STAR: Next-Scaled Autoregressive Model for Time Series Forecasting

Minh-Tuan Tran, Xuan-May Le, Ling Luo, Uwe Aickelin, Dinh Phung, Trung Le

Keywords: Time Series, Time Series Forecasting, Next-scaled, Autoregressive Model

TL;DR: STAR is a novel time series forecasting model that captures both global trends and fine-grained details, achieving state-of-the-art accuracy across benchmarks.

Abstract: We present the Next-Scaled Time-Series Autoregressive Model (STAR), a novel and effective method for time series forecasting that captures both the global structure and local details of temporal data. The architecture of STAR consists of two core components that work in tandem to improve the accuracy of the forecast. First, a transformer-based module models the coarse, long-term dynamics of the target series at a reduced scale, effectively capturing extended temporal dependencies. Second, a next-scale autoregressive module progressively refines forecasts from coarse to fine scales. It iteratively improves predictions using information from the preceding coarser scale, enabling precise reconstruction of fine-grained temporal dynamics. We conducted extensive experiments on seven widely used benchmark datasets in the time series forecasting domain. The results consistently demonstrate that STAR achieves state-of-the-art performance, significantly outperforming existing diffusion-based and transformer-based forecasting models across multiple evaluation metrics. Our code is available at \url{https://anonymous.4open.science/r/STAR-TSF/}.

Primary Area: learning on time series and dynamical systems

Submission Number: 20160