GPverdict
Choose genomic prediction methods for selection — from your own cross-validation, in your browser.
Which metric can rank methods?
Checked on your predictions: only metrics that ignore order-preserving changes and respond to order reversal represent within-environment selection.
Which method, and is its lead real?
Methods ranked by the within-environment rank and Pearson correlations, with the methods your trial cannot tell apart from the leaders.
What would RMSE have chosen?
The method an error-magnitude ranking picks, and an out-of-sample test of how much selection gain each rule recovers.
How large must a trial be?
Genotype–environment cells needed to resolve a given accuracy gain.
One row per environment, genotype and method, with columns environment, genotype, observed, predicted, method (also accepted: Env, k, y, p). Predictions should come from cross-validation.
Your data stay on this computer: every computation runs locally in the browser (Pyodide); nothing is uploaded.
Forward evaluation over target years
Predictions of each target year made from earlier years only. Each method is compared with a reference method within environments, the differences are pooled over target years (DerSimonian–Laird with Hartung–Knapp alongside), and a difference counts as resolved only if it is at least 3/√N and its interval excludes zero (Lv and Gu 2026, forward benchmark, manuscript in preparation).
Columns environment, year, genotype, observed, predicted, method; optional dataset, and new (1 = new line) or first_year to report new and old lines separately.