Machine learning in trading: theory, models, practice and algo-trading - page 367

 
Maxim Dmitrievsky:


2 pieces with different periods, and one rsi.

But I want an ensemble of self-training NSs or one but very cool one... constantly using an optimizer and hoping for a miracle is not an option

If complex - you can put it on Google's servers, or azure, teach it there, and then just send a bot to the server queries and take the result ... and that's it, no stress on your computer or VPN ... the idea is a million . I.e. we are training a normal neuronet in a suitable cloud, and the terminal is just used for trading and getting results.

I'll try to feed it to my NS and calculate the correlation with the target one, just for fun). Maybe today I'll have time... Then I'll write how correlated they are.
 
Dimitri:


Then the second option - stick in NS everything you have. But there are two BUTs:



Well, why all?

There is a method that has been tested for several centuries. Not even a method, but a science with a capital letter.

It's called astrology.

It's all scientific, all by the book. You stick it into neural networks, and they are sure to bring a lot of money. The main thing is to be foolproof.

 
elibrarius:
I'll try to feed it to my NS, and calculate the correlation with the target, for interest) Just need to code the calculation of the correlation of inputs with outputs. Maybe I can do it today... ...then I'll write down how correlated they are.

I think it is possible to find a reg period when they become quite well correlated
 
elibrarius:
I'll try to feed it to my NS, and calculate the correlation with the target, just for fun) Just need to code the calculation of the correlation of inputs with outputs. Maybe I can do it today... ...then I'll write down how correlated they are.


Why are you picking on the correlation?

Dmitry told you a few tools above. This is not his invention. Predictor selection is one of the most important pieces of dataminig.

Don't be fooled.

Take caret. It's all shelved. There are three functions. They work perfectly.

 
SanSan Fomenko:


Why are you picking on correlation?

Dimitri gave you some tools above. This is not his invention. Predictor selection is one of the most important pieces of dataminig.

Don't be fooled.

Take caret. It's all shelved. There are three functions. They all work fine.

I just do everything in MT5 and NS from Alglib, by the way in Alglib there is a correlation calculation with matrix, so I don't bother with it too much.) But I should choose what to delete first and what to remove later...
 
SanSan Fomenko:


Why are you picking on correlation?

Dimitri gave you some tools above. This is not his invention. Predictor selection is one of the most important pieces of dataminig.

Don't be fooled.

Take caret. It's all shelved. There are three functions. They all work fine.


Have you earned a lot with such an approach? ;) With your karets and datamining you should know the honor). If to use data mining it is more or less intuitive for more or less experienced trader, and you almost do not need data maker, maybe sometimes for some not at all obvious things, but for the rest it is obvious... It is like using calculator all the time and you don`t know how to do it, when you get to the end of the day it is over and you don`t even know what to do. It's like using a calculator all the time or just learning the multiplication table. It's like paperwork with no result, just a process for the sake of the process... if you don't know where to look you'll NEVER find it with a dataminer. This has already been confirmed by many people here who have never found good predictors.
 
Maxim Dmitrievsky:

Have you earned a lot with this approach? ;) With your carriages and datamining... one should know the honor)). If to understand that data mining is an intuitive thing for more or less experienced trader, and you almost do not need data maker - maybe sometimes for some not obvious things, but for the rest - everything is obvious... It's like using calculator or just learning a table. It's like using a calculator all the time or just learning the multiplication table. It's like paperwork without any result, just a process for the sake of the process... if you don't know where to look you'll NEVER find it with a dataminer. This has already been confirmed by many people here who never found good predictors.
I think SanSanych meant not to bother with writing his own codes, but to use ready-made functions from R
 
elibrarius:
I think SanSanych meant not to suffer with writing your own codes, but to use ready-made functions from R


I have such a null hypothesis so far just - datamining is about nothing and is not necessary if you don't know what you want... now you don't know what you want from a neural network... what is the probability of success in this case?

Or do a FULL search of all kinds of predictors by means of datamining or what? without even the slightest genetics... well, the result, again, is obvious...

 
Maxim Dmitrievsky:

I have such a null hypothesis so far just - datamining is about nothing and is not needed at all if you don't know what you want... here you are now don't know what you want from a neural network... what is the probability of success in this case?
The first thing I want - to get the existing NS to work properly, and then I will look for the right predictors, and the ultimate goal - to earn, as well as everyone here)
 
elibrarius:
The first thing I want - to make the existing TS work properly, and then I'll look for the right predictors, and the ultimate goal - to make money, like everyone here).

And supposedly it should be like this: I already have TS, but I can't formalize it correctly and select parameters for it, I'll assign it to NS :)) Or just some simple classifier, by the example of Reshetovsky, for starters