AURORA: An Auditable Causal AI Framework for Detecting Statistical Anomalies in A-Share Listed Companies

Man Yuan Lee Email to :ht3100z@hotmail.com ORCID: 0009-0003-5770-5634 This study proposes AURORA (Auditable Unified Reasoning for Oversight & Risk Analysis)—a novel framework that integrates structural causal models with multi-source knowledge graphs to enable early detection, causal attribution, and auditable evidence generation for statistical anomalies in enterprise reporting. Built upon China’s National Standard “Enterprise Integrated Reporting System” (EIRS), AURORA operates through a three-channel architecture: perception, reasoning, and audit. Using comprehensive public data from all A-share listed companies (2010–2025), including EIRS annual reports, regulatory filings, enforcement actions, and industry benchmarks, we demonstrate that AURORA issues effective warnings an average of 6.3 months before major anomaly events (e.g., financial restatements, qualified audit opinions, CSRC investigations), achieving a recall rate of 82% and a false positive rate of only 5.1%—significantly outperforming XGBoost, LSTM, and traditional statistical models. Critically, AURORA outputs structured, traceable evidence chains compliant with national statistical regulations, directly supporting regulatory inquiries, due diligence, and compliance reviews. We further argue that nationwide implementation of the National Standard for Enterprise Statistical Electronic Ledgers is essential to ensure high-quality source data, mitigate systemic data risks, and modernize corporate governance.