League table
How to read this table, and its limits
- One measure among many. Accuracy has no single definition. This page uses average errors in the daily high and low and a simple wet/dry call; other measures - the size of the worst misses, how much rain fell, how well a model's probabilities match outcomes, skill relative to the seasonal average - can rank the same models quite differently.
- What is scored. For each day, each model's forecast made the chosen number of days earlier is compared with what actually happened: the error in the day's high and low temperature (taken from hourly values), and whether it called the day wet (1 mm of rain or more) or dry. Temperature ranks by the average of the high and low errors.
- "What actually happened" is ERA5, not a weather station. ERA5 is the Copernicus reanalysis: a reconstruction of past weather from observations, on a grid of roughly 9 to 25 km. It smooths out very local effects - a sea breeze, a frost hollow, a city's warmth - so the errors here are model against grid, not model against your thermometer. It is also produced by ECMWF, which may slightly favour ECMWF's own model.
- The last week is missing. ERA5 is published about five days after the fact, so the period ends roughly a week ago.
- A month is a small sample. Models marked too close to call are not clearly worse than the leader: their daily errors overlap its errors too much to separate them. Click or tap the label to see the numbers - on how many days each did better, the average gap, and the gap that would be needed to call it. Small differences, a tenth of a degree, are noise.
- Rain is the roughest score. Reanalysis rainfall is less reliable than its temperature, and a wet/dry call ignores how much rain fell. In dry spells nearly every model scores well.
- Accuracy here, not everywhere. The ranking is for this place, this period and this lead time. It changes with the season and from one town to the next, and a model's past month does not guarantee its next one. Several models switch to a regional high-resolution version near home (for example UKMO over the British Isles, ICON over Europe), so they can do much better there than elsewhere.
- A model is left out if Open-Meteo has fewer than ten scoreable days of its forecasts for the period.
Last month in London, New York and Sydney
The same scoring for three cities, forecasts made 3 days ahead, ranked by temperature. Computed when this page was last published; choose a place above for live numbers.
London
Most accurate here, 3 days ahead: ECMWF IFS (Europe). Its forecast highs and lows were off by 0.9 °C on average. Too close to call with it: ICON, ECMWF AIFS, JMA, ARPEGE, UKMO and GEM. Bottom of the table: GFS (1.4 °C).
| # | Model | Temperatureavg error | Highsavg error | Lowsavg error | Tends to run | Rainwet/dry right | Days |
|---|---|---|---|---|---|---|---|
| 1 | ECMWF IFSEurope | 0.9 °C | 0.9 °C | 0.9 °C | about right | 70% | 29 |
| 2 | ICONDWD, Germany | 1.0 °C | 1.0 °C | 0.9 °C | about right | 78% | 29 |
| 3 | ECMWF AIFSEurope, AI | 1.0 °C | 0.8 °C | 1.3 °C | 0.9° too cold | 74% | 29 |
| 4 | JMAJapan | 1.2 °C | 1.2 °C | 1.2 °C | about right | 67% | 29 |
| 5 | ARPEGEMétéo-France | 1.2 °C | 1.1 °C | 1.4 °C | 0.6° too warm | 78% | 29 |
| 6 | CMA GRAPESChina | 1.3 °C | 1.5 °C | 1.1 °C | 0.9° too cold | 70% | 29 |
| 7 | UKMOMet Office, UK | 1.3 °C | 1.1 °C | 1.5 °C | 0.8° too warm | 81% | 29 |
| 8 | GEMECCC, Canada | 1.3 °C | 1.1 °C | 1.5 °C | about right | 78% | 29 |
| 9 | GFSNOAA, US | 1.4 °C | 1.6 °C | 1.2 °C | 1.0° too warm | 78% | 29 |
Forecasts made 3 days ahead, checked against ERA5 for 29 days, 29 Aug – 26 Sept 2026; 6 of them wet (1 mm or more).
New York
Most accurate here, 3 days ahead: ECMWF AIFS (Europe, AI). Its forecast highs and lows were off by 0.8 °C on average. Too close to call with it: GEM and ECMWF IFS. Bottom of the table: UKMO (1.9 °C).
| # | Model | Temperatureavg error | Highsavg error | Lowsavg error | Tends to run | Rainwet/dry right | Days |
|---|---|---|---|---|---|---|---|
| 1 | ECMWF AIFSEurope, AI | 0.8 °C | 0.9 °C | 0.8 °C | about right | 93% | 29 |
| 2 | GEMECCC, Canada | 0.9 °C | 1.1 °C | 0.7 °C | about right | 81% | 29 |
| 3 | ECMWF IFSEurope | 1.0 °C | 1.0 °C | 1.0 °C | 0.4° too warm | 78% | 29 |
| 4 | CMA GRAPESChina | 1.2 °C | 1.2 °C | 1.2 °C | about right | 74% | 29 |
| 5 | ICONDWD, Germany | 1.3 °C | 1.3 °C | 1.2 °C | 0.8° too warm | 70% | 29 |
| 6 | GFSNOAA, US | 1.5 °C | 1.4 °C | 1.6 °C | 1.2° too warm | 81% | 29 |
| 7 | ARPEGEMétéo-France | 1.5 °C | 1.5 °C | 1.6 °C | 0.9° too warm | 74% | 29 |
| 8 | JMAJapan | 1.8 °C | 1.3 °C | 2.3 °C | 0.9° too warm | 78% | 29 |
| 9 | UKMOMet Office, UK | 1.9 °C | 1.9 °C | 1.9 °C | 1.8° too warm | 81% | 29 |
Forecasts made 3 days ahead, checked against ERA5 for 29 days, 29 Aug – 26 Sept 2026; 13 of them wet (1 mm or more).
Sydney
Most accurate here, 3 days ahead: GEM (ECCC, Canada). Its forecast highs and lows were off by 0.8 °C on average. Too close to call with it: ECMWF IFS and JMA. Bottom of the table: ICON (2.2 °C).
| # | Model | Temperatureavg error | Highsavg error | Lowsavg error | Tends to run | Rainwet/dry right | Days |
|---|---|---|---|---|---|---|---|
| 1 | GEMECCC, Canada | 0.8 °C | 0.7 °C | 0.8 °C | about right | 93% | 29 |
| 2 | ECMWF IFSEurope | 0.9 °C | 0.9 °C | 0.8 °C | about right | 89% | 29 |
| 3 | JMAJapan | 1.1 °C | 1.0 °C | 1.1 °C | 0.3° too warm | 89% | 29 |
| 4 | ECMWF AIFSEurope, AI | 1.3 °C | 1.2 °C | 1.4 °C | about right | 85% | 29 |
| 5 | ARPEGEMétéo-France | 1.3 °C | 1.7 °C | 0.9 °C | 0.9° too warm | 93% | 29 |
| 6 | UKMOMet Office, UK | 1.4 °C | 1.2 °C | 1.5 °C | 0.8° too warm | 89% | 29 |
| 7 | CMA GRAPESChina | 1.4 °C | 1.2 °C | 1.6 °C | 0.5° too cold | 89% | 29 |
| 8 | GFSNOAA, US | 1.6 °C | 2.0 °C | 1.1 °C | 0.9° too warm | 93% | 29 |
| 9 | ICONDWD, Germany | 2.2 °C | 2.7 °C | 1.6 °C | 0.7° too warm | 85% | 29 |
Forecasts made 3 days ahead, checked against ERA5 for 29 days, 29 Aug – 26 Sept 2026; 3 of them wet (1 mm or more).
Computed on 2 October 2026.
Questions
Which weather forecast is most accurate?
There is no single answer: it depends on where you are, the season, what you measure and how far ahead. Official global verification has long ranked ECMWF's IFS at or near the top for forecasts several days ahead, but at one particular place another model - often one that blends in a regional high-resolution model - can do better. The table on this page scores nine models for the place you choose.
Is the European model (ECMWF) more accurate than the American model (GFS)?
Averaged over the whole globe, ECMWF's IFS usually verifies better than NOAA's GFS for forecasts several days ahead. At a single place over a single month the order often flips, and the gap between them is frequently too small to call. Choose your place to see how they compare there.
How is the accuracy measured?
For every day in the period, each model's forecast made 1, 2, 3, 5 or 7 days earlier is taken from Open-Meteo's archive of previous model runs and compared with ERA5, the Copernicus reanalysis of what actually happened. The score is the average error in the daily high and low temperature, plus how often the model correctly called the day wet (1 mm of rain or more) or dry.
Why does the table stop about a week ago?
ERA5, the record of what actually happened, is published about five days after the fact. Using a faster source would mean scoring the models against one of the models themselves, so the most recent days are left out until ERA5 covers them.
Why not use weather station readings?
This page runs entirely in your browser, with no server or database, and free station data with worldwide coverage is not available that way. ERA5 is available for every point on Earth and treats every place the same way. It is a reconstruction from observations on a grid of roughly 9 to 25 km, so it smooths out very local effects, and it is produced by ECMWF, which may slightly favour ECMWF's own model.
Is this an official measure of forecast accuracy?
No - it is a bit of fun. Weather services and forecast centres verify their models with many measures at once, such as typical and worst-case errors, skill compared with the seasonal average, hit and false-alarm rates, and scores for probability forecasts, against weather stations, radar and analyses, over years and whole regions. Different measures and different periods can put the models in a different order, so a model topping this table for one month at one place is not a verdict on which model is best.
Is anything sent or stored?
Nothing is tracked and there is no account. The coordinates of the place you choose go to Open-Meteo to fetch the forecasts and the ERA5 record, and anything you type in the search box goes to Photon to look the place up. Your choices are stored in this browser only.