Detecting market manipulation in stock market data

Abstract

Anomaly Detection is an extensively researched problem that has diverse applications in many domains. Anomaly detection is the process of finding data points or patterns that do not conform to expected behavior within a dataset. Solutions to this problem have used techniques from disciplines such as statistics, machine learning, data mining, spectral theory and information theory. In the case of stock market data, the input is a non-linear complex time series that render statistical methods ineffective. The aim of this thesis, is to detect anomalies within the Standard and Poor and Qatar Stock Exchange using the behavior of similar time series. Many works on stock market manipulation focus on supervised learning techniques, which require labeled datasets. The labeling process requires substantial efforts. Anomalous behavior is also dynamic in nature. For those reasons, the development of an unsupervised market manipulation detection technique would be very interesting. The Contextual Anomaly Detector (CAD) is an unsupervised method that finds anomalies by looking at similarly behaving time series and uses them to predict expected values. When the predicted value is different from the actual value in the time series by a certain threshold, it is considered an anomaly. This thesis will look at the Contextual Anomaly Detector (CAD) and implement a different preprocessing step to improve recall and precision.