Machine learning in trading: theory, models, practice and algo-trading - page 1268

[Vizard_]

Alexander_K2:

:))) Let my grandson Kesha and his pochekan Aliosha think and tell the whole story here, like at the Last Judgment. And I'll just convert their commandments into currency. Beautiful!

Internet - it's a big village, where the word of mouth quickly spreads, so the frequent references to cumulative and
(cumulativeative) led to creation of turkey by Hindu and whining of our Messiah about saving souls with rl didn't leave indifferent author of this thread)))
medium.com/@alexeybnk/improving-q-learning-agent-trading-stock-by-adding-recurrency-and-reward-shaping-b9e0ee095c8b

[mytarmailS]

Vizard_:

medium.com/@alexeybnk/improving-q-learning-agent-trading-stock-by-adding-recurrency-and-reward-shaping-b9e0ee095c8b

likes

[Aleksey Vyazmikin]

Vizard_:
Aleksey Vyazmikin
Alesha, not grid and parameters, but hyperparameters by grid, randomly or pr. But you have to think how to validate,
How to check (if necessary), not just how and what, otherwise the game is not worth the trouble...

Sir, what is the difference between parameters and hyperparameters in this terrarium? The name of the library to report would be appropriate...

I have a goal to test GPU performance in python and command line for small model sizes - 10-30 catbust trees.

[[Deleted]]

elibrarius:

Yes. And in DFSplitR duplicate, so that the regression forests also have the same functionality

put different values

qcnt=15;

qmin=1;

qmax=5;

etc., the file size does not change, the error does not seem to have much effect either

Maybe I don't understand it well, because I have no time

[[Deleted]]

Competent addition of noise in RL evens out the result on the trace with OOS, adding noise to the trace section, too, of course. Following the example of that article with DQN, but I have implemented it even earlier

https://habr.com/ru/post/436628/

Of course he went too far with sinusoid, too simple phrase for learning, but for searching errors in the logic, it's ok

I wonder how to add LSTM cells "with my own hands", I'll have to brainstorm

[Vasily Perepelkin]

The RNN and LSTM are in vogue now, has anyone tried them on Metatrader? They are supposed to be useful, because they work with sequences, which represent price time series, while conventional regression, which is the " workhorse of econometrics" just works with a normal distribution with Gaussian point clouds.

[Aleksei Kuznetsov]

Maxim Dmitrievsky:

set different values

qcnt=15;

qmin=1;

qmax=5;

etc., the file size does not change, the error does not seem to have much effect.

I may have understood it too badly, since I have no time

You can also serialize not in text files, but in binary files through FileWriteStruct. I think the files will be more compact and the processing will be faster.
Before writing the data, you can convert it to Float, if you don't need the double precision (in my opinion, you don't need it).
I'll probably do it myself, when I'll need it.

[Vladimir Perervenko]

Vasily Perepelkin:
I have a good idea to use RNN and LSTM, have someone tried them in Metatrader? By idea they must be useful, because they work with sequences, which are exactly the price time series. The ordinary regression, which is the "working Horse of econometrics" only works with normal distribution with Gaussian point clouds.

Use the R.mqh library and keras/tensorflow libraries, if you want from R, if you want from Python. No problem and full functionality and lots of examples to suit all tastes are available.

Good luck

[Renat Akhtyamov]

Vladimir Perervenko:

Use the R.mqh library and the keras/tensorflow libraries want from R want from Python. No problem and full functionality is available.

Good luck

Vladimir, if you have monitoring, please send me the link in a personal note.

[Vladimir Perervenko]

Renat Akhtyamov:
Vladimir, if you have monitoring, please throw the link in your personal message.

I don't monitor my own, I don't know about other people's. In the article quoted above, there is not enough information to reproduce and the code is too complicated. I think everything can be implemented with standard package layers, without using R6.

Good luck