As everyone in AI jokes right now, open = close. By symmetry, close = open.

After enough ski friends nudged me, I built CloseSnow (close-snow.clifan.com), a 100% free, zero-paywall 15-day mountain snow forecaster covering all 120 Ikon and Epic Pass resorts.

CloseSnow live Alta forecast showing ±1σ error bars

Why single-number snow bars lie

Most snow apps show a single crisp integer on every day of a 10- or 15-day chart. To a skier planning PTO, 8" on Wednesday and 8" next Thursday look identical. Physically, they aren't even the same species of number:

  1. Snow-to-liquid ratio (SLR) swings 4× across storms. Half an inch of liquid water (0.50" SWE) falling through a saturated -15°C dendritic growth zone in Little Cottonwood Canyon stacks up to 9" of cold smoke (18:1 SLR). That same 0.50" SWE in a 31°F atmospheric river at Palisades Tahoe settles into 4" of sierra cement (8:1), or turns to rain at base lodge elevation.
  2. Ensemble spread explodes with lead time and marginal temperatures. At Day 2, high-resolution convection-allowing models like NOAA HRRR (3 km) and NBM v4.2 often agree within an inch. By Day 9, or whenever wet-bulb temperature sits within 1.5°C of freezing, global models routinely span 0" to 14". Hiding that variance behind a single mean tricks people into chasing phantom storms.

What CloseSnow computes instead

CloseSnow pulls live pressure-level profiles (850, 700, 600, 500 hPa) across six NWP models—NOAA HRRR 3km, NBM v4.2, ECMWF IFS 9km, ECMWF AI AIFS, DWD ICON, and GFS 16-day—at Base, Mid-Mountain, and Summit elevations for 120 resorts:

  • 3D Cobb/Alcott microphysics: Weights vertical motion through the -12°C to -18°C dendritic growth zone, penalizes 50+ mph ridge-top wind riming, and applies resort-specific 700 hPa canyon alignment multipliers.
  • Color-matched ±1σ I-beam error bars: Every bar displays a calibrated ±1σ range (5 ± 1", 7 ± 3", 6 ± 4"). Uncertainty grows monotonically with lead time and 0°C wet-bulb phase risk while staying physically non-negative.
  • Live USDA SNOTEL ground truth: 41 US mountain resorts pair directly with hourly USDA NRCS SNOTEL ultrasonic snow-pillow telemetry so you can check what actually fell over the past 15 days.

Out-of-sample SNOTEL & HRRR backtest

Before sharing it publicly, I evaluated the model out-of-sample (live Accuracy Lab):

  • Atmospheric sounding test (N = 7,863 held-out HRRR 3km soundings, 2,829 mountain events ≥1,500m, trained strictly on 31,178 ERA5 soundings): CloseSnow reached 3.23 SLR MAE on mountain snow events, compared to 3.91 for Kuchera, 5.12 for OpenSnow's temperature heuristic, and 5.43 for fixed 10:1.
  • Real 48-hour mountain storms (N = 221 held-out USDA SNOTEL storm cycles from Jan 16 – Apr 15, 2025 across 20 Ikon/Epic resorts):
    • CloseSnow ($0): 2.37" MAE (52.9% within ±2", 72.4% empirical ±1σ storm coverage)
    • Bergfex / MeteoBlue Pro ($25/yr): 3.21" MAE
    • OpenSnow ($49.99–$99.99/yr): 4.01" MAE
    • pow.fyi / Fixed 10:1 ($0): 4.14" MAE

Honest scope limits

Extended Days 11–15 reflect synoptic storm-window probabilities across ECMWF IFS, AI AIFS, and GFS rather than exact inch counts—which is why their ±1σ whiskers stretch wide. Also, SNOTEL snow pillows cover 41 US mountain resorts and sit in sheltered clearings; non-US resorts use Open-Meteo reanalysis for past-15-day history.

It is a solo, zero-ad, zero-login public-good build. Expect rough edges and bugs—if you spot a resort with weird canyon physics or mobile layout quirks, send feedback my way. Fixes ship between laps.