Permutation importance with correlated factors

Two factors contain almost the same information. Why can single-feature permutation importance underestimate their joint usefulness?

Permuting one factor leaves its correlated substitute intact, so model performance may change little even if the pair is important. Independent permutation can also create unrealistic combinations. Consider grouped permutation, conditional methods with explicit assumptions, and refit-based ablations. Measure importance on held-out periods with dependence-aware uncertainty. Importance describes a fitted model and evaluation distribution, not an invariant causal contribution.