Numerical weather prediction (NWP)

Numerical weather prediction forecasts future weather by taking a snapshot of the current atmosphere, temperature, pressure, wind and moisture at millions of grid points, and stepping the physical equations of fluid motion and thermodynamics forward in time on a computer. It is the physics-based core behind essentially all modern weather forecasts, from a phone app's hourly forecast to hurricane track guidance.

Model output depends heavily on the quality of the starting snapshot, called the initial condition, which is built by blending observations from satellites, weather stations and balloons through a process called data assimilation; small errors in that starting point grow over the forecast, which is why skill degrades sharply beyond roughly a week. This differs from reanalysis data, which reruns the same kind of model over historical periods to produce a consistent record of the past rather than a forecast of the future, and from purely statistical time series forecasting, which learns patterns from historical data without simulating atmospheric physics.

Because a single NWP run is highly sensitive to small initial-condition errors, forecasters run ensemble forecasting to quantify uncertainty, and short-range, high-resolution variants called nowcasting fill the gap for the next few hours. Pitfalls include mistaking a single deterministic run for a certain outcome, and comparing forecasts from models with different resolutions or physics without accounting for those differences.