Each model is placed at (β − 1, α − 1), where β = μm/μo is the bias ratio and α = σm/σo the variability ratio. The origin is the perfect model.
Eαβ = √[ (α−1)² + (β−1)² ]
E = 1 − KGE = √[ (ρ−1)² + (α−1)² + (β−1)² ]
Each model appears as a circle at (β−1, α−1) , filled when ρ ≥ 0 and hollow when ρ < 0, whose distance from the origin is Eαβ, the error induced by bias and variability. A triangle sits on the same ray at distance E: the connecting segment’s length E − Eαβ is the error contributed by imperfect correlation.
Circles mark constant total-error levels (a mark inside the 10% circle has error below 0.10). Quadrants tell you at a glance whether a model over- or under-estimates the mean (left–right) and the variability (down–up).
Demo dataset
Thirty synthetic years of total annual precipitation for a Southern-Italy-like climate (E-OBS-style reference, μ ≈ 750 mm), with six model series constructed to hit exact bias, variability and correlation targets, including one negatively correlated model so you can see the hollow endpoint marker. Replace it with your own EURO-CORDEX / CMIP / hydrological series.