An Almost Perfectly Predictable Process with No Optimal Predictor

Abstract

A novel kind of a negative result is presented for the problem of computable prediction. A non-stationary binary stochastic process is constructed for which almost surely no effective method of prediction achieves the infimum of prediction errors defined as the normalized Hamming distance between the sequence of predictions and the realization of the process. Yet it is shown that this process may be effectively predicted almost surely up to an arbitrarily small error since the infimum of prediction errors is zero.