Effect of six-hourly forecasting on kilometre-scale skill

Determine whether bypassing sub-six-hour atmospheric timescales in the six-hourly Varda-single forecaster degrades the forecaster’s performance or indirectly limits the temporal downscaler’s performance, compared with architectures that jointly predict multiple hourly steps.

Background

Varda-single advances its primary forecast autoregressively in six-hour increments and uses a separate temporal downscaler to reconstruct intermediate hourly states. The authors argue that many kilometre-scale processes evolve on timescales shorter than six hours and may therefore be skipped by the primary forecaster.

The paper proposes jointly predicting multiple hourly steps as an alternative design, but leaves unresolved whether the six-hourly temporal structure is itself harmful to forecast quality or to the information available to the downscaler.

References

Most processes for which kilometre-scale information is useful take place and remain predictable only at timescales much shorter than 6\,\unit{h}. Because the forecasting model bypasses these timescales entirely, it has little incentive to incorporate that information in its latent space. We speculate that this might be detrimental to the forecaster's performance as well as, indirectly, to that of the temporal downscaler.

— Varda-single-1.0: deterministic data-driven weather forecasting at 1 km resolution over Switzerland's complex topography  (2610.01835 - Pennino et al., 1 Oct 2026) in Section 6, Conclusions