Genome-wide prediction is a promising approach to boost selection gain in hybrid breeding. Our main objective was to evaluate the potential and limits of genome-wide prediction to identify superior hybrid combinations adapted to Northwest China. A total of 490 hybrids derived from crosses among 119 inbred lines from the Shaan A and Shaan B heterotic patt were used for genome-wide prediction of ten agronomic traits. We tested eight different statistical prediction models considering additive (A) effects and in addition evaluated the impact of dominance (D) and epistasis (E) on the prediction ability. Employing five-fold cross validation, we show that the average prediction ability ranged from 0.386 to 0.794 across traits and models. Six parametric methods, i.e. ridge regression, LASSO, Elastic Net, Bayes B, Bayes C and reproducing kel Hilbert space (RKHS) approach, displayed a very similar prediction ability for each trait and two non-parametric methods (random forest and support vector machine) had a higher prediction performance for the trait rind penetrometer resistance of the third intode above ground (RPR_TIAG). The models of A+D RKHS and A+D+E RKHS were slightly better for predicting traits with a relatively high non-additive variance. Integrating trait-specific markers into the A+D RKHS model improved the prediction ability of grain yield by 3％, from 0.528 to 0.558. Of all 6328 potential hybrids, selection of the top 44 hybrids would lead to a 6％increase in grain yield compared with Zhengdan 958, a commercially successful hybrid variety. In conclusion, our results substantiate the value of genome-wide prediction for hybrid breeding and suggest dozens of promising single crosses for developing high-yielding hybrids for Northwest China.