Fig. 2: Gridded global forecast performance for selected variables.

a–h, Latitude-weighted RMSE using ERA5 (ref. ³⁴) reanalysis data as the ground truth, on the held-out test year (2018), for the four surface variables: 2-m temperature (a; T2M), 10-m eastward wind (b; U10), 10-m northward wind (c; V10) and mean sea level pressure (d; MSLP), as well as four headline upper-atmosphere variables: temperature at 850 hPa (e; T850), eastward wind at 700 hPa (f; U700), specific humidity at 700 hPa (g; Q700) and geopotential at 500 hPa (h; Z500) as a function of lead time t. At lead time t = 0, Aardvark predicted the initial atmospheric state from observational data alone. The error at t = 0 corresponds to the error in the initial state. Note that HRES has a non-zero error at t = 0 compared to ERA5 reanalysis. The HRES forecasts³³ we used have been conservatively re-gridded to prevent aliasing, and we performed the same operation on the GFS forecasts⁴⁹. We report the mean performance of each system together with 98% confidence intervals in our estimate of the mean performance.