Abstract
In this paper, we propose a so called Quasi-ARMA model, for financial time series prediction which is to model the history data of an existing series, and forecast the future unknown values. The Quasi-ARMA model contains two parts: the first is a macro-part, which is a linear interface with linear ARMA like structure and embeds the complexity into the coefficient; The second part is a kernel-part, which is a nonlinear model using a neurofuzzy networks, which is used to parameterize the coefficients. The model we proposed shows better performance on prediction accuracy, and gives more consistent and reliable forecasting results than conventional neural networks (NNs) in modeling high noisy financial time series. Computer simulations are carried out and confirms the effectiveness of the proposed model.
