Reproduce “Deep Time Series Models: A Comprehensive Survey and Benchmark” on OpenResearch

About this paper

Yuxuan Wang, Haixu Wu, Jiaxiang Dong, Yong Liu, Chen Wang, Mingsheng Long, Jianmin Wang

The authors survey deep time-series methods and introduce TSLib, a benchmark for comparing 41 models across 30 datasets and five tasks. They distinguish reusable design modules from model architectures and evaluate selected models under shared protocols. Across ten UEA classification datasets, TimesNet averages 73.67% accuracy versus iTransformer’s 70.06%; these task-specific results do not establish a universal best model.

Key insights

• Deep time series models can be effectively understood through a dual perspective, analyzing both their fundamental building blocks (e.g., handling non-stationarity, decomposition) and their primary deep learning architectures.
• The TSLib benchmark reveals specific strengths and weaknesses of different model architectures across various tasks, highlighting the general effectiveness of Transformer-based models for capturing complex dependencies.
• Zero-shot forecasting with Time Series Foundation Models benefits from increased input context length, but an optimal balance is crucial, as excessively long inputs can sometimes introduce noise.

Summary by alphaXiv.