How accurate is Ilma's forecast?
Measured over the last 73 days at 14 Finnish weather stations, for forecasts 1–7 days ahead: 768 686 matched forecast hours against the Finnish Meteorological Institute. The table shows how many days each comparison covers. Updated 04 Oct 2026.
- Temperature: Ilma's error was lower than FMI's (−7 %) and Foreca's (−8 %).
- Wind speed: Ilma's error was lower than FMI's (−12 %) and Foreca's (−14 %).
- Rain: Ilma caught rain hours better than FMI (hit score 0.217 vs 0.184) and Foreca (hit score 0.217 vs 0.187).
Each of these differences is statistically significant after correction for multiple comparisons.
| Compared with | Variable | Ilma | Them | Verdict |
|---|---|---|---|---|
| Finnish Meteorological Institute (FMI) | Temperature (°C average error) | 1.36 | 1.46 | Ilma better (73 d) |
| Finnish Meteorological Institute (FMI) | Wind speed (m/s average error) | 1.15 | 1.30 | Ilma better (73 d) |
| Finnish Meteorological Institute (FMI) | Rain or no rain (hit score, CSI) | 0.217 | 0.184 | Ilma better (73 d) |
| Foreca | Temperature (°C average error) | 1.36 | 1.48 | Ilma better (73 d) |
| Foreca | Wind speed (m/s average error) | 1.15 | 1.34 | Ilma better (73 d) |
| Foreca | Rain or no rain (hit score, CSI) | 0.217 | 0.187 | Ilma better (73 d) |
| Temperature (°C average error) | 1.39 | 1.33 | No significant difference (39 d) | |
| Wind speed (m/s average error) | 1.15 | 1.22 | Ilma better (39 d) | |
| Rain or no rain (hit score, CSI) | 0.231 | 0.223 | No significant difference (39 d) | |
| Yr | Temperature (°C average error) | 1.17 | 1.26 | Ilma better (39 d) |
| Yr | Wind speed (m/s average error) | 1.02 | 1.07 | Ilma better (39 d) |
| Yr | Rain or no rain (hit score, CSI) | 0.347 | 0.323 | No significant difference (39 d) |
| ECMWF AIFS (best single model) | Temperature (°C average error) | 1.36 | 1.51 | Ilma better (73 d) |
| ECMWF AIFS (best single model) | Wind speed (m/s average error) | 1.15 | 1.28 | Ilma better (73 d) |
| ECMWF AIFS (best single model) | Rain or no rain (hit score, CSI) | 0.217 | 0.187 | Ilma better (73 d) |
Where Ilma is not better, the table says so.
Temperature error by forecast day (°C, lower is better)
| Service | Day 1 | Day 2 | Day 3 | Day 5 | Day 7 | |
|---|---|---|---|---|---|---|
| Ilma | 0.93 | 1.06 | 1.19 | 1.50 | 1.89 | |
| Finnish Meteorological Institute (FMI) | 1.02 | 1.14 | 1.27 | 1.67 | 1.94 | |
| Foreca | 1.14 | 1.23 | 1.35 | 1.59 | 1.90 | |
| 0.89 | 0.98 | 1.11 | 1.53 | 1.91 | ||
| Yr | 1.06 | 1.12 | 1.23 | 1.61 | 1.98 | |
| ECMWF AIFS (best single model) | 1.11 | 1.21 | 1.31 | 1.68 | 2.03 |
How this is measured
Every 5 hours Ilma stores what each service forecast for the next two weeks. After the weather has happened, each forecast hour is compared with the Finnish Meteorological Institute's measurement at the same station and hour. Only hours that every compared service forecast count (matched pairs).
Uncertainty comes from a block bootstrap over consecutive days (neighbouring days are not independent). A competitor is called beaten only when the difference stays significant after Holm correction for all comparisons made, and only with at least 20 days of data.
Limits: the record is one late summer and autumn. From day 7 onward there is no reliable difference between the services. Ilma's forecast itself excludes Foreca and Google; they are compared, not used.