Conclusions

In this poster, we presented the postprocessing of a multi-model ensemble forecast (Poor Man's Ensemble; PME) using two different approaches. One approach applies non-homogeneous Gaussian regression or Ensemble Model Output Statistics (EMOS), and the other applies Distributional Regression Neural Networks (DRN). From the Results shown, the following conclusions arise:

• Using any of the postprocessing methodologies implies an improvement over the PME raw forecast.
• Among the methodologies used, DRN-Members, which uses the members of the PME as input features instead of mean and standard deviation, obtains the lowest mean CRPS and also shows the flattest rank histogram for 0-48 h lead times and a slightly positive bias for the 49-72 h lead time range.
• IMPROVER achieves better performance compared to DRN-Mean in terms of CRPS, but the latter exhibits better performance regarding rank histograms as it shows a slight positive bias while IMPROVER shows clear underdispersion.