thuml/time-series-library

★ 12,248PythonAudience · researcherComplexity · 4/5Setup · hard

Mindmap

mindmap
  root((time-series-library))
    What it does
      Model benchmark suite
      Five task types
      Leaderboard included
    Tasks
      Long-term forecasting
      Short-term forecasting
      Anomaly detection
      Imputation
      Classification
    Tech Stack
      Python
      PyTorch
      Docker
    Audience
      ML researchers
      Data scientists
    Notable Models
      TimesNet
      iTransformer
      TimeMixer

mindmap root((time-series-library)) What it does Model benchmark suite Five task types Leaderboard included Tasks Long-term forecasting Short-term forecasting Anomaly detection Imputation Classification Tech Stack Python PyTorch Docker Audience ML researchers Data scientists Notable Models TimesNet iTransformer TimeMixer

Click or tap to explore — scroll the page freely

Things people build with this

USE CASE 1

Run a fair benchmark comparison of published time-series models on your own dataset.

USE CASE 2

Reproduce published forecasting results for models like iTransformer or TimesNet.

USE CASE 3

Detect anomalies in sensor or energy data using a pre-implemented deep learning model.

Tech stack

PythonPyTorchDocker

Getting it running

Difficulty · hard Time to first run · 1h+

Some models like Mamba require separate pip installs, and GPU is expected for training experiments.

In plain English

This is a Python library for researchers who work with deep learning models applied to time series data. "Time series" means any data that comes in order over time, such as sensor readings, financial prices, energy usage, or health metrics. The library provides a collection of model implementations and a consistent framework for running experiments, so researchers can compare models against each other using the same setup. The library covers five types of tasks: forecasting future values over long time spans, forecasting over short time spans, imputation (filling in missing values in a series), anomaly detection (flagging unusual data points), and classification (assigning a label to an entire time series). The README includes a leaderboard showing which of the included models performed best on each task category as of early 2024, with models like TimesNet, iTransformer, and TimeMixer appearing at the top depending on the task. The collection of models is large. Dozens of published research models have been implemented and added over the years, each linked to its original paper and code. More recent additions include Mamba (a sequential model that needs a separate pip install) and support for Large Time Series Models that can perform zero-shot forecasting, meaning they can produce predictions on a new dataset without being trained on it first. The maintainers posted a note in early 2026 that they will not actively add new features going forward. They also note that some of the existing benchmarks, which were designed when the library launched three years ago, may no longer be the most relevant measures of model quality given how quickly the field has moved. The baseline implementations are described as still correct. A tutorial notebook for one of the included models (TimesNet) is provided for people new to the library, and a Docker-based setup was contributed by a community member for easier deployment. The project originated as an extension of an earlier repository called Autoformer, and a related companion project called OpenLTM explores a pre-training and fine-tuning approach for large time series models.

Copy-paste prompts

Prompt 1

Using the Time Series Library, set up an experiment to benchmark iTransformer on a CSV of daily electricity usage data.

Prompt 2

How do I add a new custom model to the Time Series Library benchmark framework?

Prompt 3

Run TimesNet on a multivariate forecasting task with a 96-step prediction horizon using the library's default config.

Prompt 4

Set up the Time Series Library with Docker so I can run experiments without installing Python dependencies manually.

Open on GitHub → Explain another repo

← thuml on gitmyhub — every repo by this author, as a profile.

Verify against the repo before relying on details.