What it does TSLib is a unified Python codebase for evaluating deep-learning models on five time-series tasks: long- and short-term forecasting, imputation, anomaly detection, and classification. It bundles implementations of 15+ papers—from vanilla Transformer (2017) through Informer, Autoformer, iTransformer, PatchTST, TimeMixer, and even a Mamba variant—behind a single experimental harness.

The interesting bit The maintainers are strikingly honest: they announced in April 2026 that they lack bandwidth for new features and that “many of its benchmarks may no longer be meaningful for evaluating the effectiveness or progress of current research.” They still vouch for the correctness of the baseline implementations. That candor is rare in a field addicted to incremental leaderboard gains.

Key highlights

• One repo, five tasks: forecasting (both horizons), imputation, anomaly detection, classification
• 15+ model implementations checked in, from Reformer to TimeXer to KAN-AD
• Leaderboard splits long-term forecasting by look-back length (96 vs. search-tuned) to expose a common methodological fudge
• Added zero-shot forecasting support in late 2025 for Large Time Series Models
• Docker deployment and a beginner tutorial (TimesNet notebook) available

Caveats

• Maintainers explicitly warn that benchmarks are aging; they recommend seeking newer ones
• No active feature development going forward due to limited maintainer bandwidth
• Some newer baselines (MambaSL, TimeFilter, KAN-AD) are added but not yet fully evaluated for the leaderboard

Verdict Grab this if you need a clean, reproducible starting point for time-series research or want to compare your shiny new architecture against a solid historical baseline. Skip it if you are chasing state-of-the-art on the latest datasets—the maintainers themselves will point you elsewhere.