Quartile-Based Seasonality Decomposition for Time Series Forecasting and Anomaly Detection

The timely detection of anomalies is essential in the telecom domain as it facilitates the identification and characterization of irregular patterns, abnormal behaviors, and network anomalies, contributing to enhanced service quality and operational efficiency. Precisely forecasting and eliminating predictable time series patterns constitutes a vital component of time series anomaly detection. While the state-of-the-art methods aim to maximize forecasting accuracy, the computational performance takes a hit. In a system composed of a large number of time series variables, e.g., cell Key Performance Indicators (KPIs), the time and space complexity of the forecasting employed is of crucial importance. Quartile-Based Seasonality Decomposition (QBSD) is a live forecasting method proposed in this paper to make an optimal trade-off between computational complexity and forecasting accuracy. This paper compares the performance of QBSD to the state-of-the-art forecasting methods and their applicability to practical anomaly detection. To demonstrate the efficacy of the proposed solution, experimental evaluation was conducted using publicly available datasets as well as a telecom KPI dataset.