Obtaining Consistent Probabilistic Predictions through Statistical Post-Processing of Ensembles

With the demand for reliable weather forecasts out to medium-range lead times increasing, forecasters often face the problem that a forecast revision, e.g. every time the NWP model is run, can potentially introduce jumpiness, in that two predictions for the same point in time contradict. This is not only frequently noted by the public, causing them to lose confidence in weather forecasts, but also reduces the value of the actual forecast.

The issue is somewhat addressed by using ensemble forecasts instead of relying on just one deterministic model, but ensembles are still subject to deterministic and probabilistic biases. By applying statistical ensemble post-processing methods like Ensemble Model Output Statistics or Bayesian Model Averaging, skillful probabilistic forecasts can be produced, but usually every model run and/or lead time is calibrated separately and the consistency between forecast revisions is not being considered.

We are looking to extend and enhance existing statistical post-processing in such a manner that both forecast skill and forecast consistency are improved, as well as suggesting a decision-theoretic approach to reduce jumpiness, if only deterministic forecasts are issued. This new approach has the potential to increase the value and usability of weather forecasts, especially over longer lead times.