Title: AdaWeather: Logarithmic Regret in the Adaptive Fusion of Probabilistic Weather Forecasts

Abstract:

Recent breakthroughs in machine learning have yielded probabilistic weather forecasting models that rival the performance of leading numerical weather prediction systems. However, no single model maintains a consistent advantage across all spatial and temporal contexts, as their relative efficacy varies significantly depending on the specific situation. This variability underscores the need for adaptive strategies that integrate multiple forecasts to enhance both accuracy and robustness. Although ensemble forecasting has been explored in prior studies—typically relying on supervised learning or prediction-with-expert-advice frameworks—our work introduces AdaWeather, a novel adaptive framework. This approach synthesizes numerous probabilistic forecasts by leveraging machine learning techniques alongside a mixture-of-experts architecture to generate a superior, unified probabilistic prediction.

While conventional expert-based methods typically establish regret bounds relative to the single best-performing expert observed in hindsight, we expand both the algorithm and its theoretical analysis. We demonstrate that our method achieves logarithmic regret when compared against the optimal static mixture of experts in hindsight. In empirical evaluations focused on temperature forecasting, AdaWeather demonstrates measurable improvements over current state-of-the-art methods.

Source: arXiv Generated at: 2026-06-03 00:00:00 UTC