Time-Series-Library

Why we included this project

Anyone working with time series who wants a proven starting point rather than reimplementing model code from scratch will find this library useful. It packages the reference implementations behind many of the field's best-known deep learning papers, with 40-plus models spanning five common tasks: long- and short-term forecasting, missing-value imputation, anomaly detection, and classification, all behind one uniform code base. That consistency is the real payoff, since researchers and teams can swap models, benchmark them on the same datasets (ETT, M4, electricity series), and get reproducible numbers without rebuilding each baseline. The maintainers are upfront that some older benchmarks may no longer be decisive for today's research, but the baseline code itself remains correct and useful. For practitioners it is also a clear reference for how modern architectures like Transformers and TCNs are actually wired for time series data.