Why Alta is hard, and what the model does about it
Alta averages more than 500 inches of snow a year, often far more than valleys a few miles away. Storms are forced up the Wasatch's abrupt terrain, the Great Salt Lake adds lake-effect bands, and snow density swings wildly from storm to storm. Global models smear all of that across grid cells several kilometers wide.
1 · Statistical post-processing, not a model from scratch
The forecast starts from NOAA's GFS. A gradient-boosted model learns how GFS output at Alta, in the Salt Lake Valley and over the Great Salt Lake maps to the snow and water NOAA's observer actually measured at Alta. This is model output statistics (MOS), the same idea the National Weather Service has used for decades, with a modern learner.
2 · Predictors that carry Alta's signal
- Upslope wind: wind projected onto north-westerly flow, the direction that loads Little Cottonwood.
- Mountain-minus-valley precipitation: how much orographic lift GFS already resolves.
- Lake–700 mb temperature difference: the instability that drives Great Salt Lake–effect snow.
- Freezing level, CAPE, humidity, temperature, and season.
The model is constrained so that more GFS precipitation can never lower the forecast. That physical guardrail measurably improved day-3 storm-day skill.
3 · Uncertainty that means what it says
The “likely range” comes from the model's own held-out errors. Past forecasts of a similar size show the spread of what actually fell, and the 10th–90th percentile of that spread is the range shown, so it should capture about 8 in 10 days. The Model skill page tracks how often it does.
4 · From water to snow
Alta is famous for light, dry snow, and the density swings a lot from storm to storm, so we model it instead of assuming it. A separate model learns Alta's snow-to-liquid ratio (how many inches of snow each inch of water became) from the observer's reports. It trains on the five winters since 2021 that have matching GFS archives, though the observer's record goes back to 1990. We forecast snowfall two ways: directly, and as forecast water × forecast ratio. The site shows the average of the two, which held up best in testing. The ratio is also why each day shows a density label, from “heavy, wet” to “blower”.
Raw GFS does this conversion with a fixed, dense ratio. At Alta it barely ever forecasts a 6″ day, which is the biggest single reason the ML beats it.
5 · By first chair
The models forecast 24-hour totals, but skiers want to know what falls between last chair and first chair. Each day's ML total is spread across its hours following GFS's timing, so the amount is the ML's and the timing is GFS's. That timing is the least-tested part of the forecast.
6 · Storm analogs
For each stormy day ahead, we find the eight past days whose GFS setup looked most alike: storm size, 700 mb wind and temperature, lake-effect potential, and freezing level. Then we show what Alta actually got on those days. The setup label (NW flow, SW–W flow, lake-effect, warm storm) comes from the same fields. The analog pool only covers 2021 onward for now, so record storms have no true analog.
7 · Cloud seeding conditions
Little Cottonwood Creek is in the Six Creeks program, which North American Weather Consultants runs for Salt Lake City. Crews burn silver iodide at generators in the foothills when storms look seedable. No public feed shows when those generators are running, so this status estimates it from the weather. It isn't a report of operations.
An hour counts when GFS at Alta has precipitation, moist air at 700 mb (about 10,000 ft), a temperature there of −5 to −15 °C, and wind from the southwest through north. That follows the operator's published criteria as closely as these fields allow. It can't see cloud base or low-level stability, and it doesn't know about suspensions, budgets or a crew's judgment.
We checked the rule once against the 2024-25 Six Creeks log, using short-range GFS:
- Unlikely (no hours): seeded on 2 of 100 days.
- Possible (1–5 hours): seeded on 5 of 17 days.
- Likely (6+ hours): seeded on 37 of 48 days.
That's one season, and forecasts days ahead will do worse than short-range GFS. The log gives dates rather than clock times, so this works by day and can't say whether seeding is happening right now. Seeded days don't always reach Alta either: the Little Cottonwood generator ran in 10 of that season's 31 seeded storms. The log also leaves out Snowbird's separate seeding, and this season's exact program dates aren't public, so we show the status from November 1 to April 15.
The operator estimates seeding adds about 6% to snowpack in the target area. That's its own estimate, and our forecast makes no separate adjustment for seeding. Read the 2024-25 report.
Data
- NOAA cooperative observer at Alta (GHCN-Daily USC00420072, 8,710 ft): daily new snow and melted precipitation, the training target. Each “day” is the 24 hours ending at the observer's 6pm MST reading.
- NOAA GFS 0.25° via Open-Meteo: an archive of short-range forecasts since March 2021 for training, plus true day-1/2/3 forecasts since January 2024 for verification.
Raw forecast and verification JSON: forecast.json, verification.json.