Prediction of white noise time series using artificial neural networks and asymmetric cost functions

Abstract

Artificial neural networks in time series prediction generally minimise a symmetric statistical error, such as the sum of squared errors, to learn relationships from the presented data. However, applications in business elucidate that real forecasting problems contain non-symmetric errors. The costs arising from suboptimal business decisions based on over-versus underprediction are dissimilar for errors of identical magnitude. To reflect this, a set of asymmetric cost functions is used as objective functions for neural network training, deriving superior forecasts even for white noise time series. Some experimental results are computed using a multilayer perceptron trained with various asymmetric cost functions, evaluating the performance in competition to conventional forecasting methods on a white noise time series extracted from the popular airline passenger data.