Sub-seasonal forecast
Brightband-SUBS
Brightband has found signal in the noise of weather forecasts at weeks 3 and 4. Brightband-SUBS is a global AI ensemble, run daily out to 32 days.
Start a free 2-week pilotForecast skill
Model specifications
- Coverage
- Global
- Resolution
- 1.5° (about 167 km at the equator)
- Ensemble
- 64 members
- Update frequency
- Once daily, 00z run
- Forecast horizon
- 32 days, daily steps
- Output
- Daily means (average of four 6-hourly states: 00z-18h, 00Z-12h, 00Z-6h, 00z)
- Variables
- 18 pressure levels up to 2 hPa and Surface variables Full list
- Delivery
- Zarr files in Earthmover Arraylake / Google Cloud Storage
- Backtest
- Held-out test set: 3 winters (2023–2026)20 winters of re-forecasts (October to March, 2003–2026)
Who it's for
-
Energy and commodity traders
A weekly temperature signal for weeks 3–4, delivered daily as a full ensemble you can feed into your own models.
-
Utilities and grid operators
An earlier view of demand and cold or heat risk for fuel, maintenance and crew planning.
-
Agriculture and logistics planners
Regional anomalies for any weather-exposed operation, weeks before the medium-range forecast can see them.
Forecast accuracy
Weekly-mean 2m temperature, weeks 3–4, 30–60°N land, winters 2023–26.
Request a pilot for a full video walkthrough of the results.
Versus climatology
The historical average for each date and place.
~30% lower error than climatology
Versus ECMWF EC46
ECMWF's extended-range forecast, debiased.
1.5–2× the skill of ECMWF EC46
Backtest data available for testing
October to March, 2003–2026.
- 20 reforecast winters
- 3 held out from training
AI Weather Quest
ECMWF's AI Weather Quest competition has showcased exciting progress on sub-seasonal skill. We ran Brightband's sub-seasonal forecast, BB-SUBS, through the AI Weather Quest evaluation code for 2m temperature and found that it is a significant step beyond every Weather Quest model. In separate evaluations we also found that BB-SUBS outperformed other commercial models.
| Model | Type | Week 3 | Week 4 |
|---|---|---|---|
| BB-SUBS | Brightband | 0.112 | 0.072 |
| LP | AIWP | 0.089 | 0.057 |
| MicroEnsemble | AIWP | 0.086 | 0.055 |
| S2S Multi-Model Mean | NWP | 0.086 | 0.049 |
| CLINT | AIWP | 0.080 | 0.050 |
| AIFS | AIWP | 0.076 | 0.040 |
| CMAandFDU | AIWP | 0.068 | 0.047 |
| EC46 | NWP | 0.062 | 0.036 |
| UWAtmosNVIDIA | AIWP | 0.033 | 0.009 |
| CliMA | AIWP | 0.020 | 0.026 |
How we measured it
We hold SUBS to the same standard we built Operational WeatherBench on: clear baselines, transparent methodology, and data you can use to check our work.
View Operational WeatherBench- Baselines
- Climatology (the historical average for each date and place) and ECMWF's subseasonal-range forecast (EC46, debiased).
- Target
- Weekly-mean 2m temperature over land, 30–60°N, with separate results for CONUS and Europe.
- Weekly skill, daily output
- Evaluation results are for weekly means. Daily values are provided in the backtest and operational product so you can build your own aggregates.
Try it before you buy
- 01
Understand the performance
A video walkthrough of the model evaluation results.
- 02
Run a 2-week pilot, with no pilot fee
Test Brightband-SUBS on held-out winters from 2023–2026 and the additional 20-winter reforecast archive (2003–2023).
- 03
Go live
With a daily operational feed.