SEASONAL ANALYSIS OF ECONOMIC TIME SERIES

A brief review of existing methods for the seasonal decomposition of economic time series is followed by an overview of an approach to the problem using Bayesian statistical modelling. Within this framework, a class of models based on fractional Gaussian processes is introduced and compared with the better-known ARIMA class of models. A detailed outline is given of computer algorithms developed to utilize these models for applied seasonal analysis. New techniques are developed for maximum likelihood estimation of unknown parameters and for sensitivity analysis of final inferences about seasonal decompositions to uncertainties in the parameter estimation. A number of case studies on simulated and actual time series are presented in detail, leading to the principal conclusion that there is substantial sensitivity in desired inferences to aspects of model structure that may be difficult to estimate from the data.