Tukuk-OSStep One: Observe Everything, Everywhere, at Once
Forecasting begins with the hardest simple question: what is the atmosphere doing right now, at every point on Earth? The answer comes from a machine the whole world maintains together. Twice daily, at the same moment (00 and 12 UTC), hundreds of weather stations worldwide launch radiosondes — instrument balloons that radio back temperature, pressure and humidity all the way up to the stratosphere. Add satellites sounding infrared and microwave profiles, ships and buoys, aircraft reporting winds, radar seeing precipitation directly, and ground stations. This global snapshot is called an analysis, and it is why every country's meteorological service shares data through the World Meteorological Organization for free: a hole in one nation's observations degrades every other nation's forecast. You can watch the current state of that snapshot city by city at tukuk.org/weather, and the reading of the synoptic maps it produces — high and low pressure, fronts, troughs — is covered in how to read a weather map.
Step Two: Solve Equations Nobody Can Solve Exactly
The atmosphere obeys fluid dynamics — the Navier–Stokes equations plus heat, moisture and rotation. In principle, given the snapshot, the future follows. In practice, the equations cannot be solved exactly for anything as messy as a planet, so models chop the world into a 3D grid — tens of kilometres across, a few hundred layers deep on current leading systems — and step time forward second by second, computing what each grid box does to its neighbours. The trouble is that the important things — clouds, rain, turbulence, convection — are smaller than the grid boxes, so they have to be guessed from statistics inside each box. These "parameterisations" are the part that keeps forecasters humble: whole schools of meteorology exist to argue about how well they work.
Step Three: Cheat Honestly — Ensembles and Chaos
In 1963 Edward Lorenz discovered that rounding a number in a weather model produced a different world two weeks later: the atmosphere is a chaotic system, where tiny errors grow exponentially. No analysis is perfect, so the honest response is not one forecast but dozens: run the model many times with slightly nudged starting conditions (and sometimes different physics) and count how often it rains in week one, or day four. That count is the probability you are quoted — "70% chance" usually means ensemble members that rain. It also explains the accuracy curve: about one extra day of skill per decade since 1980, a five-day forecast now beating what a three-day forecast did in 1980 — and a hard ceiling near ten to fourteen days, because past that chaos wins no matter how big the computer.
Step Four: The Human and the Local
Supercomputer output is not a forecast — it is raw physics that must be interpreted, bias-corrected for your coast or valley, and wrapped in judgement: when models disagree about where a low will track, a forecaster weighs which error mode matches today's setup, exactly as explained for the clouds themselves in cloud types and what they mean and for the bigger picture in weather versus climate. The final public product is deliberately conservative: warnings are issued for impacts, probabilities for the uncertain. The next time a forecast says 40% and you get sun, you were not wronged — you were inside a system that measured the whole planet, ran it on one of the fastest machines on Earth, and honestly admitted what it could not know. That is a far better deal than a confident lie.