Advisors on neural networks, sharing experiences. - page 4

[Roman Shiredchenko]

https://www.mql5.com/ru/forum/8158/page10#comment_334926

On. Experience.
Read it all. Including the references.
Especially the last one.
https://www.mql5.com/ru/users/better
Neuro is out.
http://forex-pamm.com/

:-)

[Serqey Nikitin]

Roman Shiredchenko:
You have to be able to prepare the input data.

It's the same as identifying patterns in the market, in which case the grid is definitely not needed.

Still, training the grid in a particular area is a disadvantage, and it will have to be retrained periodically, despite all the cooking skills ...

[Maxim Romanov]

Maxim Dmitrievsky:

Yes, this seems to be called the curse of dimensionality :) In fact, in my case, a bunch of identical oscillators on the input, all this should be removed and one should be left.

Another question - when normalizing data for inputs, is it better to normalize all vectors simultaneously, in one cycle, considering max and min values from the whole set, or normalize for each input separately, considering max and min of each particular vector?

A non-network will not give an advantage unless you make it generate trading ideas. You have a bunch of indicators and they already have hard parameters that you have chosen. Think about what you want from the non-network? That you create a wrong theory and it corrects it?

A non-network must create a trading model by itself and be trained not only to trade but also to create models, check them and reject knowingly losing ones. This cannot be achieved by a simple network. Ideally it should create trading ideas in real time, And in real time weed out the wrong ones. and there should be several models traded simultaneously to create redundancy.

[valeriy odintsov]

There's a Hlaiman neural network lying around.

I am working on gold, which gives me a chronic loss.

[Andrey Dik]

I used to make neuronics-based EAs. And I dedicated several years of hard research to them.

My opinion is that neural networks are not suitable for financial markets.

There are much easier and more reliable ways to make money.

But as a hobby and an elective - why not...

[Maxim Dmitrievsky]

Roman Shiredchenko:
https://www.mql5.com/ru/forum/8158/page10#comment_334926

On. Experience.
Read it all. Including the references.
Especially the last one.
https://www.mql5.com/ru/users/better
Neuro is out.
http://forex-pamm.com/

:-)

Yeah, I've read it... lots of philosophy... few realizations... that is, people like to philosophize more than to do something real :) :) :) Just kidding, of course. But I would like to see something that would stably trade at + on neural networks. Without code and disclosure of algorithms, just as a motivating factor :)

[Maxim Dmitrievsky]

Joo Zepper:

I used to make neuronics-based EAs. And I dedicated several years of hard research to them.

My opinion is that neural networks are not suitable for financial markets.

There are much easier and more reliable ways to make money.

But as a hobby and an elective - why not...

What were the inputs, if it is not a secret?

[Maxim Dmitrievsky]

Maxim Romanov:
A non-network will not give an advantage until you make it generate trading ideas. You have a bunch of indicators and they already have hard parameters that you have chosen. Think about what you want from the non-network? That you create a wrong theory and it corrects it?

A non-network must create a trading model by itself; it must be trained not just to trade, but to create models, check them and weed out ones that are known to be unprofitable. This cannot be achieved on a simple network. Ideally it must create trading ideas in real time, and reject incorrect ones in real time.

My neural network works, yes. Well, minus, zero, it doesn't matter. The main thing is that it works. I look trades on the chart and what I see - it predicts the zigzag direction (or it tries to do so) and you can even see that sometimes the trades are very accurate. But the problem is that it does not create anything itself, but tries to do exactly what it is trained to do. And since the market itself is non-linear, it turns out that a non-linear neural network function tries to fight a non-linear market that has infinite freedom of various patterns. The patterns, sometimes, may be similar, but it does not follow that they predict the same consequences. (Or does it follow, from fractal theory??) Therefore, what specific pattern creation can we talk about here? We can only talk about maintaining some kind of dependence beyond the training sample that is found and used. But there is noise in the market, chaotic oscillations that build up and take the system out of equilibrium. And even the smallest deviations accumulate and completely change the system's behavior and there is nothing the neural network can do about it. And how can it sift something out in real time?

If we talk about modelling in general, there is a multifractal model of return on assets Mandelbrot described by Weierstrass-Mandelbrot function. I've been interested in this topic, the f-y constructs charts very similar to market charts. So, if the grid is trained using it in some way, perhaps it will learn to "catch" those very chaotic deviations and produce an interesting forecast. In other words, it will be a real market model. But it is all very complicated and imaginary, and I certainly will not do it.

[Maxim Dmitrievsky]

valeriy odintsov:

There's a Hlaiman neural network lying around.

I am working on gold, which gives me a chronic loss.

(Standard question, what have you tried to learn? ))

[Stanislav Korotky]

Have done some things in the past. You can read it on the blog under the forecasting tag.