Adding TimeBase model to the library (tiny linear NN model with quite interesting results), code available

[uladribia]

Recently, a group of Chinese researchers put forward a tiny model based on low dimension projection of patterns that is quite effective across a variety of datasets.

Their code is open source in this python repo: https://github.com/uladribia
And here is the paper: https://icml.cc/virtual/2025/poster/45815

I have built a (agent coded) version that is fully compatible with NeuralForecast, including the Auto* versions. My interest was on CPU usage (because those are tiny models), but it could be easily extended to GPU (have not tested them).

The code is open source, and I was wondering if Nixtla would consider adding it onto the library.

Here is the link: https://github.com/uladribia/timebaseula

The repo is available here and is fully documented across two main branches: A clean library implementation containing the classes and another branch that includes some benchmarks.

Feel free to tell me if it is worth doing a PR or similar.