Prediction of highly non-stationary time series using higher-order neural units

Abstract

Adaptive predictive models can use conventional and nonconventional neural networks for highly non-stationary time series prediction. However, conventional neural networks present a series of known drawbacks. This paper presents a brief discussion about this concern as well as how the basis of higher-order neural units can overcome some of them; it also describes a sliding window technique alongside the batch optimization technique for capturing the dynamics of non-stationary time series over a Quadratic Neural Unit, a special case of higher-order neural units. Finally, an experimental analysis is presented to demonstrate the effectiveness of the proposed approach.