How Far Ahead Can We Forecast the Weather?

This simple relationship explains the whole history of forecasting. If errors double every day and a half, halving the starting error adds a day and a half of useful forecast. A tenfold improvement adds about five days. Nothing short of a perfect description of every gust and eddy could extend it indefinitely, and that is impossible. Worse, small features like thunderstorms grow their errors in hours, and those errors cascade upwards to larger weather systems. That cascade is why theory puts a ceiling of about two weeks on forecasting day-to-day weather in the mid-latitudes, however good the models become.

The same arithmetic is encouraging, too. Today’s forecasts still sit several days short of that ceiling, so there is real room left, and every gain in observations and assimilation converts directly into usable days.

A thunderstorm, a weather system and a season have different clocks

There is no single expiry date for a forecast. A local thunderstorm is sensitive to small-scale details that become uncertain within hours. The track of a large pressure system remains useful for days. Slower signals from the ocean, soil moisture and large-scale circulation can support probabilities for wetter or warmer weeks beyond the familiar two-week horizon. ECMWF explicitly distinguishes these scales. A forecast of increased seasonal flood risk can be useful without predicting which street will flood on a particular afternoon. ECMWF’s forecast error guide; ECMWF’s forecast skill horizon study

The logarithm in the model explains diminishing returns. If errors double every 1.5 days, making the starting error ten times smaller buys 1.5 × log₂(10), or 4.98 days. It does not buy ten times the forecast range. This is an illustration of error growth, not a universal law: doubling times depend on scale and flow, while errors generated inside the model also matter. Near the intrinsic limit, small-scale uncertainty spreads upward even when the initial large-scale picture is unusually accurate. Selz et al., Journal of the Atmospheric Sciences, 2022

Roughly two weeks describes detailed midlatitude weather under particular skill measures. It is not a ceiling on all probabilistic, subseasonal or seasonal information.

Hurricanes show the gains in human terms

In 2024 the U.S. National Hurricane Center’s track forecasts were the most accurate in its history at every lead time from 12 hours to five days. Its intensity forecasts have improved more slowly: storms that strengthen rapidly are still hard to anticipate, though forecasters’ average underestimate of such storms fell from 26 knots in 2010–14 to 16 knots in 2020–24. Better track forecasts mean smaller evacuation zones, fewer unnecessary evacuations, and more trust when an evacuation is ordered.

Africa’s surface weather stations are about an eighth as dense as the World Meteorological Organization recommends, and more stations report from Germany than from the whole African continent. That starves every model of data over the regions where the population is most exposed to droughts, floods and heat. The UN-backed Systematic Observations Financing Facility aims to fix part of this, supporting the least developed countries and small island states to meet basic observing standards, which would multiply internationally shared balloon data more than tenfold.

Warnings are improving fast. In 2025, 119 countries, 60% of the total, reported having a multi-hazard early warning system, more than double the number ten years earlier. The WMO estimates that investment in early warnings could deliver benefits of at least $162 billion a year, about ten times their cost, and that better forecasts and warnings could save around 23,000 lives a year.

Cheap forecasts still need expensive observations and trusted warnings

ECMWF put its deterministic AIFS into operational service on 25 February 2025, reporting roughly a thousandfold reduction in energy for generating a forecast. That boundary matters: it excludes the entire observing network, preparation of the initial atmospheric analysis and training the model. AI learns from reanalysis built with physics and observations and still needs an up-to-date starting state. It makes one part of the forecasting chain dramatically cheaper; it does not remove the rest. ECMWF’s AIFS operational announcement

GenCast’s published evaluation beat ECMWF’s comparison ensemble on 97.2% of 1,320 evaluated targets. Those targets and the historical test period define the result. A practical warning system also needs calibrated probabilities, local rainfall detail and dependable performance on unusual events. Cheap ensembles can help estimate rare-event risk, but more members do not fix a shared bias. Public value comes from a chain: observe the hazard, estimate its probability, connect it to local exposure, communicate an understandable warning, and make evacuation or protective action possible. Price et al., Nature, GenCast; WMO’s 2025 early-warning report

The delivery chain is editorial synthesis. A country reporting an early-warning system is not evidence that every resident can receive and act on a warning.

Who is building what

Forecast centres, AI labs and observation programmes. Search the record, or filter by role.

Benchmark scores compare models against reanalysis on chosen variables and periods. Operational value is shown by sustained performance in real time, including on rare extremes.

Four numbers to watch

First, the lead time at which forecasts lose useful skill, tracked each year by major centres. Second, the performance of AI models on record-breaking extremes, not just averages. Third, the number of weather balloon and surface stations reporting from the least observed regions. Fourth, the share of people covered by early warnings, which measures whether better forecasts reach anyone.