Robust Methods for the Prediction of Customer Demands based on Nonlinear Dynamical Systems

This paper investigates the forecasting accuracy of different prediction methods based on the theory of nonlinear dynamical systems and phase space reconstruction. In particular, a locally linear regression method is described. Different regularization methods to achieve better predictions in case of short time series as well as to increase the robustness against noise are specified. An evaluation by means of synthetic and real time series indicates high forecasting accuracy compared to established methods used as benchmarks.