Forecast Combination and Bayesian Model Averaging: A Prior Sensitivity Analysis

In this study the forecast performance of model averaged forecasts is compared to that of alternative single models. Following Eklund and Karlsson (2007) we form posterior model probabilities - the weights for the combined forecast - based on the predictive likelihood. Extending the work of Fernández et al. (2001a) we carry out a prior sensitivity analysis for a key parameter in Bayesian model averaging (BMA): Zellner's g. The main results based on a simulation study are fourfold: First the predictive likelihood does always better than the traditionally employed 'marginal' likelihood in settings where the true model is not part of the model space. Secondly, and more striking, forecast accuracy as measured by the root mean square error (rmse) is maximized for the median probability model put forward by Barbieri and Berger (2003). On the other hand, model averaging excels in predicting direction of changes, a finding that is in line with Crespo Cuaresma (2007). Lastly, our recommendation concerning the prior on g is to choose the prior proposed by Laud and Ibrahim (1995) with a hold-out sample size of 25% to minimize the rmse (median model) and 75% to optimize direction of change forecasts (model averaging). We finally forecast the monthly industrial production output of six Central Eastern and South Eastern European (CESEE) economies for a one step ahead forecasting horizon. Following the aforementioned forecasting recommendations improves the out-of-sample statistics over a 30-period horizon beating for almost all countries the first order autoregressive benchmark model.(This abstract was borrowed from another version of this item.)