PRTSM: Pattern recognition-based time series modeler

Abstract

In this paper, a new approach using pattern recognition techniques is suggested for time series modeling which means identification of a time series into one of autoregressive moving-average models. Its main recipe is that pattern is derived from a time series and classified into a suitable model via a notion of pattern matching. The pattern is obtained from extended sample autocorrelations of the time series. The pattern recognition techniques used are learning and decision tree classifier. Learning is used in combination with linear discriminants whose goal is to discriminate one pattern from another. Decision tree classifier divides decision procedures involved in time series modeling into simpler and local decisions at each node of a decision tree. To facilitate complex tree search, knowledge-based approach is used. To implement the idea, a scheme of decision support system is employed to develop a prototype system named PRTSM (Pattern Recognition-based Time Series Modeler). Experimental results with several examples show that a pattern recognition-based approach can yield a promising solution to the time series modeling. © 1989 Kluwer Academic Publishers.