Neural network (deep learning) EA & some manual trades

Hi,
I don't want to compete with what is done here, rather curious what others are doing to compare it with my own ideas. Above all my primary goal was to learn Data Science using trading to be able to practice and then apply learnings in real life applications...

In summary I am using Deep learning framework of h2o from R. I am just using deep learning function for that. Process is not super integrated with MT4/MQL (I can't do backtest in MT4) but I can do some tests in R to verify model.

The idea is explained below:

1. generate financial data from MT4 and write it to the files
2. read data in R, pre-process it to train model (I am using indicator pattern as predictor and shifted price change as label)
3. train model to predict either price change direction (classification) or in absolute term (regression)
4. test model directly in R using independent dataset
5. use model for predictions (typically to predict what will happen in 2, 20, 200 hours)
6. write predictions to the files to be consumed by Expert Advisor

I have tested this system and the results are very interesting. Unfortunately rules are so much restricted here: I have no possibility yet to post my own threads but I hope to read these threads more in detail one day

Your on the correct path , 2018 champion was a neural network , cleaned up all the rule based systems.

Keep motivated orangutangas

I too am running a neural network based robot, built using Googles Tensorflow Neural network toolbox with python connecting it to MT4.

Having good results and feeling optimistic about its performance.

Trained on 10 years worth of daily data, and retrained each week to react to new market movements.

So the best results are obtained using predictions on the daily charts?

Sorry for asking such details - in my case I can generate predictions for next hour, day or week; but I am not really sure which predictions to use exactly. My initial idea was to use long-term prediction, confirm it with medium and short terms, however I afraid that using multiple models may 'multiply' probabilities of errors. What I found most relevant is to 'filter' entries by entering the trade only when predictions is more than XX pips...

Is this the right approach?

Thanks

How to create AI system using mt4 or connect with ai system (like python ?)

This is a good question!

I have created one and even teaching how to do that. However I am not sure about the rules here so I don't want to be banned from this forum by moderators

My approximate path [using R and h2o] was:

MT4 side:

1. learning intuition to code in MT4, knowing how to debug, write functions, etc
2. learning how to read/write files from EA to the sandbox [side note1] create automated periodic data update

R/h2o side:

1. have a valid idea of what is needed to achieve (e.g. predict future price change after X bars by using Y bars in the past and certain bunch of historical data)
2. knowing basic idea on how deep learning regression modelling works some hints here: https://github.com/h2oai/h2o-tutoria...ning/README.md
3. learning basic data manupulation 'tidyverse' etc to prepare data for Deep Learning - probably not a trivial task!
4. training deep learning and saving the model, most likely this will be very iteractive process finding grid search, diffent parameters... CPU will be hot
5. have an idea on how to test your model on independent data (sounds hard to achieve but for me it was an easy journey)
6. once model selected one will need to save it and periodically test it or challenge it by trying to find new model (see point above)
7. easiest part is to use the model to predict using fresh data, I am writing files to the sandbox from R see [side note1]

MT4 side:

1. find a logical way to use predictions in Expert Advisor template. In my case I am still working on that by live testing as there are very limited back testing possibilities see because of [side note1]

Probably there are other approaches including those mentioned in the forums however I am choosing this one because of much more user-friendly coding environment [R]. Above all this approach has came to me naturally because I am using R/h2o for work so any time effort spent on this project was all very interesting learning experience

[side note1] - I don't have enough time/skills to learn DLL or other ways to learn how to do advanced connections

Hello. I have published in MQL5, an evolution of this type of Eas, which is more evolved, the recurrent neural network, there you will see how it works:
Holy Grail Deep Neural Network Recurrent, (for Metatrader 4).

Hello. I have published version II, in MQL5 of the Ea based on recurrent neural networks, if someone is interested in testing it, publishing their option, it would help me a lot to improve it. Thanks

If you have time, to have fun for a while, do a backtestig of this Ea, in an aggressive way from May 2003 to date, I tell you this date because from here I began to learn the Ea. (4 hours, EurUds), initial box = 300 dollars, laverage = 1/100. (tick to tick 99%) Thanks.

and after all this years how is going your trading Orangus? i'm curious about your story,
wish ya the best

Googled two research papers attached here. Hope this thread will summon academic professionals to explore more