Bayes models and forecasts of Australian macroeconomic time series
Peter C. B. Phillips*
This paper provides an empirical implementation of some recent work by the author and Werner Pioberger on the development of Bayes models for time series. The methods offer a new data-based approach to model selection, to hypothesis testing and to forecast evaluation in the analysis of time series. A particular advantage of the approach is that modelling issues such as lag order, parameter constancy, and the presence of deterministic and stochastic trends all come within the compass of the same statistical methodology, as do the evaluation of forecasts from competing models. The paper shows how to build parsimonious empirical Bayes models using the new approach and applies the methodology to some Australian macroeconomic data. Bayes models are constructed for thirteen quarterly Australian macroeconomic time series over the period 1959(3)-1987(4). These models are compared with certain fixed format models (like an AR(4) + linear trend) in terms of their forecasting performance over the period 1988(1)-1991(4). The Bayes models are found to be superior to these forecasting exercises for two of the thirteen series, while at the same time being more parsimonious in form.
Not all econometric models are designed as instruments for forecasting Nevertheless, the capacity of one model to forecast adequately in comparison with competing models is an important element in the evaluation of its overall____________________