i'm getting baited into reviewing this but it doesn't seem finished yet... mark as WIP @glemaitre ? ;)

Ups sorry. This is WIP :)

You can have a look at the API thought. I change examples and I am writing proper test now that I did my example-based development :)

@NicolasHug @thomasjpfan I think the PR is ready to get some attention.
Open to any suggestion regarding the API

It seems that error plots are really just a matter of calling predict and scatter so it's already quite simple.

I agree with these statements. Basically, I wanted my arguments were the following:

• it is a useful plot that you want to use often in teaching (we need to it during developing our MOOC) or even in practice.
• the predict and scatter are actually the easy part. What I most dislike with some plots that I saw around: (i) the wrong diagonal is shown and (ii) the aspect ratio of the plot is not squared.

So I agree that this is more of a convenience function.

And I like its convenience! I am not sure we need to limit our plotting utilities to things that are difficult to achieve without.

I addressed the issue pointed out by @ogrisel

I think that using kind="residuals" as the default is indeed a good idea because the class name PredictionErrorDisplay is about "prediction error" which better "residuals vs predicted" than "actual vs predicted".

Let's go !

We can still improve some wordings later if needed.

Thanks @glemaitre !