Entrepreneurship in Action, 21-24 Nov, online. Build a working product with AI

How Are Weather Forecasts Made? | How It Works

12 min read

On this page

What is the short answer?

A weather forecast is a calculation. Forecasting centres gather millions of measurements every day from balloons, satellites, ships, aircraft and ground stations.

They blend those measurements into one best picture of the atmosphere now. A computer model then applies the laws of physics to work out how that picture will change. Because small errors grow, centres run the model many times from slightly different starts. The spread becomes probabilities, which forecasters check before issuing the forecast the student reads on a phone.

Definition: Numerical weather prediction (NWP) forecasts the weather by solving the equations of physics for the atmosphere on a computer. The analysis is the model's best estimate of the atmosphere at the start time. An ensemble is a set of forecasts run from slightly different starting conditions.

Where do the observations come from?

ECMWF, the European Centre for Medium-Range Weather Forecasts, lists weather stations, aircraft, radiosondes, ships and buoys among the in situ sources, which measure the air where they are. Remote sensing comes mainly from satellites and from ground-based weather radar, which sends out pulses and measures what bounces back.

A radiosonde is an instrument package lifted by a weather balloon. It measures pressure, temperature and humidity as it rises, and tracking its GPS position gives the wind. Most launches worldwide happen twice a day, at 00 and 12 UTC. NOAA counts about 1,300 upper-air stations worldwide; the United States' National Weather Service runs 92.

ECMWF receives around 800 million observations a day and makes around 60 million quality-controlled ones available to its model, most from about 90 satellite instruments.

In 2019, the World Meteorological Organization (WMO) approved a Global Basic Observing Network, with standards such as hourly surface observations and twice-daily balloon flights. WMO members are committed to sharing real-time data worldwide, though in some places observations are not made or not shared.

How do scattered observations become a starting picture?

ECMWF notes that observations are unevenly spread and carry errors, so alone they cannot describe the whole globe. Data assimilation fills the gaps.

Each cycle begins with a recent short-range forecast, called the background. New observations correct it, with background and observations weighed by their uncertainty, so a trusted measurement pulls harder than a doubtful one. The result is the analysis, the starting point for the next forecast.

The Met Office in the United Kingdom repeats this every six hours for its global model and every hour for its UK model, taking in tens of millions of observations per global cycle. ECMWF uses 4D-Var, which fits a short-range forecast to observations spread across a time window, such as twelve hours, while respecting the laws of physics.

How does a computer model step the weather forward?

The model divides the atmosphere into a three-dimensional grid: columns over the surface, stacked in layers. Each point holds values such as temperature, wind and humidity. Equations for the motion of air, heat and water give how each value changes over a short time step. The model takes the step and repeats.

Grid spacing is the distance between neighbouring points. ECMWF's physics-based Integrated Forecasting System (IFS) has used 9 km for all its medium-range forecasts since 27 June 2023, when its ensemble moved from 18 km. NOAA's Global Forecast System (GFS) in the United States uses 13 km and 127 vertical layers, and forecasts up to 16 days ahead.

In India, the Bharat Forecast System, developed by the Indian Institute of Tropical Meteorology and launched on 26 May 2025, raised the resolution from 12 km to 6 km. The Met Office runs its UK model at 1.5 km.

No grid sees what falls between its points. ECMWF notes that radiation, convection, clouds and turbulent mixing often act at scales smaller than the grid. Parametrisation represents their combined effect using the values the grid does hold. ECMWF's supercomputer has over one million processors.

Why do forecasts lose skill, and what is an ensemble?

In the early 1960s, the MIT meteorologist Edward Lorenz found that small differences in a system such as the atmosphere can grow into large, unexpected results. This sensitive dependence on initial conditions lies at the heart of chaos. A 1972 talk of his asked whether a butterfly's wings in Brazil could set off a tornado in Texas, giving the idea its name, the butterfly effect.

Every analysis holds small errors. The Met Office explains that a chaotic atmosphere amplifies them, so detail can only be predicted so far ahead. MIT's account of Lorenz's work puts it at about two or three weeks.

ECMWF's ensemble runs 15 days ahead with 51 members: one unperturbed control forecast and 50 perturbed members, each started from a slightly altered analysis.

The Met Office designs members to be equally likely, and estimates the chance of an event by counting the share of members that forecast it. If six of 24 predict more than 5 mm of rain at a place, the chance is 1 in 4, or 25%. Agreement among members means high confidence.

How does model output become the forecast people see?

A model can be consistently a little too warm or too windy at one place. Post-processing corrects such habits with statistics.

In the United States, the National Weather Service's Model Output Statistics (MOS) uses regression to relate model output to the weather later observed, correcting certain systematic biases. Its National Blend of Models combines guidance from NWS and other models into calibrated guidance, a starting point for the forecast. The Met Office calibrates ensemble probabilities using past performance.

The forecaster then judges. The Met Office notes that forecasters often study individual ensemble members, while users who need probabilities choose a threshold that balances useful alerts against false alarms.

The result is often a probability of precipitation. The National Weather Service office for Atlanta explains that a 40 percent chance of rain means a 40 percent probability that the forecast point gets at least 0.01 inch, about 0.25 mm.

How did forecasting by calculation begin, and what failed in 1987?

Lewis Fry Richardson tried first. Serving with an ambulance unit in France in the First World War, he calculated a six-hour pressure change by hand, publishing the method in 1922.

He also imagined a forecast factory of 64,000 human computers. His answer, a pressure change of 145 hPa in six hours, was wildly unrealistic. Peter Lynch later traced the error to an imbalance between the pressure and wind fields in the starting data, not the method.

In April 1950, Jule Charney's group made the first one-day computer forecast on ENIAC in the United States. After breakdowns, it took more than 24 hours.

On the night of 15 to 16 October 1987, the Great Storm struck the United Kingdom with gusts of up to 115 mph, killing 18 people. The models had suggested the weather would pass south of England.

The storm formed over the Bay of Biscay, where few observations were taken, and very few came from satellites. Computing limits held the global model's grid boxes to 150 km. Afterwards, the Met Office set up the National Severe Weather Warning Service and placed deep-ocean buoys around the British Isles.

How good are forecasts now, and where do machine-learning models fit?

One headline measure of skill at ECMWF is the anomaly correlation, which compares a forecast's departures from normal with the departures that occurred. ECMWF tracks the day at which it falls to 60%. In the early 1980s that came at about day 5. In February 2010, the monthly average stayed above 60% for all 10 forecast days for the first time.

ECMWF called this steady improvement of around one day per decade. The Met Office says its four-day forecast now matches its one-day forecast of 30 years ago.

Machine-learning models are newer. ECMWF's Artificial Intelligence Forecasting System (AIFS) is trained on how the weather has evolved in the past, instead of coding the laws of physics as the IFS does. It became operational on 25 February 2025, on a 28 km grid, from the same initial conditions as the IFS. A 51-member ensemble version followed on 1 July 2025.

ECMWF reports that the AIFS beats physics-based models on many measures, by up to 20%, using about 1,000 times less energy. It calls the two systems complementary, and says the physics-based ensemble remains indispensable for high-resolution detail.

How does this work with real numbers?

Both examples are illustrative, except the named grid spacings.

How does an ensemble give a chance of rain?

Take 51 equally likely members. For one town on Saturday afternoon, 17 give at least 1 mm of rain and 6 give more than 10 mm.

Table: Illustrative probabilities from a 51-member ensemble

EventMembersCalculationProbability
At least 1 mm1717 ÷ 51≈ 33%
More than 10 mm66 ÷ 51≈ 12%

Each member is worth 1 ÷ 51 ≈ 1.96 percentage points, so one more wet member adds about 2.

How fast does the number of grid points grow?

Take an illustrative square 900 km on each side, with an illustrative 100 vertical levels.

Table: Grid points over an illustrative 900 km square

Grid spacingColumnsPoints with 100 levels
150 km (Met Office global model, 1987)6 × 6 = 363,600
9 km (ECMWF since 2023)100 × 100 = 10,0001,000,000
4.5 km (half of 9 km)200 × 200 = 40,0004,000,000

Halving the spacing doubles the points along each side, so the columns grow four times. Moving from 150 km to 9 km multiplies them by 10,000 ÷ 36 ≈ 278.

Where can the student go deeper in World 101?

This page follows one forecast, from a balloon launch to the chance of rain on a phone. The World 101 subject Climate Zones and Weather Systems covers the wider field: the atmosphere and its large-scale wind patterns, how weather differs from climate, and the careers built on both, from meteorologist to climatologist.

Three related subjects take these ideas further. Chaos Theory and the Butterfly Effect explores why tiny differences grow, the idea behind every ensemble. Modelling and Simulation shows how equations become computer models tested against reality. Satellite Imagery and Remote Sensing explains how instruments in orbit measure the Earth and its air.

Sources

  • NOAA JetStream, "Radiosondes", updated 16 September 2025.
  • WMO, "New Global Basic Observing Network gets go-ahead", 7 June 2019.
  • ECMWF, "Fact sheet: Earth system data assimilation", 15 July 2025; satellite data article, 3 September 2026; "Key facts and figures", December 2025.
  • Met Office, "How the Met Office uses data assimilation to produce its forecasts", 5 November 2025.
  • ECMWF, "Model upgrade increases skill and unifies medium-range resolutions", 27 June 2023.
  • ECMWF, "Atmospheric physics", and Forecast User Guide, "Generation of the Ensemble", updated 24 May 2026.
  • NOAA Environmental Modeling Center, "Global Forecast System", read 8 November 2026.
  • Press Information Bureau, Ministry of Earth Sciences, Bharat Forecast System release, 26 May 2025.
  • MIT News, Edward Lorenz obituary, 16 April 2008 (his 1972 talk).
  • Met Office, ensemble pages and "How accurate are our public forecasts?", read 8 November 2026.
  • NOAA Meteorological Development Laboratory, MOS and National Blend of Models pages, and NWS Atlanta, PoP FAQ, read 8 November 2026.
  • P. Lynch, "Richardson's Forecast-factory", 1993, and "Richardson's Marvellous Forecast", 1999, on Richardson's 1922 book.
  • L. Morone and C. Gillis, "The History of Numerical Weather Prediction", NOAA, December 2007 (ENIAC, April 1950).
  • Met Office, "Lessons and legacy of the Great Storm of 1987", read 8 November 2026.
  • ECMWF, "Landmark in forecast performance", 12 March 2010.
  • ECMWF, "ECMWF's AI forecasts become operational", 25 February 2025, and ensemble version, 1 July 2025.

Common errors and misconceptions

  • Misconception: A 40% chance of rain means rain over 40% of the area. Correct: The NWS office for Atlanta explains it as a 40% probability that the forecast point gets at least 0.01 inch.
  • Misconception: A big enough computer could forecast detailed weather months ahead. Correct: Chaos amplifies tiny errors in the analysis, setting a limit of roughly two or three weeks.
  • Misconception: Forecasts come from satellite pictures alone. Correct: Data assimilation blends satellites with balloons, stations, ships, aircraft, buoys and radar.
  • Misconception: A weather model calculates every cloud. Correct: Clouds and turbulence often act below the grid spacing, so parametrisation represents their effect.
  • Misconception: Machine-learning models have replaced physics-based ones. Correct: ECMWF runs its AIFS beside the IFS, from the same initial conditions, and calls them complementary.

Key takeaways

  • Forecasts start from observations by satellites, radiosondes, stations, ships, aircraft, buoys and radar, shared between countries through the WMO.
  • Data assimilation corrects a short-range background forecast with new observations, weighed by their uncertainty, to produce the analysis that starts each forecast.
  • Models step the physics forward on a 3D grid; ECMWF uses 9 km and the GFS 13 km, and smaller processes are parametrised.
  • Chaos makes small starting errors grow, so ECMWF runs 51 members, and the share of members showing an event gives its probability.
  • Post-processing such as MOS corrects certain systematic model biases, and forecasters weigh ensemble members and calibrated guidance before issuing warnings.
  • ECMWF's skill has improved by about a day per decade, and its machine-learning model now runs alongside its physics-based model, not instead of it.

Test yourself

What does data assimilation start from, and what does it produce?

It starts from a short-range background forecast, corrects it with new observations weighed by uncertainty, and produces the analysis.

Why must clouds be parametrised?

Clouds, convection and turbulence often act at scales smaller than the grid spacing, so the model cannot calculate them directly.

Why is the chance of at least 1 mm about 33% in the example?

Seventeen of the 51 equally likely members give at least 1 mm, and 17 ÷ 51 ≈ 0.33.

What happens to the grid columns when the spacing halves?

The points along each side double, so the columns grow four times, from 10,000 to 40,000.

Why was Richardson's hand forecast so unrealistic?

Peter Lynch traced the 145 hPa change to an imbalance between the pressure and wind fields in the starting data.

Give two reasons the 1987 storm was missed.

Few observations came from the Bay of Biscay or from satellites, and computing limits held the global model's grid boxes to 150 km.

Go deeper in World 101: Climate Zones and Weather Systems

Write about it

Turn a question from this chapter into a paper.

The Writing Studio has a structure guide, a rubric and citation tools for 15 formats, from essays to research papers. A free account keeps up to five drafts. Learnacy Essential adds feedback from an AI writing coach and submission to The OYI Review.

Publication depends on review and is not guaranteed.