The short answer
Understanding how weather forecasts work starts with a simple idea: a forecast is a calculation. Forecasters measure the atmosphere as it is right now, feed those measurements into a computer model that applies the laws of physics, and let the model work out what the atmosphere will do next. Satellites, radar, weather balloons, aircraft, ships, buoys and thousands of ground stations all supply the measurements. Supercomputers run the models. Human forecasters then check the model output, compare several models, and turn the result into the forecast you see on your phone — the temperature, the rain chance, the wind, and the warnings. The process is powerful and, by the standards of a generation ago, remarkably accurate. It is also limited by the atmosphere itself: tiny errors in today’s measurements grow as the model looks further ahead, which is why a forecast for tomorrow is far more reliable than a forecast for next week.
How weather forecasts work: step one — measuring the sky
Before any forecast can be made, forecasters need to know the current state of the atmosphere, and that means observations from everywhere, taken around the clock. The Met Office, the UK’s national weather service, describes the main sources as weather satellites, balloons, land-based instruments, ships, buoys and aircraft, with observations made 24 hours a day all over the world. Satellites measure radiation from the Earth’s surface and the atmosphere, and they are the only way to cover the oceans and remote regions where there are no ground stations. Radar systems measure the reflection of radiation from raindrops, which shows where rain is falling now and how fast it is moving.
The US National Weather Service draws on the same mix: satellite data, aircraft observations, weather balloons, stream gauges and even information from citizens with personal weather stations can all be fed into its models. The key point is that no single source is enough. A satellite sees the big picture but not the fine detail; a ground station gives precise local readings but says nothing about the ocean a thousand miles away. A good forecast needs all of them, stitched together into one picture of the atmosphere as it is now.
Step two: turning observations into a starting point
Observations alone are not a forecast — they are a snapshot, and a patchy one. Some places have many measurements, others have almost none, and every instrument has small errors. So the next step is data assimilation: combining the observations with a short-range forecast of what the atmosphere should look like, to produce the best possible estimate of the current conditions. The Met Office describes this as turning observational data into a numerical representation of the current atmospheric conditions. The National Weather Service’s own explanation of its models puts it the same way: the critical starting point for all models is a reliable and expansive network of observations, and once the current atmosphere has been established in data form, it can be fed into the model.
This step matters more than most people realise. If the starting picture is wrong — because a storm over the ocean was poorly measured, for example — the forecast built on it will be wrong too, and the error will grow with time. That is one reason forecast centres invest so heavily in observations, and why new satellites and better assimilation methods are a constant priority.
Step three: the models do the physics
The heart of a modern forecast is numerical weather prediction: a computer model that divides the atmosphere into a three-dimensional grid and solves the equations of physics at each grid point, step by step, into the future. The equations describe how air moves, how moisture changes, how temperature rises and falls, and how all of these interact. Once the calculations are done for one short time step, the result becomes the starting point for the next, and the model repeats the process many times — producing forecasts that run from a few hours to many months ahead.
The National Weather Service runs an extensive suite of models. Its Global Forecast System, or GFS, covers the entire globe and runs into the future about one to two weeks. Regional models such as the North American Model (NAM) cover a smaller area at higher resolution, and the High-Resolution Rapid Refresh (HRRR) model, with grid spacing of around 3 km, is used for short-term, local forecasts such as severe weather. Global models typically have grid spacing of roughly 10 km, while higher-resolution models run from around 1 km to 20 km. Smaller grid spacing means finer detail — a model can represent a single thunderstorm rather than a whole region — but it also demands far more computing power, which is why forecast centres cannot simply run everything at 1 km resolution.
Other countries run their own models. The European model, run by the European Centre for Medium-Range Weather Forecasts (ECMWF), the UK Met Office’s Unified Model, Canada’s GEM and Australia’s ACCESS are the best known, and forecasters in each country compare them against their own. In the United States, the National Weather Service also uses the National Blend of Models, a statistical blend that combines many models into one calibrated, high-resolution forecast, and it is the official starting point for the Service’s public forecasts.
Why forecasters run many models at once
No single model run gives the whole story, because each run starts from a slightly uncertain picture of the atmosphere. The answer is ensemble forecasting: running the same model many times with the starting conditions tweaked slightly, to produce a range of possible outcomes. The differences between the runs show how uncertain the forecast is. The American ensemble, GEFS, is made of 30 ensemble solutions, against the GFS’s single deterministic solution. The European ensemble has 50 solutions and the Canadian ensemble has 20, and together they form a Grand Ensemble of nearly 100 members. A forecaster who sees all 100 runs pointing the same way can be confident; one who sees them spread widely knows the forecast is uncertain and says so.
The National Weather Service has moved towards this blended, ensemble approach in recent years, preferring a multi-model technique over relying on any single model. That is why modern forecasts so often come with a probability — a 40% chance of rain, for example — rather than a simple yes or no. The percentage is not a guess about whether it will rain on 40% of your street. It is a statement about the share of model runs, or the share of similar past situations, in which rain fell. A 40% chance of rain means rain is possible but not the most likely outcome, and it is worth carrying an umbrella if getting wet would ruin your plans.
Step four: the human forecaster
Models do the heavy calculation, but they do not write the forecast you read. Human forecasters check and improve the model output, using their knowledge of local geography, known model biases and experience of similar weather patterns. The Met Office puts it plainly: human forecasters check and improve our forecasts, and the service’s forecasters work around the clock, 365 days a year. A model may struggle with a valley that traps cold air, a coastline that changes the wind, or a mountain range that forces rain out of a cloud. A forecaster who knows the area can correct for that.
For very short-range forecasting — the next few hours — forecasters use nowcasting, which maps the current weather, especially from radar, and projects it forward. It powers the “rain starting in 20 minutes” alerts on weather apps, and it is much more accurate than a model run for such short periods, because it starts from what is actually happening right now.
Why forecasts change
Everyone has watched a forecast change: sunshine on Monday becomes rain by Wednesday. There are two main reasons, and neither means the forecasters were careless. First, the atmosphere is chaotic. Small differences in today’s conditions grow larger as the model runs forward, so a forecast for seven days ahead is inherently less certain than one for tomorrow. The Met Office lists this as one of its key facts about forecasting: the atmosphere is chaotic, and small errors grow. Second, new observations arrive constantly, and each new batch improves the starting picture, so forecast centres update their forecasts several times a day. A forecast that changes is usually a forecast that is being updated with better information, not a forecast that was wrong in the first place — which is why the forecast for next weekend should be treated as a guide, not a promise.
How to read a forecast like a forecaster
A few habits make forecasts much more useful. First, look at the probability, not just the icon. A 20% chance of rain and an 80% chance of rain are very different days, even if both show a cloud. Second, check the timing, not just the day: rain at 3 a.m. matters less than rain at 3 p.m. if you are planning a walk. And third, pay attention to the source: a forecast from your national weather service is built on that country’s models and observations, and it is the source behind the warnings that matter.
Watches, warnings and what to do
When hazardous weather is expected, the National Weather Service issues alerts in tiers, and the words matter. A watch means conditions are favourable for a hazardous event to develop: the risk has increased significantly, but whether it happens, and where and when, is still uncertain. A watch is the time to be prepared — review your plan, charge your phone, and know where you would go for safety. A warning means the hazardous event is occurring, is imminent, or has a very high probability of occurring, and it poses a threat to life or property. A warning is the time to take protective action immediately. An advisory sits between the two: the event is expected, but it is likely to cause significant inconvenience rather than a threat to life. The Met Office issues its own warnings in the UK on a colour scale — yellow, amber and red — with red reserved for the most dangerous weather. Whatever the system in your country, the rule is the same: a warning means act now, not later.
The bottom line
Weather forecasts are calculations built on observations from satellites, radar, balloons, aircraft, ships and ground stations, run through physics models on supercomputers, and checked by human forecasters who know the local weather. Ensemble forecasting — running the model many times with slightly different starting points — is how forecasters measure uncertainty and turn it into the probabilities you see. Forecasts change because the atmosphere is chaotic and new data keeps arriving, and a forecast that changes is usually a forecast that is improving. Check the probability, check the timing, and when a warning appears, act on it. For storm preparation, our guide to how to prepare for a hurricane explains what to do before, during and after a storm, and our guide to winter storm preparedness covers cold-weather emergencies in the same step-by-step way.
Sources verified 15 Nov 2026: NWS about weather models for numerical weather prediction, ensemble modelling, the GFS, NAM and HRRR, and observations from satellites, aircraft, balloons and personal weather stations; NWS how models are used in forecasting for the 30-plus model systems, the GFS, ECMWF, CMC and UKMET global models, the GEFS 30-member, EPS 50-member and CMCE 20-member ensembles, and the Grand Ensemble of nearly 100 solutions; Met Office weather forecasting guide for observations, data assimilation, numerical models, human forecasters, nowcasting and the chaotic atmosphere; NWS watches and warnings for watch and warning definitions. No live forecast for any place or date is included, and no accuracy figure was stated from memory.

