Abstract
Modeling and estimating the volatility of financial time series data has been of great importance in financial analysis over the last decade. The early developments are attained by the introduction of autoregressive conditonal heteroskedasticity and its variants. The main purpose of this article is to develop new forecasting models for time series data with heteroskedastic variance, using artificial neural networks. The neural network systems under consideration contain sub-module to estimate the future volatility from the financial data. Finally, in empirical test based on the real Japanese stock market data, we show that the proposed models are superior on their predictive abilities to ordinary neural networks and statistical autoregressive models.
