On the accuracy in high-dimensional linear models and its application to genomic selection