Discussion of article "Deep Neural Networks (Part VIII). Increasing the classification quality of bagging ensembles"

MetaQuotes 2018.09.28 14:33

New article Deep Neural Networks (Part VIII). Increasing the classification quality of bagging ensembles has been published:

The article considers three methods which can be used to increase the classification quality of bagging ensembles, and their efficiency is estimated. The effects of optimization of the ELM neural network hyperparameters and postprocessing parameters are evaluated.

The figure below provides a simplified scheme of all calculations: it shows the stages, the used scripts and data structures.

Fig. 11. Structure and sequence of the main calculations in the article.

Author: Vladimir Perervenko

Aleksandr Masterskikh 2018.07.19 21:43 #1

Thanks to the author for the interesting work.

There is only a problem that has nothing to do with the author, but is a problem of analysis in general:

the concept of "noise" and "not noise" in the price dynamics of financial instruments is a very subjective thing, as in the methods available in the analytics industry,

there is no unambiguous definition of the concepts of "noise" and "trend" (for example, in the impulse equilibrium theory, this issue has been worked out at a new level).

This article shows some private solutions within the framework of traditional analytical approaches, but "with a twist" - noise sets, calculation of thresholds. Therefore - good work!

Discussion of article "Using Discussion of article "MTF Discussion of article "How

Vladimir Perervenko 2018.07.20 12:21 #2

Aleksandr Masterskikh:

Thanks to the author for the interesting work.

There is only a problem that has nothing to do with the author, but is a problem of analysis in general:

the concept of "noise" and "not noise" in the price dynamics of financial instruments is a very subjective thing, as in the methods available in the analytics industry,

there is no unambiguous definition of the concepts of "noise" and "trend" (for example, in the theory of impulse equilibrium, this issue has been worked out at a new level).

This article shows some private solutions within the framework of traditional analytical approaches, but "with a twist" - noise sets, calculation of thresholds. Therefore - good work!

I agree, jargonisms are often misleading. I specifically put the term "noise" examples in quotes and give a simplified definition of what is meant by it. The important thing is that this approach can yield positive results.

Good luck

Switch - declare variable Zero sample correlation does FIR filter with minimum

Evgeniy Zhdan 2018.07.20 13:15 #3

Very interesting material. Only it is not clear how much money all this makes and whether it makes money at all?

Vladimir Perervenko 2018.07.21 11:18 #4

Evgeniy Zhdan:
Very interesting material. Only it is not clear how much money all this makes and whether it makes money at all?

Check it out in practice. Everything you need for an expert in the article/yah is there.

Good luck

Yevhen Chystiukhin 2018.09.05 14:47 #5

Evgeniy Zhdan:
Very interesting material. Only it is not clear how much money all this makes and whether it makes money at all?

It makes money, but you need to be a bit of a programmer, not that everything will work the first time as it should ), and tests on a cent account before investing large sums of money

learn how to earn Electra EA ** Released Discussion of article "How

marzhinator 2018.10.20 15:19 #6

Vladimir, thank you very much for your wonderful articles!

Thanks to them, I have started learning R. Of course, for a "non-programmer" this article is not the right place to start programming and trading, but I'm already involved)))

I understand that I need to feed new data from the terminal into the block "#---test-aver--------". I thought about the GetThreshold function. It peeks at correct answers during tests to determine the optimal threshold of separation of continuous ensemble predictions.

Do you think it is necessary to use the thresholds obtained during training or recalculate them taking into account "combat predictions" minus the last one (there is no correct answer for it yet).

While I was going through these troubles, I came across the following peculiarity that if you redesign the cycle, you can get predictions several times faster. I think it will come in handy when testing the EA.