Comparative Evaluation of Automated Data Consistency Detection Methods in Corporate SEC Filings

Abstract

Corporate disclosure documents submitted to the U.S. Securities and Exchange Commission contain extensive financial data that requires verification across multiple periods and document sections. Manual verification of data consistency in SEC filings presents significant challenges due to document length and complexity. This research conducts a systematic comparative evaluation of three automated detection approaches: rule-based logic verification, statistical analysis-based anomaly identification, and neural network-based semantic contradiction discovery. The study examines 248 SEC reports from publicly traded companies, identifying 1,847 manually annotated data inconsistencies across numerical discrepancies, logical conflicts, and semantic contradictions. Performance evaluation demonstrates that rule-based methods achieve 94.3% precision with 78.6% recall for numerical inconsistencies, statistical approaches attain 87.9% precision with 82.4% recall for time-series anomalies, and neural methods reach 91.7% precision with 85.2% recall for semantic contradictions. Cross-document validation capabilities reveal substantial performance variations, with detection accuracy declining by 23-31% when analyzing inter-report relationships. The findings provide empirical evidence supporting the deployment of hybrid detection frameworks combining complementary strengths of each methodology. These automated tools offer practical value for regulatory oversight and corporate compliance management.