Reproduce “A Generalised Signature Method for Multivariate Time Series Feature Extraction” 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
James Morrill, Adeline Fermanian, Patrick Kidger, Terry Lyons
Researchers at the University of Oxford and Sorbonne Université developed a unified framework for the signature method, a feature extraction technique for multivariate time series, and established a canonical pipeline that performs comparably to state-of-the-art models on 26 datasets while offering computational efficiency.
Key insights
- Invariance-removing augmentations (time with basepoint or invisibility-reset) consistently improve performance by preserving parametrization and translation information in the series.
- Hierarchical dyadic and expanding windows are significantly more effective than sliding or global windows for capturing multi-scale temporal information.
- The signature transform consistently outperforms the logsignature transform, and increasing signature depth generally correlates with improved performance up to an optimal point.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.