Learning Ensemble Mixtures from Observations and Adapting Online with Theoretical Bounds

Neural weather models have comparable aggregate performance to numerical weather predictors, but with heterogeneity in performance since no single model dominates benchmarks. This naturally motivates combining models for improvements, but existing methods are usually simplistic averages on coarse grids with little online adaptation and no formal error bounds. We learn the best model based on historical gridded forecast and ground \textit{station} data, and adapt online with expert advice-based algorithms with novel bounds for the final forecast error relative to the best-in-hindsight. We get 40 \% improvements in key metrics relative to standard bias correction methods. This inexpensive framework can be extended to any point observation set to create bespoke predictions customized to different uses cases with formal guarantees, vital for climate change adaptation particularly in data sparse and compute constrained areas.