Machine learning in trading: theory, models, practice and algo-trading - page 3623

New comment
 

Over a cup of coffee today.

if I'm not mistaken, I'll go out on the 1st :) I'll sabmise later.


 
Maxim Dmitrievsky #:

If I haven't got it wrong, I'll go out on the 1st :) I'll sabmise later.

https://hub.crunchdao.com/competitions/causality-discovery/leaderboard


If successful, it will be an excellent demonstration of the superiority of quality over quantity.
 
Maxim Dmitrievsky #:

Today, over a cup of coffee.

if I'm not mistaken, I'll go out on the 1st :) I'll sabmise later.

It's stressful.

 
mytarmailS #:

strains

Sure, but a separate test sample would be hard to fool. They're having some errors on the server today, I can't download. Tension's rising.

 
Maxim Dmitrievsky #:

Sure, but a separate test sample would be hard to fool. They're having some errors on the server today, I can't download. Tension's rising.

It's not a big deal.

When you don't realise an idea that will "definitely work" based on previous experience of the contest, but 5 hours to the end of the contest, that's where the tension is)).

 
mytarmailS #:

It's no big deal.

When you don't realise an idea that will "definitely work" on the basis of the previous experience of the contest, but 5 hours to the end of the contest, that's where the tension is)).

Found a catch. I trained on parts of the dataset, to test. There everything is good.

But the main dataset is very large, the invariance of features is not enough, training starts to "bump" on the spot. In the end it can't reach the same high speed.

Maybe we can fix it. For example, divide the sample into subdatasets, and then stack the models.

And without it the same magic 0.39-0.40 on validation will be there. Dataset is cool, it screws up in all directions.

 
Maxim Dmitrievsky #:

Maybe we can fix it. For example, divide the sample into subdatasets, and then stop the models.

Well, you can cluster the dataset and train a model for each cluster, although I don't believe this will do anything.

 
mytarmailS #:

Well, you can cluster the dataset and train a model for each cluster, although I do not believe that this will do anything

Not today. Yes.
 
mytarmailS #:

Well, you can cluster the dataset and train a model for each cluster, although I do not believe that this will do anything

It's a tree, it will make the clusters it needs internally.
 
Maxim Dmitrievsky #:

if I'm not mistaken, I'll go out on the 1st.

It is very difficult, there is most likely in the first place the famous expert Alexander Molak, he has a book on this topic

1...361636173618361936203621362236233624362536263627362836293630...3699
New comment