Séance Séminaire

Séminaire de Probabilités et Statistique

Bayesian investigations on ensemble probabilistic forecasting.

Ensemble member prediction consists in altering initial conditions and parameters of a physically based dynamic deterministic model (such as regional meteorological models, climate models or rainfall-runoff models ...) in the hope of taking into account the uncertainty due to the initial system state (e.g. atmosphere temperature, soil moisture, etc.) or stemming from the uncomplete knowledge of parameters. When propagating limiting condition replicates through the complex integro-differential numerical code mimicking the system behavior, one obtains a collection of hypothetical trajectories of the variable to predict (e.g. airport temperature, rain, streamows...) also known as members. To what extend can the information conveyed by the various members of the ensemble help construct a probabilistic forecast of the quantity of interest? The statistical literature offers a wide range of successful ready-to-use methods known as "postprocessing" techniques and we will review whether they address the very specific features of ensemble members. Grounding on Bayesian decision theory as a unifying framework, we reinterpret the most common postprocessing techniques and suggest further developments. The performance of the proposed model is checked on simulated exchangeable ensemble data and compared to other postprocessing techniques for air temperature forecasting at the French meteorological station of Vouglans. Keywords: Bayesian model averaging; Bayesian processor of output; Ensemble post-processing; Ensemble prediction system; Hierarchical Bayesian model;