Forecasting the volatility of stock index returns: A stochastic neural network approach

Abstract

In this paper we are concerned with the volatility modelling of financial data returns, especially with the nonlinear aspects of these models. Our benchmark model for financial data returns is the classical GARCH(1,1) model with conditional normal distribution. As a tool for its nonlinear generalization we propose a Stochastic neural network (SNN) to the modelling and forecasting the time varying conditional volatility of the TUNINDEX (Tunisia Stock Index) returns. Such specification also helps to investigate the degree of nonlinearity in financial data controlled by the neural network architecture. Our empirical analysis shows that out-of-simple volatility forecasts of the SNN are superior to forecasts of traditional linear methods (GARCH) and also better than merely assuming a conditional Gaussian distribution. © Springer-Verlag Berlin Heidelberg 2004.