Machine learning in trading: theory, models, practice and algo-trading - page 411

 
elibrarius:

Here's the RNN experiment

https://www.mql5.com/ru/blogs/post/697984


P.S.
An ordinary neural network solved the XOR+ noise problem better, but with a non-zero error on the training examples. Although 0.1% error is ridiculous to consider as error)) So the regular MLP did better on unfamiliar data.

Average error on the training (75.0%) plot =0.001 (0.1%)
Average error on validation (25.0%) plot =0.034 (3.4%)

Training plot

0.00000000, 0.00000000, 0.00000000, 0.00106781
0.00000000, 0.00000000, 1.00000000, 0.99912817
0.00000000, 1.00000000, 0.00000000, 0.99887599
0.00000000, 1.00000000, 1.00000000, 0.00146612
1.00000000, 0.00000000, 0.00000000, 0.00015709
1.00000000, 0.00000000, 1.00000000, 0.99958727

Validation:
1.00000000, 1.00000000, 0.00000000, 1.00136207
1.00000000, 1.00000000, 1.00000000, 0.04871333

 
Alexander Ivanov:

I've been watching for a long time... waiting for the results.

And you're being clever and achieving nothing.

I have to make a NEURO-BOT.


You will not get much out of it. Maximum exchange of ideas, then everyone sits and implements it himself...
 
Mihail Marchukajtes:

There is no one to ask. The maximum is an exchange of ideas, and then everyone sits down and puts them into action...


This is what we have. But I need you to work here as in the workshop. To work together to create a bot.

Suppose I appoint Dmitrievsky as chief designer, he is highly trained )) and all of you will follow his orders and create his block diagrams.

That's how we built the atomic bomb in Russia - all together.

 
Alexander Ivanov:


That's what it is. But I need you to work here as a workshop. To create the bot together.

Suppose I appoint Dmitrievsky as chief designer, he is highly trained )) and all of you will follow his orders and create his block diagrams.

That's how we built the atomic bomb in Russia - all together.


Then what is your role????
 
Alexander Ivanov:


Question, how's it going with the neuro-robot? Any progress? Or here you are constantly getting clever and there are no achievements. ))

We should all take up the structure of a neuro-bot. Then maybe there will be a result.


I'm sick and tired of these neural networks.)

It's hard to say yet, I can't get 20000% every 3 months for at least 5 years, maximum 10-15% per month for such a period

 
Mihail Marchukajtes:

Then what is your role????


My role? - It's the demand!) Without demand all your dreams will collapse. Where there is demand and money, there is work. So far, I'm offering the DEMAND ))

Demand is a huge thing. Without it there is no market and no movement.

 
Maxim Dmitrievsky:


I'm just sick and tired of these neural networks)

It's hard to say yet, I can't get 20000% every 3 months for at least 5 years, maximum 10-15% per month for such a period


O))) So you already have a robot. And when will you give it to us at a "brotherly" price? ))
 
Alexander Ivanov:


my role? - is demand !)) Without demand all your dreams will collapse. Where there is demand and money, there is work going on. For now I suggest the demand.)

Demand is a huge thing. Without it there is no market and no movement.


All right, if you form demand, then form it in terms of money. How big is your demand???? And then the supply will appear...
 
elibrarius:

P.S.
An ordinary neural network solved the XOR+ noise problem better, but with a non-zero error on the training examples. Although an error of 0.1% is ridiculous to consider as an error)) So the regular MLP did better on unfamiliar data.

Average error on the training (75.0%) plot =0.001 (0.1%)
Average error on validation (25.0%) plot =0.034 (3.4%)

Training plot

0.00000000, 0.00000000, 0.00000000, 0.00106781
0.00000000, 0.00000000, 1.00000000, 0.99912817
0.00000000, 1.00000000, 0.00000000, 0.99887599
0.00000000, 1.00000000, 1.00000000, 0.00146612
1.00000000, 0.00000000, 0.00000000, 0.00015709
1.00000000, 0.00000000, 1.00000000, 0.99958727

Validation:
1.00000000, 1.00000000, 0.00000000, 1.00136207
1.00000000, 1.00000000, 1.00000000, 0.04871333


this is already interesting... it means it will be necessary to carry out additional tests, maybe there is no sense in this rnn, as it was thought before

The only advantage is the selection of weights in the optimizer simultaneously with other parameters of the system

 
Mihail Marchukajtes:

All right, if you form demand, then form it in terms of money. How big is your demand???? Then you'll see, and supply will appear...


At least I'm the only one asking. Monetary = well.... I don't know yet.