Neural Network For Modeling Nonlinear Time Series: A New Approach

Abstract

Abstract. Nonlinear modeling with neural networks offers a promising approach for studying the prediction of a chaotic time series. In this paper, we propose a neural net based on Extended Kalman Filter to examine the nonlinear dynamic proprieties of some financial time series in order to differentiate between low-dimensional chaos and stochastic behavior. Kalman filtering, because it can deal with varying unobservable states, provides an efficient framework to model these non-stationary exposures. A controlled simulation experiment is used to introduce the issues involved and to present the proposed approach. Measures of forecast accuracy are developed. The pertinence of this model is discussed from the Tunisian Stock Exchange database.