Time Series Modeling and Prediction Using Postfix Genetic Programming

Abstract

Traditional techniques for time series modeling can capture linear behavior of data and lack the ability to identify nonlinear patterns in time series. Therefore, machine learning techniques like Neural Network or Genetic Programming (GP) are used by practitioners for modeling nonlinear and irregular time series. GP is preferred over other techniques because it does not presume model structure a priori. This paper introduces the use of Postfix-GP, a postfix notation based GP, for real-world nonlinear time series modeling problems. The Postfix-GP uses linear genome representation and stack based evaluation to reduce space-time complexity of GP. The Postfix-GP is applied on two real time series modeling problems: sunspots and river flow series. Performance of evolved Postfix-GP models on training data and out-of-sample data are compared with those obtained by others using EGIPSYS. The obtained results indicate that Postfix-GP offers a new possibility for solving time series modeling and prediction problems.