This preprint introduces Counterfactual Predictive Sufficiency (CPS), a domain-agnostic, metatheoretical criterion for selecting observable features that retain all counterfactually relevant predictive information about an outcome. CPS formalizes the intuition that a feature set S is sufficient for prediction if, for every intervention or structural perturbation within a specified class, the conditional distribution of the outcome given S is invariant. Unlike purely observational sufficiency concepts, CPS explicitly links sufficiency to counterfactual stability — the capacity of predictions to remain valid when data-generating mechanisms change in ways that do not alter the S→Y mapping. This work develops: (1) a concise formal statement of CPS and its minimality variants, (2) operational tests and falsifiable criteria that practitioners can apply without committing to full structural knowledge, and (3) illustrative examples demonstrating CPS across classification, time-series, and policy-sensitive settings. This work argues that CPS provides a unifying lens for robust feature selection, principled domain adaptation, and certifying model reliability under realistic distributional shifts. Written deliberately at the meta-theory level, this abstract establishes priority for the CPS concept and preserves flexibility for downstream frameworks, algorithms, and applications that refine or operationalize CPS in specific contexts.
Murad Ahmadov (Fri,) studied this question.