Median Absolute Percentage Error missing inside scikit-learn/metrics

[paolobighignoli]

As already pointed out in the title, I see that the MDAPE is missing inside scikit-learn/metrics/_regression.py, and I was wondering if there is a particular reason behind that? If could be helpful I can work on it.

[adrinjalali]

@lorentzenchr has a better overview of our metrics I think.

[lorentzenchr]

With a grain of salt, let‘s bet: If someone can say with reference to a proof what property/functional of the distribution of the response variable y is minimized by this score, then I‘ll implement it myself😉

[paolobighignoli]

Ok so I guess its because is not so useful as I thought 😄.

[paolobighignoli]

Anyway, with a bit of delay, I would love to implement it. As you already pointed out is a metric that don't minimise any distribution of the response variable, but in my opinion, and from my experience, could be a useful metric to be compared with the more common MAPE.
If I implemented it do you think it would be useful? @lorentzenchr

[lorentzenchr]

I‘d say no.
Reason: We would like to focus on metrics that favor the best/optimal model predictions. Often, one wants to predict the mean or a quantile (conditional on X). With your proposed metric, it is completely unclear what the „best“ model would be.

Also where to stop? You can evaluate your metric (APE) for each sample point and you could be interested in a lot of quantities of that distribution, eg the 90%-quantile.