Machine learning in trading: theory, models, practice and algo-trading - page 1384

New comment
 
Vitaly Muzichenko:

It's a pity you can't give a "Like".

You can just buy money )) just kidding

 
elibrarius:

Yuri does well with simple increments, too

I don't use gradients.)
 
Maxim Dmitrievsky:

the price in the market reflects the balance of supply and demand, mostly at different historical moments

there is another problem: how much history should be analyzed in MO?

if we use some constant Bars = 1000

won't it be unreliable data for training?

 
Maxim Dmitrievsky:

x.append((SD.history[i-j][c.c]/SD.history[i][c.c]-1)*1000)

it makes no sense, each subsequent feature contains half of useful information from previous feature, i.e. they, 1: strongly correlate, 2: the feature with the biggest lag contains all variance, which is contained in the previous features, i.e. they do not give any information increment

The result will be as follows: importance of the return with the longest lag will be the biggest (more variance, more information gain), and this return contains all variance of other features

In a long trend = yes. And the importance, the farther away, the stronger and the correlation is, as everyone grows in the same direction.

And in this situation:


The 20th bar is at the same level as the 0th, but the 5th and 10th bars have more information than the 20th. And there is a correlation except for the 2-3 neighboring ones.

There are many situations and it is necessary to analyze all the bars.

As an option, you can thin out the series as the creator of this branch did (in his blog).

 
Maxim Dmitrievsky:

x.append((SD.history[i-j][c.c]/SD.history[i][c.c]-1)*1000)

it makes no sense, each subsequent feature contains half of useful information from previous feature, i.e. they, 1: strongly correlate, 2: the feature with the biggest lag contains all variance, which is contained in previous features, i.e. they do not give any information increment

the result will be: the return with the biggest lag will have the biggest import (more variance, more information gain), and this return contains all the variance of other features

Creepy.))
The analogue of what I do is a photo.
 
Igor Makanu:

There is another problem: how much history should be analyzed in the MO?

If we use some constant Bars = 1000

won't it be unreliable data for training?

I assume that if we break the price into levels, then we can calculate the average depth of history by levels, starting from when the price came to it and finishing when it left it

 
Yuriy Asaulenko:
I don't use increments.)

SD.history[i-j][c.c]/SD.history[i][c.c]

They are relative increments. You just call them by other names.

 
elibrarius:

In a long trend = yes. And the importance, the longer the trend, the stronger the correlation is, as everyone is growing in the same direction.

And in this situation:


The 20th bar is at the same level as the 0th, but the 5th and 10th carry more information than the 20th. And there is a correlation except for the 2-3 next to it.

There are many situations and you have to analyze all the bars.

As an option - it is possible to thin series, as the creator of this branch did (in his blog).

it means that with increasing samples the correlation will be maximal, if we average

locally no one is interested.

 
Yuriy Asaulenko:
Creepy.))

Well, calculate the correlation between your predictors, for the whole sample

and then throw them all out)

 
Maxim Dmitrievsky:

Well, calculate the correlation between your predictors, for the whole sample

and then throw them all out)

You're wrong. This is the only way to do it.
1...137713781379138013811382138313841385138613871388138913901391...3591
New comment