It is obvious, that recurrent neural networks are able to model high dimensional nonlinear dynamical systems. But let´s refocus our interest on the data: why should it be natural to model the data in the way they are observed. It might be an advantage to change the coordinate system to find an easier dynamical representation. Thus we have to find an optimal coordinate system in which the data should be presented to the forecasting. Finally we have to combine the coordinate transformation step and the forecasting step.
