This paper describes a novel application of Statistical Learning Theory (SLT) for motion prediction. SLT provides analytical VC-generalization bounds for model selection; these bounds relate unknown prediction risk (generalization performance) and known quantities such as the number of training samples, empirical error, and a measure of model complexity called the VC-dimension. We use the VC-generalization bounds for the problem of choosing optimal motion models from small sets of image measurements (flow ). We present results of experiments on image sequences for motion interpolation and extrapolation; these results demonstrate the strengths of our approach.