Prediction Sets for Counterfactual Decisions: Coverage, Optimality, and Conformal Prediction

What the math gives to ML

This paper supplies a decision-aware uncertainty interface for settings where each action changes the outcome distribution, rather than treating prediction uncertainty as independent of the policy. The key transferable asset is policy-coupled coverage: construct action-conditional outcome sets, select the action whose worst-case utility over its own set is largest, and calibrate coverage on the outcome actually induced by that selection rule. This can become a conformal robust-policy layer for contextual bandits, offline reinforcement learning, or high-stakes classification with abstention, with calibration performed against the deployed decision rule rather than against each potential outcome separately.