// the find
thuml/Time-Series-Library
A Library for Advanced Deep Time Series Models for General Time Series Analysis.
TSLib is a research benchmark suite from Tsinghua covering five time series tasks — forecasting, imputation, anomaly detection, and classification — with ~30 model implementations all runnable through a single CLI entry point. It's aimed at ML researchers who want to compare architectures on standardized benchmarks without reimplementing baselines from scratch. The maintainers themselves now warn (April 2026) that many benchmarks have saturated and may no longer be meaningful for evaluating progress.
The unified experiment layer (exp/ + run.py) means swapping a model into any task takes one argument change — no refactoring required. Shell scripts under scripts/ capture exact paper configurations, so reproducing a published number is a single bash command. The layers/ directory extracts attention and frequency components as shared primitives, so implementing a new model typically means composing existing blocks rather than writing from scratch. Zero-shot LTSM support (Chronos2, TiRex, Sundial) was added recently and is rare in benchmark repos of this age.
The maintainers' own April 2026 note is a red flag: they've frozen active development and the ETT/ECL benchmarks used throughout are widely considered saturated — delta improvements there don't tell you much anymore. Configuration is entirely CLI flags with no config file support, so 'experiment' really means a bash script with 20+ positional arguments; this gets painful fast when you want to do a proper hyperparameter sweep. There's no native support for custom datasets beyond dropping a CSV in the right folder and wiring a data_loader subclass — users with anything non-standard will fight the data_factory pattern. Mamba support is Linux-only due to the mamba_ssm wheel constraint, which silently breaks on Windows/macOS Docker environments.