Which Methodology is Better for Combining Linear and Nonlinear Models for Time Series Forecasting?

Abstract

Both theoretical and empirical findings have suggested that combining different models can be an effective way to improve the predictive performance of each individual model. It is especially occurred when the models in the ensemble are quite different. Hybrid techniques that decompose a time series into its linear and nonlinear components are one of the most important kinds of the hybrid models for time series forecasting. Several researches in the literature have been shown that these models can outperform single models. In this paper, the predictive capabilities of three different models in which the autoregressive integrated moving average (ARIMA) as linear model is combined to the multilayer perceptron (MLP) as nonlinear model, are compared together for time series forecasting. These models are including the Zhang's hybrid ANNs/ARIMA, artificial neural network (p,d,q), and generalized hybrid ANNs/ARIMA models. The empirical results with three well-known real data sets indicate that all of these methodologies can be effective ways to improve forecasting accuracy achieved by either of components used separately. However, the generalized hybrid ANNs/ARIMA model is more accurate and performs significantly better than other aforementioned models.