Machine learning tools enhance financial fraud detection accuracy

Research published in the International Journal of Information and Communication Technology suggests that machine learning tools might be used to detect and so combat financial fraud.

According to Weiyi Chen of the Monitoring and Audit Department of the Financial Shared Center at the National Energy Group Qinghai Electric Power Co., Ltd. In Xining, China, financial fraud is a constant challenge for capital markets, especially in developing economies where regulatory systems are still not fully mature. Fraudsters use sophisticated techniques to outpace conventional detection methods, which can leave investors exposed to potentially devastating risks beyond the everyday risks of investments.

Chen’s work offers a promising new approach to fraud detection by combining machine learning and deep learning to bridge the gap between financial data and the information found in corporate reports.

Financial fraud has long afflicted markets, distorted investment decisions, and weakened public trust in financial systems. Manual audits and statistical models can detect some fraudulent activities, but they can be inefficient when faced with increasingly complex fraud in the digital age. The problem is especially obvious in developing markets, including China, where financial fraud is widespread, and the regulatory structures have not necessarily kept pace with the fraudsters.

Machine learning can analyze vast datasets more quickly and accurately than traditional methods. However, it struggles with the non-linear aspects of financial data and, in particular, textual rather than numeric information. As such, applying advancements in deep learning could bolster machine learning and allow qualitative text found in corporate reports, such as the Management Discussion and Analysis (MD&A) section to be “understood” by fraud-detecting algorithms that might then spot the telltale signs of problematic corporate activity.

Chen’s dual-layer approach brings together financial data analysis and sentiment analysis. The use of bidirectional long short-term memory (BiLSTM) networks allows the system to interpret sequences of data, while a parallel network refines the key financial indicators using a convolutional neural network (CNN). Inconsistencies between the sentiment and the financial data can then be revealed.

Tests showed a fraud-detection accuracy of 91.35%, with an “Area Under the Curve” of 98.52%. This surpasses traditional fraud-detection methods by a long way, Chen’s results suggest.