Robust estimation of confidence interval in neural networks applied to time series

Abstract

Artificial neural networks (ANN) have been widely used in regression or predictions problems and it is usually desirable that some form of confidence bound is placed on the predicted value. A number of methods have been proposed for estimating the uncertainty associated with a value predicted by a feedforward neural network (FANN), but these methods are computationally intensive or only valid under certain assumptions, which are rarely satisfied in practice. We present the theoretical results about the construction of confidence intervals in the prediction of nonlinear time series modeled by FANN, this method is based on M-estimators that are a robust learning algorithm for parameter estimation when the data set is contaminated. The confidence interval that we propose is constructed from the study of the Influence Function of the estimator. We demonstrate our technique on computer generated Time Series data. © Springer-Verlag Berlin Heidelberg 2003.