Are ARIMA neural network hybrids better than single models?

Abstract

Hybrid methods comprising autoregressive integrated moving average (ARIMA) and neural network models are generally favored against single neural network and single ARIMA models in the literature. The benefits of such methods appear to be substantial especially when dealing with non-stationary series: nonstationary linear component can be modeled using ARIMA and nonlinear component using neural networks. Our studies suggest that the use of a nonlinear component may degenerate the performance of such hybrids and that a simpler hybrid comprising linear AR model with a TDNN outperforms the more complex hybrid in tests on benchmark economic and financial time series.