Before implementation of genomic selection, evaluation of the potential accuracy of prediction can be obtained by cross-validation. In this procedure, a population with both phenotypes and genotypes is split into training and validation sets. The prediction model is fi tted using the training set, and its accuracy is calculated on the validation set. The degree of genetic relatedness between the training and validation sets may infl uence the expected accuracy as may the genotype × environment (G×E) interaction in those sets. We developed a method to assess these effects and tested it in cassava (Manihot esculenta Crantz). We used historical phenotypic data available from the International Institute of Tropical Agriculture Genetic Gain trial and performed genotyping by sequencing for these clones. We tested cross-validation sampling schemes preventing the training and validation sets from sharing (i) genetically close clones or (ii) similar evaluation locations. For 19 traits, plot-basis heritabilities ranged from 0.04 to 0.66. The correlation between predicted and observed phenotypes ranged from 0.15 to 0.47. Across traits, predicting for less related clones decreased accuracy from 0 to 0.07, a small but consistent effect. For 17 traits, predicting for different locations decreased accuracy between 0.01 and 0.18. Genomic selection has potential to accelerate gains in cassava and the existing training population should give a reasonable estimate of future prediction accuracies.