Fig. 6

Boxplot of predictive abilities for hybrids across different training set optimization and modelling iterations. The grid displays combinations of traits and test set years (top) and optimization methods (right). Only the best performing optimization methods are shown, and the training set size used was the optimal one found by Min_GRM except for Tail_GEGVs_sd1, which concurrently optimizes size and composition, and Entire_CS, using all available data without optimization. The candidate set considered comprised data from all years preceding the test set. Test set hybrids are categorized as T0, T1, T2 or Common, based on number of common parents in training and test sets. The dashed horizontal line represents the average predictive ability for all hybrids in each scenario. The percentage below each box denotes the proportion of the total test set comprised by the corresponding hybrid type