Evaluating Conformal Prediction in Weather Forecasting

Thomas Mortier, Cas Decancq, Diego G. Miralles

Track: Type E (Late-Breaking Abstracts)

Keywords: Machine Learning, Conformal Prediction, Weather Forecasting

Abstract: Machine learning offers a promising alternative to traditional numerical weather prediction models, providing faster and more cost-effective forecasts. However, a major challenge remains the lack of reliable uncertainty quantification, which is essential for intepretable and trusthworthy predictions in complex climate systems. Existing Bayesian and frequentist approaches often face scalability issues or rely on restrictive assumptions. Conformal prediction (CP) offers a flexible, model-agnostic framework that yields valid prediction intervals without requiring assumptions about the underlying data distribution. Yet, its application to weather forecasting is limited by spatial and temporal dependencies that violate CP's exchangeability assumption. In this work, we assess the potential of CP for weather forecasting, discuss its main limitations, and highlight recent advances and future directions.

Submission Number: 96