Comparison of GARCH and neural network methods in financial time series prediction

Abstract

Many researchers have already published huge number of papers comparing Autoregressive (AR) model, a model based on Box-Jenkins methodology, and Back Propagation Artificial Neural Network (BPANN) in financial time-series forecasting. Among them, some compared SVMs and BPs taking AR as a benchmark in forecasting the six major Asian stock markets. They showed that both the SVMs and BPs outperform the traditional models, ARs. They did prediction of transformed data, but not level data. They did not take account of GARCH model, specially developed to model financial time series. Generalized Autoregressive Conditional Heteroskedastic (GARCH) model is needed to capture high volatility for better forecasts. This article applies GARCH model instead AR or ARMA model to compare with standard BP in forecasting of the four international including two Asian stock markets indices. Our fitted GARCH models give better forecasts than the fitted standard BP models in forecasting of the four international markets indices except one market.