The Effect of Virtual Term Generation on the Neural-based Approaches to Time Series Prediction

Time series prediction is the process of predicting future measurements by analyzing the nonlinear relation among past values and a current one. Many literatures have proposed the neural approaches, such as back propagation, RBF (Radial Basis Function), and recurrent network, to time series prediction instead of statistical approaches: AR(Auto Regressive), MA(Moving Average), ARMA(Auto Regressive Moving Average), and Box-Jekins Model. The reason is that neural-based approaches outperform the statistical ones in the performance of predicting future measurements. In the case of neural-based approaches replacing statistical ones to time series prediction, the sufficient data as samples should be prepared. In order to mitigate the requirement in the neural based approaches, it is proposed that the generalization performance of the neural network be improved by generating the artificial training patterns derived from the original ones and adding them to the exited training patterns.