Machine learning in trading: theory, models, practice and algo-trading - page 3658

New comment
 
Maxim Dmitrievsky #:
Split the data into several similar groups, for example. Applicable, why not. Through clustering you can look for patterns. Each group will contain similar observations.
So, "turnkey" it doesn't work?

You have to "manually" analyse what "it" has scattered there into groups?
 
Ivan Butko #:
So, "turnkey" it doesn't work?

You have to "manually" analyse what "it" has scattered there in groups?
As part of the overall algorithm it can work hands off


Just one example.
 
Ivan Butko #:

1. Andrey, if it is not too much trouble, describe how you imagine the implementation of "without teacher" on forex, in my topic"What to feed to the input of neural network"

2. If you have a working machine without a teacher - then in private

1. there are the same "teachers", and threads with self-moderation are not expected in the near future.

2. In private.

Think about one thing - what you are offered by "teachers" are deterministic methods. It's like a hairbrush, you'll never get the hair style you want, only the way the "comb" method does.

 
Ivan Butko #:

There have been several articles on Kohonen's maps here

 
Rorschach #:

There have been several articles on Kohonen's maps here

The maps are not particularly common due to their tendency to overtraining. They are never in popular packs :)

I would call k-mins a timeless classic.
 
It seems that lgbmöklmn is the fastest bousting in terms of learning. There is also an option to use linear trees, which should speed things up. Then I will check it on minutes and ticks.

But in terms of response of the trained model it is the slowest, which may not be suitable for hft trading. Therefore, the model is then converted to catboost via ONNX, because it has the fastest response. Such manipulations.
 
Maxim Dmitrievsky #:
Sit down, col
+

He's far from the very basic, it's funny, and he's teaching something else😀😀😀😀
 
Maxim Dmitrievsky #:
Maps are not particularly common due to their tendency to overtrain. They are never in popular packs :)

I would call k-mins a timeless classic.
Umap + dbscan is the modern top...
Works without retraining
 
mytarmailS #:
Umap + dbscan is a modern top...
Works without retraining

hdbscan is ok, yes. Umap hasn't tried it )

 
Maxim Dmitrievsky #:

hdbscan is fine, yes. I haven't tried yumap )

Yumap is a modern non-linear PCA.
Hdbscan is a modern kmens but without parameters on the number of clusters and with better cluster separation.
1...365136523653365436553656365736583659366036613662366336643665...3741
New comment