Anomaly detection in time series data

Abstract

Fraudulent activities have been affecting the world’s day-to-day business, resulting losses over millions to its users. Due to the advancement of technology, fraudsters have many platforms to stay undetected and fraud detection has become critical for organizations and stakeholders such as bankers, investors, credit card users, financial auditors and insurance agents. In addition to detecting fraud, anomaly detections such as diagnosing certain sicknesses, identifying climate change disasters, early warnings on floods and volcano activations are important for many stakeholders. Time series data, in any field can provide significant information when it comes to detecting anomalous activities. Bayesian network, neural network, logistic method, K-nearest neighbor and regression are the most popular data mining approaches in anomaly detection that are being introduced so far. Finding anomalies among time series data is the key approach of detecting fraudulent activities and identifying striking real-world scenarios such as economic recessions. The existing methods and approaches have limitations such as manual thresholds to determine anomaly points, lack of theoretical foundation and lack of attention to class imbalance problem and multivariate data. Approaches such as interquartile rang (IQR) methods assumes that time series data are normally distributed. This assumption does not comply with practical data as they could be positively or negatively skewed. Therefore, to overcome these limitations, this research project introduced a framework to detect anomalies in both univariate and multivariate time series data using extreme value theory. Results of this framework demonstrates that this method’s succeeds identifying the anomalies better than the methods implemented in the two R software packages, anomalize and timetk for detecting anomalies in univariate timeseries data. Further, this framework visualizes the identified anomalies which can be utilized by all stakeholders in for both univariate and multivariate anomalies.