Mean-preserving online shrinkage for prediction-assisted finite-population auditing: replication package

<p dir="ltr"><b>Replication package for the manuscript “Mean-preserving online shrinkage for prediction-assisted finite-population auditing” by Julian Hoxha and Marius Panxhi.</b></p><p dir="ltr">This repository contains the complete computational materials used to reproduce the statistical experiments, tables, and figures reported in the manuscript and its technical supplement. The study considers sequential estimation of a positive-label proportion in a fixed finite population when predictions are available for every item but reference labels are acquired sequentially by simple random sampling without replacement.</p><p dir="ltr">The principal proposed procedure is <b>mean-preserving online shrinkage (MPOS)</b> coupled with an explicit binary-design evidence factor and integer-total confidence-sequence inversion.

MPOS adaptively contracts item-level working probabilities toward their remaining prediction mean using only previously acquired reference labels. The procedure is designed to preserve time-uniform coverage while controlling excessive prediction confidence.</p><p dir="ltr">The package includes implementations and results for MPOS, the binary-design optimizer, integer-total confidence sequences, integer uncapped comparators, reference-only methods, finite-population prediction-powered inference (PPI) comparisons, native PPI/PPI++ fixed-budget comparisons, ridge-sigmoid and smoothed-isotonic calibration comparisons, record-disjoint benchmark evaluations, and the supporting OPC and RVO investigations described in the supplementary material.</p><p dir="ltr">The evaluation includes recurring public PPI benchmarks and additional fixed datasets, with matched acquisition orders used for method comparisons.