Our forecast accuracy
Every day we archive the forecast we show for the coming seven days. Once the day is over we compare it with what SMHI's stations actually measured — but only for days where every hour was observed. Here is the result, whatever it looks like.
Nationwide, last 30 days
We are collecting enough data for a fair evaluation. Figures appear once archived forecasts can be checked against days with full hourly coverage.
How we measure
The forecast is archived once a day with its lead time: 1 day ahead means the forecast shown the day before, 7 days ahead the one shown a week earlier. Archived forecasts are never changed afterwards.
The answer key comes from SMHI stations every hour and is stored at the observation's own time — temperature as an instantaneous value on the hour, precipitation as the sum for the hour ending then. Each station value keeps SMHI's quality flag; fetching the same observation again adds no row, and a corrected value replaces the old one without double counting. The town's hourly value is a distance-weighted blend of stations within a few tens of kilometres, with the station selection and weights stored so the value can be reproduced.
The evaluation day is the local calendar day in Swedish time, with 23, 24 or 25 hours around the DST changes. A day only counts when every hour has a value from at least two stations. A missing hour is unknown — never zero — and excludes the day for that metric. Flags G (checked) and Y (not yet fully checked) are kept apart; R (suspect) is not used. The share of hours relying only on Y values is reported.
The temperature error is the mean deviation of the day's highest and lowest hourly observation, in degrees. It is the min and max among hourly values, not necessarily the day's true extreme. A rain day has at least 0.2 mm in the hourly sums. Rain right is how often forecast and outcome agreed on rain or no rain. Share of rainy days we missed uses rainy days as the denominator; share of dry days where we forecast rain uses dry days; share of rain forecasts that were false alarms uses rain forecasts. The three answer different questions.
Counts in the tables are location-days — the same calendar day in 60 towns gives 60 location-days. The period and number of calendar days are therefore always stated; a short period can hold little variation in the weather however many location-days it contains. Stations are a few kilometres from the town centre, so the answer key is not a measurement at your address. This is method version 3; older outcomes without time-stamped evidence are not shown.
The source comparison on town pages uses a common sample: the same town, day and lead time for every source, archived within an hour of each other. We do not compare ourselves with services we do not archive the same way.