Optimized ARIMA-ANN hybrid model for time series analysis

Abstract

In recent years, hybrid models, using more than one models together, are presented in the field of time series analysis. One of the most important hybrid model classes is ARIMA-Artificial Neural Networks (ANN) hybrids. ARIMA is a linear model according to its structure, whereas ANN models are quite successful on modeling nonlinear time series. Time series encountered in real life usually carry linear and nonlinear characteristics together, which causes high forecasting performance of ARIMA-ANN hybrid models. In this study, a novel optimization based ARIMA-ANN hybrid model is presented. Proposed Optimized ARIMAANN (OptAA) hybrid model assumes time series is the sum of linear and nonlinear two series. In the first stage of the two staged model, ARIMA and ANN models with real time series pass through a least squares optimization process to obtain linear and nonlinear components. In the second stage, error values of the linear component are transferred to nonlinear component, nonlinear component is revised and remodeled with ANN. To determine the forecasting performance, Wolf's sunspot, Canadian lynx and GBP/USD exchange rate data sets, which are applied often in the literature, are used. Results obtained from the application show that OptAA hybrid model has higher performance than other models especially in relatively short term forecasting and is a very powerful methodology in time series analysis field.