How accurate are our forecasts?
Every day we forecast the weather at 30 weather stations, from ski towns to islands, next to other forecasts for the same spots. Once each day has passed, every forecast is scored against what the station measured. All the numbers on this page come from that scoreboard and update daily.
Loading the scoreboard…
Forecast vs what happened
Day-ahead forecasts at one station (issued the day before) and the station's measured temperature.
Temperature error by days ahead
Average miss of the daily high and low, in °F, across all stations. Shorter bars are better.
Rain and snow chances
The Brier score is the average squared miss of a chance: 0 is perfect and lower is better; always saying 50% scores 0.25. A yes/no forecast is scored the same way, as 0% or 100%.
Reliability: on days a forecast gave about a given chance, how often it rained or snowed. A reliable forecast sits on the dashed line: when it says 30%, it rains about 30% of the time.
Head to head
Only the days and numbers both forecasts gave.
Backtest
How we score
- Stations.
- No home advantage. Our forecast corrects each weather model with nearby stations' reports, but never with the station being scored: the score is what you get at a spot with no station of its own.
- Measured weather. The day's high and low come from the station's hourly reports (at least 18 hours that day). A day counts as rain or snow if any report had rain, drizzle, showers, thunderstorms or snow, and as snow if any had snow, sleet or ice pellets.
- The forecasts. Veabird: our point forecast (all major models at the spot's own elevation, corrected and blended, with chances from 82 ensemble members). Typical app: one weather model (Open-Meteo's best match), like many weather apps show; its rain and snow are yes/no. National Weather Service: the official forecast at the same spot (US stations only); its snow chance is its precipitation chance on days it forecasts snowfall.
- Open data. The full scoreboard is point-skill.json; the forecast is free in ChatGPT and Claude (Veabird for AI assistants).
Scores cover the stations and days shown, not every place or season; one season's results can change in the next. Forecasts are guidance, not guarantees.