Machine learning in trading: theory, models, practice and algo-trading - page 1426

 
Ivan Butko:

Is the neural network a fit with history? Or it is necessary only to change something (to poke where), that something similar to profitable went.

The neural network performed only the task you gave it, if you wanted to get a "market formula" - it calculated this formula for you on historical data

Keep reading Wiki

Imho, the question is not about what you were poking and poking at, and what you were planning to poke and where.

 
Igor Makanu:

The neural network performed only the task you gave it, if you wanted to get a "market formula" - it calculated this formula for you on historical data

Keep reading here on Wiki.

I.e., instead of 3 layers with 20 neurons each (as in example) it makes no sense to use 10 layers with 100 neurons each? I already wanted to load my kor i7 today.

 
Ivan Butko:

So, it makes no sense to use 10 layers with 100 neurons each instead of 3 layers with 20 neurons each (as in the example)? I already wanted to load my cor i7 today

Here you have asked a specific question, but alas there is no specific answer, I've watched over a hundred videos about NS and googled this question specifically - the configuration of NS is a matter of chance, there are no specific recommendations - just general principles

 
Vizard_:

The misunderstood Teachers continued to dive... over the cuckoo's nest...


Are there any strategies without drawdowns?

Another thing is that Maxim's method implies constant retraining, so it is interesting to see how the chart will crawl if they retrain every week.

 
Ivan Butko:

Can you tell me what's wrong with the neural network? I read on local site about porting neural network from NeuroPro program to Expert Advisor. I did everything as described and got the same result

...
Although, my Expert Advisor has no trading logic, except for the set of hundreds of real numbers - neuronet weights and summing formula.

In my opinion, this is the problem - there is no logic, which means complete freedom for the best approximation to the price history. For a stationary process this would be the right approach, but in our case it is not complete, and therefore not stationary. Use predictors with ideas - in my opinion this is more correct.

By the way, how many input neurons and layers with neurons on each layer in the network?

 
Ivan Butko:

So, it makes no sense to use 10 layers with 100 neurons each instead of 3 layers with 20 neurons each (as in the example)? I already wanted to load my cor i7 today.

NS is the same fit as an optimizer, even to a greater extent.
The number of layers and neurons in them is determined by the specific task.
Without any theoretical background, it's better not to work on NS.
 
Aleksey Vyazmikin:

Are there any strategies without drawdowns?

Another thing is that Maxim's method involves constant retraining, so it's interesting to see how the chart will crawl if I retrain every week.

I purposely did not retrain for the 2nd week, so I would "see" what would happen.

So far, nothing much. And no one wrote that it's a grail, it's a cool fast learning machine that works well on "good" pieces of the chart

Right now I'm rewriting new features from python, it's pretty complicated, but they should improve TC. I'll write a separate article about them in MB.


 
Grail:

What about the dataset?

I have akurasi 57.3% loglos 0.688

https://c.mql5.com/3/275/market66.zip

Accuracy:65% Logloss:0.61382

ZY pure dataset is not cool, it would be good with every thing bid\ask so you could run it on tester, well and semantics of output values, return, vola, etc.

 
Yuriy Asaulenko:
NS is the same as an optimizer, even to a greater extent.
The number of layers and neurons in them is determined by the specific task.
Without any theoretical training it is better not to deal with NS.

Unfortunately, many people don't understand this and continue to deal with it. And to understand it is so simple and obvious.

No MO or neural networks are useful in trading.

 
Ivan Butko:

Can you tell me what is wrong with the neural network? I read the article on local site about neuronet transfer from NeuroPro into Expert Advisor. I did everything as described and got the same result

But as soon as I get to the forward or vice versa, as soon as I go beyond the specified period, pure chaos begins, not a single sign of life, profits either immediately go to the bottom or fly off, without any signs of systematic profit increase, at least a little.

Is the neural net a fit to the story? Or do I just need to change something (poke where) to make it look like a profitable one. But then what exactly needs to be changed (within the program or maybe a different program to use).

It's just extremely surprising, there is no intentional tweaking in the algorithm, as it happens in tester grails, that is, the EA does not look into the future, but the funny thing is that when you set "do not open trade if forecast less than 50 points" in the settings, then on retesting ... The only trades left are those opened right before the very news candlesticks, right in the desired direction and of the desired profit size (the entire momentum). However, there's no trading logic in the Expert Advisor except for the set of hundreds of real numbers - weights of the neural network and summing formula.

It turns out that you have found the Holy Grail everyone is talking about?

Just don't forget to share the profit with Victor Genadievich Tsaregorodtsev(creator of NeuroPro), they say he lives in poverty in the city of ecological disaster Krasnoyarsk, making a living by tutoring at $ 5-10 per hour and writing essays.

NeuroPro - нейросети, анализ данных, прогнозирование и классификация
  • neuropro.ru
Это сайт Виктора Царегородцева − здесь я предлагаю услуги по анализу данных, созданию интеллектуальных компьютерных программ для решения разнообразных задач обработки данных, изображений и видео, а также описываю результаты моих научных исследований по теории и практике изучения и применения нейросетей. Круг задач обработки информации...