2.1 Bayesian Model Averaging and DICast Forecasts - Ensuring Consistency

Monday, 11 January 2016: 1:30 PM
Room 226/227 ( New Orleans Ernest N. Morial Convention Center)
James Cowie, NCAR, Boulder, CO

Uncertainty in weather forecasting has traditionally been quantified by examining the behavior of ensembles which are run at national centers or locally on supercomputers or large clusters. But running these ensembles can be resource-intensive and beyond the reach of some researchers. In addition, while the ensembles exhibit a beneficial spread-skill relationship, they often tend to be underdispersive and hence limit their usefulness with regard to forecasting weather extremes. A simpler approach which overcomes some of these issues is Bayesian Model Averaging. With BMA, histories of several different deterministic model runs (or possibly ensemble members) can be used to create the predictive distributions for each of the input components. Probabilistic forecasts can then be generated using these predictive distributions as applied to current model runs outside the training data set. These forecasts have been shown to be well calibrated with a good spread-skill relationship.

While probabilistic forecasts are useful, most consumers of forecast information are used to forecasts in a deterministic format. The Research Applications Laboratory (RAL) at the National Center for Atmospheric Research (NCAR) developed the Dynamic Integrated Forecast system (DICast) over a decade ago to produce optimized, tuned consensus forecasts in a deterministic form. These forecasts are currently used by a wealth of societal sectors, including the general public, transportation, agriculture and energy.

When combining these probabilistic and deterministic forecasting systems together, a desired goal is that they show some form of correlation between the two systems, that is, the deterministic forecast should be close to or equal to the mean probabilistic forecast. Since the systems both use similar input data, this is often the case, however, they can diverge frequently enough to obviate the need to adjust one or the other forecast to bring them into agreement. A method to adjust the forecasts into agreement is discussed.

- Indicates paper has been withdrawn from meeting
- Indicates an Award Winner