Machine learning in trading: theory, models, practice and algo-trading - page 3624

New comment
 
Evgeni Gavrilovi #:

It's very difficult, there is most likely the famous expert Alexander Molak at the top of the list, he has a book on the subject

Oh, wow. That's cool if you do. I've been reading bits and pieces.
 

You can still read the leaderboard there, because it reflects the results on the test sample, which is available. But at the end of the contest the results on another unknown sample will be counted.

Most likely Molak doesn't participate there, just random dudes.

 
Interesting thoughts on risk https://youtu.be/KyePP4oxums
 
Yesterday fumbled an approach that gave 0.43 scor not by accident, but steadily. I haven't poured it in yet. But I am still lagging behind a lot, because I switched on late :) there are still some thoughts on how to improve.
 
Most likely, the dataset has quotes for about 2000 stocks encoded in it. He is in charge of financial markets. That is, participation in the contest may have interesting consequences in terms of time series classification, some experience.
 
Maxim Dmitrievsky #:
Most likely, the dataset has quotes for about 2000 stocks encoded in it. He is in charge of financial markets. That is, participation in the contest may have interesting consequences in terms of time series classification, some experience.

Especially for the organisers. They will be able to use the models later, because they know how to create this dataset. But the participants do not.

 
Forester #:

Especially for the organisers. They can then use the models, because they know how to create this dataset. But the participants don't.

That's what the contest is all about, head!
 
Forester #:

Especially for the organisers. They can then use the models, because they know how to create this dataset. But the participants don't.

That's why I expressed my dissatisfaction with the prize fund :) because a good model can be used to get even more benefits.

Some stinking blogger earns so much from releasing a couple of videos.
 
Evgeni Gavrilovi #:

It's very difficult

There's a problem with labels 1, 3, 4, 6, 8. If you go down the list of their labels.

Each node K (with the exclusion of X and Y) can be in one of these 8 categories:

  1. Confounder : K→X, K→Y, X→Y

  2. Collider : X→K, Y→K, X→Y

  3. Mediator : X→K, K→Y, X→Y

  4. Independent : X→Y (no links to X or Y)

  5. Cause of X: K→X→Y

  6. Consequence of X: X→K, X→Y

  7. Cause of Y: K→Y, X→Y

  8. Consequence of Y: X→Y→K

If by number in the dataset, you have to subtract one from all, because in the dataset the numbering starts at zero.

If we look at the balance of classes, there are also very few of them compared to the remaining ones. Accordingly, the "balanced accuracy" metric, by which the result in the contest is evaluated, gives a low scor.


So, a hint if you haven't given up yet :)

They make the main confusion when learning.

 
Maxim Dmitrievsky #:

Okay, a hint if you haven't given up yet :)

They make the main confusion when learning.

Interesting. Certainly haven't given up, I'm experimenting with hyperparameters in catbusta

1...361736183619362036213622362336243625362636273628362936303631...3699
New comment