RBF neural network is more suitable for function approximating than modeling in nonlinear time series analysis

This paper considers the application of the radial basis function neural network to nonlinear time series modeling. By introducing the RBF to approximate the coefficients of the general state-dependent autoregressive model, a RBF-based state-dependent AR model namely RBF-AR is derived. Then an empirical comparison between the RBF network and the RBF-AR model for complicated time series modeling is investigated. The results show that the RBF-AR model is of much better performance than the RBF network either for its lower predicting error or for the great improvement of the curse of dimensionality which is known as one of the main potential difficulties in RBF network modeling.