Post-processing techniques are nowadays frequently used in order to reduce the impact of errors in ensemble forecasts of meteorological variables. Ensemble model output statistics (EMOS) are a widely spread post-processing approach built on a heteroscedastic linear regression model. After
replacing unknown parameters with suitable estimates, an estimative EMOS distribution function for prediction is obtained. However, it is well known that forecasts based on estimative EMOS may lack calibration, particularly when the number of ensembles is large compared to the number of historical
observations. Here, we suggest overcoming this drawback by applying in the EMOS context a predictive approach based on the concept of confidence distribution. The result is a new predictive distribution that takes the form of a variance correction of the classical estimative EMOS distribution.
The performance of the confidence EMOS distribution is tested on a real-data application for temperature
forecasting. It can be seen that our proposal performs better than the classical estimative EMOS, both in terms of coverage probabilities and log-score.
