Using Generative Adversarial Networks for Detecting Stock Price Manipulation: The Stock Exchange of Thailand Case Study

We implemented an automated system that uses unsupervised learning to detect stock price manipulation events. Generative adversarial networks (GANs) were trained with regular market transactions in a limit order book format. GANs can recognize normal trading behaviors of good governance stocks with the various price ranges, trading volume, and market capitalization. Stocks that were traded differently were assumed to be suspicious, thus required further manual investigation. We tested the system with 6 real manipulation cases that had been prosecuted from the stock exchange of Thailand. The proposed system can identify 5 out of 6 cases correctly with a very low false-positive rate.