2025-01-18 TabPFN + nnetsauce = probabilistic time series forecasting

[thierrymoudiki]

Hi,

Following up from LinkedIn post: https://www.linkedin.com/posts/frank-hutter-9190b24b_tabpfn-neurips-tabpfn-activity-7284926409796419584-6YKX?utm_source=share&utm_medium=member_desktop

Here's a notebook with examples of use of external regressors:
https://github.com/Techtonique/nnetsauce/blob/master/nnetsauce/demo/thierrymoudiki-2025-01-18_tabpfn_nnetsauce_mts.ipynb

About scalability, I've seen that TabPFN scales quadratically. Don't know your architecture precisely, but could architectures like Mamba (scales linearly) be envisaged?

[liam-sbhoo]

Hi @thierrymoudiki

Thanks for sharing your demo :)

About the demo, I'm curious about couple of things if you could help:

• have you had a chance to add the time-series features we discussed? If so, did they help with your problem?
• have you compared the probabilistic forecasting output of TabPFN to TabPFN + nnetsauce?

About the scalability and architecture, I'm not exactly an expert in TabPFN architecture, but perhaps you could leave a comment in the TabPFN repo?

On a side note, my colleagues have worked on using Mamba for forecasting: https://github.com/automl/Mamba4Cast, perhaps it could be of your interest!

[thierrymoudiki]

Hi @liam-sbhoo,

• I saw your function indeed: the principle is exactly the same, i.e similar to what I did in the notebook. You'd just need to provide your features' data frame, and (important) add an index to it, the same as the time series input data frame.
• When you say comparison, do you mean benchmarking? I didn't. Interested in seeing anything you may have/suggest.

Thanks for the link!