Reproduce “Tsururu: A Python-based Time Series Forecasting Strategies Library” 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

Alina Kostromina, Kseniia Kuvshinova, Aleksandr Yugay, Andrey Savchenko, Dmitry Simakov

Developed by Sber AI Lab, Tsururu is a Python library offering a modular framework for time series forecasting that enables flexible combinations of preprocessing, multi-series approaches, multi-step strategies, and models. The library's capabilities led to the discovery of highly effective yet rarely adopted techniques, such as the LastKnownNormalizer, and demonstrated the importance of exploring non-obvious combinations for optimal forecasting performance.

Key insights

  • The LastKnownNormalizer (LKN) preprocessing technique demonstrably improves forecasting accuracy over standard methods, despite being rarely used in existing libraries.
  • The effectiveness of features is context-dependent, with ID features generally improving accuracy, while date features can paradoxically degrade performance for both neural and GBDT models.
  • Optimal multi-step-ahead forecasting strategies are highly dependent on the specific model and dataset, with hybrid recursive and global approaches often outperforming commonly assumed optimal strategies like pure MIMO or multivariate settings.

Summary by alphaXiv.

Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.