#30 · Primary category: Time Series Machine Learning

TimeMixer

deep-learning machine-learning time-series time-series-analysis time-series-forecasting

[ICLR 2024] Official implementation of "TimeMixer: Decomposable Multiscale Mixing for Time Series Forecasting"

Project last updated:10/05/25

GitHub Stars

2.0K

Forks

238

Contributors

7

License

Apache-2.0

Why we included this project

TimeMixer is the official ICLR 2024 release of a forecasting architecture built on decomposable multiscale mixing. The core idea is to break a series into several time scales and recombine them so the model captures both short-term fluctuations and longer trends, which makes it a reasonable choice when you need something stronger than a plain LSTM or Transformer baseline. The repo ships training scripts and benchmark results, so you can reproduce the reported numbers and then point the same setup at your own data to see how it holds up. There is also support for TimeMixer++, a newer variant that extends the approach to a wider set of predictive tasks beyond plain forecasting. If you are comparing modern forecasting models, this is a useful reference implementation to keep around.

Articles for this project

No articles for this project yet.

To suggest a topic or contribute an article, contact us.

Related projects in this category