Epistemic vs. aleatoric mapping of CP hyperparameters

Prove or refute that the analyst-controlled Conformal Prediction hyperparameters—the confidence level α and the choice of non-conformity score ψ ∈ 𝔽—primarily encode epistemic uncertainty, while aleatoric uncertainty arises from the intrinsic variability of the underlying exchangeable data-generating distribution 𝔓.

Background

In discussing uncertainty quantification, the authors note that region-size is ordinal and propose using credal sets to endow CP with cardinal metrics. Within this framing, they state a conjecture about how CP design choices relate to epistemic versus aleatoric uncertainty.

Validating this conjecture would clarify how CP’s parameters contribute to different uncertainty types, with practical implications for tuning and interpreting CP procedures.