Computational approaches for detecting manipulations in capital markets

Abstract

Financial markets abuse activities aim to artificially manipulate the prices or trading volumes of financial instruments to make profit. Among three primary market abuse formats, information-based, action-based and trade-based manipulation, trade-based manipulation has attracted much attention from the regulators around the world especially since the flash crash in 2010. This thesis presents novel algorithms for detecting trade-based market manipulations, specifically price manipulation, volume manipulation and cross-market manipulation.

Two detection models, termed the static model and the dynamic model, are proposed for detecting different aspects of the price manipulation activities. The static model considers each trading action as a single object without contextual relations and uses a "non-stationary feature reduction" transformation together with support vector machine (SVM) and k-nearest neighbour (k-NN) as the detection model. The dynamic model considers the temporal contextual relationship among sequential trading behaviours and uses a hidden Markov model based algorithm to detect the complex price manipulation behaviours. This thesis also proposes a directed graph and dynamic programming based two-step algorithm for detecting wash trade, major format of the volume based manipulation. In addition, this thesis proposes a Kalman filter (KF) and one-class support vector machine (OCSVM) based hybrid model for cross-market manipulation detection. This model monitors the trading actions in each possible exchange market as well as the correlations among those markets through a consolidated order book.

 All the proposed models have been extensively evaluated against existing models on a variety of datasets, including synthetically generated datasets, publicly available real market datasets and real cross market order book datasets donated by collaborators of financial companies. Experimental results demonstrate that Static Model and Dynamic Model can effectively detect the price manipulation activities, i.e. spoofing trading and quote stuffing and outperforms the selected bench mark models; the directed graph and dynamic programming based two-step algorithm on wash trade detection is also effective under different parameter configurations; and the KF and OCSVM hybrid model performs better than seven constructed KF or OCSVM based models on detecting manipulations across multiple markets.

Date of Award	Mar 2015
Original language	English
Supervisor	Yuhua Li (Supervisor)