With the main function to detect any abnormal trading activities, the stock market surveillance system plays a key role to enhance market integrity and efficiency which are essential to the success of a stock exchange. To detect abnormal trading activities, the existing market surveillance systems heavily rely on the surveillance rules that are predefined by market regulators and domain experts. As the surveillance rules mainly focus on the high level of market movements in terms of trading price and volume, they cannot identify those malicious trading activities associated with complex manipulation schemes and also likely generate too many false positive alerts. Therefore, there is a need for effective technologies that investigate the low level of trading activities directly.
This thesis studies the problem of discovering actionable abnormal behaviour patterns in investors’ trading activities. To deal with this problem, a methodology, namely market microstructure pattern analysis, is proposed. This methodology provides a framework to discover actionable patterns in market microstructure data for market surveillance. In order to obtain the actionable knowledge that actually supports market surveillance, domain knowledge, organisational factors and business interest are involved in the whole knowledge discovery process. On the basis of market microstructure pattern analysis, three case studies are carried out to address the issues of detecting insider trading and market manipulations. The first case study develops a method for insider trading detection by identifying turning points of market movement within announcement pre-disclosure. To address market manipulation detection, the second case study presents an approach to discover exceptional patterns in microstructure order sequences. The last case study also deals with market manipulations. It further investigates the multiple coupled trading activity sequences and proposes a method to identify abnormal trading patterns in them by an adaptive coupled Hidden Markov Model. The experiments conducted on real datasets in the above three case studies demonstrate the potential of proposed methodology to support effective market surveillance.
The main contributions of this research include the following aspects. Firstly, to the best of our knowledge, this work is the first systematic research effort studying trading behaviour patterns using intra-day data for market surveillance. Secondly, a methodology namely market microstructure pattern analysis is proposed for discovering actionable behaviours patterns in market microstructure data. Thirdly, a method for detecting insider trading is proposed. Fourthly, two methods for detecting market manipulations are proposed. Fifthly, three agent-based data mining systems for the above three methods are developed respectively.
In summary, this study is a cross-disciplinary research, by combining latest methodologies and techniques in data mining such as domain driven data mining and in agent mining with domain knowledge including market microstructure theory, to build the methodology of microstructure behaviour pattern analysis. It targets the actionability of the identified knowledge, which can support decision-making actions in business.