Identification of stock market manipulation using a hybrid ensemble approach

Anomaly detection in time series data is a complex data mining issue with many useful, real-world applications. Anomalies in datasets represent deviations in the expected behaviour of a system and can indicate rare but significant events that require intervention. Market manipulation is a serious issue in financial jurisdictions worldwide, with financial regulators such as the SEC constantly trying to prevent it and prosecute those guilty of it. This paper makes use of state-of-the-art deep learning techniques as well as more classical statistical techniques in order to detect anomalies in five real-world datasets. The predictions of these models are then aggregated in two different ensemble models. The results of the individual models as well as the ensemble models, are evaluated, and F1-Score measures performance. Nine individual models, consisting of three models based on LSTM with Dynamic Thresholding, three ARIMA models and three Exponential Smoothing models, were used to generate predictions of anomalies based on daily trading volumes. The individual predictions of these models were then aggregated, with two different ensemble methods being used, namely the majority voting ensemble method and the ensemble averaging aggregation method. While both performed well, the majority voting ensemble method was seen to be the superior method in this study, with an average F1Score of 0.494, compared to an F1Score of 0.414 for the ensemble averaging aggregation method.