Prediction of Mixtures

The problem of predicting time series originating from mixtures of signals from independent dynamical systems is considered. We show that the problem of nding representations for the dynamics of such systems is hard if the mixing structure of the system is not taken into account. If, on the contrary, the sources can be unmixed in a prepro-cessing step the complexity of system identiication may be drastically reduced. This is demonstrated using chaotic maps. It is shown that applications of methods for blind separation of sources can substantially improve both: prediction performance and prediction horizon.