Detection of Point-Wise Anomalies in Large-Scale Hierarchical Time Series: Methods and Applications

Anomaly detection in multivariate time series data is a critical challenge in various domains, including finance, healthcare, cyber security, retail, and many other fields. The ability to identify unexpected patterns or deviations in large-scale datasets has significant implications for decision-making and operational efficiency. This thesis focuses on the detection of outliers in hierarchical, high-dimensional time series data. The dataset presents a unique challenge due to its hierarchical structure of multiple stores and departments with interdependent sales patterns over time, making it a complex time series problem. In the first part of this thesis, a comprehensive review of theory and anomaly detection models is conducted, assessing their suitability for large scale hierarchical time series data. Based on the findings from the research phase, anomaly detection methods are selected and implemented on the dataset. The second part of the thesis applies these approaches to the data, comparing their effectiveness in identifying significant outliers. The results are then analyzed to assess their practical implications for real-world applications. This study contributes to the field by identifying robust methodologies for outlier detection in hierarchical time series data and providing actionable insights for businesses dealing with large-scale sales forecasting. The findings can support more informed decision-making, reducing financial losses and enhancing operational strategies. Point-wise anomaly detection, large-scale hierarchical multivariate time series, decomposition, unsupervised models, interpretability, scalability, evaluation, precision, hierarchical reconciliation.