See what changed
Compare the latest forecast with earlier runs and other models to understand how the outlook is shifting.
From forecast to foresight
Western EuropeHeat wave · May 22–31, 2026
See the heat build across Europe
7.25days
Sustained agreement with the regional peak temperature
WeatherMesh settled within 1°C of the May 26 regional temperature peak more than a week ahead, and stayed there in every later forecast—12 hours before AIFS.
Across the UK study region, Germany, France and Spain, WeatherMesh had lower temperature error than IFS at six reported leads from one to fifteen days.
Key takeaways
- A sustained signal for the regional peak from the May 19, 12 UTC forecast.
- Lower error than IFS across the reported 1–15 day leads; AIFS led at six days.
Forecast differences, 7 days ahead
2 m air temperature · Valid May 26, 2026 · 12 UTC
Drag to compare · Use arrow keys when focused
2 m air temperature (°C) · −5 to 40
Forecast − ERA5 (°C) · −6 to +6
Blue: cooler than analysis · Red: warmer than analysis
Forecast issued May 19, 2026 · 12 UTC · 168-hour lead
Source difference maps are smoothed (σ 56 km). Colors saturate beyond ±6°C.
Study details & evidence
About this study
The heat wave's peak is taken as the combined study region's temperature at 18 UTC on May 26 (26.6°C in ERA5). From the May 19, 12 UTC run onward, about a week before the peak, every WeatherMesh forecast stayed within 1°C of it. AIFS reached that accuracy 12 hours later. IFS and GFS did not reach it in any of the runs tested.
Temperature errors cover forecasts valid at 12 UTC on May 22–31. Only initialization and valid-time pairs available for every model are used. Land points are sampled on a common 0.5° grid and weighted by latitude. The UK region also includes Ireland and a small part of northern France.
WeatherMesh, IFS and AIFS are ensemble means; GFS is a single deterministic run. The accompanying map is a separate seven-day forecast, issued May 19 at 12 UTC and valid May 26 at 12 UTC, six hours before the headline peak.
This is a single heat wave, with 7–10 matched valid times per lead time. AIFS was the stronger model at many medium-range lead times in the UK and Germany. ERA5 is a reanalysis used as the reference, not direct station observations. The midday and early-evening samples are not daily maximum temperatures.
Want to explore the results in more detail? Talk to our team ↗
TexasFern · Jan 24–29, 2026
A clearer view of the Texas freeze
50%
Lower temperature error than IFS at 5–7 days
WeatherMesh averaged 2.37°C of temperature error versus 4.78°C for ECMWF’s IFS ensemble, verified against observations at 10 Texas airports.
The length of the freeze mattered, too. Around three days before the event, WeatherMesh forecast 51 freezing hours in San Antonio versus IFS’s 19. Observations recorded 69 hours.
Key takeaways
- Lower temperature error than IFS, AIFS and GFS in both the 5–7 and 7–10 day bands.
- A closer estimate of San Antonio’s prolonged freeze, relevant to planning for sustained heating demand.
Temperature error, 5–15 days ahead
2 m air temperature · Mean absolute error vs. observations
- WeatherMesh
- 2.37 °C
- WeatherNext-2
- 2.73 °C
- ECMWF AIFS
- 3.78 °C
- ECMWF IFS
- 4.78 °C
- GFS
- 4.16 °C
Lower error is better. WeatherMesh led at 5–7 days.
Ensemble means except GFS. WeatherNext-2 uses archived research forecasts.
Study details & evidence
About this study
Temperature accuracy is measured at 10 Texas airports over January 24–26, against quality-controlled METAR observations. All models are compared on the same forecast runs, valid times and stations.
The long-lead WeatherMesh results come from a retrospective backtest of the ensemble mean. A separate freeze-duration analysis covers January 24–29, with lead times measured to the start of that window.
The maps show seven-day forecasts issued January 18 and valid January 25 at 00 UTC, compared against ERA5 reanalysis. ERA5 is used only for the maps. The temperature-error chart is scored against station observations.
This is a single, selected event. WeatherNext-2 locked onto the peak cold earlier and held it, and it led on freeze duration at leads of 1.5–5 days. At 10–15 days, AIFS had slightly lower temperature error, and all models were too warm.
Want to explore the results in more detail? Talk to our team ↗
North AtlanticMelissa · Oct–Nov 2025
A more accurate track, four days ahead
WeatherMesh82km
NHC Official395km
Mean track error four days ahead
North of Bermuda, WeatherMesh averaged 79% less track error than NHC Official across five matched four-day forecasts.
A separate retrospective study near Jamaica revealed another benefit: some ensemble members showed a deep low 15 days ahead, while the averaged forecast showed little signal.
Key takeaways
- Track: a closer view of Melissa’s path as it moved into the North Atlantic.
- Ensemble: see stronger-storm possibilities that an averaged forecast can hide.
Forecast tracks vs. observed path
Position error in km · Times in UTC
Model comparison
Observed
Models: ensemble means, except NHC Official. Tracks end when no closed storm center is identified. Observed squares: post-tropical phase. Headline results cover only the track north of 32°N.
128 membersAveraged forecast mapERA5 analysis
- Strongest member’s low
- 973.7 hPa
- Averaged map’s low
- 1008.2 hPa
- ERA5 analysis low
- 981.7 hPa
23 of 128 members at or below 995 hPa · Highlighted dots at the selected lead.
The 0.25° grid does not resolve Category 5 intensity. Member counts are not calibrated storm probabilities.
Study details & evidence
About this study
At a four-day lead, WeatherMesh's mean track error was 82 km, compared with 395 km for the NHC official forecast. Both are verified against five HURDAT2 best-track positions north of Bermuda on October 31–November 1. Individual forecast cycles can be explored on the map.
The ensemble analysis looks at October 28, when Melissa was near Jamaica, using all 128 members from forecasts issued 1 to 15 days ahead. The chart compares the regional minimum pressure in each member, in the ensemble-mean field and in the ERA5 analysis.
The track and ensemble analyses cover different phases of the storm. At 0.25° resolution, the pressure comparison cannot resolve Category 5 intensity and does not show better intensity accuracy than other models. Spread across members is not a calibrated probability for the storm.
Want to explore the results in more detail? Talk to our team ↗
Delivery methods
Weather intelligence that fits the way you work.
Contact us for a solution tailored to your team.
Ask questions. Plan your next move.
Bring forecasts into your own systems
forecast.pyPython SDK / example
from windborne import get_point_forecasts
# By coordinates
point_forecasts = get_point_forecasts(
coordinates="37.77,-122.42")
# By station
point_forecasts = get_point_forecasts(
stations="panc;kjfk")Your forecasts. Your locations. One view.
Proven in industry
Focus on your markets
Track temperature, wind, and rainfall across the regions and stations that matter to your trading decisions.
Go from regional to local
Explore the wider weather pattern, then look closer at individual locations to see where conditions differ.
Uncertainty is expensive
When weather shapes your bottom line, better forecasts help you plan ahead and act with greater confidence.
Aviation
Plan for changing winds and weather to support efficient routes, fuel planning, and reliable operations.
Energy trading
Anticipate weather-driven shifts in demand and renewable generation to inform your next trading decision.
Shipping
Choose safer, more efficient routes around wind, waves, and ocean currents.
