The forecasting recipe in three steps
Modern forecasts follow a basic pattern. First, instruments measure the current state of the atmosphere. Second, computers combine those measurements into a complete picture of the air at one moment. Third, computer programs called numerical weather prediction models calculate how that picture will change over the following hours and days. Human forecasters then interpret the results, check them against experience and local knowledge, and write the forecast.
Step one: observing the atmosphere
No single instrument can describe the whole atmosphere, so forecasters rely on many sources:
- Surface stations on land and automated buoys and ships at sea measure temperature, pressure, humidity, wind and rainfall.
- Weather balloons carry small instrument packages called radiosondes upward through the atmosphere, usually launched at fixed times from stations around the world, and report conditions with height.
- Weather satellites photograph clouds, measure temperature and moisture at different levels, and track winds by following cloud movement. Some orbit in step with Earth’s rotation to watch one region continuously, while others pass over the poles.
- Radar sends out radio pulses and detects the echoes from raindrops, snow and hail, showing where precipitation is and how it is moving.
- Aircraft report wind and temperature in flight, adding data at cruising altitudes.
The coverage is uneven. Oceans, polar regions and remote land areas have far fewer measurements than populated regions, which is one source of error.
Step two: building the starting picture
Raw observations are scattered in space and time and contain small errors, so they cannot be fed directly into a model. A process called data assimilation blends them with a recent short forecast from the previous model run to produce a consistent, gridded snapshot of temperature, pressure, wind and moisture across the globe. This starting snapshot is called the initial condition.
Step three: running the model
The atmosphere obeys the laws of physics. Models apply the equations for how fluids flow, which are based on the laws of motion, together with the rules for energy conservation, the behavior of heat and moisture, and how air and water change phase. The model divides the atmosphere into a three-dimensional grid of boxes and steps the equations forward in time, calculating how each box affects its neighbors.
The first serious attempt at this idea was made by the English scientist Lewis Fry Richardson, who calculated a forecast by hand and published his method in 1922. The result was not accurate, and practical forecasting had to wait for electronic computers in the mid-twentieth century. Today’s models run on some of the most powerful supercomputers in the world. Smaller regional models use finer grids to capture detail, such as storms and the effects of mountains, while global models cover the whole planet more coarsely.
Features smaller than a grid box, like individual clouds, are not directly resolved. Models use simplified representations called parameterizations for these, and they are another source of uncertainty.
Why forecasts have limits
Two things make perfect forecasts impossible. One is that the starting picture is never exactly right. The other is that the atmosphere is chaotic, meaning that tiny differences in the starting conditions can grow into large differences later. The meteorologist Edward Lorenz identified this behavior in the early 1960s. It does not mean the weather is random. It means that predictability fades with time, and the practical limit for detailed day-by-day forecasts is often described as roughly one to two weeks.
Accuracy also depends on what is being predicted. The broad pattern of highs and lows is usually easier than exactly when and where a thunderstorm will occur. Temperature forecasts tend to be better than precipitation forecasts. The path of a large storm such as the ones described under hurricane formation is typically more predictable than its exact intensity.
Ensembles and probabilities
To account for uncertainty, forecast centers run the same model many times with slightly different starting conditions and settings. This group is called an ensemble. When most runs agree, confidence is higher. When the runs spread widely, the forecast is less certain, and forecasters can say so.
This is also the origin of “chance of rain” percentages. A 40 percent chance of rain generally means there is a 40 percent probability that measurable rain falls at a given point in the forecast area during the stated period. It does not mean it will rain for 40 percent of the day, although the exact definition can vary a little between agencies.
What happens after the model run
Raw model output usually has known biases, so forecasters and automated programs apply corrections based on how the model has performed in the past at specific locations. Forecasters also issue warnings for hazards such as floods, severe storms and heat, using radar and recent observations to update their judgment on short timescales. Very short-range predictions of the next hour or two, sometimes called nowcasts, lean more on current radar and satellite data than on full model runs.
New tools
Researchers are also testing forecasting systems based on machine learning, which learn patterns from decades of past weather data rather than solving the physical equations step by step. Early results are promising for some measures, though the methods are still being evaluated and are generally used alongside physical models. In the meantime, the central idea stays the same: observe carefully, compute carefully, and be honest about how much is uncertain.
The physical background for everything the models try to capture is laid out in the discussion of weather systems.