Detecting Financial Fraud in Listed Companies via a CNN-Transformer Framework

Financial fraud in listed companies has grown increasingly complex, challenging traditional detection methods. This paper introduces the Transformer-style Convolutional Neural Network (CNN-Transformer), a unified architecture that integrates CNNs' capability for localized feature extraction with Transformers' strength in capturing long-range contextual dependencies. CNN-TRANSFORMER employs large-kernel convolutions and a self-attention mechanism to enhance feature interaction while maintaining computational efficiency via a linear-complexity Hadamard product approach. Evaluated on a dataset from China's A-share market, CNN-TRANSFORMER outperforms models like ResNet, MLP, and standalone CNNs or Transformers in accuracy, precision, recall, and AUC. The model's capacity to simultaneously extract localized features and model comprehensive contextual relationships enables robust detection of sophisticated fraudulent patterns. This research underscores the potential of CNN-Transformer hybrids in financial fraud detection, offering a scalable and interpretable AI solution. Future work may focus on integrating unstructured data, improving transparency, and optimizing efficiency for broader applications.