How Weather Forecasts Are Made, Step by Step

A weather forecast isn't a guess. It's the output of a pipeline that starts with millions of observations, feeds them into physics-based simulations, and ends with a human (or algorithm) interpreting the results. Here's how that pipeline works.

Step 1: Collect observations

Every forecast starts with data. Before a model can predict what the atmosphere will do next, it needs to know what the atmosphere is doing right now. This is called the initial state, and it comes from a mix of sources:

• Surface stations (METARs): Airports and weather stations report temperature, wind, pressure, humidity, visibility, and clouds every hour. There are roughly 10,000 stations worldwide issuing METAR reports.
• Radiosondes: Weather balloons launched twice daily from ~900 sites. They measure temperature, humidity, and wind from the surface up to about 30 km altitude. This vertical profile data is critical for model accuracy.
• Satellites: Geostationary satellites (GOES, Himawari, Meteosat) provide continuous imagery. Polar-orbiting satellites (JPSS, MetOp) measure atmospheric temperature and moisture profiles from space.
• Aircraft (AMDAR): Commercial aircraft report temperature, wind, and turbulence during flight. This fills gaps between radiosonde launch sites, especially over oceans.
• Ocean buoys and ships: Sea surface temperature, pressure, and wind from moored buoys, drifting buoys, and voluntary observing ships.
• Radar: Doppler radar detects precipitation intensity, type, and motion. Dual-polarisation radar can distinguish rain from hail, snow, and sleet.

All of this data flows into national meteorological centres (NOAA, ECMWF, Met Office, JMA, and others) within minutes of being recorded.

Step 2: Data assimilation

Raw observations are messy. Stations report at slightly different times. Instruments have biases. Some regions have dense coverage; others (oceans, poles, deserts) have almost none.

Data assimilation solves this. It's the process of combining observations with a short-range forecast (the "first guess") to produce a complete, physically consistent snapshot of the atmosphere at a single point in time. The most common technique is called 4D-Var (four-dimensional variational assimilation), used by ECMWF. NOAA's GFS uses a related method called EnKF (Ensemble Kalman Filter).

The output is a 3D grid of the atmosphere: temperature, humidity, wind, and pressure at every grid point, from the surface up to the stratosphere. This grid becomes the starting point for the forecast model.

Step 3: Run the numerical weather prediction model

This is the core of modern forecasting. Numerical Weather Prediction (NWP) models solve the equations of fluid dynamics, thermodynamics, and radiative transfer on a 3D grid covering the entire globe.

The two most widely used global models:

Each model run takes the initial state from Step 2 and steps forward in time. At every time step (a few minutes of simulated time), the model calculates:

• How air moves (wind, pressure gradients, Coriolis force)
• How heat transfers (solar radiation, infrared emission, conduction)
• Where moisture condenses (cloud formation, precipitation)
• What happens at the surface (evaporation, snow melt, soil moisture)

A single GFS run produces forecasts for every hour up to 120 hours out, then every 3 hours out to 384 hours. That's terabytes of output.

Sub-grid processes

At 13 km grid spacing, the model can't resolve individual thunderstorms, cumulus clouds, or turbulent eddies. These are handled by parameterisation schemes: simplified equations that estimate the net effect of small-scale processes on the grid-scale variables.

This is one of the biggest sources of forecast error. Two models using different convective parameterisations can produce very different precipitation forecasts from the same initial state.

Step 4: Post-processing

Raw model output has systematic biases. The GFS might consistently over-predict temperatures in certain regions, or under-predict wind speeds near coastlines. Post-processing corrects for this.

Common techniques:

• MOS (Model Output Statistics): Statistical relationships between model output and observed weather, trained on years of historical data. NOAA publishes MOS guidance for thousands of US stations.
• Bias correction: Simple adjustments based on recent model performance. If the model has been 2 °C too warm at a station over the past week, subtract 2 °C.
• Ensemble averaging: Running the model multiple times with slightly different initial conditions produces an ensemble. Averaging the members smooths out noise and gives a better central estimate. The spread between members tells you how confident the forecast is.
• Downscaling: Global models with 13 km grids can't capture local effects like valley cold pools, sea breezes, or urban heat islands. Regional models (HRRR at 3 km, AROME at 1.3 km) nest inside the global model to add local detail.

Step 5: Human interpretation

At national weather services, forecasters review model output and add their own judgement. They look for:

• Where models disagree (GFS vs ECMWF vs regional models)
• Whether the latest observations match the model's predictions
• Known model biases for their specific region
• Mesoscale features the models might miss (lake-effect snow, orographic rainfall, frontal positioning)

For routine weather, the model output is usually close enough. For high-impact events (severe storms, tropical cyclones, winter storms), the human forecaster adds significant value by recognising when models are struggling.

Step 6: Delivery

The final forecast reaches you through a chain of products:

1. Gridded data: NWP output in formats like GRIB2, accessible through APIs (NOAA's NOMADS, ECMWF's CDS)
2. Text forecasts: Written by local forecast offices (NWS in the US)
3. Apps and websites: Pull from model data, MOS, and sometimes proprietary post-processing
4. Alerts: Watches, warnings, and advisories issued when thresholds are exceeded

By the time you see "High of 28 °C" on your phone, that number has been through observation collection, data assimilation, a fluid dynamics simulation, statistical post-processing, and possibly human review.

Where things go wrong

Forecasts fail for predictable reasons:

• Bad initial data: Missing observations in data-sparse regions (open ocean, polar areas) degrade the initial state.
• Model resolution: A 13 km grid can't resolve a thunderstorm that's 5 km wide. It has to guess.
• Chaos: Small errors grow exponentially. Beyond 7-10 days, individual day-by-day forecasts become unreliable. Ensemble spreads widen and skill drops toward climatology.
• Parameterisation failures: When the simplified physics breaks down. Convective parameterisation is notorious for this, especially in the tropics.
• Rare events: Models are trained (implicitly) on common weather patterns. Novel or extreme events can fall outside the range where the model performs well.

How DailyHigh fits in

Our prediction engine takes a different approach to the daily high specifically. Instead of running a full NWP model, we:

1. Start with the GFS forecast for the station's grid point
2. Nudge it toward real-time METAR observations as they come in throughout the day
3. Apply bias correction based on the model's recent track record at that station
4. Update the prediction every time a new METAR arrives

This means the prediction gets more accurate as the day progresses. Early morning, it leans heavily on the GFS. By midday, it's mostly grounded in what the station has actually recorded.

You can see this in action on any station page: the predicted high tightens toward the observed max as the day goes on.

Further reading

• How to Read METAR Weather Data: the observation format at the start of the pipeline
• METAR vs TAF: observations vs forecasts, side by side
• NOAA's Environmental Modeling Center for GFS documentation
• ECMWF's IFS documentation for the European model