Weather forecasting has improved measurably and continuously for decades, and it is among the clearer examples of scientific progress producing everyday benefit.
The measurement
Forecast skill is assessed against outcomes, which allows improvement to be quantified.
Which shows that a forecast today for a given lead time is roughly as accurate as a forecast for a substantially shorter lead time decades ago.
The gain has been approximately one day of lead time per decade over a long period.
Observation
Surface stations, weather balloons, aircraft reports, ships, buoys, radar and satellites.
Which together produce the initial state that models require, and the quality of that state limits forecast quality directly.
Satellite observation transformed coverage of oceans and remote regions where surface observation was sparse.
Numerical prediction
Solving the physical equations governing the atmosphere numerically.
Which was proposed before computers existed and required them to be practical.
The first computed forecasts took longer to produce than the period they predicted, which made them useless operationally and demonstrated the principle.
Chaos
Small differences in initial conditions produce large differences in outcome after sufficient time.
Which was discovered through weather modelling and became foundational in a broader field.
It sets a fundamental limit on deterministic forecasting regardless of computing power or observation quality.
Ensemble forecasting
Running many forecasts from slightly different initial conditions.
Which produces a range of outcomes rather than a single prediction, indicating confidence.
Where ensemble members agree, confidence is high. Where they diverge, it is not.
This is the basis for probabilistic forecasts, and communicating probability remains genuinely difficult.
Computing
Higher resolution requires more computation, and forecast centres operate among the largest computing installations in their countries.
Which allows finer grids resolving features that coarser models cannot represent.
Convection-permitting models, resolving individual storms, are comparatively recent and have improved short-range forecasting of severe weather substantially.
Machine learning
Models trained on historical data have recently matched or exceeded physics-based models on some measures.
Which is a substantial development, and they are far cheaper to run once trained.
Their behaviour in conditions unlike the training data is the open question, which matters most for the extreme events forecasts are needed for.
Warnings
Forecast improvement only produces benefit if warnings reach people and prompt action.
Which is a communication and behavioural problem, studied separately.
Impact-based warning, describing consequences rather than only meteorological thresholds, has been adopted by several services on this reasoning.
International cooperation
Weather does not respect borders, and forecasting depends on observations from everywhere.
Which produced one of the oldest and most successful international scientific cooperations, with free exchange of observational data.
That arrangement has come under pressure as some data became commercially valuable.
Seasonal forecasting
Predicting conditions months ahead, which relies on slowly varying influences including ocean temperature patterns.
Which produces useful skill for some regions and seasons and very little for others.
It is probabilistic by nature and is frequently reported as though it were deterministic.
Climate against weather
Weather forecasting predicts specific conditions at specific times. Climate projection describes statistical distributions over decades.
Which are different problems, and the chaos limit affecting one does not affect the other in the same way.
Confusion between them is a persistent feature of public discussion.
Commercial forecasting
Private providers serve sectors including energy, aviation, agriculture and retail.
Which generally build on public model output, and the relationship between public provision and commercial use is a live policy question.
Warnings and behaviour
Accurate forecasts only help if warnings reach people and prompt action.
Which is a communication problem studied separately from the meteorology.
False alarm rates affect subsequent response, which is why warning thresholds balance sensitivity against credibility.
Naming storms
Adopted by several services to improve communication of a single system.
Which has been evaluated with findings that named storms receive more attention and produce more protective action.
Historical records
Long observational records allow trends to be identified and allow models to be validated.
Which makes the continuity of observation stations valuable, and station closures break records that cannot be recovered.
Digitisation of historical paper records has extended usable datasets substantially.
Nowcasting
Very short range forecasting using radar and satellite observation rather than full model runs.
Which supports warnings for rapidly developing severe weather where model lead times are too long.
It has improved substantially with higher resolution observation and automated pattern recognition.
Access to data
Model output and observations are published freely by several services.
Which supports research, commercial application and public interest in a way that restricted data would not.
The economic value generated by free public data has been estimated and is substantial.
Verification
Forecast performance is measured and published by national services.
Which allows improvement to be demonstrated and allows services to be compared, and it is unusual among public services in this respect.
The measures used are technical, and headline accuracy figures conceal substantial variation by variable, region and lead time.