In plant breeding, genomic selection has been mainly used to predict untested single crosses and testcrosses. The objectives of this study were to assess the efficiency of prediction of untested testcrosses and the significance of the factors affecting the prediction. We simulated 20,000 testcrosses from two groups of 5000 related doubled haploid lines (DHs) and two unrelated elite inbred lines. The DHs were genotyped for 20,000 single nucleotide polymorphisms (SNPs) and the testcrosses were phenotyped for grain yield. The average SNP density was 0.1 cM. We assumed genetic control by 400 quantitative trait loci (QTLs) and heritability of 25, 50, and 75% for the assessed testcrosses. The training set sizes were 10 and 30% of the available DHs. The process of random sampling of the field assessed and predicted testcrosses were replicated 50 times. We computed the prediction accuracy and the coincidence index, a measure of the selection efficacy of the nonassessed testcrosses. The results evidenced that genomic selection is an efficient process for selecting superior nonassessed testcrosses, if there is sufficient relatedness between the available DHs, linkage disequilibrium in the DHs and genotypic variance between testcrosses, a training set size of at least 10% of the available DHs, and a SNP density of at least 1 cM. The efficacy of selecting the superior nonassessed testcrosses ranged between 0.1 and 0.9, proportional to the heritability of the assessed testcrosses (0.25–1.00). It is important to highlight that the parametric coincidence ranged from 0.1 to 0.8. Furthermore, genomic prediction is much more efficient than pedigree-based best linear unbiased prediction.