Good average forecast skill does not imply good extreme-event skill. Pick a year and lead day to compare every model we have results for on both axes at once.
Regular RMSE is only available at lead days 1/3/5/7 (WeatherBench2's own evaluation grid uses lead hours, not the daily grid the extreme-event scores use) — other lead days show extreme-event columns only. "Ours (ablation)" rows are training-objective variants explored during development, kept for transparency rather than filtered out. Click a column header to sort.
Ours Published Ours (ablation) — hover a point for the exact model name.
Case study: one model, with and without a tail-aware correction
The SR sharpening head (paper Sect. 6): a frozen FourCastNet-architecture backbone with a small trainable head that corrects 2 m temperature and 10 m wind, trained on TailWeather's intensity score. Same rollout scored with and without the correction — isolates the effect from any cross-architecture confound.
Browses the ~275 MB illustrative subset of the full TailWeather archive: 50 named historical case-study events, regional (Nordics/Germany) time series, and one full global sample year (2003, the historic European heatwave year). Images below are pre-rendered (this is a static page — no live NetCDF parsing in the browser).