Reference
Evaluate multiple models with all metric families A–E. Returns scores, rankings, recommendation, and optional bootstrap CIs.
Family A — Weighted marginal sum of RMSE (continuous) + Log-Loss (binary). Marginal baseline, blind to dependence.
Family B — Multivariate Energy Score. Strictly proper, sensitive to joint distribution. Requires normalised inputs.
Family C — Variogram Score (p = 0.5). Proper, directly penalises pairwise covariance errors. Robustness check.
Family D — Marginal-Dependence Decomposition. Separates S_marg from S_dep via PIT correlation matrices.
Family E — E-CVWMD. CV-derived variable weights, residual-correlation dependence score, adaptive α* via distance correlation test.
RMSE per column (named vector).
MAE per column (named vector).
Binary cross-entropy Log-Loss for predicted probabilities.
Classification accuracy at threshold 0.5.
Non-parametric percentile bootstrap for 95% CIs. Pairwise CI overlap test for exactly two models.
Four types: bars, forest, heatmap, decomp. Returns a ggplot2 object.
bars
forest
heatmap
decomp
Column-wise z-score normalisation using training-set parameters. Required before Metrics B and C.
Rule-based metric recommendation. Called automatically by evaluate_models().