Machine learning in trading: theory, models, practice and algo-trading - page 981

 
elibrarius:


I wonder if you can't compile the code which is written by yourself at the interpreter into byte code. For example if it's formatted as a package.

It is possible and not necessarily a package. The functions are the most interesting. It seems to be possible to write chunks of code which can then be executed by eval


cmpfun(f, options = NULL)
compile(e, env = .GlobalEnv, options = NULL)
cmpfile(infile, outfile, ascii = FALSE, env = .GlobalEnv,
        verbose = FALSE, options = NULL)
loadcmp(file, envir = .GlobalEnv, chdir = FALSE)
disassemble(code)
enableJIT(level)
compilePKGS(enable)
getCompilerOption(name, options)
setCompilerOptions(...)
 

Doesn't anyone have any pictures?

Would anyone trade in something like that?


 

The drain on the deposit will be fast and unrestrained. On January 1, 2018, we can clearly see the knee. This is bad.

On the face of it - training with overfit from mid-March to the last date, and choosing the best model by OOS (January - mid-March 2018).


Here's a picture for you, I can do that too. But in spite of the beauty - there will be no profit on new data, profitable models are made and look quite different.


 
Dr. Trader:

The drain on the deposit will be fast and unrestrained. On January 1, 2018, we can clearly see the knee. This is bad.

On the face of it - training with overfit from mid-March to the last date, and choosing the best model by OOS (January - mid-March 2018).


Here's a picture for you, I can do that too. But in spite of the beauty - there will be no profit on new data, profitable models are made and look quite different.


So you have no test.

how do profitable ones look? and if you re-train virtually all the time, is it worth it?

 
Dr. Trader:

But in spite of the beauty - there will be no profit on new data, profitable models are made and look quite different.


How profitable models are made and look, if it is not a secret?

 

Is it like this, to sit around for a whole year and dance?

Hmmm... is this a learning experience?



 
Alexander Ivanov:

Is it like this, to sit around for a whole year and dance?

Hmmm... is it learning?

It is training to the right of the red line, then it works in the past for a while, and then it already wiggles.

i don't have a good basic strategy.

 
Maxim Dmitrievsky:

The training to the right of the red line, then works in the past for a while, and then it already wobbles.

I have no good basic strategy.

i don't have a good basic strategy. is it possible to accelerate training?

 
Alexander Ivanov:

Is there any way to speed up the training?

even a day

You can't find a real pattern through MOE, you can only teach something that doesn't change. And this is something you should look for separately.

 
Maxim Dmitrievsky:

The training to the right of the red line, then works in the past for a while, and then it already wobbles.

This is essentially a pure model, only prices are fed to the input and outputs are selected automatically. since I don't have a normal basic strategy

Why is it such a fashion to test on older data than the training?

I propose to work together with my basic trend strategy - look for entry and exit is predefined - trawl using Doncian's channel.