How to build a EA using LSTM model?

 
Is it able to build a EA using LSTM RNN model in Metatrader? 

I have done some research and found something called FANN, however it seems to work with simple neural network only. 
 
What do you mean by powerful enough ?
 
Marco vd Heijden:
What do you mean by powerful enough ?
Whether the language supports
 
Supports what.
 
Marco vd Heijden:
Supports what.
so can one actually build a EA using lstm in mt4 
 
luckyvictor: Is it able to build a EA using LSTM RNN model in Metatrader? 

I have done some research and found something called FANN, however it seems to work with simple neural network only.

so can one actually build a EA using lstm in mt4 

Short Answer:

Yes

Long Answer:

Only if you are a very experienced and knowledgeable coder. There are no libraries or open source code for it in MQL. You will have to develop it yourself.

Observation:

Given that you are asking the question, it is obvious that you have limited knowledge of the MetaTrader environment and coding in MQL, and therefore it would be highly improbably that you would be able to create such an EA in the short or even medium term.

Also, judging from the quality displayed by the vast majority of coders available for hire in the Freelance section, you will also have a tough time being able to farm out the development of such an EA and even if you do find a suitable coder to help you, it will not be cheap.

Suggestion:

Keep it simple! Don't complicate things. Highly complex EAs are not a guarantee to achieve more profitability. In fact, many times it can actually make it worst.

 
luckyvictor:
Is it able to build a EA using LSTM RNN model in Metatrader? 

I have done some research and found something called FANN, however it seems to work with simple neural network only. 

In the context of EA building, LSTM is a recurrent model for forecasting the market based on mixing short and long term in process memory.

You already have results on the availability of such types of memory in market price series and their relationship to each other?

If not, then it's better to start with simple neural networks, for example is may be FANN and conduct their model to train on short then long historical data periods to check separately the corresponding types of process memory.

It should be serious and massive experiments and I'm also doing it now in the russian part of the forum, if you want, can to join.

https://www.mql5.com/ru/forum/270216

Машинное обучение роботов
Машинное обучение роботов
  • 2018.08.02
  • www.mql5.com
Привет всем, я занимаюсь машинным обучением (МО) советников и индикаторов и решил вынести на всеобщее обсуждение свои эксперименты...
 

Fernando Carreiro:

Short Answer:

Yes

Long Answer:

Only if you are a very experienced and knowledgeable coder. There are no libraries or open source code for it in MQL. You will have to develop it yourself.

Observation:

Given that you are asking the question, it is obvious that you have limited knowledge of the MetaTrader environment and coding in MQL, and therefore it would be highly improbably that you would be able to create such an EA in the short or even medium term.

Also, judging from the quality displayed by the vast majority of coders available for hire in the Freelance section, you will also have a tough time being able to farm out the development of such an EA and even if you do find a suitable coder to help you, it will not be cheap.

Suggestion:

Keep it simple! Don't complicate things. Highly complex EAs are not a guarantee to achieve more profitability. In fact, many times it can actually make it worst.

Hi Fernando


Your answer is really what I expected to be honest, and I agree with your observation.


I coded a few indicators and EAs, using normal strategies and tweak their entry and exit. However I consider them to be static, they dont "understand" the dynamic of the market, so performance is not good. Therefore I am hoping that the power of neural network can make it dynamic, especially LSTM, which is particularly good for time series data, can help to filter out bad signal.


I guess the workaround is to implement it in python then interface to MT4.

 
Ivan Negreshniy:

In the context of EA building, LSTM is a recurrent model for forecasting the market based on mixing short and long term in process memory.

You already have results on the availability of such types of memory in market price series and their relationship to each other?

If not, then it's better to start with simple neural networks, for example is may be FANN and conduct their model to train on short then long historical data periods to check separately the corresponding types of process memory.

It should be serious and massive experiments and I'm also doing it now in the russian part of the forum, if you want, can to join.

https://www.mql5.com/ru/forum/270216

I dont have the result of their relationship yet, and to be honest, i dont know how to view this relationship from the output of LSTM yet.


Have you worked with FANN? How is ur experience with it? Does it generally give a boost to your strategy? 

 
luckyvictor:

I dont have the result of their relationship yet, and to be honest, i dont know how to view this relationship from the output of LSTM yet.


Have you worked with FANN? How is ur experience with it? Does it generally give a boost to your strategy? 

FANN is a good and fast library, but I prefer to use my unique development P-NET.

My network, for example, in comparative tests with TensorFlow DNN, conducted on python scripts, works out similar results a thousand times faster.

To create EA's, I also use this network together with other machine learning models as part of my automatic EA generator. https://www.youtube.com/watch?v=NXrQ0KXWO5o

In the forum thread about which I wrote above there are examples of ready-made advisers, you can drop own template there and I will generate an EA for you to try.

Testing the artificial neural network P-net, included in the Expert Advisor for the Forex market
Testing the artificial neural network P-net, included in the Expert Advisor for the Forex market
  • 2018.04.02
  • www.youtube.com
The purpose of this video is to show the advantages of the new artificial neural network of the P-Net type, compared to the multi-layer artificial neural net...
 
The fall of RNN / LSTM – Towards Data Science
The fall of RNN / LSTM – Towards Data Science
  • 2018.04.13
  • Eugenio Culurciello
  • towardsdatascience.com
We fell for Recurrent neural networks (RNN), Long-short term memory (LSTM), and all their variants. Now it is time to drop them!