Bayesian forecasting with the structural damped trend model

Abstract

In this paper we consider the structural damped trend model which is standard in the arsenal of forecasting analysis. We consider both the multiple sources of error (MSOE) as well as the single source of errors (SSOE). Relative to existing research, we propose Bayesian analysis for estimation and forecasting based on Markov Chain Monte Carlo techniques and, especially, the Gibbs sampler with data augmentation. Monte Carlo and empirical applications (from the M3 competition as well as data from the Bank of International Settlements) show the superior performance of the MOSE versus the SSOE model. We also document superior performance of the Bayesian MSOE model versus its sampling-theory counterpart. Additional evidence is provided by a Bayesian optimal model pool approach which determines optimal weights in combining predictive posterior distributions.