Machine Learning Approaches for Detecting Irregular Financial Activities in China

Abstract

Irregular financial activities (IFAs) pose serious challenges to regulators, especially in China where high-profile scams have highlighted gaps in oversight. This study develops a machine learning framework to identify such risks using a dataset of 540 financial cases from 2014 to 2024. Activities are classified as irregular or normal, and the performance of 18 algorithms—including traditional machine learning, ensemble methods, and deep learning models—is compared. Ensemble learning models demonstrate superior performance in detecting IFAs, balancing high accuracy with practical applicability. In particular, Bagging and LightGBM achieve the highest accuracy and robust F1-scores among all tested methods. These findings offer novel insights and technical tools for early warning of IFAs, contributing to the literature on financial risk detection and informing policy design. This comparison is among the first systematic evaluations of diverse machine learning algorithms for IFA detection in China, bridging a gap in the literature on regulatory technology and risk management. The proposed approach provides regulators with real-time, data-driven tools to identify irregularities before substantial losses occur.