Machine learning in trading: theory, models, practice and algo-trading - page 2074

 
Maxim Dmitrievsky:

no

What's the difference?

=========

Listen, do you remember what is the name of this forecast when you forecast one step ahead, then you forecast with this point, etc.

 
mytarmailS:

what is the difference?

read ) I don't know the first thing about math.

it's more like MSUA, but works through convolution of features, not polynomials
 
mytarmailS:

Listen, do you remember what is the name of this forecast when you forecast one step forward, then forecast further but with this point, etc.

many to many in RNN

 

I signed up for a course on convolutional nets %) 72 hours. well, the first part is learning python, all sorts of stuff (which i already know)


 
Maxim Dmitrievsky:

I gave a link to ROCKET for a reason - it's kind of a cool feature converter. Creates a lot of uncorrelated features from original ones, increases quality of classification.

It is recommended to use it with linear models (because it makes a lot of features).

I'll have to check it out.

Beautifully drawn. To understand how to prepare the data. After all, normalizing a number removes a lot of unnecessary data.

 
Maxim Dmitrievsky:

Many to many in RNN

Just there? How to search, you can key phrases. When will you watch tcn?

 
Maxim Dmitrievsky:

I signed up for a course on convolutional nets %) 72 hours. well, the first part is learning python, all sorts of stuff (which I already know)


It will not be superfluous)

 
Ilnur Khasanov:

Just there? How to search, can you key phrases. When are you going to look at tcn?

I don't know what you mean, it's the only place I know

When I take a course on coils. I don't know how to use them yet.

 

This is a question for the residents.

Have anyone tried to describe any indicator by a network? Not to predict, but to describe/copy/restore/create some phantom...

It is not very useful, but as a result of reconstruction we can judge about the quality of data processing/normalization and their quality, and it is not bad to know that grid doesn't work not because it is stupid, but because we didn't represent data well or vice versa

 

Many people have tried it. Me included. Simple MA can be done, and complex bandpass filters can be done.
Everything that can be built from bars can be easily duplicated by NS/forest.

That is why there is little sense in feeding indicators based on bars.


But ZZ, for example, is hardly reproducible. Or what Alexey Nikolaev suggested, hardly anyone would feed 60 or more bars for a depth of 2 months.