Stock Price Manipulation Detection using Generative Adversarial Networks

We implemented Generative Adversarial Networks (GANs) for detecting abnormal trading behaviors caused by stock price manipulations. Long short-term memory (LSTM) was used as a base structure of our GANs, which learned normal market behaviors in an unsupervised way. After the training, the discriminator network of GANs was used as a detector to discriminate between normal and manipulative trading. Our work is different from the previous work in that we did not use manipulation cases to train the neural networks. Instead, we used normal data to train them, and simulated manipulation cases were only used for testing purposes. The detection system was tested with the trading data from the Stock Exchange of Thailand (SET). It can achieve 68.1% accuracy in detecting pump-and-dump manipulations in unseen market data.