Universal distribution of the empirical coverage in split conformal prediction

When split conformal prediction operates in batch mode with exchangeable data, we determine the exact distribution of the empirical coverage of prediction sets produced for a finite batch of future observables. This distribution is universal, being determined solely by the batch size, the nominal miscoverage level, and the calibration sample size. The exact distribution of the almost sure limit of the empirical coverage as the batch size goes to infinity is also identified, leading to a criterion for choosing the minimum required calibration sample size in applications.