Interactions and partial dependence

Why can a partial dependence plot be misleading when two financial predictors are strongly dependent?

Partial dependence averages predictions while varying one feature and retaining other observations' feature values. This can create combinations rarely or never observed, forcing interpretation of unsupported regions. Examine joint support, conditional or local alternatives and interaction effects. Restrict conclusions to plausible combinations and compare across time. No explanation plot substitutes for out-of-sample evaluation or causal identification.