Where Is Weather Predictable and Which Models Get It Right? Global Assessment of Conventional and AI-Based ForecastModels

Accurate medium-range weather guidance is essential, yet end-users lack clear, location-specific evidence on where forecasts are predictable and which freely available global systems perform best. We present a global evaluation of three conventional numerical weather prediction (NWP) models (ICON, IFS, GEFS) and an AI forecast model (AIFS) for 1-10-day lead times, focusing on four near-surface variables (2-m air temperature, precipitation, 10-m wind speed, and 2-m relative humidity). Using 00 UTC cycles over 1 September 2024 to 30 November 2025, we resampled forecasts to a 1° grid and assessed day-to-day variability (correlation and root mean square error), mean bias, variance bias, and lead-time dependence (drift) against multiple references (primarily JRA-3Q, with additional evaluation against ERA5, station data, and additionally IMERG-Late for precipitation). AIFS achieves the highest skill for temperature, precipitation, and wind at all lead times (relative humidity is unavailable from AIFS); at 3-day lead it explains, on average, 53\% more variance in daily precipitation globally than the next-best model (ICON), and 232\% more variance than GEFS in the tropics. Among the conventional systems, ICON is generally most skillful, while GEFS ranks lowest overall. Mean-bias drift is negligible across models, but variance drift is evident for several variables, most notably increasingly attenuated AIFS precipitation variability with lead time. Model correlation rankings are robust across reference datasets, although precipitation and humidity show greater reference sensitivity than temperature. We also map global predictability using 3-day lead daily temporal correlation of the locally best-performing model, showing highest predictability for temperature and wind in mid-to-high latitudes and markedly lower predictability for precipitation in the tropics. Our study provides actionable guidance on where global forecasts can be trusted and establishes a baseline for future AI and NWP model assessments.