🎯 Summer Night · Skill

How well did the forecast do? We log the forecast for each night at every lead time, then compare it to what was observed — for hours ≥ 20° and hours ≥ 25° per night.

Loading…

Forecast error by lead time — all cities

For every night with both a logged forecast and an observation, the error is (forecast − observed) hours. MAE is the mean absolute error (lower = better); bias is the mean signed error (positive = the forecast over-predicts hot hours). “Lead” is how many days ahead the forecast was made (0 = same evening).

Lead
days ahead
Pairs
nights × cities
MAE hours ≥ 20° Bias hours ≥ 20° MAE hours ≥ 25° Bias hours ≥ 25°

Night by night

Observed vs the forecast at 1 day and 3 days ahead. Error shown in brackets; green = within 1 hour, amber = within 3, red = further off. Nights still to come show the current forecast with no observation yet.

Night Observed
≥20° / ≥25°
Forecast 1 day ahead
≥20° / ≥25° (error)
Forecast 3 days ahead
≥20° / ≥25° (error)

How this works

Every time the data updates (about every 3 hours), we record the forecast number of overnight hours at or above 20° and 25° for each upcoming night, tagged by how many days ahead it is. When a night has fully passed (09:00 local), we record what was actually observed. The tables above compare the two.

This measure accumulates. Logging began recently, so there are few completed forecast-vs-observation pairs at first — a night forecast 3 days out only becomes a pair once those 3 days have elapsed. The sample grows daily; come back in a week or two for a well-populated skill table. “Observed” values come from Open-Meteo's recent analysis, the same series that feeds the grey line on the everyday chart, not raw station data.