Abstract Statistical model significance, sampling and forecast errors are compared between linear regression models developed from preselected and ordered Empirical Orthogonal Function (ROF) predictors and those selected by a forward stepwise screening technique. As a particular application, grid-point height prodictors are used to forecast tropical storm displacements in a storm-heading oriented coordinate system. Critical correlations for model significance and upper bounds on expected sampling errors are derived from a Monte Carlo method. It is found that dependence among predictors selected by screening reduces expected sampling errors below those for the same number of independent screened predictors. For the given application, expected forecast errors for screened predictors are only slightly greater than those for EOFs.