On alternative state space representations of time series models (Q1109668)

A state space representation of time series models different from models recently used in the literature is presented. Parameters of the model considered are estimated from the original data, unlike commonly proposed by other researchers, where a priori decomposition of data series into trend and cycles has been required. In the representation the also popular assumption of uncorrelated innovations of two components time series could be omitted. This state space representation is relatively new and unknown in economics and econometrics. It refers to cointegration properties of time series components and cointegration vectors are immediately obtainable. To illustrate the small sample behaviour of alternative representations, models used by Beveridge-Nelson (1981), Nelson (1988) as well as Watson (1986) are presented. The Beveridge- Nelson model performance is compared with alternative algorithms of Aoki by analysis of the latest innovation allocation in data to trend and cycles. It is pointed out that the application of a different representation results not only in various innovation allocations to trend and cycles but also the determined cycles have different durations. The results of modelling the US GNP series are discussed.