SKYBENCHWeather models.
Reputation through evidence.
SKYBENCH measures how well weather models predict the real world. Submit a forecast before its deadline. Compare it with observations. Build a record anyone can inspect.
The current field
Weekly forecasts. Shared observations.
Loading the current season…
From prediction to proof.
A fixed target, a common observation, a record you can check.
Commit your forecast
Connect a wallet and register a model version. Seal its temperature prediction at least 24 hours before the target. The accepted value is final.
Measure what happened
Each model faces the same airport observations. Missing data stays visible, and every ranked entry is compared on the same set of tasks.
Inspect the evidence
After the deadline, predictions and source evidence become public. Open a model’s record, inspect the errors, and verify its receipt hashes.
A common test.
An open record.
The weekly league starts with a focused question: how accurately can a model predict tomorrow’s temperature?
Read the competition rules- Forecast target
- Air temperature · °C
- Observation time
- 12:00 UTC, every day
- Stations
- New York · London · Tokyo
- Weekly coverage
- 21 tasks across 7 dates
- Ranking measure
- Mean absolute error · lower is better
- Public evidence
- Forecasts, observations and hash receipts
Put your model to the test.
Enter through the workbench or connect your forecast pipeline through the API.
SKYBENCH currently uses server-held timestamps and verifiable content hashes. Model identity is self-declared; records are not yet anchored on-chain.