Clear Tonight?
Your weather app's cloud forecast, corrected by an open model that learned from the forecast's own mistakes.
Loading…
Top row: the cloud cover your weather app forecast (lower is better). Bottom row: Clear Tonight's chance that the hour is actually clear (cloud ≤ 25%).
Does the correction actually help?
How many nights of history does it need?
Brier score (lower is better), averaged over 6 cities × 6 months, training only on the last K nights before each month. TabPFN v2 is the only 12-feature model that beats the raw forecast from the first week; the 12-feature logistic regression and gradient boosting need two to four months before they're even as good as the raw forecast.