Data-driven seasonal climate predictions via variational inference and transformers

Most operational climate services providers base their seasonal predictions on initialised general circulation models (GCMs) or empirical statistical techniques. GCMs are widely used but require substantial computational resources, limiting their capacity. In contrast, statistical methods often lack robustness due to the short historical records available. Recent works propose machine learning methods trained on climate model output, leveraging larger sample sizes. Yet, many of these studies focus on prediction tasks that may be restricted in spatial or temporal extent, thereby creating a gap with existing operational predictions. Others fail to disentangle the sources of skill in the context of climate change, where strong trends provide spurious estimates. This study combines variational inference with transformers to predict global and regional seasonal anomalies of temperature and rainfall. The model is trained on output from CMIP6 and tested using ERA5 reanalysis data. Temperature predictions demonstrate skill beyond the climatology and climate-change trend and even outperform the numerical state-of-the-art system SEAS5 in some ocean and land areas. Precipitation forecasts show more limited skill, with fewer regions outperforming climatology and fewer surpassing SEAS5. Furthermore, the consistency found in both teleconnections and skill spatial patterns against SEAS5 suggests that both systems build on similar sources of predictability.