Evaluating ensemble density combination - forecasting GDP and inflation

Forecast combination has become popular in central banks as a means to improve forecasts and to alleviate the risk of selecting poor models. However, if a model suite is populated with many similar models, then the weight attached to other independent models may be lower than warranted by their performance. One way to mitigate this problem is to group similar models into distinct `ensembles'. Using the original suite of models in Norges Bank's system for averaging models (SAM), we evaluate whether forecast performance can be improved by combining ensemble densities, rather than combining individual model densities directly. We evaluate performance both in terms of point forecasts and density forecasts, and test whether the densities are well-calibrated. We find encouraging results for combining ensembles.