Exploring phase space concepts in the forecasting of time series with artificial neural networks

摘要

The authors study the performance of feedforward artificial neural networks in forecasting future values of several different time series. They explore both short- and long-term prediction of several periodic time series. They find that a significant source of error in long-term prediction of time series is introduced by a phase shift between the network output and the time series. They explore the origin of this phase shift and suggest strategies for minimizing its effect. They find that the phase diagrams of the time series and the neural network forecast contain useful diagnostic information.