How weather forecasts are made, and how accurate they really are

A weather forecast starts with data, not guesswork. According to the US National Oceanic and Atmospheric Administration (NOAA), supercomputers gather billions of observations every day from weather balloons, satellites, ocean buoys, radar, aircraft sensors and ground stations, converting each reading into a numerical value for temperature, air pressure, moisture and wind.

From observations to a model run

Those numbers become the starting point, or 'initial conditions', for a numerical weather prediction model: a set of mathematical equations describing how air moves, heats and holds moisture. NOAA says its supercomputers run at 2.8 quadrillion calculations per second from data centres in Virginia and Florida, stepping the model forward in short time increments, repeatedly recalculating the atmosphere's state until the forecast period ends, currently out to 16 days for NOAA's global model.

Forecasters do not rely on one model. NOAA notes that several global and regional models are compared for each situation; when they agree, confidence in the forecast rises, and when they diverge, forecasters communicate that uncertainty publicly.

Why forecasters run dozens of versions at once

Because the atmosphere is chaotic, a tiny error in the starting data can grow into a large error days later, as the European Centre for Medium-Range Weather Forecasts (ECMWF) explains below. To manage this, ECMWF runs an ensemble: the same model launched many times from slightly different starting points. Its first operational ensemble in December 1992 had 33 members; by 2023 the extended-range ensemble had grown to 101 members, with medium-range ensembles run at roughly 9-kilometre resolution. The spread between the ensemble members shows how confident a forecast should be.

How accurate are forecasts today

NOAA's satellite and information service, NESDIS, states that a 5-day forecast is accurate about 90% of the time, a 7-day forecast about 80% of the time, and forecasts of 10 days or longer are right only about half the time because models cannot access future data and must extrapolate from current conditions.

Forecasts have also improved sharply over recent decades. The Met Office, the UK's national weather service, says its four-day forecast today is as accurate as its one-day forecast was 30 years ago. The Met Office also states that its model is verified against World Meteorological Organization standards and that it is consistently ranked among the top two forecast providers worldwide. Separately, Our World in Data reports that three-day forecasts, citing ECMWF, have been consistently accurate since the 1980s and now reach about 97%.

Resolution, computing power and the newest change: AI models

Part of the improvement comes from finer model grids. ECMWF says its horizontal resolution has gone from about 210 kilometres to 18 kilometres over 30 years, and that the lead time of its skilful ensemble predictions has grown by roughly one and a half days per decade on average. Our World in Data adds that the Met Office's own forecasting grid has moved from about 90-kilometre spacing to 1.5-kilometre spacing, closing gaps between observation points.

The newest shift is machine learning. In its evaluation of 2025 performance, ECMWF reported that its AI-based forecasting system, AIFS, produced errors 5–15% lower than its traditional physics-based model for medium-range forecasts out to 15 days, and that AIFS substantially outperformed the older system at predicting tropical cyclone track, though it under-predicted storm intensity. ECMWF also said 2025 brought its largest single-year skill improvement in over a decade, driven by system upgrades introduced at the end of 2024, including a roughly 10% reduction in errors for near-surface temperature and wind speed in its ensemble forecasts.

Not every forecast is equally reliable

Accuracy is not distributed evenly around the world. Our World in Data reports that low-income countries spend 15 to 20 times less per person on weather services than high-income countries, a gap that affects both the quality of local forecasts and the reach of early-warning systems for extreme weather.