[API] `scikit-learn`-like API using `scikit-base` scaffolding

[fkiraly]

I think it would be best if pyportfolioopt drove closer to scikit-learn, based on scikit-base. This would also enable more natural and direct API integration with time series AI frameworks such as sktime, optimization frameworks like hyperactive, and, vice versa, use of time series covariance estimator components from pyportfolioopt or similar in sktime etc.

Key assumptions about these frameworks are dataclass-like __init__, and strategy pattern applied throughout; while data input does not happen in __init__. Currently, some of these assumptions are violated in different parts of the code base, arising most likely from ML de Prado's very erratic software design, in his famous book (good for methods but not good for API design...).

One topic to consider is also downwards compatibility - we do not want to break user code while moving closer to scikit-base.

[Yashasvikhatri9]

Hi! I'm interested in helping with this enhancement.

Aligning PyPortfolioOpt more closely with the scikit-style API (via scikit-base) sounds like a great direction, especially for interoperability with libraries like sktime and other optimization frameworks.

I’d be happy to contribute to the refactoring effort. Before starting, I wanted to ask:

• Are there specific modules that should be prioritized for this transition?
• Should new components directly follow the scikit-base estimator interface from the start?
• Would it make sense to begin with smaller steps (e.g., refactoring base classes like BaseOptimizer) to maintain backward compatibility?

Happy to start with a small PR if that helps move things forward.

[fkiraly]

@Yashasvikhatri9, the first item on this issue would be coming up with a good design. That is not a pull request but a short design document. However, you can also open a pull request as a supplement.