Every model's 1- to 7-day forecasts are checked hour by hour against the last
30 days of weather and scored 0 to 100, with the same formulas as the VeriSky app.
- Temperature and wind are 70% Accuracy + 30% Extremes
(open a row to see the split).
Accuracy is how often the forecast was close: within 2 °C, or within
5 km/h or 20% for wind.
Extremes (SEDI) checks
whether the model caught the unusual hours for that place, and counts false
alarms against it. In a month too calm to judge, Accuracy takes its weight.
- Rain scores each hour's rain/no-rain call (F1 at 0.1 mm/h). It only
counts when the month actually had some rain; otherwise every model would ace
a dry month.
- Sharpness (in the expanded row, not in the score) shows whether the
forecast swings as much as the real weather. 1.0× is right; much lower means
the model is hedging toward an average.
Skill combines the lead days, with tomorrow counting 7× as much as day 7.
Models that stop before day 7 are scaled down a little: a perfect next-day-only
model tops out around 85, a perfect 7-day model reaches 100.
Skill by lead time shows the score per lead day. In the grid, read down a
column for "who do I trust three days out". The colour bands are fixed, and lower
for rain, whose scores naturally run lower. Blank cells are days a model doesn't
forecast. The curves view colours four models and leaves the rest gray.
Rain W–L and form dots look only at tomorrow: a win is a correct rain/no-rain
call for the day, a filled dot a day with next-day skill of 70 or more.
▲▼ compares today's rank with a week ago.
Airport boards (✈) score each model at the city's main airport, against real
METAR station readings for temperature and wind and radar for rain where available.
Gaps fall back to the model analysis. The Compare and labs charts still use the
analysis, so they can differ slightly from the standings, and some airports have a
shorter history so far.
Why not average error (RMSE)? It rewards models that blur toward a smooth
average and punishes sharp forecasts that are slightly off in timing twice
(Ben Bouallègue et al. 2024).
This board used it until August 2026; the expanded row still shows each model's old
rank, and the long-term trends page still charts it.
Data: forecasts come from Open-Meteo's
previous runs archive
(what each model actually said 1 to 7 days earlier). Because the reference is a model
analysis rather than a station record, it can slightly flatter related models at short
leads. The latest, incomplete day is left out. Public data is cached for about
6 hours, commercial summaries for 15 minutes.