How accurate is the model?

raceday.ski continuously compares the calculated snow temperature with real measurements — directly at the measuring stations: 132 IMIS stations of the WSL Institute for Snow and Avalanche Research SLF in the Swiss Alps plus 145 stations of the Tyrol Avalanche Warning Service. Every 2 hours an automatic run computes the model exactly at each station position (station altitude, the station's horizon profile) and pairs it with the measured value — automatically and without preselection. How the calculation works is explained on the methodology page.

Overall picture

Forecast metric (9 pairs)
RMSE 2.54 °C
Bias -0.78 °C · MAE 1.86 °C
Last 7 days (3 pairs)
RMSE 3.77 °C
Bias -2.84 °C · MAE 2.84 °C

Data window: 13/07/2026 to 23/08/2026 · as of: 25/08/2026, 04:35

13 hourly pairs · 3 in melt regime · 0 forecasts resolved · 30 open · 5 stations

By mode and lead time

 nBias °CMAE °CRMSE °C
Forecast, current hour9-0.781.862.54
Forecast +24 h0
Forecast +48 h0
Analysis back-hours4*-5.555.555.71
Melt regime (excluded)3*2.192.524.09

By time of day

 nBias °CMAE °CRMSE °C
00–068-2.753.003.96
06–124*-0.872.823.50
12–180
18–241*-3.693.693.69

By weather situation

 nBias °CMAE °CRMSE °C
clear10-3.093.574.29
variable2*0.561.351.46
overcast1*0.600.600.60

By measuring network

 nBias °CMAE °CRMSE °C
LWD Tyrol (Austria)9-0.781.862.54
SLF IMIS (Switzerland)4*-5.555.555.71

By measuring station

 nBias °CMAE °CRMSE °C
RosskarLWD-SFES2 · LWD Tyrol (Austria)5-1.232.793.34
Ils PlaunsPUZ2 · SLF IMIS (Switzerland)4*-5.555.555.71
Mosesgipfel SchneestationLWD-IMOS2 · LWD Tyrol (Austria)3*-0.020.670.72
SchöckbichlLWD-NSCH1 · LWD Tyrol (Austria)1*-0.790.790.79

* Groups with fewer than 5 pairs are statistically not very robust.

Metrics — how they are calculated

  • Error = calculated − measured snow surface temperature (°C), compared at the exact station location (station altitude, flat field, the station's horizon profile — no spatial distance blur). The comparison uses the pure physics calculation of the 3-layer model — before residual correction and before the Kalman station adjustment. Deliberately so: the station correction that is additionally active in the product learns from the very same measurements; including it would flatter the accuracy.
  • Bias = mean of the errors (systematic deviation; negative = the model runs too cold). MAE = mean absolute error. RMSE = root mean square error (penalises outliers more strongly).
  • Forecast vs. analysis: The headline metric uses only pairs in forecast mode — weather inputs as a user receives them at calculation time. Hours already past are computed with the retrospective weather analysis and reported separately as "analysis back-hours". The skill for +24 h/+48 h comes from archived forecasts that are resolved once the measurement arrives.
  • Melt regime: If measurement AND model sit at the melting point (≥ −0.5 °C), the pair carries no error information — it is counted, but excluded from all metrics and shown separately.
  • Pairing: Every 2 hours an automatic run compares the model at each active measuring station (SLF IMIS Switzerland, LWD Tyrol Austria) for each full hour. Stations without snow (surface-temperature plausibility check, snow depth below 20 cm) or with anomalous sensor behaviour (automatic quarantine against the network median) are not included.
  • Dimensions: time of day in four 6-h windows (local time, identical to the correction windows of the Kalman filter); weather situation from cloud cover (clear < 25 %, variable 25–75 %, overcast > 75 %). The diagnostic tables (time of day, weather, network, station) count all hourly pairs — forecast and analysis; the headline metric only forecast.

Measurement data: WSL Institute for Snow and Avalanche Research SLF (www.slf.ch), licence CC BY 4.0. In Austria, additionally surface-temperature measurements from the Tyrol Avalanche Warning Service — Datenquelle: Land Tirol - data.tirol.gv.at (CC BY 4.0). On Swiss and Austrian ski resort pages, raceday.ski shows the live comparison with the nearest station.

To the wax advisor — or to the methodology: how the 3-layer model calculates and where its limits are.