Detection and Simulation of Malicious Behaviour in Financial Markets

Detecting malicious behaviour in financial markets is an unsupervised machine learning problem, using unbalanced data with an unknown distribution. That is, respectively, there are no labels available when malicious behaviour occurred, the total volume of malicious trades is unknown, and there is no complete list of the types of manipulation. Our research will develop a tool which can detect malicious behaviour, by combining deep learning methods with domain knowledge from the financial field and anomaly detection knowledge grounded in particle physics. Due to the lack of true labels, verification will be performed in a simulated environment. Study 1 will search for contextual anomalies in the raw data, thereby splitting the order book dataset into individual events for further analysis. Study 2 will use domain knowledge about financial markets to create a set of known properties about manipulation, and link them with the individual events gathered in study 1. Study 3 will setup a virtual agent-based simulation, replicating a financial market. Study 4 will develop a malicious agent, participating in the virtual market from study 3, and verify whether the manipulation detection tool is able to extract these actions.