Research on Financial Statement Anomaly Risk Prediction Based on Generative Adversarial Networks

Financial statements are crucial for reflecting the operational status and financial health of a company. However, some companies may manipulate their financial records or conceal key information, leading to distorted data and systemic risks. Traditional audit and ratio analysis methods struggle to identify such risks in high-dimensional financial data with limited anomalous samples. To address this, this paper proposes a financial statement anomaly detection method based on Generative Adversarial Networks (GANs) with a small number of anomaly labels: the generator simulates the data distribution of “normal companies” to generate realistic fake samples, while the discriminator learns the feature differences between normal and anomalous data by incorporating known anomalous samples during adversarial training. This approach enhances the ability to identify anomalies while reducing false negatives. Experimental results show that, compared with machine learning approaches including Random Forest, KNN, SVM, and Deep Neural Networks, the proposed method performs better, making it more suitable for scenarios with scarce anomaly indicators. It also aligns well with the business needs of auditing and regulatory bodies to “minimize false negatives.” This research provides a feasible path for auditing institutions and regulatory departments based on deep learning, and offers new insights for intelligent risk forecasting and anomaly detection in highdimensional, heterogeneous financial data.