On the usefulness of cross-validation for directional forecast evaluation