Collusion set detection within the stock market using graph clustering & anomaly detection

Abstract

Manipulations that happen within the financial markets directly affect the stability of the market. Therefore detection of manipulation ensures fair market operation. Most of these manipulations occur in the guise of collusion. Collusion in financial markets involves a group of market participants trading amongst themselves to execute a manipulative trading strategy. Most existing models do not consider the seemingly rare yet normal transactions into account when proposing collusive groups. Neither have they considered the effect of time within collusion. This work proposes a model to detect collusion in stock markets through the application of graph mining and anomaly detection. Creation of investor graphs denoting the relationships between investors and timely sampling of these graphs using Graph mining allows this research to consider the effect of time in collusion, subsequent anomaly detection allows for the filtering of results to avoid misnaming normal behaviour within the stock market. This research presents that Graph mining techniques such OPTICS and Spectral clustering perform consistently well to extract meaningful collusive groups, while the Local Outlier Factors work well as an Anomaly detector to filter out results received from Graph Clustering. The combination of these methods creates a pipeline which can outperform existing methodologies.