Determining the order of an autoregressive progress is an ubiquitous problem which arise in many fields. The overall objective of this thesis is to examine whether it is possible to develop an automatic procedure for the discrimination of autoregressive models capable of implementation on a microcomputer. It should be appreciated, however, the author is not proposing that such a complex task can be completely divorced from the analytic capabilities of human researchers.
As parameter estimate techniques form an integral part of many order identification proposals, the principle methods for estimating the parameters of a stochastic linear system are reviewed. Then Box-Jenkins and related schemes for determining the order of an autoregressive process are considered. Drawing upon the theoretic concept of entropy, a frame-work by which econometric and time series models can be linked in outlined and information theory based criteria are examined.
For time series analysis the information theory approach appears to be the most suitable base for an automatic procedure which aims not only to identify the most appropriate model order, but a1so to estimate the model parameters. However, an empirical investigation of both the traditional and information theoretic approaches clearly indicates the strengths and weaknesses of each technique. Informational criteria are the more robust model identification procedures and are undoubtedly superior to the more traditional statistical testing procedures. At the same time, however, ordinary least squares clearly emerges as the most robust parameter estimation method.