Use of Discriminant Analysis in Time Series Model Selection

Abstract

Authors: Agada Joseph Oche and Ugwuowo, Fidelis Ifeanyi

A systematic approach to time series model selection is very important for reduction of the uncertainties associated with highly subjective and inaccurate method currently being used. Information criteria as a measure of goodness of fit have been criticized because of its statistical inefficiency. In this paper, we develop a rule using discriminant analysis for classification of a time series model from a finite list of parsimonious ARMA (p,q) models. A discriminant function is developed for each of the six alternative ARMA(p,q) models using fifty sets of simulated time series data for each model. An algorithm is developed for the evaluation of discriminant scores and model selection. The selection rule is based on the highest discriminant score among the six alternative models. The method was applied to a real life data and thirty sets of simulated data. The real life application resulted in correct model selection while the simulated data gave 93% correct classification.