The one-day comparison
Scores & comparison details
Across the station network
Where the errors differ
Explore the local differences, from the coast to the mountains.
Lower Yr error Lower WeatherNext error Equal error No paired data
Zoom in to separate nearby stations: use +/−, scroll or pinch. Colours show measured average error differences, not statistical winner verdicts. An outer ring marks no recent measurements in the scored history.
Select a station on the map.
Weather that matters
Does it catch the weather that matters?
Looking ahead
Who is closer, by how far ahead
Show as table
Notable days
Storm replays
What each service said 7 days, 5 days ... 6 hours before a notable day, next to what was measured. Heavy rain is a day with at least 20 mm, strong wind at least 15 m/s at a land station (at the offshore platforms that is an ordinary day), a cold snap at most −5 °C.
More analysis: trends, bias, forecast changes and uncertainty
Between forecast runs
How steady are the forecasts?
Take one hour at one station. How much did each service's forecast for it change between being made, say, 2 days ahead and 1 day ahead? Smaller changes mean a steadier forecast.
Show as table
Direction of error
Which way do they miss?
Show as table
The daily record
Over time
Daily average error, with a 7-calendar-day average when at least 3 days have data. Gaps stay visible.
Show daily values as a table
Weather and terrain
Patterns to watch
How the error changes with the weather and the terrain, for forecasts made ahead.
Month by month
Earlier months
A service is only named ahead where the 95% test below says so, and not before 7 days of data in the month. Rain compares WeatherNext's average and its median with Yr.
How to read this fairly and data attribution
- Same moment, same head start. Both forecasts are saved in the same run. "24 h ahead" means what you could have seen on Yr and WeatherNext 24 hours before the weather happened.
- Yr has a home advantage. For the first 2–3 days, Yr's forecasts are corrected using measurements from Norwegian weather stations, including many of these ones. WeatherNext 3 also trains its temperature output against station observations, which differs from Yr's local forecast corrections. This compares the products people receive, rather than isolating the underlying models.
- WeatherNext averages many possible outcomes. For temperature and wind we score its average, which smooths out extremes and tends to lower the average error. Its median, almost identical so far, is in the score details and the table. For rain we lead with the median because averaging many scenarios spreads light drizzle across nearly every hour.
- Reading the station map. Colours show observed error differences, not statistically tested station winners. The lists show stations with at least 7 paired days and at least 10% smaller error; a grey ring marks missing measurements in finalized history.
- "Too close to call" means the 95% uncertainty range of the difference still includes zero. Days are the unit: errors from nearby stations on the same day are not independent.
- "Same as yesterday" sets the bar. This naive guess predicts the value measured at the same time of day on the latest day already known when the forecast was made: yesterday for forecasts up to 1 day ahead, 2 days back for 2 days ahead, and so on. A useful forecast should beat it, and both do by a wide margin.
- Station measurements vs forecast grids. A forecast describes an area; a station measures one spot. Both models face the same challenge.
- Past scores keep retired stations. The active collection count above can be smaller than the historical scoring network. Earlier station measurements and forecasts stay in the record.
- Grid height. WeatherNext's temperature is for the average terrain height of its ~5 km grid area, not the station's height. In steep terrain the area can be a few hundred metres higher than a station in a fjord or valley, so WeatherNext reads colder there (about 0.65 °C per 100 m). Stations where this matters are marked in the station details. The "height-adjusted" values are our own estimate, a standard correction to each station's height, and is not Google's product; the published values stay the main score and every station is kept. On 26 Sep we also switched to reading WeatherNext exactly as Google does (interpolated between grid points); earlier days sometimes used a neighbouring grid point.