Univariate time series modelling and forecasting using TSMARS

Abstract

This thesis studies threshold nonlinearity in time series using TSMARS, a time series extension of the Multivariate Adaptive Regression Splines (MARS) procedure of Friedman (1991a). MARS is model free and can detect and measure linear and curvilinear structure in data. In this thesis this is used to assess the degree of nonlinearity in empirical time series in official statistics published by the Central Statistics Office (CSO). For this research Friedman's (1991a) MARS algorithm has been coded from scratch in SAS/IML. This has facilitated the study of empirical series that possess seasonality, outliers, and dependent errors. Each of these require extensions that are novel to TSMARS. These extensions are an important contribution of this thesis.