Prediction of mixtures

Abstract

The problem of predicting time series originating from mixtures of signals from independent dynamical systems is considered. We show that the problem of finding 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 preprocessing step the complexity of system identification 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.