The time-series benchmark library its maintainers are retiring

What it does TSLib is a unified evaluation harness and model zoo for deep time-series research. It bundles implementations of about twenty architectures—from vanilla Transformers to iTransformer, Mamba, and KAN-AD—under one roof so you can benchmark them on long- and short-term forecasting, imputation, anomaly detection, and classification without rewriting boilerplate. The maintainers describe it as a “neat code base” extended from their earlier Autoformer work.

The interesting bit The most unusual feature is the leaderboard split by look-back window length, an explicit attempt to fix the scientific rigor problem caused by papers using inconsistent historical lengths. More strikingly, the maintainers themselves warn that many of these benchmarks have become saturated and “may no longer be meaningful for evaluating the effectiveness or progress of current research,” even though the baseline code remains correct.

Key highlights

• Implements established models such as TimeXer, TimeMixer, PatchTST, TimesNet, DLinear, and FEDformer, plus newer arrivals like MambaSL and KAN-AD.
• Covers five tasks: long-term forecasting (split into Look-Back-96 and Look-Back-Searching), short-term forecasting, imputation, anomaly detection, and classification.
• Recently added zero-shot forecasting support for Large Time Series Models.
• Includes a Docker deployment and a tutorial notebook aimed at beginners.
• Backed by a comprehensive survey paper and the “Accuracy Law” proposal for identifying saturated datasets.

Caveats

• The maintainers have announced they will not actively add new features and recommend seeking newer benchmarks for current research evaluation.
• The leaderboard was last updated through March 2024, so recent model rankings may not be reflected.

Verdict Researchers who need a quick, fair head-to-head comparison of classic and recent deep time-series architectures will find this invaluable. If you are chasing state-of-the-art on modern benchmarks, the maintainers themselves suggest looking elsewhere.