Adaptive Seasonal-Trend Decomposition for Streaming Time Series Data with Transitions and Fluctuations in Seasonality

Abstract

Seasonal-trend decomposition is useful for breaking down time series data into trend, seasonal, and residual components. However, the process requires knowing the season length, which corresponds to the duration of a complete cycle of the seasonal component. This requirement limits the method’s applicability to streaming data, the season length of which may change. An inappropriate parameter will cause the seasonal component to be determined inaccurately. To overcome this limitation, we propose a novel method that integrates season length estimation (SLE) into the decomposition process. The proposed method improves computational efficiency and ensures robust decomposition when faced with changing season lengths in data streams. By leveraging the sliding discrete Fourier transform, the computational cost of our SLE is O( N ), where N denotes the sliding window size, thus outperforming current SLE methods with computational costs of O $(N \log N)$ ( N log N ) . By adjusting the season length, we treat seasonality transitions and fluctuations in data streams. The proposed method demonstrates its accuracy, smoothness, and adaptability for online decomposition in synthetic and real-world datasets. The code is available at https://sites.google.com/view/astd-ecmlpkdd/.