How Weather Forecasts Work

Every morning, billions of people check a forecast before leaving the house. Most of the time, it’s right. Sometimes it’s wildly wrong. And rarely, it’s catastrophic. What most people don’t realize is that the prediction they see on their phone is the output of an enormous global machine one hundred times more powerful than anything that existed 50 years ago.

The short answer

A weather forecast starts with observations: millions of measurements from satellites, weather stations, buoys, aircraft, and balloons, all fed into a numerical weather prediction (NWP) model. These models divide the atmosphere into a 3D grid and use the laws of physics to calculate how air, heat, and moisture will move. The model runs on supercomputers, producing dozens of simulations that get blended into the forecast you see. The challenge is that the atmosphere is chaotic, meaning tiny errors in initial observations grow into large errors over time, which is why forecasts become unreliable beyond about a week.

The full picture

The observation network: sensing the atmosphere

Before any computer can predict the weather, humans need to know what the weather is doing right now. This requires observing the atmosphere at as many points as possible.

The backbone of this system is the Global Observing System, a coordinated network of instruments that measure temperature, pressure, humidity, wind speed, and more. It includes over 10,000 surface weather stations, thousands of upper-air stations that launch weather balloons twice daily, moored and drifting buoys in the oceans, commercial aircraft that report conditions during flights, and a fleet of weather satellites.

Weather balloons are particularly important. At about 1,300 locations worldwide, meteorologists release balloons with instruments called radiosondes attached. These ascend through the atmosphere, measuring conditions at different altitudes and transmitting data back to the ground. The balloons burst around 100,000 feet and the instruments fall back to earth, often recovered and reused. This twice-daily ritual, described by the National Weather Service, provides the vertical slice of the atmosphere that models desperately need.

Satellites have revolutionized observing. Geostationary satellites (orbiting at the same speed as Earth’s rotation) provide constant coverage of large regions, tracking cloud patterns and large-scale weather systems. Polar-orbiting satellites circle the Earth from pole to pole, providing detailed data as they pass over different locations. Together, they see storms forming, wildfires spreading, and heat waves building in near-real-time.

Numerical weather prediction: turning physics into forecasts

Once you have observations, the next step is turning them into predictions. This is done through numerical weather prediction (NWP), a method that uses computers to solve the equations that govern atmospheric motion.

The atmosphere obeys fluid dynamics. Air flows from high pressure to low pressure, warms and rises where heated, cools and sinks where it loses heat, and carries moisture that condenses into clouds and precipitation. These processes are described by equations originally derived in the 19th century: the Navier-Stokes equations for fluid motion, the first law of thermodynamics for heat, and equations for water vapor and other gases.

The challenge is that these equations are impossible to solve analytically for the real atmosphere. Instead, meteorologists discretize the problem: they divide the atmosphere into a 3D grid and calculate the state of each grid cell based on its neighbors. This requires supercomputers capable of performing quadrillions of calculations per second.

The European Centre for Medium-Range Weather Forecasts (ECMWF) runs one of the most respected models, known as IFS (Integrated Forecasting System). Its high-resolution forecast breaks the atmosphere into grid boxes about 9 kilometers across globally, with higher resolution (about 5 kilometers) for Europe. The U.S. runs the Global Forecast System (GFS), with similar resolution. These models run four times a day, producing forecasts that extend 10 to 16 days into the future.

Ensemble forecasting: hedging against chaos

A single forecast is useless because small errors in initial observations grow. Instead, meteorologists run ensembles: dozens or hundreds of slightly different versions of the model simultaneously.

Each ensemble member starts with slightly perturbed initial conditions. Some assume the temperature might be 0.1°C warmer. Others assume humidity is slightly lower. These perturbations are within the uncertainty of the observations, but they produce different outcomes.

If all ensemble members agree, the forecast is confident. If they diverge wildly, the forecast is uncertain. This is why you might see a forecast that says “70% chance of rain” not as a guess, but as the percentage of ensemble members that predicted rain for your location.

The ECMWF ensemble runs 51 members every six hours, producing 51 different possible futures that get averaged and analyzed. This approach has significantly improved forecast skill over the past few decades, extending the useful range of medium-range forecasts by several days.

From model output to your phone: human expertise

Raw model output is not a weather forecast. It’s a starting point.

Meteorologists, called forecasters, take model output and layer in their expertise. They know local terrain effects that models miss. They understand which model biases exist. They can interpret satellite imagery for cloud types that models struggle with.

This human-in-the-loop approach is crucial. WhenHurricane Sandy approached the U.S. East Coast in 2012, the models initially showed it curving harmlessly out to sea. Forecasters recognized the pattern from historical analog events and issued warnings that proved lifesaving, as described in post-event analyses by the National Weather Service. The models were wrong. The humans were right.

Most forecasting done for local TV news and apps involves taking model output and applying statistical post-processing. This corrects for known biases, like how models systematically overpredict high temperatures in summer or underpredict precipitation amounts. The final forecast you see is a blend of model physics and statistical correction.

The limits of prediction: chaos and uncertainty

There’s a hard limit to how far ahead we can meaningfully predict the weather. The atmosphere is a chaotic system, a concept first identified by mathematician Edward Lorenz in the 1960s. Chaotic systems are deterministic (they follow physical laws) but unpredictable (small differences in starting conditions produce vastly different outcomes).

This is why a 10-day forecast is only about 50% accurate, while a 5-day forecast is closer to 85-90%. Beyond about two weeks, the atmosphere essentially “forgets” its initial state and becomes unpredictable. Long-range forecasts (months to seasons) can only predict general trends, like whether a region will be warmer or cooler than average, not specific daily weather.

The chaotic nature of the atmosphere also means we can never fully “solve” weather prediction with more computing power alone. We will always be limited by the precision of our initial observations, which can never be perfect.

What forecasting gets right (and where it struggles)

Short-range forecasts (0-72 hours) are remarkably good. A 24-hour temperature forecast is typically within 2-3°F of what actually happens. Precipitation forecasts are accurate about 80-90% of the time for the first 24 hours.

Where forecasting struggles most is with convective storms: thunderstorms, tornadoes, and flash floods. These small-scale phenomena form quickly and depend on tiny variations in temperature and moisture that are hard to observe. Predicting exactly when and where a thunderstorm will form hours in advance remains one of the hardest problems in meteorology.

Hurricane track forecasting has improved dramatically. A 48-hour hurricane track forecast today is as accurate as a 24-hour forecast was 30 years ago. But intensity forecasting, predicting whether a storm will strengthen from Category 2 to Category 4, remains difficult.

Common misconceptions

Weather forecasts are getting worse. They’re actually getting much better. Five-day forecasts today are as accurate as three-day forecasts were 30 years ago. Ten-day forecasts continue to improve as models and computing power advance.

Climate and weather are the same thing. Climate is the long-term average of weather (typically 30 years). Weather is what’s happening right now. Climate change makes individual weather events more extreme, but doesn’t make every day warmer.

The forecast is always wrong. Most of the time, the forecast is right. A 70% chance of rain means it rains about 70% of the time such a forecast is issued. People remember the wrong forecasts more vividly because they’re more surprising.

Weather apps all use different data. Most weather apps use the same underlying model data (GFS, ECMWF, or their derivatives). Differences come from how each app post-processes the data and presents it to users.

Why it matters

Weather is the single biggest external factor in daily human life, affecting everything from what you wear to whether you can travel to whether your crops survive. The economic impact of weather forecasting runs into billions of dollars annually, from agriculture to aviation to retail.

More subtly, forecasting is a test case for how we understand complex systems. The atmosphere is a single connected system, and predicting it requires coordinating observations globally, computing at extreme scale, and applying physical understanding developed over centuries. The same techniques used for weather are being applied to climate prediction, ocean currents, and eventually, perhaps, to other complex systems we want to understand.