Analysis of time series data with heteroskedastic variance via neural networks; Neural network wo mochiita fukinitsu bunsan wo motsu jikeiretsu data no kaiseki
Masaaki Hatakeyama, Tadashi Dohi, Shigeki Osaki · OSTI OAI (U.S. Department of Energy Office of Scientific and Technical Information) · 1998
Recurrent network with stochastic volatility (R.N.N.S.V.) wherein variance of time series data volatility is employed for the forecasting evaluation is proposed and compared with the conventional time series analytical models and recurrent networks (R.N.N.). Two kinds of R.N.N.S.V. are investigated in which variance relative to volatility time is taken into consideration. Firstly, a network system is studied wherein R.N.N., which is used for the estimation of future volatility value from the historical volatility, and R.N.N., which outputs forecast value when stock market data are fed, are arranged in series. A network system of hybrid structure is then introduced wherein statistic auto-regressive model is included as the volatility estimation module. Necessary learning algorithm and estimation method for parameters are described, and finally the forecasting performance of R.N.N.S.V. is evaluated based on the actual stock market data to verify the effectiveness. 19 refs., 9 figs., 2 tabs.