Efficient design of neural networks for time series prediction
R. Drossu, Zoran Obradović · 1997
The objective of this study is to investigate the possibility of rapidly designing an appropriate neural network (NN) model for time series prediction. Experiments on both a complex real life prediction problem (entertainment video traffic series) as well as on an artificially generated nonlinear time series on the verge of chaotic behavior (Mackey-Glass series) indicate that stochastic analysis can provide some initial knowledge regarding an appropriate data sampling rate and NN architecture, as well as regarding the choice of initial values for the NN parameters. Although not necessarily the optimal, such a rapidly designed NN model performed comparable or better than more elaborately designed NNs obtained through expensive trial and error procedures. Keywords: time series, neural network modeling, ARMA modeling, prediction horizon. 1 Correspondence: Z. Obradovi'c, phone: (509) 335-6601, Fax: (509) 335-3818 2 Research sponsored in part by the NSF research grant NSF-IRI-9308523. ...