A recurrent neural network for nonlinear time series prediction-a comparative study

The performance of recurrent neural networks (RNNs) is compared with those of conventional nonlinear prediction schemes, such as a Kalman predictor (KP) based on a state-dependent model and a second-order Volterra filter. Simulation results on some typical nonlinear time series data indicate that the neural network can predict with accuracies on a par with the KP. It is noted that a higher-order extended Kalman filter or a Volterra model might provide a better performance than the ones considered. The network requires very few sweeps through the training data, though this will be computationally much more intensive than that required by conventional schemes. The authors discuss the advantages and drawbacks of each of the predictors considered.<>