Proceedings of the International Workshop on Applications of Neural Networks to Telecommunications

Bootstrap Methods in Neural Network Time Series Prediction

Jerome T. Connor

Bell Communications ResearchMorristown, N.J. 07960 [e-mail]

Abstract

Bootstrapping is a common technique for estimating quantities of a statistical model that are mathematically intractable. In this paper, we take a close look at the application of bootstrap techniques to the estimation of time series prediction error by neural networks. Bootstrap methods are shown to be useful in both prediction and the assessment of confidence intervals and prediction bias. In particular, we present an example involving access line data where bootstrap techniques are demonstrated.

One of the big problems in statistics is that a function, h(θ, ), of a true parameter, ...