Artificial neural networks.

[Dmitry]

Neural networks are IMHO the hardest part of algorithmic trading. There are no newbies here. I, too, only after 8 years of forex, having spent a lot of money, having tried all possible strategies and indicators and all possible combinations of them, have decided that this is the only trading strategy I have not touched. Why have I not tried it before? I thought it was too complicated and required special programs like "neurosolutions". It turns out I was wrong. The most important thing is to clearly understand how neural networks work, and the rest is a technical matter.

I invented my first prototype neuronet in about a day, wrote it in 15 minutes using pure MQL5, trained it using optimization. And oh my miracle, the results are encouraging... tank robustness is present for several years, tank tests are stable. I'm now almost completely convinced that if the Grail exists, it's hidden somewhere in the bowels of neural networks. Its potential is almost unlimited, I can add there any number of indicators with any number of parameters... and it can be done in pure MQL5. After thinking and trying them, all my previous efforts to create an Expert Advisor out of standard indicators appear to be child's play.

In this thread I propose to share my experience in the field of neural networks. It would be interesting to learn the principles of designing and training networks in MQL5. I will be waiting for your comments.

[Ilnur Khasanov]

Why start a topic all over again? There are already topics and people eating neural networks for breakfast, e.g. joo.

[Igor Makanu]

07041982:   

blah...blah...blah....

In this thread I propose to share experiences in the field of neural networks

start sharing - watch closely

[Dmitry]

IgorM:
start sharing - watch closely

That's the thing, there's not much experience, practically none, one simple neural network written in 15 minutes...

[Igor Makanu]

07041982: That's the point is that I have little or no experience, I have only one simple neural network which I wrote in 15 minutes...

experience with NS is not important (I'm not pointing fingers, but Mr. R. has been dealing with NS for a long time and after watching his posts for a long time I came to the conclusion that time and experience are not important - it's the case that counts...)

OMG, your Grail code will be posted and we'll try to work it out in the forum.

I've got a lot of ready-to-use mql4 TS, but i've come to the conclusion that it's no better than standard indicators.

[Vladimir Kazakov]

"Explain to me, please, why it is necessary to artificially fabricate Spinoza when any woman can give birth to it whenever she wants..." (c) ;)

[secret]

07041982:

In this thread I propose to share my experience in the field of neural networks. It would be interesting to learn the principles of building and training networks in MQL5. I'm looking forward to your comments.

Try to read both forums and articles as part of "learning the principles", maybe you won't need this thread.

[Dmitry]

I don't want to post the code, because I don't want any unnecessary flub about my peculiar style of writing code without using standard classes and libraries, etc. It's difficult to understand it, there are a lot of variables. I'll explain it better in a nutshell.

I've taken ten neurons; the first neuron calculates what is higher the MA or the current price, the second one - the MA on the previous bar or the price on the previous bar, etc. for all ten neurons. Each neuron decides whether price will go up or down and all 10 opinions are summed up to give us an answer, e.g. price will go up with 90% probability (9 out of 10 neurons voted for it and 1 against). Each neuron contains a condition: if price is below the MA it outputs 1 or -1, which depends on parameters being optimized. A 1 means that it votes for the price going up and a -1 for the price going down. Optimization is the process of training the neural network. That's it.

[Igor Makanu]

07041982:

I don't want to post the code, because I don't want any unnecessary flub about my peculiar style of writing code without using standard classes and libraries, etc. It's difficult to understand it, there are a lot of variables. I'll explain it better in a nutshell.

Maybe you can rewrite your TS with the new articlehttps://www.mql5.com/ru/articles/497 , the code there is not complicated, it should suit your purpose, it seems you also have a committee of networks of one neuron

But if we discuss it, we should discuss the subject and not the topic of the thread, since there is no subject to discuss.

[Dmitry]

IgorM:

Then maybe you can rewrite your TS using the new articlehttps://www.mql5.com/ru/articles/497 , the code there is not complicated, it should suit your purposes, it seems that you also have a committee of networks from a single neuron

But if we discuss it, we should discuss the subject and not the topic of the thread, since there is no subject to discuss.

I understand that this is a forum for programmers not traders, IMHO it's more interesting to discuss the trading strategy than examples of its implementation in a programming language. What is interesting about discussing the bare code?

[Andrey Dik]

aharata:
So why start the topic all over again? There are already threads and people eating neural nets for breakfast, like joo.

Thanks, but there are a lot of people on this forum who understand grids much better than I do. Would be more than happy to read them here.

And in general, a good understanding of the structure and properties of networks is not enough to create a profitable in real life (rather than on stories) system. This requires original ideas in the selection of inputs for the grid, as for me - I'm still searching.