Anomaly detection approaches for stock price manipulation detection

Stock price manipulation uses illegitimate means to artificially influence market prices of several stocks. It causes massive losses and undermines investors' confidence and the integrity of the stock market. It is evident from the literature that most existing research focused on detecting a specific manipulation scheme using supervised learning but lacks the adaptive capability to capture different manipulative strategies. This begets the assumption of model parameter values specific to the underlying manipulation scheme. In addition, supervised learning requires the use of labelled data which is difficult to acquire due to confidentiality and the proprietary nature of trading data. This thesis presents novel manipulation detection models that can generally detect all of the targeted manipulative schemes independent to the need of varying parameters for specific schemes. This thesis contributes five different detection algorithms for stock price manipulation in unsupervised domain that are categorised into three major models: decomposition based, artificial immune inspired and deep learning based. Decomposition based models transform stock price trades into orthogonal and principal components whilst preserving the original information of the input data. The transformed components are then subjected to a proposed multi-dimensional binary clustering techniques for manipulation detection. Two decomposition based algorithms have been proposed in this category that efficiently improved detection rates with reduced computational complexity. Immune inspired detection model translates the natural immune system approach into market manipulation treating a manipulative instance as a pathogen. The proposed approach is adapted for scaling down the dimension of the input data set to a set of only three outputs that are then clustered using KDE clustering. This avoids the need for assigning different threshold parameters as in a conventional DCA, hence automating the detection process. One of the main advantages of using this approach is the significant reduction in false positive rates while further improving the detection rates from the decomposition models. Deep learning based models can further simplify the problem by providing a set of features that can be used for training a model avoiding the need of designing features using an expert. Two deep learning algorithms are presented in this category: one model exploits the relationship among trading instances in the form an affinity matrix and later train an autoencoder based upon it. The second model presents a novel idea to reduce the false positives by detecting the overlap among normal and abnormal trades using a defined context. It proposes to jointly train a temporal convolutional network (TCN) and a generative adversarial network (GAN) together under the context extracted from the input data. Additionally, an updated similarity metric is explored using the feature representations learned by the GAN’s discriminator as the basis for reconstruction. All of the proposed research models are comprehensively assessed on multiple datasets of some highly traded stocks and outperforms some of the selected state-of-the-art models in anomaly detection. The robustness of the proposed models is further evaluated by comparing the results with selected benchmark models in stock price manipulation detection. Further a series of experiments on multiple datasets are also performed including when two or more manipulative activities occur within a short duration of each other and by varying the window length of the dataset fed to the model to evaluate the effectiveness of the models. The results show a significant performance enhancement in terms of the AUC, F-measure values while a significant reduction in false alarm rate (FAR) has been achieved