STLformer: Exploit STL Decomposition and Rank Correlation for Time Series Forecasting
Zuokun Ouyang, Meryem Jabloun, Philippe Ravier · 2023
The challenge of time series forecasting has been the focus of research in recent years, with Transformer-based models using various self-attention mechanisms to uncover long-range dependencies. However, complex trends and nonlinear serial dependencies presented in some specific datasets may not always be captured properly. To address these issues, we present STLformer, a novel Transformer-based model that utilizes an STL decomposition architecture and the rank correlation function to improve long-term time series forecasting. STLformer outperforms four state-of-the-art Transformers and two RNN models across multiple datasets and forecasting horizons.