Reproduce “Time-SSM: Simplifying and Unifying State Space Models for 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
Jiaxi Hu, Disen Lan, Ziyu Zhou, Qingsong Wen, Yuxuan Liang
Researchers from the Hong Kong University of Science and Technology (Guangzhou) introduced Time-SSM, an efficient State Space Model-based foundation model for time series forecasting. Grounded in a novel Dynamic Spectral Operator theory, Time-SSM achieves superior performance across diverse datasets with one-seventh the parameters of Mamba.
Key insights
- State Space Models can be formally understood as "Dynamic Spectral Operators," adaptively projecting time series into optimal spectral spaces for learning complex temporal dependencies.
- Contrary to findings in other sequence modeling domains, robust and finite-measure SSMs (e.g., S4D-real) demonstrate superior performance for time series forecasting due to the nature of TSF data with noise and limited lookback.
- The HiPPO-LegP basis, utilizing piecewise Legendre polynomials, provides stronger dynamic representation capability for multi-scale time series compared to existing bases.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.