Machine learning in trading: theory, models, practice and algo-trading - page 347

[Yuriy Asaulenko]

elibrarius:

i.e. we need =RNN(a0,a1,a2,a3);

I do not understand whether RNN is RNN, or ReshetovNNN (RNN), which is not RNN.

Now it's about one, now it's about the other. Already can not understand in what context.

[Forester]

Yuriy Asaulenko:

I do not understand, RNN is RNN, or ReshetovNNN (RNN), which is not RNN.

It's about one or the other. I already do not understand in what context.

Judging by the code there is no memory in the literal sense. The indicator data for older bars are just fed to the inputs, for example bars 0,1,2,3 or 0,2,4,8 or 0,10,20,30 - as you choose. So the memory is not in the NS itself, but from an external source.

[Yuriy Asaulenko]

elibrarius:
Judging by the code, there is no memory there in the literal sense. The inputs are just data from the indicator for older bars, for example bars 0,1,2,3 or 0,2,4,8 or 0,10,20,30 - whatever you choose. So the memory is not in the NS itself, but from an external source

And this is not RNN. RNN is only with feedback - with recursion. Someone needs to be renamed.)

And, strictly speaking, it's not a network, but a single neuron.

[Forester]

Yuriy Asaulenko:

And this is no longer RNN. RNN is only with feedback - with recursion. Someone needs to be renamed)).

And, strictly speaking, it's not a network, but a single neuron.

Add another one, and then feed the outputs of these 2 to the 3rd - there's the grid.) Only the weights can't be optimized, if the 1st neuron has 100 inputs, the second has 100 inputs, and the third has 2.

Here it would be better not to calculate from all 202 inputs, but exactly from training commands... In other words, don't calculate balance optimization in optimizer, but go through e.g. 1000 training points and calculate coefficients from them.

[Yuriy Asaulenko]

elibrarius:
Add another one of these, and then feed the outputs of these 2 to the 3rd - that's the grid) Only the weights will be unrealistic if the 1st neuron has 100 inputs, the second has 100 inputs, and the third has 2.

Reshetov's solution is interesting. Something similar is probably going to be done. Only from the context you often do not understand what we are talking about - which RNN.

[Maxim Dmitrievsky]

elibrarius:

Because there is an error in RNN3 -
data is requested for 5 points, while the probability is calculated for 4 points, and you don't start with the current bar a0, but with an offset for a period, i.e. with a1. Probability is calculated not for the current moment, but for a1 point - hence the results are bad for a0))

// Преобразование показаний технического осциллятора RSI в диапазон от 0 до 1 включительно
   double a0 = rsi[0] / 100.0;
   double a1 = rsi[p] / 100.0;
   double a2 = rsi[2*p] / 100.0;
   double a3 = rsi[3*p] / 100.0;
   double a4 = rsi[4*p] / 100.0;



// Вычисление вероятности
   double results=RNN(a1,a2,a3,a4);

i.e. we need =RNN(a0,a1,a2,a3);

Well, in order to count for 5 points, you should also increase the log matrix to 25 rules, i.e.⁵². And so on, if more points/inputs are needed. If there are 10 inputs, it means 100 input variables ))))). I wonder, is MT5 able to optimize it?

Nope, that's his original as well, I didn't trade it.

For 10 entries it would be already problematic to count even through the cloud) But I'll try to make 3 such expert systems, fed to the input 4 ) If the opening prices are not tested for a very long period then fine

[Maxim Dmitrievsky]

Yuriy Asaulenko:
Reshetov has an interesting solution. Something similar is probably going to be done. Only from the context you often do not understand what we are talking about - what RNN.

It's not RNN, it's an expert system, as he calls it... why it is called RNN I do not know, maybe by the last name :)

[Yuriy Asaulenko]

Maxim Dmitrievsky:

Nope, this is also his original, I did not trade

For 10 entries it would already be a problem to count even through the cloud) But I'll try to make three of these expert systems, submitted for the entry of the 4th ) If the opening prices are not tested for a very long period, then fine

What if we divide it into frames with a small overlap? We obtain 2-3 partially overlapping Expert Advisors and then merge them.

[Maxim Dmitrievsky]

Yuriy Asaulenko:
What if we divide it into frames with a small overlap? We get 2-3 partially overlapping experts, and then we merge them.

as a variant, yes

[Forester]

I think it is unrealistic to calculate something profitable with only 3 to 5 inputs using such a matrix. I agree that it covers all possible variations.

But if for example we make a network with 5 inputs, it will be 32 coefficients for calculations. Genetic algorithm usually converges in 10,000 passes, i.e. inputs will be enumerated as -1.0-1 on average.
With 3 inputs maybe a pattern can be calculated, but 3 inputs is not enough, in my opinion.

A neural network in R or even from ALGLIB you can build any and quickly calculate them. Their internal structure will not be so complete, but the strongest dependencies will be found during training.