One of the main aspects of brain activity is the ability to predict. Large efforts has been made in nonlinear dynamics to creaie predicting systems for dynamics of complex objects. One of the main tools for making such predictors is the application of multilayer neural networks. The methods based оп the chaos theory prove to be less efficient. and in lact work only for low-dimensional model systems. From our peint of view, the problems here are not technical, but related with the applicability of the approach of low-dimension nonlinear dynamics to real systems. Since the brain and some of its very simple models are able to make predictions in real situations, we propose to unify the ideas of nonlinear dynamics and neural networks. From our point of view, in complex real situations it may be possible to find low-dimensional projections, for which the approaches of noulinear dynamics can be applied, but with serious restrictions. Most concepts, like attractor, its dimension, Lyapunov exponents etc. become inapplicable, and the observed phase space splits into predictable parts {«channels») and non-predictable ones («jokers»), where probabilistic description may be more appropriate. We propose some mathematical basis for this idea and its possible application for time series anaiysis.