16The many quantitative traits of interest to plant breeders are often genetically correlated, which 17 can complicate progress from selection. Improving multiple traits may be enhanced by 18 identifying parent combinationsan important breeding stepthat will deliver more favorable 19 genetic correlations (rG). Modeling the segregation of genomewide markers with estimated 20 effects may be one method of predicting rG in a cross, but this approach remains untested. Our 21 objectives were to: (i) use simulations to assess the accuracy of genomewide predictions of rG 22 correlations, meanwhile, are common and often the bane of the breeder. In crop improvement, 53 notorious examples include grain yield and grain protein content in wheat (Triticum aestivum L.; 54 Simmonds 1995), grain yield and plant height in maize (Zea mays L.; Chi et al. 1969), and seed 55 protein and oil content in soybean (Glycine max L.; Bandillo et al. 2015). The directions of such 56 correlations imply an unfavorable response in one trait when selecting on another (Falconer and 57 Mackay 1996), and the underlying cause will impact the prospects of long-term improvement. 58 5 Selection on traits with shared, antagonistic genetic influence is functionally constrained, but 59 correlations induced by linkage disequilibrium are transient and can be disrupted by 60 recombination (Falconer and Mackay 1996; Lynch and Walsh 1998). 61 Genomewide selection has become popular among plant breeders as a method of 62 predicting the merit of unphenotyped individuals using genomewide markers and a phenotyped 63 training population (Meuwissen et al. 2001). Typical prediction models are univariate (i.e. one 64 trait), but multivariate models have recently been explored as a means of borrowing information 65 from genetically correlated traits and improving the prediction accuracy of both traits (Calus and 66 Veerkamp 2011; Jia and Jannink 2012). Selection on multiple traits using predicted breeding 67 values would proceed as if using phenotypic values, relying on procedures such as tandem 68 selection, independent culling levels, or the construction of a trait index (Bernardo 2010), with 69 most studies of multi-trait genomewide selection using the latter (Combs and Bernardo 2013; 70 Beyene et al. 2015; Sleper and Bernardo 2018; Tiede and Smith 2018). 71These models and selection methods implicitly assume that the breeding population has 72 already been developed from selected parents. Therefore, the genetic variance of each trait, and 73 the genetic correlation between traits, both of which determine the direct or correlated response 74 to selection (Falconer and Mackay 1996), are fixed parameters of the population. In addition to 75 more accurate selection within an established population, multi-trait genetic gain could be 76 increased by developing better populations through deliberate selection of parent combinations 77 with a more ideal mean, larger genetic variance, and more favorable genetic correlation. 78Typically, breeders select parents using the expect...