Abstract
Anomaly detection in time series data is a critical task across a broad spectrum of industries, including finance, healthcare, cybersecurity, and industrial process control. Effective identification of anomalous events is pivotal for proactive decision-making processes and fast, effective incident responses. Time series anomaly detection (TSAD) has been fastidiously researched for decades and there is no shortage of innovation in the field today. While laudable for moving TSAD forward and, in fact, necessary to leverage new technology, constantly looking forward risks obscuring sound methodologies from the past. To assume that current research always illuminates the best path does a disservice to people who need effective, efficient, and explainable solutions.This thesis provides a framework to help TSAD practitioners identify their time series anomaly problems and choose appropriate solutions from a broad range of options. First, understanding anomalies and data characteristics are essential prerequisites for finding anomalies. Second, research, identify, and select several candidates from a broad range of TSAD algorithms. Third, methodically and empirically test the selected algorithms using tools available to the TSAD community. Last, analyze the results and synthesize the test data into practical recommendations for decision making. This work progresses through each of these steps while simultaneously defining them and demonstrating their value.