The term “disruptive trading behaviour” was first proposed by the U.S. Commodity Futures Trading Commission and is now widely used by US and EU regulation (MiFID II) to describe activities that create a misleading appearance of market liquidity or depth or an artificial price movement upward or downward according to their own purposes. Such activities, identified as a new form of financial fraud in EU regulations, damage the proper functioning and integrity of capital markets and are hence extremely harmful. While existing studies have explored this issue, they have, in most cases, either focused on empirical analysis of such cases or proposed detection models based on certain assumptions of the market. Effective methods that can analyse and detect such disruptive activities based on direct studies of trading behaviours have not been studied to date. There exists, accordingly, a knowledge gap in the literature. This paper seeks to address that gap and provides a hybrid model composed of two data-mining-based detection modules that effectively identify disruptive trading behaviours. The hybrid model is designed to work in an on-line scheme. The limit order stream is transformed, calculated and extracted as a feature stream. One detection module, “Single Order Detection,” detects disruptive behaviours by identifying abnormal patterns of every single trading order. Another module, “Order Sequence Detection,” approaches the problem by examining the contextual relationships of a sequence of trading orders using an extended hidden Markov model, which identifies whether sequential changes from the extracted features are manipulative activities (or not). Both models were evaluated using huge volumes of real tick data from the NASDAQ, which demonstrated that both are able to identify a range of disruptive trading behaviours and, furthermore, that they outperform the selected traditional benchmark models. Thus, this hybrid model is shown to make a substantial contribution to the literature on financial market surveillance and to offer a practical and effective approach for the identification of disruptive trading behaviour.
- machine learning
- one-class support vector machine
- joint Gaussian mixture model
- hidden Markov model