Stochastic Modelling Hints for Neural Network Prediction.pdf

Sto chastic Mo delling Hints for Neural Network Based Time Series Predictions Radu Drossu Zoran Obradovi c rdrossueecswsuedu zoraneecswsuedu Scho ol of Electrical Engineering and Computer Science Washington State University Pullman Washington Abstract The ob jective of this study is to investigate the relationship b etween sto chastic and neural network approaches to time series mo delling Exp eriments on b oth a complex real life prediction problem entertainment video trac series as well as on an articially generated nonlinear time series on the verge of chaotic b ehavior MackeyGlass series indicate that the initial knowledge obtained through sto chastic analysis provides a reasonably go o d hint for the selection of an appropriate neural network architecture Although not necessarily the optimal such a rapidly designed neural network architecture p erformed comparable or b etter than more elab orately designed neural networks obtained through exp ensive trial and error pro cedures Keywords time series nonstationary pro cess ARMA mo delling neural network mo d elling prediction horizon 1 Corresp ondence Z Obradovic phone Fax 2 Research sp onsored in part by the NSF research grant NSFIRI INTR ODUCTION A time series x can b e dened as a random or nondeterministic function x of an t indep endent variable t Its main characteristic is that its future b ehavior can not b e predicted exactly as in the case of a deterministic function of t However the b ehavior of a time series can sometimes b e anticipated by describing the series through probabilistic laws Common