Using Smoothing Splines in Time Series Prediction with Neural Networks

摘要

The smoothing spline based neural network is used for prediction of a trend from complex and noisy time series. First, the time series is smoothed by a cubic spline and then multilayered feedforward neural networks are applied to predict the parameters of the spline and by this the next values of the smoothed time series. The level of smoothing can be chosen by the smoothing parameter. We show that in the case of a complex time series like the bike tire sale, prediction of a trend with the smoothing spline based neural network gives us more reliable information than a classical prediction with the multilayered feedforward neural network.