Machine learning in trading: theory, models, practice and algo-trading - page 974

 

It seems that if you are looking for research software, Python would be the choice, as it is the most evolved.

My software is not even in the basement of the list.) But the next tasks it obviously can no longer hold.

There's another rapidly evolving -RapidMiner. I will have to see what kind of a tool it is.

 
Dr. Trader:

But now they finally did what they promised back in 2017 - reputation-based payouts. In the sliding window for the last 20 tournaments they give a point for each live<0.693. The number of points is reputation, max=20. Current record in my opinion = 9. Contestants with the highest reputation will be paid prize money even without stacking.

Yeah, i think i saw an email from them i remember it was 0.1NMR per round (live<0.693) if i'm not mistaken...

 
Yuriy Asaulenko:

It seems that if you are looking for research software, Python would be the choice, as it is the most evolved.

My software is not even in the basement of the list.) But the next tasks it obviously can no longer hold.

There's another rapidly evolving -RapidMiner. I'll have to see what kind of a bug it is.

RapidMiner is not a free application, either.

 
Vladimir Perervenko:

RapidMiner is an application and it's not free.

Yeah, I already looked it up. It's like LabView from NI or VisSim. Paid - minimum configuration is 10 thousand GEL per year.

 

I wonder why this can happen?
I use DNN Darch, backpropagation, dropout.

Epoch: 16 of 100
Classification error on Train set: 41.69% (2668/6400)
Epoch: 17 of 100
Classification error on Train set: 31.87% (2040/6400)
Epoch: 18 of 100
Classification error on Train set: 32.09% (2054/6400)
Epoch: 19 of 100
Classification error on Train set: 39.55% (2531/6400)
Epoch: 20 of 100
Classification error on Train set: 38.02% (2433/6400)
Epoch: 21 of 100
Classification error on Train set: 49.89% (3193/6400)
Epoch: 22 of 100
Classification error on Train set: 50.56% (3236/6400)
Epoch: 23 of 100
Classification error on Train set: 45.56% (2916/6400)

I.e. it takes a well-trained network from step 17 and degrades it by step 21.

Got it.
dropout put a new mask on each epoch, I'll switch to each miniBatch - it should get better

Didn't help:
Classification error on Train set: 45.11% (2887/6400)
Epoch: 10 of 100
Classification error on Train set: 34.92% (2235/6400)
Epoch: 11 of 100
Classification error on Train set: 30.16% (1930/6400)
Epoch: 12 of 100
Classification error on Train set: 45.28% (2898/6400)
Epoch: 13 of 100
Classification error on Train set: 39.56% (2532/6400)
 

SanSanych Fomenko:

Unlike the R-rating, there are several nuances to understand in the Python ranking:

  • there are many versions of this language, the two main ones are NOT compatible with each other, in search queries stackoverflow.com it is necessary to separate one version from another, moreover there are additional versions of this language, which should also be separated in search queries(Jython, Pypy, Iron-python, etc.
  • When you use rankings, you should understand WHAT the ranking shows: google queries, education, job search, ... The main thing is the subject area. Python is more versatile than R and at the expense of outsiders to statistics it ranks above R.
  • The competitors for R in the field of statistics and related visualization is SAS, which is far behind it.

The situation with Python 2.7 is the same as with MQL4. So for a new project there is no version choice - you take Python 3.

Jython, Pypy, Iron-python, etc. - are Python interpreters for other programming languages, not versions of Python itself. It is because of these interpreters that Python is now shoved everywhere.

The versatility is a big plus because you don't have to depend on a specific domain.

 
Roffild:

The situation with Python 2.7 is the same as with MQL4. So for a new project, there is no choice of version - you take Python 3.

Jython, Pypy, Iron-python, etc. - are Python interpreters for other programming languages, not versions of Python itself. It is because of these interpreters that Python is now shoved everywhere.

The versatility is a big plus because you don't have to depend on the area.

I've made my choice quite consciously, because I consider Python a mere gadget in comparison with R.

Moreover, I support such a desire to use Python, since Python is necessary for implementation of decision-making blocks in Expert Advisors, which (blocks) are difficult/complicated/impossible to implement in µl. But I am very interested in expanding such a community, especially if the testing of EAs which use Python will be really shown here. That's not a problem with R.

 

I made counter-trend targets, i.e. we look to see if there will be a profit on opening above/below the MA when the price touches the MA.

The results outside the training sample.

Well, 30% of trends are apparently confirmed, so it is not clear how to earn.

 
SanSanych Fomenko:

Especially if they really show here at least testing EAs that use Python. There is no such a problem with R.

I have shown testing of such TSs several times. Not in Python, but in other software. With Python, if there is a TS, there will also be no problem.

No problem with R either - I made TC in R and tested it. But R itself has not stuck with me, it turned out to be more convenient other software - SciLab (it's not only in the Top, but also in the basement is not)).

Imho, you are dramatizing the situation, reducing it to the irreplaceability of R. By the way, I have nothing against R either, and even use it occasionally.

 
SanSanych Fomenko:

I've made my choice, and quite consciously, because I consider Python to be an underdevelopment compared to R.

But I'm not against Python - let those who want to learn it, moreover I support such a desire to learn Python, because Python is needed to implement decision blocks in EAs, which (blocks) are difficult/complicated/impossible to implement in µl. But I am very interested in expanding such a community, especially if the testing of EAs which use Python will be really shown here. With R this problem does not exist.

A programming language is only a tool to solve problems. A programmer must not be tied to a certain language. Tools are constantly changing, evolving, and disappearing.

For me, to learn a new language at the entry level is to download "reference language" or "language specification" (normal languages have this documentation on the official website), and implement a simple task in the studied language. Only after that you can assess the syntax and capabilities of the programming language.

There are a lot of Python+MQL5 bindings on Github. Perhaps I will create my own...

A lot of people started with PHP. When you know only one language, it seems to be good. For a long time I could not understand why professional programmers do not like PHP, until I learned C# and Java. Now I can definitely say that PHP is terrible in syntax and library.

Right now Python is better than R. If tomorrow R will be improved, I'll go back to it, but not yet...