A Reproducibility Audit of Public Financial Misstatement Data: Chronological Benchmarks and Measurement Diagnostics

This study audits the reproducibility and measurement assumptions of a public financial misstatement benchmark. After removing 214 surplus positive records, a report-period split yields 2,695 evaluation observations in 2019–2021, including 93 enforcement-associated positives from 26 firms. With twelve processed ratios, logistic regression achieves average precision of 0.0447, compared with 0.0345 for a constant score and 0.0328 for a class weighted random forest. A post hoc paired firm-cluster interval for logistic improvement is −0.0020 to 0.0381. The forest leads during validation but falls from 0.1270 to 0.0328 in evaluation. The audit identifies overlapping report-period ranges in the supplied split and unresolved histories of predictor values, control sampling and training-label availability. Supplementary checks examine selected SEC accessions and annotation disagreements. The contribution is methodological: an executable account of the assumptions behind the benchmark and evidence of weak retrospective ranking. An untouched cohort and historically reconstructed inputs remain necessary for a prospective prediction study. Preprint, not peer reviewed. Replication code and documentation: https://doi.org/10.5281/zenodo.24027492