In this paper the LS-SVM is used to resolve the problem of fading channel prediction

thus a novel prediction algorithm is proposed. In this algorithm

the learning samples are constructed by the observed values of fading channel coefficients

and the prediction is implemented by resorting to the learning and decision ability of the LS-SVM. Performance evaluation of the proposed algorithm is carried out on Jakes fading channels. The results illustrate the feasibility and efficiency of the algorithm. Additionally

the experiments discuss the influences of embedding dimension on the prediction accuracy. The method of optimal embedding dimension selection is given.