5 Comments
User's avatar
Joan's avatar

Great post as usual.

I have a question though. It seems to me that you are running different models for each quantile and therefore the quantile crossing might occur. What about the idea of Meinshausen 2006 https://www.jmlr.org/papers/volume7/meinshausen06a/meinshausen06a.pdf.

This is just running 1 model, but still being able to calculate quantiles. Any feeling of pros and cons of one approach vs the other?

Many thanks for the great work.

Joan

Expand full comment
Christoph Molnar's avatar

Running one model can have many advantages, like avoiding quantile crossing, having just one model instead of 3. For my latest project the reason I dind't use the quantile forest was simply that it couldn't handle missing values, to be honest.

Adavantage of using (3) xgboost models is that you use a quite standard ML algorithm and benefit from all the ecosystem around that.

Expand full comment
Dhruv Nigam's avatar

Great, pratical stuff as always. Sharing with my entire DS team.

Expand full comment
Frank Corrigan's avatar

Potentially dumb question 😬 Would you use quantile regression to replace the news vendor problem optimization approach? Or as a compliment? I have a sense they are complimentary, but wondering how you would describe it. Thanks for sharing this :)

Expand full comment
Christoph Molnar's avatar

Honestly didn't know the news vendor problem before 😅 Need to look it up

Expand full comment