How Weather Forecasting Works: From Data Collection to Your Forecast App

Every forecast shown on a phone app or television broadcast traces back to the same basic pipeline: observations go in, a supercomputer solves atmospheric equations, and a forecaster or algorithm turns the output into a prediction. The U.S. National Oceanic and Atmospheric Administration (NOAA) describes this process simply as weather prediction built on math — equations that represent the physical behavior of the atmosphere, solved repeatedly as time moves forward.

The six steps behind a forecast

1. Observations are collected worldwide — from weather balloons, satellites, buoys, radars, sensors on commercial aircraft and ships, river and coastal gauges, and ground stations — and converted into numerical values for temperature, pressure, moisture and wind, according to NOAA.
2. Data assimilation blends these fresh observations with the atmosphere's most recent short-range forecast. The European Centre for Medium-Range Weather Forecasts (ECMWF) says hundreds of millions of observations are filtered and combined with recent short-range forecasts every day through this statistical process, producing the best estimate of the atmosphere's current state.
3. A numerical model divides the globe (or a region) into a three-dimensional grid of points. NOAA notes that finer grids give more local detail but demand sharply more computing power — doubling the resolution in each direction roughly quadruples the calculations required for the same area.
4. Supercomputers solve the physics equations at every grid point for short time steps, then feed the result back in as the starting point for the next step. NOAA's machines, located in Virginia and Florida, perform about 2.8 quadrillion calculations per second; ECMWF's system uses more than one million processors delivering roughly 30 petaflops.
5. The model is run many times with slightly different starting conditions or physics settings, called an ensemble. NOAA says that when the runs agree, forecasters have higher confidence; when they diverge, that spread itself communicates uncertainty to the public.
6. Human forecasters and automated systems interpret the raw model output — adjusting for local terrain, known model biases, or recent observations — before issuing the public forecast, watch or warning.

Global vs. regional models

NOAA distinguishes between global models, which cover the entire planet at coarser resolution, and regional models, which cover a smaller area in finer detail. ECMWF runs its global model twice a day, at midnight and noon, with two additional intermediate runs that supply boundary conditions to member states' own regional models. The agency notes that a global model is necessary because weather disturbances on one side of the world can influence conditions elsewhere days later.

Why forecasts lose accuracy over time

The atmosphere is a chaotic system: tiny differences in starting conditions grow over time until the forecast diverges from reality. This principle, first identified by MIT meteorologist Edward Lorenz, is why no model — however powerful — can forecast indefinitely into the future.

A study led by Penn State researcher Fuqing Zhang tested the two leading global forecasting systems — ECMWF's model and the U.S. next-generation global prediction system — against real events, including a European cold surge and Chinese flooding, using near-perfect initial data. The team concluded that dependable day-to-day forecasts are achievable about 9 to 10 days ahead for the mid-latitudes, where most of the world's population lives, and that this could extend to roughly 14 to 15 days as data collection, algorithms and computing power improve.

NOAA's own technical guidance on model output gives a more conservative figure for raw, unprocessed grid forecasts, noting that models remain generally accurate out to about four days, after which run-to-run variation increases significantly — one reason ensembles and human forecasters remain part of the pipeline for the medium range.

What's next for forecasting technology

ECMWF says future gains depend on finer-resolution ensemble forecasts, better representation of Earth system processes (atmosphere, land and ocean together), and improved assimilation of weather observations — all of which require more computing power. The centre's Scalability Programme is preparing for exascale computing (a billion billion calculations per second) and is also exploring machine-learning techniques as part of its forecasting systems, according to its own fact sheet.