Econometrics Seminar - Andres Ramirez Hassan (Monash University and Universidad EAFIT)

Title: Divergence-Based Forecast Combinations

Abstract: We propose an information theoretic framework for constructing forecasts based on linear pools. This framework allows us to construct forecasts that are specifically designed to outperform, in a user-chosen scoring rule, an existing, user-specified benchmark forecast. Furthermore, we demonstrate that this information theoretic approach to forecast combination couches several existing approaches based on maximizing, in a given scoring rule, the performance of the linear pool; examples include the approaches of Hall and Mitchell (2007), Geweke and Amisano (2011) and Opschoor et al. (2017). As such, this procedure gives useful information theoretic foundations for these often used approaches to choosing ensemble weights. In addition, our methodology enables multiple performance criteria to determine the weights, ensuring that accuracy along more than one dimension is ensured. Simulation and empirical examples are used to illustrate our proposed method.