Machine learning in trading: theory, models, practice and algo-trading - page 348

 
elibrarius:

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 are 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. Internal structure will be not so complete, but in training you will find the strongest dependencies.


Do not forget about combined strategies, where the NS may perform only a part of inputs, for example, to show the general direction, while signals, for example, are given by another system

Suppose there is a system, that pours in a flat, you can optimize NS to filter such areas, and the rest of the logic will work "as is".

 
Maxim Dmitrievsky:


Do not forget about the combined strategies, where the NS may perform only part of the inputs, for example, to show the general direction, while signals, for example, are given by another system

Suppose there is a system that pours in a flat, you can optimize the NS to filter out such sections, and the rest of the logic will work "as is".


And, by the way, here is the same Reshetov espert system, in other words :) https://www.mql5.com/ru/articles/3264 i.e. we can call it a Bayesian classifier, apparently
Наивный байесовский классификатор для сигналов набора индикаторов
Наивный байесовский классификатор для сигналов набора индикаторов
  • 2017.05.12
  • Stanislav Korotky
  • www.mql5.com
В статье анализируется применение формулы Байеса для повышения надежности торговых систем за счет использования сигналов нескольких независимых индикаторов. Теоретические расчеты проверяются с помощью простого универсального эксперта, настраиваемого для работы с произвольными индикаторами.
 

What makes you think that NS should work at all? As far as I know, NS make it easier to find an algorithm/lawfulness in a data set.

But price behavior is the sum of the positions of all participants, including MM. What kind of algorithm can there be in the behavior of the crowd? And some part of the crowd doesn't care whether the position will be profitable or not. This is the behavior of the bee that flies over the fields. The fields are the same, but it is impossible to predict which flower will sit on.

 

Let's wish luck to this monster :) It is simply to determine the limits of applicability of such an approach


 

The more neurons and inputs the more stable but less profitable system, I optimized for the last 3 months by opening prices on the minutes, 1.5 of which is a forward, then I ran it for almost a year and it showed a stable result. I have 3 neurons for each of 3 inputs and these 3 neurons enter the 4th one that gives me the final result

In the circle is the area in which the grid was optimized (approximately), followed by the forward, and the rest of the PO was not involved in the training in any way


 
Maxim Dmitrievsky:

The more neurons and inputs the more stable but less profitable system, I optimized for the last 3 months by opening prices on the minutes, 1.5 of which is a forward, then I ran it for almost a year and it showed a stable result. I have 3 neurons for each of 3 inputs and these 3 neurons enter the 4th one that gives me the final result

Here is a circle with a segment, on which the mesh has been optimized (approximately), followed by the forward, and the rest of the BC was not involved in the learning process in any way



The main thing is stability. In less than a year - 800% and if it is true some sort of a self-learning Expert Advisor on arrays similar to a neural network - I shake your hand. It's too clever for me to understand what's in it, but I shake your hand for daring to dive into this field of machine learning. I think it will stumble for the same reason - unpredictability of the market, but you apparently have a system of loss constraints there, so it's really interesting. And where does it run on a VPS or on a home PC?
 
geratdc:

The main thing is stability. In less than a year - 800% and if it really is some kind of self-training advisor on arrays similar to a neural network - I shake your hand. It's too clever for me to understand what's in it, but I shake your hand for daring to dive into this field of machine learning. I think he's going to stumble for the same reason - unpredictability of the market, but you apparently have a loss constraint system there, so that's really interesting. And where does it run on VPS or on your home PC?

Yes it's tests in the tester) It's not even a neural network but a classifier, something in the middle, I do not know how to call it... handmade ) Yes, i need to retrain periodically and to introduce some limits, for example on drawdown
 
Maxim Dmitrievsky:

The more neurons and inputs the more stable but less profitable system, I optimized for the last 3 months by opening prices on the minutes, 1.5 of which is a forward, then I ran it for almost a year and it showed a stable result. I have 3 neurons for each of 3 inputs and these 3 neurons enter the 4th one that gives me the final result

Here is a circle with a segment, on which the mesh has been optimized (approximately), followed by the forward, and the rest of the BC was not involved in the learning process in any way


Not bad!
What do you feed to the inputs?
 
elibrarius:
Not bad!
What do you feed to the inputs?

The same as before, regression and rsi, I haven't thought of anything more clever yet
 
elibrarius:
Not bad!
And what do you feed to the inputs?

By the way, you were looking for the most convenient grid - try this one https://www.mql5.com/ru/code/9002

I haven't got into it myself yet, please tell me if it's usable or not, if I don't manage to do it myself)

Ценовой прогноз с использованием нейронных сетей
Ценовой прогноз с использованием нейронных сетей
  • votes: 14
  • 2016.06.14
  • Vladimir
  • www.mql5.com
Индикатор, который использует нейронные сети для прогнозирования ближайших нескольких цен открытия. Сеть обучается методом обратного распространения ошибки. Обучение проходит автоматически, результат — самообученная сеть и самообучающийся индикатор.
Reason: