Accounting data anomaly detection and prediction based on self-supervised learning

This study proposes a Hierarchical Fusion Self-Supervised Learning (HFSL) framework to address the challenge of scarce labeled data in accounting anomaly detection, integrating domain knowledge with advanced deep learning techniques. Based on financial data from Chinese listed companies in the CSMAR database spanning 2000–2020, this framework integrates temporal contrastive learning, a dual-channel LSTM autoencoder structure, and financial domain knowledge to construct a three-tier cascaded detection system. Empirical research demonstrates that the HFSL framework achieves a precision of 0.836, recall of 0.805, and F1 score of 0.820 in accounting anomaly detection, significantly outperforming traditional methods. In terms of practical metrics, the framework attains an early detection rate of 0.726 while maintaining a false alarm rate of just 0.068, providing technical support for early risk warning. Financial feature contribution analysis reveals that core indicators such as Return on Assets (ROA), Return on Equity (ROE), and their interaction effects play crucial roles in anomaly identification. Through analysis of 2,150 samples in the test set, the study identifies five typical financial fraud patterns (revenue inflation 38.6%, expense concealment 21.7%, asset overvaluation 17.4%, liability understatement 15.2%, and composite manipulation 7.1%) and their temporal evolution characteristics. The research also finds that financial anomalies typically exhibit three evolutionary patterns: progressive deterioration (64%), sudden anomalies (22%), or cyclical fluctuations (15%), providing empirical evidence for regulatory practice. This study applies self-supervised learning to accounting anomaly detection, not only solving the detection challenges in unlabeled data scenarios but also providing effective tools for financial supervision and risk management.