Machine learning in trading: theory, models, practice and algo-trading - page 2758

New comment
 
Maxim Dmitrievsky #: Plus a non-fixed stuttering window.

What is the uncommitted stuttering window? Different number of features/columns in each row? But you should always feed the same number of columns into the model.

 
СанСаныч Фоменко #:

1. The outlier resampling is not removed. There are programmes, but we can do it the kolkhoz way: we change everything greater than +/- 0.005 of the corresponding quantile to this value. The statistics change surprisingly.

2. Extremely interesting, especially on entropy. Would love to see the result. Correlation is for stationary series, you can forget it.

Resampling through everything with Gaussians inside removes it
 
elibrarius #:

What is the uncommitted window? Different number of features/columns in each row? But you should always feed the same number of columns into the model.

I wrote above somewhere a while ago, if the modders haven't cleaned it up.
 
Maxim Dmitrievsky #:
And the above was written somewhere recently, if the mods haven't cleaned it up

Search on " unfixed window" gives only this page

 
Maxim Dmitrievsky #:
Resampling through anything with gaussians inside it - removes it

Curious, but very clever.

 
elibrarius #:

A search on " non-fixed window" gives only this page

There was a thought somewhere about fractals and stuff, that the last price doesn't always have the best predictive power. That is, sometimes it is necessary to stop the window by conditions or through another ns to fix it, so that the previous bars participate in the prediction, not the last ones. And so it should run back and forth through history.
 
СанСаныч Фоменко #:

Curious, but very complicated.

Try Gaussian mixtures, I have an article about it. It's a generative model. Works better on increments than autoencoder.
 
Maxim Dmitrievsky #:
Resampling is done to remove outliers, Gaussianise the sample

I was generally suggesting meaningful sampling by entropy or correlation. To make the chips more informative. Plus take the increments and add maximum information to them from the original series by all sorts of transformations. Plus a non-fixed stuttering window. It's a frosh approach and no one has done this. But I got some coronavirus crap and I'm resting ☺️.

Casual infernnce was supposed to help pick out informative fiches as an option, but it turned out not to be about that there

Completely unclear.

I know exactly that I need different window on different parts of the time series. But the problem is an old one: if you pick a window on some segment, then it is not necessarily the same window on the next segment, some variant of non-stationarity of window width.

 
СанСаныч Фоменко #:

It's not clear at all.

I know for sure that it is necessary to have a different window on different parts of the time series. But the problem is an old one: if you pick up a window on some section, then not necessarily this window will be on the next section, some variant of non-stationarity of window width.

My head is wooden, I'll make an example later
For example, there was a set of chips that predicted a long decline. Then there is no point in moving the window with each bar, but to submit the same features on a new bar. And so on up to a certain point, when it will be moved again. These points are also to be populated.
 
Maxim Dmitrievsky #:
There was a thought somewhere about fractals and other things, that the last price does not always have the best predictive power. That is, sometimes it is necessary to stop the window by conditions or through another ns to fix it so that the previous bars participate in the prediction, not the last ones. And so it has to run back and forth through history.

It's more of a non time fixed window. And as we feed 10 columns, so we will be 10.
I tried feeding a dozen ZZ columns (got 50%50 as usual). The time of their formation is always different and the last knee was formed from 1 to several bars ago. You could say it's a conversion of regular bars to their own. In my case the zigzag knee was forming a bar. A year ago Prado was recalled here, he suggested to rebuild bars not by time, but by traded volume for example by 100 lots.


Examples of raw prices, time, "volume" and "dollar" candles

1...275127522753275427552756275727582759276027612762276327642765...3674
New comment