Predicting a stationary process when the correlation function is convex
Jaroslav Hájek · Czechoslovak Mathematical Journal · 1958
By the method worked out in the paper [1] it is proved, that, in the case of a convex correlation function, it suffices to base the linear prediction on the last observation only, because the relative reduction of the residual variance, attainable by making use of any number of preceding observations, cannot exceed 50%.