We address the problem of online prediction for time series. We show that any universal code (or a universal data compressor) can be used as a basis for constructing asymptotically optimal methods for this problem for a certain class of stationary and ergodic processes.