Title: Identifying market manipulation \& abusive trading using anomaly detection techniques
Authors: Robert James - The University of Sydney (Australia) [presenting]
Artem Prokhorov - University of Sydney (Australia)
Henry Leung - Sydney University (Australia)
Abstract: A novel semi-supervised procedure is presented to detect instances of intraday abusive trading in financial markets, addressing the limitations present in existing rule-based expert surveillance systems, which are pervasive within the industry. It is considered that abusive trading produces highly abnormal patterns in the time series of limit order book activity, making such abusive activity detectable even in the absence of explicit assumptions regarding its form. We employ a state-of-the-art optimized implementation of the K-Nearest Neighbour Dynamic Time Warping algorithm to compute the similarity between multivariate time series sub-sequences of trading activity. A threshold defining the boundary between normal and abusive activity is constructed by applying univariate extreme value theory to the set of DTW similarity scores observed under estimated normal trading conditions. Using real world, tick-by-tick transaction data provided by a global investment bank we highlight the utility of the procedure in identifying instances of insider trading and demonstrate its competitiveness with respect to several benchmark algorithms used in related literature.