Empirical evaluation demonstrates superior fraud detection in highly imbalanced financial datasets using generative autoencoders, highlighting the predictive value of market-based stock signals.
This paper presents a generative framework for financial fraud detection in highly imbalanced settings. Specifically, it proposes a Convolutional Conditional Variational Autoencoder (CNN-CVAE) to reconstruct the temporal dynamics of financial indicators and identify potential fraud through reconstruction errors. For model input, stock price information is incorporated as a conditional signal to enhance the model’s ability to exploit market-based information. In addition, we construct 31 financial indicators based on the suspicious-information perspective and audit practice. Empirical results show that the proposed framework outperforms traditional benchmark methods, particularly under extreme class imbalance. The findings also indicate that stock price information contributes more to performance improvement than alternative auxiliary information such as corporate governance and textual sentiment and the proposed indicator provides useful suspicious signals for fraud detection.
No takes yet. Share an insight, caveat, or question.
Feng et al. (2026) studied this question.
Synapse has enriched 5 closely related papers on similar clinical questions. Consider them for comparative context: