Financial Statement Fraud Detection by Integrating Supervisory Punishment Reports Into Machine Learning Methods: Evidence From China

ABSTRACT This study proposes a novel text analysis framework for supervisory punishment reports (SPR) to enhance financial statement fraud detection (FSFD). By analysing 41,146 samples of Chinese listed companies with 11,711 SPR (2007–2021), the results show significant gains: 7.125% sensitivity, 12.85% specificity, 10.75% accuracy and 10.025% AUC improvement. The number of punishments was identified as the most critical factor. Furthermore, SPR also improves serious fraud detection accuracy and one‐year‐ahead prediction accuracy. This research provides methodological innovations and mechanistic insights for financial fraud analysis, highlighting the information value of textual regulatory disclosures in FSFD.