Enhanced Multi-SARIMA Model for Anomaly Detection in Multi-Seasonal Time Series Data
Ashton T. Williams, Soon Myoung Chung · 2024
Today, there are many critical data sources generating multi-seasonal time series data that demand efficient and accurate anomaly detection. Current existing seasonal forecasting models struggle with multiple seasonal variations. To address this, we propose an enhanced multi-seasonal Autoregressive Integrated Moving Average (SARIMA) model. Our approach significantly improves anomaly detection accuracy in multi-seasonal time series data while maintaining practical runtime efficiency. Our experimental results demonstrate superior performance compared to both the original multi-SARIMA and SARIMA models, as well as popular TBATS [5] and Hierarchical Temporal Memory (HTM) [6], [7] models.