ARMA Models for Cyclical Data

Abstract

The class of Autoregressive Moving Average (ARMA) models has long been recognised as excellent for representing the structure of stationary time series, and particularly those with marked cyclical components. This paper considers problems in the identification of ARMA models, and demonstrates semi-automatic procedure based on a least squares analysis of the predictor space. One of these procedures is found to yield promising results when applied to an artificial series and an economic series, both of which are short and cyclical. The spectra derived from the ARMA models are used for assessing the results.