Current breeding programs in animal production systems involve selection of candidate animals with superior genetic merit to serve as progenitors of the next generation. In this study, we predicted genomic breeding values (GEBV) for 2 quantitative traits from the Michigan State University Swine Resource Population using standard whole-genome prediction (WGP) models that assume homogeneous residual variance (i.e., RR-BLUP, BayesA, BayesB and BayesCπ) and their heteroskedastic counterparts. We divided the data into 5 mutually exclusive folds, such that 4 folds were alternatively used to train homoskedastic and heteroskedastic versions of each WGP model and then predict GEBV on animals on the remaining validation fold. The pseudo-Bayes factors indicated that heteroskedastic error WGP models improved model fit at all 5 crossvalidation folds. Within each fold, we then computed the Spearman rank correlation between homoskedastic- and heteroskedastic-based GEBV for top and bottom 10% individuals to compare their relative rankings. We noticed a considerable degree of re-ranking of animals with 10% top and bottom homoskedastic GEBV, particularly as the amount of residual heterogeneity in the data increased. For loin muscle pH at 45 min post-mortem, the median rank correlation of GEBV for top (bottom) 10% animals between heteroskedastic and homoskedastic models ranged from 0.52 to 0.70 (0.64 to 0.70) across data folds and WGP models. Similarly, for carcass temperature at 45 min post-mortem, the median rank correlation ranged from 0.05 to 0.38 (top 10%) and from 0.43 to 0.54 (bottom 10%). These results indicated non-negligible re-ranking of individuals with extreme genetic merit when heterogeneity of residual variances across environments is accounted for, thereby supporting potential practical implications for selection purposes in breeding programs.