A weather forecast is an estimate of how the atmosphere may evolve, built from recent observations, a model of the current state, and calculations of possible future states. Its detail and confidence depend on the weather pattern, location, variable, spatial scale, and lead time.
A newer run can change that picture. That is often new observations and a new analysis, not proof that the forecasting system failed. Read each value at the scale and valid time it supports.
From observations to an initial state
Forecast centres receive measurements from satellites, surface stations, weather balloons, aircraft, ships, buoys, and radar. These observations cover different places and times, contain measurement error, and leave gaps. A model cannot use them as a complete snapshot on their own.
How Weather Radar and Satellite Imagery Work explains what radar and satellite instruments measure, and where those maps still contain gaps.
Data assimilation combines quality-controlled observations with a short previous forecast, called the background. It weights both according to their estimated errors and produces an analysis, the best consistent estimate of the atmosphere and related parts of the Earth system at the model’s starting time. Assimilation also helps estimate uncertainty in that initial state.
New observations can change the analysis. Two model runs issued hours apart may therefore start from different estimates of the same developing weather system.
How forecast models move weather forward
Traditional numerical weather prediction models divide the atmosphere and surface into a three-dimensional grid. They solve equations for motion, heat, moisture, radiation, and other physical processes. Processes smaller than the grid, such as cloud microphysics or turbulence, must be represented through approximations called parametrisations.
Operational forecasting now also includes data-driven machine-learning models. These systems learn relationships in past analyses and forecasts, then predict how the atmospheric state will develop. ECMWF runs its AIFS alongside its physics-based IFS. AIFS still begins from an analysis produced through physics-based data assimilation. Machine learning provides another operational modelling method; it does not remove uncertainty, observation gaps, or the need to verify forecasts.
Model resolution matters, but a finer grid does not guarantee a better local forecast. Initial conditions, model physics, training data, terrain, and the type of weather all affect skill. A model value is an estimate for a grid cell, not a measurement in a particular street or garden. Terrain, coastlines, and local surfaces can make the weather at your location different from the grid value.
Ensembles show a range of outcomes
An ensemble runs multiple forecasts for the same valid time. Members may begin from slightly different initial states and may also perturb model physics, parameters, or stochastic components. These differences sample some of the uncertainty in the initial analysis and forecast model.
When members cluster around a similar outcome, confidence usually rises. When they separate, several outcomes remain plausible. Agreement is not a guarantee. Members can share the same bias, miss the same observation, or fail to represent a process. A narrow ensemble can therefore be overconfident, while a wide ensemble may still contain useful signals.
An ensemble is also different from a multi-model comparison. One ensemble can contain many runs of one model. A multi-model system combines guidance from different models and may capture additional structural uncertainty.
What happens after a model run
Raw model output often contains systematic bias and lacks detail at a specific site. Post-processing can correct known errors, calibrate probabilities against past performance, downscale fields, and combine guidance from more than one source. Calibration aims to make probabilities reliable. Over many comparable cases, events given a certain probability should occur at about that frequency.
Operational meteorologists at national services compare observations, deterministic models, ensembles, and local knowledge. They assess model conflicts and issue forecasts and hazard warnings. How to Read a Synoptic Weather Chart covers one of the pattern tools they use. Automated products can perform much of the routine processing, but a model run alone is not the entire public-warning process.
How far ahead a forecast remains useful
There is no universal table that assigns one confidence level to every lead time. Predictability depends on what is being forecast and how it is measured.
Large, slow weather patterns usually retain skill longer than the exact location and timing of a thunderstorm. Regional temperature can remain useful after precise rain timing has become uncertain. Temperature Explained and Humidity and Dew Point Explained cover those surface variables in more detail. Mountains, coastlines, convection, fog, and narrow rain-snow boundaries can reduce local skill earlier than a broad pressure pattern loses skill.
The often quoted two-week predictability limit is a useful description of early research into day-to-day atmospheric prediction, not a hard wall for every variable and scale. Some large-scale averages or recurring signals can retain skill beyond it. Precise local detail may lose useful skill much sooner.
The claim that forecasting gains one extra day of skill per decade is also a historical summary based on selected models, periods, and verification scores. It does not mean every location, weather type, or app forecast improves at that exact rate. Compare like-for-like verification rather than treating the phrase as a universal rule.
Forecasts beyond the day-to-day weather range usually shift towards probabilities and anomalies, such as the chance of a warmer or wetter period than normal. Seasonal outlooks do not specify the weather for a particular hour or day months ahead.
Why forecasts change
A newer forecast can differ because fresh observations changed the analysis, a weather feature developed differently than the previous run expected, model guidance shifted, or post-processing changed the local result. Small initial errors can also grow in a chaotic atmosphere.
A change does not prove that the earlier forecast was sound, nor does it automatically mean the forecasting system failed. The earlier forecast may have been wrong, or it may have represented the best-supported outcome before new information arrived. Check the issue time, valid period, recent trend, and official warnings instead of judging one update in isolation.
How to read probability and uncertainty
Use the forecast at the scale it supports:
- Match each value to its valid hour or day. A daily value and an hourly value answer different questions.
- Treat a precise local time farther ahead as less stable than a broad pattern.
- Look for repeated signals across updates. A single run can shift before the event.
- Use official watches and warnings for hazardous weather. They include impact thresholds and expert assessment that a general app forecast may not show.
Warning terms are not universal. In the United States, a Watch means a significant hazard is possible, while a Warning means it is imminent or occurring. The UK Met Office uses yellow, amber, and red warnings based on impact and likelihood, not the US watch-warning pair. Weather Warnings Explained covers how to read the official message.
For the US National Weather Service, probability of precipitation means the chance that at least 0.01 inch of liquid precipitation, or the liquid equivalent of frozen precipitation, reaches a point during the specified period. That amount is about 0.25 millimetres. It does not mean precipitation for that percentage of the period or over that percentage of the area. Other providers can use different thresholds, periods, or spatial definitions, so the product definition matters.
How Airpult builds and shows a forecast
Airpult currently reads forecast fields from ECMWF’s physics-based IFS, ECMWF’s data-driven AIFS Single, and NOAA’s physics-based GFS. Our pipeline applies a priority by field and availability. A lower-priority source fills hours that a higher-priority source does not provide; we do not average every model into every displayed value.
The public Airpult forecast provides 24 hours of hourly data and 10 daily rows. The web service caches a successful location forecast for five minutes. The underlying model cycles are less frequent: our current ingestion configuration uses six-hour cycles for GFS and AIFS Single and 12-hour cycles for IFS. A page refresh may therefore show the same model run. If a live upstream request fails, the service can temporarily use an older cached response for up to one hour.
The page presents one processed forecast, not an ensemble plume or confidence band. Its hourly precipitation percentage is a model-provider probability for that time step, and the page does not state a measurable threshold. Read it as a chance, not a rainfall amount or duration. For the measurement and precipitation-type details behind that value, see What Is Precipitation?.
