Machine learning in trading: theory, models, practice and algo-trading - page 3659

 
mytarmailS #:
Umap + dbscan is a modern top...
Works without retraining
I support it. the bundle works well;
 
mytarmailS #:
Yumap is a modern non-linear PCA
Hdbscan is a modern kmens but without parameters for the number of clusters and with better cluster separation.

I comparedit here.



Is there any "unconventional" ways of using UMAP, or only for reducing the number of features?

I would like to read somewhere...

 
Maxim Dmitrievsky #:

or just to reduce the number of signs?

For reduction

If 2-4 signs, then smart normalisation + hdscan and nothing else to add.
 
mytarmailS #:
To reduce

If 2-4 signs, then smart normalisation + hdscan and nothing else to add.

They say that for clustering you can use

Using UMAP for Clustering — umap 0.5 documentation
  • umap-learn.readthedocs.io
Using UMAP for Clustering UMAP can be used as an effective preprocessing step to boost the performance of density based clustering. This is somewhat controversial, and should be attempted with care. For a good discussion of some of the issues involved in this, please see the various answers in this stackoverflow thread on clustering the...
 
Maxim Dmitrievsky #:

It is written that you can use for clustering

Well, they just showed a bunch of yumap + different clusters and showed how badly kmins clusters and how coolly dbscan clusters.
 
mytarmailS #:
Well there just showed a bunch of yumap + different clusters and showed how badly kmins clusters and how coolly dbscan clusters.

No, at first hdbscan was even worse than kmins, and then after dimensionality reduction with yumap they made dbscan, but forgot about kmins :)

We can figure out how to create TCs with this engineering marvel.
 
mytarmailS #:
To reduce

If 2-4 signs, then smart normalisation + hdscan and nothing else to add.

Reduction should be done by throwing out predictors, not burying noise.

 
Maxim Dmitrievsky #:
You can figure out how to create TCs with this engineering marvel.
Classification recipe GUARANTEED without retraining

You take the class labels.

Predictors you pass through yumap then into hdscan then look for such clusters in the sets of which the same class label is often found....

That's it... :)

The classifier will work fine on new data.

 
СанСаныч Фоменко #:

https:// dzen.ru/a/Yq3r326lRUNnWscY


collect all the materials that are presented in this article.

o From the first study - "Speculative Skill - Deutero-Traders' Outcomes" - it appears that less than 1% of deutero-traders were able to earn consistently in the long term (a year or more), 0.11% could earn super profits (higher than average salaries).

o From the second study -"Can you live from deutrading?" - it appears that only 0.24% were able to earn in the long term (again, a year or more). Only 0.09% were able to earn super profits.

Notice how the two papers have similar numbers. Recall that the first study was from 1992 to 2006, the second from 2013 to 2015.

Several tens of thousands of backtests by our team (can't find the

Well, and the main thesis of this paper:

If you come to the market for profit and with rational intentions, action trading and scalping will not bring you that very profit in the long term. The probability is about 99%.

Some lazy losers wrote an article for equally lazy non-slackers....

They did a couple of thousand backtests and came to a conclusion... ahahahaha....

Backtests of what? muggy crosses and overbought stochastics.... Oh shit....
 
mytarmailS #:
Recipe for classification GUARANTEED without retraining

Take the class labels

Predictors you pass through unmap then into hdscan then look for such clusters in which the same class label is often found....

That's it... :)

The classifier will work fine on new data

Well the same works for k-mins )