Compare side by side the forecast of the three major numerical weather models for any city. Detect divergences greater than 3 °C between models.
GFS (NOAA/USA): global, ~13 km, 4 cycles/day. ICON (DWD/Germany): global ~13 km and European ~6.5 km. ECMWF IFS (European Centre): global ~9-15 km, generally the highest mean accuracy. Data served by Open-Meteo.
A numerical weather prediction model solves the equations of the atmosphere on a three-dimensional grid, starting from the observed state. The three compared on this page are the ones feeding almost every app on the market, and they are not interchangeable.
| Model | Who runs it | Global grid | Runs per day | Range |
|---|---|---|---|---|
| IFS (ECMWF) | European Centre for Medium-Range Weather Forecasts, Bonn and Reading | ~9 km | 00 and 12 UTC | 10 days |
| GFS | NOAA / NCEP, United States | ~13 km | 00, 06, 12 and 18 UTC | 16 days |
| ICON | Deutscher Wetterdienst, Germany | ~13 km | 00, 06, 12 and 18 UTC | 7.5 days |
The differences come from two places. The first is the initial conditions: no model knows the exact state of the atmosphere, only a reconstruction built from satellites, radiosondes, stations and buoys, and that reconstruction is made by different methods at each centre. The second is parameterisation: what happens at scales smaller than the grid — the convection of a thunderstorm, turbulence near the ground, cloud microphysics — cannot be resolved and has to be approximated with empirical formulas that differ from one model to another. The atmosphere amplifies those differences: it is a chaotic system, and two nearly identical initial states diverge beyond recognition within days.
The point of comparing is not to pick a winner but to measure the uncertainty. When all three agree on a 48-hour forecast, confidence is high and the forecast can be taken at face value. When they differ by more than 3 °C in temperature, or change sign on rain, what they are saying is that the situation is poorly defined: a front that may come in six hours earlier or later, a pool of cold air that may or may not cut off. In those cases the honest forecast is not a number, it is a range.
One important qualification: all three are deterministic models, a single run each. Operational centres also work with ensembles of fifty slightly perturbed simulations to quantify that uncertainty statistically, and that is where the probabilities in official forecasts come from. Comparing three deterministic runs is a cheap approximation to the idea, not a substitute.
In the verification the centres publish themselves, the ECMWF’s IFS has for years been the most reliable in the medium range, three to ten days, and is the reference model at the synoptic scale. For the detail of the first 48 hours over Iberia, limited-area higher-resolution models carry more weight, and above all so does the judgement of the human forecaster: AEMET’s warnings are not an automatic model output but an interpretation of several, and in risk situations they are always the reference.
