: how much better is this forecast than just guessing the usual amount of snow?

  1. For every station and day we check the forecast against the snow that actually fell and note how far off it was (the “miss”, in inches).
  2. We do the same for a no-skill guess: the typical snowfall for that station on that date, from decades of records (“climatology”).
  3. Powder Accuracy = how much smaller your misses are than the typical guess's misses, as a percentage.
below 0
worse than guessing the usual 0
same as guessing the usual 50
misses half as big 100
perfect

Example: guessing the usual amount misses by 2.0″ on average. Your forecast misses by 1.2″. You cut the misses by 40%, so your Powder Accuracy is 40. Higher is better. Anything above the weather models is beating the pros.

The other columns
  • Avg miss: the average error in inches (lower is better).
  • Range score: for forecasts that give a likely range (low to high), how well the ranges matched reality (lower is better).
  • 6″ alert score: for “chance of 6 inches or more” forecasts, how well those chances matched what happened (lower is better).
  • Weather model rows are the world's forecast models (ECMWF, GFS, NOAA's blends…), run automatically every day. They're the pros to beat.

Leaderboard

Ranked by Powder Accuracy: #1 is the most accurate forecaster in this league, and everyone below scored lower. Higher is better.

#Forecaster Powder Accuracy ⓘ Avg miss Range score 6″ alert score DaysCoverage

Baselines run automatically every round. Unranked teams haven’t met the eligibility bar yet (5+ rounds at ≥70% coverage). Late or invalid submissions are scored for reference but never ranked.

The race

Season-to-date Powder Accuracy, round by round. 0 is climatology; the weather-model baselines are the villains to beat. Dashed lines are automated baselines.

Recent rounds

This season’s snow

Daily fresh snowfall from this league’s ground truth, station by station — exactly the number competitors are scored against.

Season explorer

Snowpack through the season — add any station-season combinations to compare a mountain against its own past or against other mountains’ years. League stations and depth records show cumulative snowfall; the public snow-station archive reaches back decades (Spencers Creek to 1954, the Argentine SWE pillows to 1990). Series tagged · SNIH plot peak snow-water-equivalent (inches of water) — the snowpack curve, which melts down rather than only climbing.

Resorts vs. reanalysis

The receipts: what each resort reported (the Resorts league’s truth) next to ERA5 reanalysis at the same coordinates (the ERA5 league’s truth). Resort numbers usually read higher — that’s the point of keeping them in their own league. ERA5 fills in with a ~8-day archive lag.

When it snows

Climatology at a glance: average daily snowfall by month (inches/day) at every station in this league, from decades of history. Brighter is snowier — this is the no-skill prior every Powder Accuracy is measured against.

Round archive

Every resolved round in this league, newest first.

RoundBiggest 24h dumpBest team Accuracy QC

How to compete

  1. Grab a round. Rounds open daily in each league — the manifest lists the stations, horizons, and your exact cutoff. data/rounds/<league>/<date>/round.json
  2. Make your forecast. One CSV row per station-horizon: station_id,horizon_h,snowfall_in — optionally add quantiles (p10..p90) for the probabilistic track and prob_6in for Powder Alerts.
  3. Submit with one command — powderbench submit my.csv --team my-team --league stations — or open a PR adding data/submissions/<league>/<date>/<your-team>.csv before the cutoff. A bot validates and merges it in minutes.
  4. Get scored. The round resolves against QC’d truth (3–6 days, ~a week for ERA5) and the leaderboard updates. No accounts, no API keys.

Starting points: a 5-model starter ensemble, a Colab notebook that trains a post-processing model, years of QC’d truth on Hugging Face, and a training camp: powderbench hindcast 2026-01-01 2026-01-31 --submission you.csv. AI agents: start at llms.txt.

Full documentation → GitHub repo