A weather model divides the atmosphere into a three-dimensional grid of boxes, fills each box with current conditions from observations, then steps forward in time using the physics equations that govern fluids, heat, and moisture. Repeat a few million times and you have a forecast.
The critical constraint is grid resolution. A global model might use boxes 9 miles across. Anything smaller than a box, an individual thunderstorm, a sea breeze front, a lake effect band, can't be represented directly. Models approximate those with statistical shortcuts called parameterizations, which is where much forecast error is born.
The models you'll hear about
- GFS: the American global model, run by NOAA four times daily, out to 16 days. Free and public.
- ECMWF: the European model, generally the most accurate global model in verification scores, especially past day three.
- HRRR: a high-resolution US model updated hourly with 3km boxes, built specifically for short-range storm forecasting.
- NBM: the National Blend of Models, which statistically combines many models into a single calibrated forecast. This is what most official NWS point forecasts start from.
No single model is best at everything. Averaging many of them cancels individual biases and reliably beats most of the individual members, which is why the NBM exists and why forecasters lean on it rather than picking a favorite.
Ensembles and the honest version of uncertainty
Running one model gives you one future. Running the same model 30 times with tiny variations in the starting conditions gives you a spread of futures: an ensemble. When those 30 runs agree, confidence is high. When they scatter, the atmosphere is in an uncertain state, and that scatter is exactly what a probability like "40% chance" is built from.
Where models end and measurements begin
Models are indispensable for the future and unnecessary for the present. Right now, at this moment, there is no need to simulate anything, an instrument at a nearby airport already measured the temperature, wind, and pressure. That's the line: forecasts are the best available simulation of what hasn't happened yet, and observations are a record of what did.
Trouble starts when interfaces blur the two, presenting a model's estimate of current conditions with the same confidence as a reading. Knowing which you're looking at is most of weather literacy.