Explainer
how weather apps work.
Weather apps do not predict the weather. They read a prediction made on a government supercomputer and render it. Here is how that prediction is actually made, where the data comes from, why the forecast decays the further out you look — and where Vesper deliberately diverges from the pattern.
how weather apps predict the weather
No weather app predicts the weather. Almost every one of them reads a prediction made somewhere else — on a government supercomputer — and renders it. Understanding that one fact explains most of what is confusing about weather apps, including why they disagree, why they change their mind at 6am, and why the ten-day forecast is worth so much less than it looks.
The prediction itself happens in four stages, and the app is only the fourth.
Observation
Satellites, radar sweeps, weather balloons, ground stations, aircraft sensors, and ocean buoys report the current state of the atmosphere. Millions of readings per cycle, unevenly scattered, arriving at different times and with different error characteristics.
Data assimilation
The hard part, and the part nobody outside meteorology has heard of. Observations do not form a grid, and a model needs one. Assimilation blends the new readings into the model’s own previous forecast, weighting each by how much it can be trusted, to produce the single best estimate of right now. A forecast is only ever as good as this step.
Integration
The model steps that starting state forward through the equations of atmospheric physics — fluid dynamics, thermodynamics, radiation, moisture. The GFS does this on a 13 km global grid across 127 vertical layers, four times a day, out to 16 days. This is the stage people mean when they say “the supercomputer”.
Post-processing and display
Raw model output is systematically biased in known ways, so it is statistically corrected and downscaled toward local conditions. Only then does it reach an app, which picks the grid cell nearest you, rounds the numbers, chooses an icon, and writes a sentence. Everything you actually see happens here — in the layer that did no predicting at all.
why the forecast gets worse the further out you look
The atmosphere is a chaotic system: two starting states that differ by less than the measurement error will diverge into genuinely different weather given enough time. That is not a flaw in the models, and more compute does not remove it. It sets a hard ceiling on prediction somewhere around two weeks, and it is the reason a day-9 forecast is closer to climatology than to a forecast.
This is what ensembles exist to measure. Instead of running the model once, the forecast centre runs it many times from slightly perturbed starting conditions — NOAA’s GEFS uses 31 members — and reads the spread. Tight spread means the atmosphere is in a predictable mood. Wide spread means it is not, and any single number pulled out of that spread is a guess wearing a decimal point.
Almost no weather app shows you the spread. That is the single biggest piece of information the industry throws away between the supercomputer and your screen.
what changed in 2025: the models started learning
For fifty years, numerical weather prediction meant solving the physics. That is now no longer the only way it is done. ECMWF took its Artificial Intelligence Forecasting System (AIFS) operational on 25 February 2025, running alongside the physics-based IFS rather than replacing it. Machine learning models are trained on decades of past atmospheric states and learn the transition directly, which makes them dramatically cheaper to run — minutes on a handful of chips instead of hours on a supercomputer.
They are also, on many measures, better. ECMWF reports AIFS outperforming its physics-based forecasts across the majority of lead times and variables, with gains of up to 20% on tropical cyclone tracks. AIFS ensemble forecasts followed in July 2025.
What has not changed is the presentation layer. The forecast arriving in your weather app in 2026 may well have been produced by a neural network, and the app will still render it as “Partly cloudy. High of 72°F.” The prediction got better. The sentence did not.
the data pipeline
Weather forecasting starts with raw observations: satellite imagery, radar sweeps, weather balloon measurements, ground station readings, and ocean buoys. Government agencies feed this data into numerical weather prediction models that simulate the atmosphere on supercomputers.
These models produce gridded forecasts — temperature, wind, precipitation, humidity, and pressure at every point on a grid covering the globe. Weather apps then query these grids for your location and translate the numbers into the text and icons you see on your phone.
where the data comes from
NWS (GFS)
The National Weather Service Global Forecast System. A 13 km global grid on the FV3 dynamical core, 127 vertical layers, run four times a day and pushed out to 16 days. Free, public, and the backbone of most US weather apps.
ECMWF (IFS)
The European Centre for Medium-Range Weather Forecasts. Its high-resolution run uses a 9 km grid — finer than the GFS — and it is widely regarded as the most accurate global model at medium range. Used by premium apps, usually for the 3-to-10 day window.
Environment Canada (GDPS)
The Canadian Global Deterministic Prediction System. Strong for North American forecasts, less commonly used by US apps.
SunsetWx
Specializes in sunset and sunrise quality forecasting using satellite imagery and atmospheric modeling.
why most weather apps look identical
Since most weather apps pull from the same data sources, the numbers are nearly identical. What differs is the presentation layer — and most apps use the same approach: template-driven text that fills in variables.
“Partly cloudy. High of 72°F. Winds SW at 5-10 mph.” This sentence exists in some form in nearly every weather app. The template is the same; only the numbers change.
What Vesper does differently
same data, different voice.
Vesper uses the same meteorological sources as other apps. The difference is what happens next. Instead of filling a template, Vesper writes an original editorial brief — short, opinionated, and local.
And with Sunset Verify, Vesper is the only weather app that publicly grades its own predictions. Accuracy verification is rare in weather because it requires admitting when you’re wrong.
How do weather apps predict the weather?
Weather apps do not predict the weather themselves — they display a prediction made elsewhere. The forecast is produced in four stages: satellites, radar, balloons and ground stations observe the current atmosphere; data assimilation blends those readings into a single best estimate of the present; a numerical model steps that state forward through the equations of atmospheric physics on a supercomputer (the US GFS runs a 13 km global grid four times a day out to 16 days); and the raw output is then statistically corrected and handed to the app, which picks the grid cell nearest you and renders it as text and icons.
Frequently Asked Questions
How does the weather app on my phone actually work?
It fetches a forecast rather than calculating one. Your app sends your location to a weather data provider, receives the numbers for the grid cell nearest you, and renders them as text and icons. The prediction itself was made hours earlier on a government supercomputer running a numerical weather model.
How do weather apps predict rain?
For the next hour or two, mostly by extrapolating radar — tracking where the echoes are and how fast they are moving. Beyond that, the numerical model predicts it from physics, and the percentage you see is a probability of precipitation for the forecast area, not a measure of how heavy the rain will be.
What is data assimilation in weather forecasting?
The step that turns scattered observations into a usable starting point. Readings from satellites, balloons, aircraft and ground stations do not form a grid, and a model needs one, so assimilation blends the new observations into the model’s own previous forecast, weighting each by how much it can be trusted. It is the least visible stage and the one that most determines forecast quality.
Where do weather apps get their data?
Overwhelmingly from government meteorological services rather than their own instruments. The main sources are the US National Weather Service GFS, the European Centre for Medium-Range Weather Forecasts, and Environment Canada, reached either directly or through an intermediary API such as Apple WeatherKit. Weather Underground is the notable exception, adding a network of privately owned weather stations.
Why do weather forecasts change?
Because the model re-runs. The GFS produces a new forecast four times a day, each one starting from a fresh set of observations, and a small difference in the starting state can grow into a meaningfully different outcome several days out. A forecast that changes overnight is the system working, not failing.
Do weather apps use AI now?
The forecasts behind them increasingly do. ECMWF took its machine-learning forecast system, AIFS, operational on 25 February 2025, running it alongside the traditional physics-based model, and reports it outperforming that model across most lead times and variables. The apps displaying those forecasts have largely not changed — better prediction, same template sentence.
Do all weather apps use the same data?
Most weather apps pull from the same handful of sources: NWS, ECMWF, and Environment Canada. The data is largely identical — what differs is how each app presents and interprets it.
Why do different weather apps show different temperatures?
Small differences come from which model run an app uses, how often it updates, and whether it applies local corrections or smoothing algorithms.
Why do most weather apps look the same?
Because they use the same data sources and similar template-driven text. Vesper is different because it writes original editorial briefs instead of filling in templates.
How accurate are weather forecasts?
Modern 3-day forecasts are accurate about 80% of the time. Accuracy drops significantly beyond 7 days. Vesper is the only app that publicly tracks and publishes its prediction accuracy for sunsets.
What is the ECMWF model?
ECMWF is the European Centre for Medium-Range Weather Forecasts, widely regarded as the most accurate global weather model. Many weather apps use ECMWF data for their 3-to-10 day forecasts because it consistently outperforms the American GFS model at medium range. Vesper blends multiple sources including ECMWF and NWS.
Why do weather apps disagree with each other?
Apps can run different model generations, weight different sources, update on different schedules, and apply different smoothing or local-correction algorithms. Even when the underlying data is identical, those processing choices produce slightly different numbers on your screen for the same location at the same time.
Are weather forecasts getting more accurate over time?
Yes, steadily. A 5-day forecast today is roughly as accurate as a 3-day forecast was in the 1980s. Satellite resolution, model compute power, and data assimilation techniques have all improved. What has not improved much is how weather apps communicate that accuracy to users — which is what Vesper is trying to fix with editorial voice.
Why do weather apps show hour-by-hour forecasts 10 days out?
They should not. Hour-by-hour forecasts beyond 48 hours communicate false precision — the underlying model uncertainty is much larger than the cells in the grid suggest. Vesper deliberately avoids showing hour-by-hour cells days out because it would misrepresent how confident the model actually is.
Try Vesper
see the difference yourself.
Vesper Sky is on the App Store — download it and your first Vesper Brief arrives with tomorrow morning’s sky.