Three ingredients
1. The physics: ECMWF
The European Centre for Medium-Range Weather Forecasts runs a physics model of the whole atmosphere twice a day, fed by satellites, weather balloons, aircraft and ground stations. It sees weather coming that no airport has reported yet. For each airport we take its forecast for the high, low, wind, gusts, humidity, cloud and chances of rain, thunder, snow and fog tomorrow, from the latest run that was public when our forecast is issued. Its grid cells are 28 km across, not a runway, so we correct it for the airport's height and for how far it ran warm or cold there over the last 30 days. On the held-out year that alone was off by 1.13°C on the high. We also take ECMWF's AI model, AIFS, trained on decades of its analyses: a second physics opinion and our source for cloud cover.
2. Our data and model: WT-1
WT-1 is a neural network of the kind behind language models, trained from scratch on GPUs on 26 years of hourly reports from 6,047 airports worldwide, 2000 to mid-2025. It reads the last three days of an airport's own reports and the last day at 32 airports around it, out to 1,500 km, hour by hour, and learns to say what comes next: the next 36 hours and tomorrow's high, low, wind, rain, thunder, snow and fog, each with a likely range. It knows each airport's habits, its sea breezes, valley fog and city heat, which a 28 km grid cannot. On the held-out year its high was off by 1.57°C, against 1.91°C for our boosted trees.
3. Combined and calibrated
For the high, low, wind, gusts, humidity, cloud and chances of rain, thunder, snow and fog, WT-1 and ECMWF are combined with weights learned on a year neither of them was trained on. ECMWF's two models, IFS and AIFS, together get 70–86% of the weight for the high and 37–69% for the low; WT-1 gets 14–30% and 31–64%. ECMWF gets less in the evening, when the latest airport reports say more. The chances are combined the same way, and the likely ranges are widened just enough to hold the measured value 80% of the time.
Our statistical models
The reports (called METARs) are what every airport publishes, usually each hour. A report is short: temperature, dew point, pressure, wind, visibility, cloud layers and any rain, snow, fog or thunder at the time. We read 2,722,598 forecast situations built from these reports since January 2021, four per airport per day, and turn each one into about 150 measurements.
About a third of those measurements describe the neighbourhood rather than the airport itself. For each airport we look at up to twelve others within 1,500 km: how warm, humid and cloudy it is to the north, east, south and west; what the air arriving from upwind looks like; and the shape of the pressure field, estimated by fitting a tilted plane through the neighbours' pressure readings. That slope and its curvature say where lows and highs sit, without anyone drawing a weather map.
Five statistical members
Normal for the date
Each airport gets its own calendar of normals, fitted as smooth yearly and daily cycles. Everything else is measured as a departure from these normals, so a warm day in Oslo and a warm day in Dubai mean the same thing to the model.
Fading memory
Weather that is unusual today tends to stay a little unusual tomorrow. This member learns how much of today's departure survives one day, separately for each time of day it is issued.
Network map
A single linear map from the whole neighbourhood's state today to tomorrow's departure from normal. It treats the station network as one system and learns the one-step rule that moves it forward.
Analog days
For each airport, the 30 past situations that looked most like today, at the same time of day and season. What happened the day after those situations becomes a forecast.
Boosted trees
Thousands of small decision trees, each correcting the last, find combinations of conditions that the simpler members miss.
One blended forecast
The members are combined with weights learned on months the members had not seen. Chances are calibrated so that a 40% chance of rain happens on about 40% of such days. The likely ranges for the high and low are widened just enough that they contain the true value about 80% of the time.
Finally, each airport's forecast is nudged by its own recent track record. A small filter follows how far the forecasts there have run warm or cold over recent weeks, and removes part of that drift. Tuned on the first half of the held-out year and tested on the second, it cut the remaining error by 0.3% to 2.1% depending on the quantity.
How accurate it is
To test the forecaster honestly we held back a full year, November 2025 to October 2026, that it never saw while learning. On that year the forecast high was off by 0.98°C on average, 67% better than assuming tomorrow will be like today and 65% better than the normal for the date. It landed within 2° of the high the airport measured on 89% of days. The overnight low was off by 1.01°C.
These figures average forecasts issued at four times of day (00, 06, 12 and 18 local time). The evening-before forecast, the one every airport's track record and the Compare page use, is a little better: 0.92°C on highs. "Off by" is the average size of the miss, up or down, in degrees Celsius.
What "the high" means
Our highs and lows are the warmest and coldest of the airport's hourly reports for the local day, and every forecast on this site is scored against them. An airport's official daily high and low come from a thermometer that records continuously, so they differ slightly. At 90 US airports, over the held-out year, official highs averaged 0.36° above ours and official lows 0.46° below, and were 1° or more apart on 10% and 18% of days. Scored against the official values our evening forecast was off by 1.05° on highs and 1.07° on lows, instead of 0.98° and 0.95°. Forecasts that aim at the official value lose about that gap when scored on ours, unless they are corrected for each airport as on the Compare page.
ECMWF forecasts: this service is based on data and products of the European Centre for Medium-Range Weather Forecasts (ECMWF), source www.ecmwf.int, licensed under CC BY 4.0. We interpolate them to each airport, correct their bias and combine them with our own model.