Discussion of article "Advanced resampling and selection of CatBoost models by brute-force method" - page 9

[Deleted]  
Overall, something is improving. But having 50 models on every tick in production is overkill
[Deleted]  

Here's an option for a nice improvement


[Deleted]  

Here's the dataset after resampling.


[Deleted]  

Train everywhere 5 months, validation 2 years, examination 5 years

all charts for 5 years

[Deleted]  

One more time.

Yeah, it makes sense sometimes.

A little more staking. Yeah, it makes sense to stack. It's still an open question as to how much.

 
Maxim Dmitrievsky:

Once again.

Well, yeah, it makes sense sometimes.

A little more staking. Yeah, it makes sense to stack. It's still an open question as to how much.

Here we go. )

That's why Breiman made a random forest, not a better one.

[Deleted]  
elibrarius:

Here we go. )

That's why Bryman made a random forest instead of a better one.

Curiously, the effect appears when using Accuracy in training

It's not so obvious with other metrics.

by the way, there are a lot of them in catbusta, and I hesitate to ask which one is better to choose
 
Maxim Dmitrievsky:

Curiously, the effect appears when Accuracy is used in training

with other metrics, it's not so obvious.

By the way, there are a lot of them in catbusta, and I hesitate to ask which is the best one to choose

If Accuracy improves balance the best, apparently it is the best.

[Deleted]  
elibrarius:

If Accuracy improves balance the best, apparently it is the best.

No, there are some that do a better job.

but there's no meaningful understanding.

 
Maxim Dmitrievsky:

No, there are some that work better.

Which ones?