Neural network moving averages for time series prediction
Bruce E. Rosen · Proceedings of SPIE, the International Society for Optical Engineering/Proceedings of SPIE · 1993
ARMA (autoregressive--moving average) time series methods have been found to be effective methods of forecasting and prediction. Using AR (autoregression) methods, predictions rely on regressing previous time series input values, while in MA (moving average) methods, predictions are calculated by regressing previous forecasting errors. We can improve ARMA type forecasts with backpropagation by nonlinear regression of both the inputs and the previous forecasting errors. The new predictions are calculated by adding a feedforward neural network that accepts the previous forecast and previously generated forecast errors as inputs and produces new forecasts having smaller prediction errors. The accuracy of these forecast can exceed that of ARMA, or backpropagation forecasts alone. The improved predictions of AR and backpropagation network forecasts are shown using the Mackey-Glass chaotic time series.