A Measure of Nonlinearity in Time Series Using Neural Network Prediction Model

A measure of nonlinearity in time series is presented. We use a feedforward neural network which has several nonlinear units in its hidden layer as a time series predicting model. The measure of nonlinearity is calculated from weights of the trained network. With this measure one can determine whether a linear or a nonlinear analysis is needed for the given time series. As examples, we measured the nonlinearity of sunspot series and a carp's electroencephalogram (EEG). On the basis of the results of a statistical test, it was concluded that the sunspot series is nonlinear but the carp's EEG is linear.