Sufficiency: Predictive Parity and Calibration

Sufficiency requires that a score or prediction mean the same thing in every group; it is satisfied by calibration within groups, is often a by-product of accurate learning, and is incompatible with equal error rates whenever base rates differ.