Machine learning in trading: theory, models, practice and algo-trading - page 3226

New comment
 
Maxim Dmitrievsky #:

Your TC is not looking for all patterns that may exist. That's why you have to match it

Then we face the "chicken and egg" problem:

  • The MO will retain patterns if it is given data from a working TC.
  • On the original tsvr, the MO must find a working TC.

Otherwise through another approach, but it will be long on ticks and later )

All such developers are far from trading. Instead of Average it should be Median.

 
Aleksey Vyazmikin #:

Maybe we should try to solve the problem of generating new data through approximation?

Take a window and try to describe a numerical series in it with different accuracy, while this approach will allow to globally preserve the dynamics of price movement, including taking into account daily fluctuations.

And, it will be enough to save the history in the form of approximators coefficients.

Sounds good. When I look at the articles on the topic of IDC research, I almost immediately begin to doubt their approaches, if there is a search for regularities under the assumption that they are round-the-clock.

 
Maxim Dmitrievsky #:

dependence with a length of 1000 ticks

And 5,000 ticks, on top of that.

With ticks such a choice of window is strange. It is logical to bind to timestamps, not to tick indices.

Forum on trading, automated trading systems and testing trading strategies

Machine Learning in Trading: Theory, Models, Practice and Algorithm Trading

fxsaber, 2023.09.09 04:40 pm

There is an error in this place: it should be time_msc. But it has no effect on the results after the post.

Please restart so that timestamps are displayed correctly for the next generations.
This is in case you decide to bind.
 
Maxim Dmitrievsky #:

dependence with a length of 1000 ticks

https://disk.yandex.ru/d/6F8FdUGthpnk3A

.

 
mytarmailS #:
You are not trying to find a single set of parameters, but some kind of clump / cluster of close sets of parameters.

Look how different the curves are in the Sample interval (between the blue lines).

This suggests that you get distant rather than close parameter sets. If in terms of a noisy target function, results near different hills are obtained.

It is clear that if one does not interrupt the GA, but waits until completion, a peak will be found. And all 20 best passes will be from there - Sample curves will be almost the same. This is of no use.

 
fxsaber #:

If you do not interrupt the GA, but wait until it is finished, a peak will be found. And all 20 best passes will be from there - Sample-curves will be almost the same.

I checked this statement on the same VDC without interrupting the GA.

It is well seen that there are dissimilar sets among this 20. Rather, this suggests that the regular GA has failed to do its job. More precisely, it interrupted itself by placing the results from different peaks in the top 20.

 
fxsaber #:

Sounds good. When I look at articles on the topic of tsvr research, I almost immediately start to doubt their approaches if there is a search for patterns assuming they are round-the-clock.

Well, I can tell you that not for all predictors a time split will give a meaningful probability bias, at least for me. So I tend to think that time is a significant factor, but other more significant factors can affect the outcome if they are in the active phase.

I also think that in the window you can simply take a different quantisation grid for the price, with a large number of intervals (for more accurate preservation of the structure), and test on such a "compressed signal" - it will already have deviations. Or you can use a small number of intervals plus random noise in the reference range between two intervals.

You can even fix the grid and store only the offset for the first reference. Then it will take up little space at all and the transformation should be fast.
 
Maxim Dmitrievsky #:

And with a 5k length, to top it off

https://disk.yandex.ru/d/1ypCrzYKk82XdA


Seems like an optimisation graph can show how hard the search process is. So here we go.


 
Aleksey Vyazmikin #:

Well, I can say that for not all predictors a time split will give a meaningful bias in probability, at least for me. So I tend to think that time is a significant factor, but other more significant factors can affect the outcome if they are in the active phase.

I also think that in the window you can simply take a different quantisation grid for the price, with a large number of intervals (for more accurate preservation of the structure), and test on such a "compressed signal" - it will already have deviations. Or you can use a small number of intervals plus random noise in the reference range between two intervals.

You can even fix the grid and store only the offset for the first reference. Then it will take up little space at all and the transformation should be fast.

Unfortunately, these are all hypotheses that require implementation and testing.

@Maxim Dmitrievsky is trying his variants, I am trying mine.
 
fxsaber #:

It seems like an optimisation graph can show how hard the search process is going. So here we go.

Graph of finding a pattern in the original series.

Honestly, I don't see a noticeable difference. This graph doesn't seem to tell us anything interesting.

1...321932203221322232233224322532263227322832293230323132323233...3516
New comment
Reason: