PC-RACP: New conformal prediction method boosts decision utility under uncertainty

Predictions increasingly guide high-stakes decisions like treatment selection or policy making, but imperfect models need uncertainty quantification. Conformal prediction builds prediction sets with coverage guarantees, but standard validity doesn't ensure optimal decision-making—especially when outcomes depend on the chosen action. In this preprint, Yurui Zheng and Ying Jin propose a decision-theoretic framework that introduces 'policy-coupled coverage': coverage of the realized outcome under the action induced by the prediction sets themselves. This concept serves as an optimal interface between uncertainty and action, justifying a max-min decision rule that is minimax-optimal even under distributional ambiguity.

The authors derive the explicit form of population-optimal prediction sets and propose a two-stage procedure, Policy-Coupled Risk-Averse Conformal Prediction (PC-RACP), which approximates these optimal sets with rigorous finite-sample coverage guarantees. Through simulations and a real email-marketing experiment, PC-RACP consistently delivers higher utility than existing conformal prediction approaches while maintaining valid coverage. The results highlight that ignoring the counterfactual structure of decision problems is suboptimal for both validity and utility, offering a practical advance for deploying AI in settings where actions affect outcomes.