Abnormal Detection of Financial Fraud in Listed Companies Based on Deep Learning

Abstract

This study aims to improve the accuracy of detecting financial fraud in listed companies by applying various deep learning algorithms. First, we comprehensively reviewed the current state of research on financial fraud theory and identified 67 recognition features to create a new recognition indicator system. Then, we collected data samples of all A-share listed companies from 2010 to 2022, preprocessed them, and generated a basic dataset. To address the unbalanced dataset, we used the Borderline-SMOTE algorithm. Empirical analysis results show that this algorithm can significantly improve the recognition performance of the model. Finally, we conducted experiments on the new dataset using three types of deep learning algorithms. The results show that the model constructed using the Long Short-Term Memory (LSTM) algorithm has the best prediction performance, with an accuracy rate higher than that of the DCRN, autoencoder, and other models. In addition, the classification effects of all deep learning algorithms are better than basic models and ensemble models. This research provides a powerful tool for the regulatory authorities of listed companies, helping them more effectively monitor and prevent financial fraud. We have three innovations in this study: (1) Development of a comprehensive recognition indicator system with 67 features; (2) Utilization of the Borderline-SMOTE algorithm to handle data imbalance; (3) Demonstration of the superior performance of the LSTM algorithm compared to other deep learning, basic, and ensemble models.