A decision-based criterion and a resolution threshold for ranking genomic prediction models within trials

Abstract

Breeding programmes choose genomic prediction models by accuracy rankings, and the metric behind a ranking decides which model wins. Selecting the best genotypes within a trial depends only on the order of the predictions, so a metric that guides the choice must be unchanged by any order-preserving change of the predictions and respond when their order is reversed. Of the 22 official metrics of the Genomes to Fields maize competition, one meets both conditions, the within-environment rank correlation; the within-environment Pearson correlation is protected against rescaling only; and twenty fail, including the root mean squared error that decided the 2022 ranking. Switching between the competitions two official metrics reverses 27 % of team pairs, and an exact decomposition shows why pooled metrics mislead: in three constructed panels, 94-95 % of the attainable square of pooled correlation is carried by the environment means. The accuracy difference a trial can resolve falls as about 3/{surd}N with its N genotype-environment cells: at least 0.03 at the competitions design and 0.04-0.12 once calibrated on real panels, against 0.003 between the first two teams; tests that two correlations differ declare far smaller differences, which do not reliably change selection. In soybean, where the two most accurate methods stand clear of most others by more than the trial resolves, ranking by either within-environment correlation selects better material, 98-99 % of the attainable selection gain in held-out environments against 13-34 % for error-magnitude rankings; in common bean and a Chinese maize population, where the leading methods lie within the resolvable gap, all rules pick material of similar value. GPverdict runs these checks on any programmes cross-validated predictions.