How weather forecasts are made and why day ten is fiction

The forecast on your phone is not a person looking at the sky. It is the output of numerical weather prediction: the atmosphere carved into a three-dimensional grid of boxes, the physics of fluids and heat applied to every box, and the whole system stepped forward in time on some of the largest computers in Britain. The idea is a century old. Lewis Fry Richardson tried it by hand in the early 1920s and needed six weeks of arithmetic to produce a six-hour forecast, which was also badly wrong. The first useful computed forecast ran on ENIAC in 1950, and everything since has been the same method with vastly more machine behind it.

A modern forecast starts with observations. Satellites measuring radiation from the top of the atmosphere supply the bulk; weather balloons launched twice daily from stations worldwide profile temperature, humidity and wind through the full depth of the air; commercial aircraft report conditions along their routes; ships, buoys and thousands of land stations fill in the surface. The Met Office ingests tens of millions of these measurements every day. None of them, individually, is what the model needs. The crucial step is data assimilation: blending the new observations with the model's own previous short-range forecast to produce a single, physically consistent estimate of the entire atmosphere at one moment. That estimate, the analysis, is the starting line.

From there the model integrates the primitive equations, the expressions of Newton's laws, thermodynamics and moisture conservation for a rotating fluid, forward in steps of a few minutes. Resolution is the expensive part. The Met Office's global model works on cells roughly ten kilometres across, while its UK-specific model, the UKV, refines that to about 1.5 km, fine enough to represent individual shower clouds rather than smearing them into an average. Processes smaller than a grid box, such as turbulence or the microphysics inside a cloud, cannot be simulated directly and are approximated by parameterisation schemes, which is where much of the craft and much of the error lives. The supercomputing contract the Met Office signed with Microsoft, worth around £1.2 billion, exists almost entirely to push these numbers.

Why one run is never enough

If the starting analysis were perfect, one simulation would suffice. It never is, and the atmosphere punishes the gap. In 1963 Edward Lorenz, rerunning a simple convection model at MIT, restarted it from a printout that rounded 0.506127 to 0.506 and watched the new run diverge until it bore no resemblance to the original. The lesson generalises: errors in the initial state roughly double every two days or so, meaning an undetectably small mistake today becomes a different weather pattern within a week. This is chaos in the strict mathematical sense, sensitive dependence on initial conditions, and it cannot be engineered away, because no observing network can measure every cubic metre of air.

The response is ensemble forecasting. Instead of one run, centres launch dozens: the European Centre for Medium-Range Weather Forecasts, headquartered in Reading, runs a 51-member ensemble, and the Met Office runs its own MOGREPS system, each member starting from slightly perturbed initial conditions and sometimes slightly varied physics. Where the members agree, confidence is high; where they fan out, the honest answer is a probability, which is exactly what "70% chance of rain" encodes. Ensembles have driven real gains: a four-day forecast today is about as accurate as a one-day forecast was in 1990, and severe-event warnings for storms now arrive days ahead rather than hours.

The ten-day wall

Skill does not decline gently forever; it hits a floor. Beyond about a week the ensemble members have usually spread so far apart that the average of them drifts towards climatology, the long-run typical weather for the date. By day ten, for specific detail such as whether it will rain in Leeds on a given afternoon, the forecast contains little information that a historical almanac would not. Lorenz's own estimate of the theoretical ceiling, roughly two weeks, has survived six decades of exponential growth in computing power, because halving the initial error buys only about two extra days and the error can never reach zero. Longer-range outlooks do exist, but they forecast shifted odds, a warmer-than-average month or a stormier-than-usual fortnight, driven by slow components such as ocean temperatures, not weather for a date. The day-ten sunshine icon in your app is a single deterministic run pushed past its sell-by date. Treat the five-day forecast as engineering and the ten-day one as decoration.