Reproduce “PhaseFormer: From Patches to Phases for Efficient and Effective Time Series Forecasting” on OpenResearch
Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.
Universal app for Apple silicon and Intel. Requires macOS 11 or later.
About this paper
Yiming Niu, Jinliang Deng, Yongxin Tong
PhaseFormer presents a novel "phase-based" Transformer architecture for long-term time series forecasting, leveraging the inherent stationarity and low-dimensionality of observations aligned across cycles. The model achieves state-of-the-art accuracy across various benchmarks while drastically reducing computational costs and parameter counts compared to existing patch-based methods.
Key insights
- Patch-level processing in current models is inefficient due to the high variability of real-world cycle patterns, leading to bloated representations and computational overhead.
- Phase tokens, representing values at identical offsets across cycles, exhibit global stationarity and form compact, stable clusters over time.
- These phase tokens reside in a significantly lower-dimensional subspace, with over 90% of their variance explained by only two principal components, enabling efficient modeling.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.