A Recurrent Neural Network for Time-series Modelling

This paper describes an architecture for discrete time feedback neural networks. Some analytical results for networks with linear nodal activation functions are derived, while simulations demonstrate the performance of recurrent networks with nonlinear (sigmoidal) activation functions. The need for models with a capacity to consider correlated noise sequences is pointed out, and it is shown that the recurrent networks can perform state estimation and entertain models with coloured noise.