Machine learning in trading: theory, models, practice and algo-trading - page 730

New comment
 
I noticed earlier that the problem of long-playing TCs is that after two weeks or a month you forget how you did it.... But not in this case. I remembered my approach, because looking at today's signals, I hope that two weeks it will work as it should. Wait and see.....
 

Question on regression/prediction.

Is it necessary to normalize/scale the inputs? Something seems to me that it is not necessary.

 
elibrarius:

Question on regression/prediction.

Is it necessary to normalize/scale the inputs? Something seems to me that it is not necessary.

Regression usually uses input-output normalization. That is, they apply coefficients in forward order at the input and in reverse order at the output.
 
The sample:

This is a long-standing topic, about the number of samples in Mikhail's datasets and the classifier of the late Yury Reshetov, Mikhail is not bowed in the matter of the number of samples, and advertising Reshetov's handiwork makes us suspect only that Mikhail and Yura are one person, but given the age of the accounts and the number of posts, it is doubtful, but Yura should be grateful to Mikhail even from heaven or underworld, for such selfless promotion, quite mediocre handiwork.

Well, mediocre craftsmanship or not, I don't know. All I know for sure is that it doesn't retrain, but it does generalize quite well....

And about the training sample. Well I can not increase it, although I would do it with pleasure. Those data that I use do not allow to do so. That's why the sample chosen is the one that allows to get a model with an acceptable quality of training.....

 
The following is a good example:

Michael and Jura are the same person, but given the age of the accounts and the number of posts, it is doubtful, but Jura must be grateful to Michael, even from heaven or hell, for such selfless promotion, quite mediocre craftsmanship.

First of all, it is not known for sure.

Secondly, it is not mediocre at all, but a simple and quite good and working example of neuron application.

 

So, all night I was playing with some crap... I wanted to write it, and then I thought, who needs it?

but maybe someone will get a useful function for themselves, although it's the only one there - neuron learning :)

and in general some very unfortunate realization of RRL (recurrent learning with reinforcement), I haven't got such nice tests on mql4 version as the authors have on python

there are sources in python and mql4

https://github.com/darden1/tradingrrl

 

I finally got acquainted with this "neuron" of Reshetov.

Far from being a neuron. It is more of a trading system based on a genetic algorithm, without using a genetic algorithm.

The idea is interesting, but I don't really understand so much attention to it.

 
Aleksey Terentev:

I finally got acquainted with this "neuron" of Reshetov.

Far from being a neuron. It is more of a trading system based on a genetic algorithm, without using a genetic algorithm.

The idea is interesting, but I do not really understand how much attention to it.

This is an ordinary expert system, you can make a lot of such modifications through fuzzy logic.

 
Maxim Dmitrievsky:

So, all night I was playing with some crap... I wanted to write it, and then I thought, who needs it?

but maybe someone will get a useful function for themselves, although it's the only one there - neuron learning :)

and in general some very unfortunate realization of RRL (recurrent learning with reinforcement), I haven't got such nice tests on mql4 version as the authors have on python

there are sources in python and mql4

https://github.com/darden1/tradingrrl

a little bit more sublimated descriptions:

http://dustwell.com/PastWork/MoodyRLTradingPresentation.pdf

https://raghavgoyal14.github.io/assets/ML_project/Final%20Report.pdf

And the test one should have bad results, everything is beautiful on the learning one... but the system itself is pretty simple, in fact it uses one sharpe ratio for learning

they also write that RRL is better than q-learning but i have my doubts because it all depends on realization (what an agent learns and what kind of rewards he gets)

 

Something I, like Maxim, am tired of trying to train the NS with the abilities of a cockroach. I'll probably go back to training the more capable thing, my own brain. Although it is subject to both emotion and fatigue. And will be bored most of the time while nothing is going on. And from boredom (when you want drive) sometimes you start to trade too, and 50% of the time in the wrong direction...
In general there are disadvantages too.

Maybe I am just melancholy in spring and want some extreme action. I may lose my money on manual trades and come back. Although it would be better to start earning.

1...723724725726727728729730731732733734735736737...3674
New comment