Analysis of Stock Market Manipulation using Generative Adversarial Nets and Denoising Auto-Encode Models

Market manipulation remains the biggest concern of investors in today’s securities market. The development of technologies and complex trading algorithms seems to facilitate stock market manipulation and make it inevitable for regulators to use Deep Learning models to prevent manipulation. In this research, a Denoising GAN-based model has been designed. The proposed model (GAN-DAE۴) consists of a three-layer encoder along with a ۲-dimension encoder as the discriminator and a three-layer decoder as the generator. First, using statistical methods such as sequence, skewness, and kurtosis tests and some unsupervised learning methods such as Contextual Anomaly Detection (CAD) and some visual and graphical methods, the manipulated stocks have been detected in the Tehran Stock Exchange from ۲۰۱۵ to ۲۰۲۰; then GAN-DAE۴ and some supervised deep learning models have been applied to the prepared data set. The results show that GAN-DAE۴ outperformed other deep learning models (with F۲-measure ۷۳.۷۱%) such as Decision Tree (C۴.۵), Random Forest, Neural Network, and Logistic Regression.

Keywords:

Anomaly detection , deep learning , Generative Adversarial Net (GAN) , Stock Manipulation Detection