Issue resolved

[Aymuos22]

Made distribution the single source of truth for probabilistic regressors: added a helper to reconcile the old aliases, rewired CyclicBoosting, ResidualDouble, and XGBoostLSS to use it, updated docs/changelog, and added regression tests to ensure aliases still work and conflicting values raise a clear error.

[joshdunnlime]

Although this is nice solution for point 2)

Allow all options: dist, distribution or dist_type etc

it adds a lot of boilerplate so I would say it needs some refactoring. Perhaps a baseclass/mixin e.g. DefinedParametricDist that handles these getters and setters. These seems like a lot of complexity for the small added gain though.

That suggests option 3

Choose an skpro default, e.g. distribution but allow package specific types

with dist or distr as the default might be a better solution. It is also favoured by @fkiraly in #653. It will be much more lightweight by just checking if the extra-package dist-like parameter is non-None.

Here is a simple example:

    def __init__(
        distribution="normal",
        ...
        dist=None,
    ):
    """
    ...
    distribution/dist : ...
    """
    self.distribution = distribution
    self.dist = dist

    def _fit():
        self.distribution_ = self.dist or self.distribution

We would also need to add deprecation warnings as Franz mentioned.

[Aymuos22]

Hi @joshdunnlime
Update:
got the context;Working on the issue

[joshdunnlime]

@Aymuos22 - Maybe also change the PR to a bit more of a descriptive name.

[fkiraly]

Here is the deprecation guide: I would say, let's use the "changing parameter name" template:
https://www.sktime.net/en/stable/developer_guide/deprecation.html

[Aymuos22]

Hi @joshdunnlime @fkiraly
I am very new to open-source

I'll do the needful
Thanks for helping