A method of linear prediction of stationary multivariate time series is discussed from the point of view of meteorological applications. Tests of significance are given and it is shown by examples that the method is practical even when the dimensionality of the series becomes quite large.