Cannot run forecast on Python 3.12 (Dependency Deadlock with scikit-learn)

[astropedrocosta]

Hello team,

I am attempting to set up the project on a fresh macOS environment with Python 3.12. I've encountered a dependency deadlock that prevents the model from running.

The Issue:

• Python 3.12 requires scikit-learn >= 1.3 (older versions do not build).
• The pre-trained gradient boosting model downloaded by gb_loader appears to have been pickled with an older scikit-learn version (<1.3).
• When loading the model with scikit-learn >= 1.5, it throws: AttributeError: Can't get attribute '__pyx_unpickle_CyAbsoluteError' on <module 'sklearn._loss._loss' ...>

Steps to Reproduce:

• Clone repo on macOS (Python 3.12).
• Manually patch pyproject.toml to allow newer libraries (since numpy and scikit-learn requirements are too old for Py3.12).
• Run a basic forecast.

We likely need to re-pickle the gradient boosting models using a newer version of scikit-learn, or provide a separate model file for Python 3.12+ users.

I am happy to work on any potential fixes!

[Raakshass]

Hi @astropedrocosta! I've been investigating this issue and made some progress.
Would you like to collaborate, or should I proceed with a PR?

[astropedrocosta]

Hi @Raakshass! Thanks for looking into this!

I would love to collaborate. Since I have the local environment that reproduces the crash, I am happy to act as a tester.

If you have a branch or a proposed fix in mind, please feel free to link your branch here. I can pull it down immediately and run the forecast locally to confirm if it resolves the dependency deadlock.

[Raakshass]

Hi @astropedrocosta! Great to have you as a tester.

Here's my progress so far:

• Successfully migrated model-0.4.0.pkl using Python 3.11
• model-0.3.0.pkl has a deeper issue (internal class state mismatch)

Branch: I'm working in my fork but need to resolve the model-0.3.0 issue first.

Question for you: When you run the forecast, which model variant are you using - the default "gb" model or "xgb"? This will help me prioritize.

I'll push a test branch shortly for you to try!

[astropedrocosta]

Hi @Raakshass !

To answer your question: I am using the default "gb" model.

Additional Context: I tried to switch to model="xgb" locally as a workaround, but it also crashed on Python 3.12 with an _estimator_type undefined error. So it seems both model variants currently fail on the latest environment.

I see you opened PR #334. I will pull that branch locally right now and report my results there!