Bayesian Model Averaging

Abstract

Bayesian model averaging (BMA) is a statistical method to rigorously take model uncertainty into account. This chapter gives a coherent overview on the statistical foundations and methods of BMA and its usefulness for forecasting, but also for the identification of robust determinants. The focus is given on economic applications. We describe the BMA framework in the context of linear models. Different model and parameter priors are discussed in detail and suitable inference methods and posterior analysis tools are presented. We illustrate the use of the presented BMA framework to study potential drivers of box office revenues and to forecast these revenues. The available data set does not only contain traditional variables such as budget or genre categorization, but also variables derived from social media content which capture sentiment and volume of Twitter messages. We discuss the impact of different model specifications and describe how results are obtained using the open-source package BMS available for the R environment for statistical computing and graphics.