Machine learning in trading: theory, models, practice and algo-trading - page 1952

[Maxim Dmitrievsky]

Rorschach:
where is that?

There is also something like shap values (a separate package), but it seems to be only for representatives of the arboretum

[Rorschach]

Maxim Dmitrievsky:

look for a bot in the mart top for mt5 that trades seasonal

And think about how to reverse. I don't get the same even, but the topic works

I have the first grid in my blog about this theme

[Maxim Dmitrievsky]

Rorschach:

I have the first grid on my blog about this topic

Have you run it in a tester?

compare your best grid with boosting or random forest, you will understand that there is not much point in MLP

The only advantage is that the response time to receive a signal will be shorter. Well, it's a fraction of a second.

[mytarmailS]

elibrarius:

1) Something seems to me that it won't help much. This is information compression. If you compress garbage, it will be compressed garbage.
2 ) If you add 1 good chip to 2500 trash, the algorithm won't notice it much, and its influence on the final result will be if more than 1/2500, then not much. Even if it's 1/100, you won't see it on the graph.
3) The only thing I expect is that high-correlated features will sort of merge into one.

1) Well "it seems" is a strong argument))

2) And who prevents you from sifting out trash bits before compression? Even though I don't do it that way, but... you have to think, you have to decide, not philosophize...

3) Dimensionality reduction algorithms can be used in different ways, for different tasks, including, but not limited to, compression

[Rorschach]

Maxim Dmitrievsky:

Have you ever raced in a tester?

Compare your best grid to a boosting or random forest, you'll see that there's not much point in MLP

The only advantage is that the response time to get a signal will be less. Well, it is fractions of a second.

Will they be able to work just on increments? Without forming or selecting features?

[Maxim Dmitrievsky]

Rorschach:

Will they be able to work simply on increments? Without formation and selection of features

You don't need normalization there, otherwise the traits are the same as for MLP

[Maxim Dmitrievsky]

Maxim Dmitrievsky:

no normalization is needed there, otherwise any signs are the same as for MLP

I recommend catboost, I have a parser of a python-trained model into mql code (for binary classification only)

thanks https://www.mql5.com/ru/users/greshnik1

[Rorschach]

Maxim Dmitrievsky:

There is no need for normalization, otherwise the signs are the same as for MLP

I understand how the grid works, and I have some ideas what to do next.

[mytarmailS]

Rorschach:

I'm in the dark about boosting.

glad it's not random )

[Rorschach]

mytarmailS:

Good thing it's not random.)

That's what I wanted to write at first)