Discussion: Allow for passing all arguments as keyword arguments

[simon-hirsch]

Hi @frazane, @sallen12

question for the v1.0.0: would you consider moving away from a position-based argument passing?

The background for the question is the following: in the current setup, it is not possible to have something like a mapping between values and parameters and do things like:

params = {"fct" : array, "obs": array}
score = sr.some_score(**params)

just for scores and observations, this might not be a big of an issue, but once one handles closed-form forecasts from different distributions, it's very convienient to do things like:

params = {"param_a" : array, "param_b" : array,  "obs": array}
score = sr.some_score(**params)

because then it's just about mapping distributions / scores / forecasts and everything else can be handled by passing keyword-value dicts.

I know the main argument for positional only is that you have API stability, that's why I'm asking now. :)

Background for this questions is that I wanted to make the scoringrules CPRPS available as optional dependency in our package for online distributional regression and had to do some weird magic to ensure that all the distribution parameters are mapped correctly all of the time https://github.com/simon-hirsch/ondil (see here) and I believe this could be benefical for wider usage.

Cheers, Simon

[frazane]

Hi @simon-hirsch, thanks for opening this issue. This has been on my mind for a while, and I am inclined to go ahead with your suggestion. I opened a draft PR: #113 feel free to check it out, test it and leave comments.

For now, I have removed positional-only arguments for all closed-form metrics, but we might do the same for ensemble metrics for consistency. What do you think?

[simon-hirsch]

Hi @frazane - thanks I'll have a look. I'm a big fan of consistency so I'd go for a positional|keyword everywhere where there are no defaults like $p$ for the VS. It resolves ambiguity and explicit is better than implicit :)