Reproduce “Time Series Forecasting Through the Lens of Dynamics” 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
Alexis-Raja Brachet, Pierre-Yves Richard, Céline Hudelot
Researchers from CentraleSupélec and Université Paris-Saclay introduced the PRO-DYN framework to systematically analyze time series forecasting models, emphasizing that explicit learnable dynamics are crucial for performance. Enhancing models with a dedicated dynamics component consistently improved forecasting accuracy in over 80% of test cases, while a PRE-DYN architecture (preprocessing followed by dynamics) generally proved superior.
Key insights
- Effective TSF models must be capable of learning the underlying 'physical dynamics' that govern time series generation, defined as the mechanism propagating past observations into future predictions.
- High-performing TSF models typically employ a 'PRE-DYN' configuration, where deep architectural blocks act as pre-processing units before a final learnable DYN block performs the prediction.
- Underperforming models often feature no or only partially learnable DYN functions, highlighting the necessity of an explicit, learnable dynamics component.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.