Comparing predictive distributions in EMOS

EMOS models are widely used post-processing techniques for obtaining predictive distributions from ensembles for future weather variables. A predictive distribution can be easily obtained by substituting the unknown parameters with suitable estimates in the distribution of the future variable, thus obtaining a so called estimative distribution. Nonetheless, these distributions may perform poorly in terms of coverage probability of the corresponding quantiles. In this work we propose the use of calibrated predictive distributions in the context of EMOS models. The proposed calibrated predictive distribution improves on estimative solutions, producing quantiles with exact coverage level. A simulation study assesses the goodness of the calibrated predictive distribution in terms of coverage probabilities and also logarithmic score and CRPS.