Enhancing intraday stock price manipulation detection by leveraging recurrent neural networks with ensemble learning

Abstract

With the rapid development of the stock markets in developing countries, determining how to efficiently detect stock price manipulation activities to protect the interests of ordinary investors is really an important problem. Previous studies have introduced machine learning techniques into stock price manipulation detection and achieved better experimental results than traditional multivariate statistical techniques. Some characteristic features show statistically significant differences between manipulated and non-manipulated stocks, but this complementary information has rarely been considered in the manipulation detection model. The main contribution of our research work is the design of a novel RNN-based ensemble learning (RNN-EL) framework that combine trade-based features derived from trading records and characteristic features of the list companies to effectively detect stock price manipulation activities. Based on prosecuted manipulation cases reported by the China Securities Regulatory Commission (CSRC), we built a specific dataset containing labeled samples with trading data and characteristic information to conduct empirical experiments. The experimental results show that our proposed method outperforms state-of-the-art approaches in detecting stock price manipulation by an average of 29.8% in terms of AUC value. The managerial implication of our work is that government regulators can apply the proposed methodology to efficiently identify suspicious trading behaviors among huge amounts of trading activities in time to take action to ensure a fair trading environment.