NeuroTrend v2
 |
| От Indicators |
This is the NeuroTrend v2 files. It contains:
Installation instructions
Description
Neural net indicator
A mqh gile that goes to the Include folder
A neural net that needed to be copied into the files folder (It is quite interesting to see that really looks like a neural net model: a set of weights)
There is a description how to install the indicators. I am posting it because it is possible to make an EA based on those ideas.
As the code is open source this can be simplified and different inputs can be selected.
The neural net software that is used is the JAVANNS. This is a great neural net simulator from the University of Tubingen.
Comments:
The neural net indicator works but you need to zoom in to the maximum in order to see the prediction.
Training:
There are three processes:
1. Exporting data into a set file
There is a script for doing that.
What is realy interesting is that collection is the data collection script. With that script you are collecting data and tranforming it into the appropriate input format that is needed by JAVA NNS to make the training.
I think that in the beginning of the file the correct input number need to be setted otherwise it can't work.
2. Neural net training
As stated you need to install the JAVA NNS.
 |
| От EA tests |
In the Java NNS when you make the architecture your input neurones need to match the number of input neurones that are coming from the data collection script.
The example there is using 17 inputs including inputs from the 1 h time frame from different indicators. You can refer to the description.
Unfortunately the last use I did of that software is two years ago, so I forgot what exactly I did. As far as I remember I used two sets one for training and one for testing.
You can experiment with different neural nets and playingas much as you want with the neural net internal parameters.
3. Importing the model
This is the easiest the neural net will be saved into a file and that file just need to be copied into the expert/files folder and that is all.
I think that this is a remarquable work.
Comments
Thanks for sharing.....nice work but works exclusively on 15 min chart.
No it is not exclusively for 15 m chart. It is possible to change the code and make it appear on different time frames.
However as it is a neural net trained exclusively on 15 m time chart it makes no sense at all to appear (make predictions) on different time frames.
That is why its creator blocked it on 15 m time frame.
Yes. A good programmer Arun Kumar. His website http://staygain.com/
Arunsideas on FXFactory.
The idea is to simplify all this and use other imputs. The neural net architecture uses too many inputs (I think so).
I tried it on 15 min live chart and I think it repaints and the cross appreas 2 bars in past .....still testing though....The indicator does not appear in strategy tester so have to watch it live.
The indi mixes up some inputs with multitime frame. That may be a reason of the problem.
And bear in mind that it has not been optimized on recent data. I would eraze most of the inputs and let 3 or 4 oscillators as inputs.
The problem of this is that for every change in the neural net architecture you need t ochange a lot of things. When I have ore time I would play with that code a little bit.
But this repainting issue as you say bothers me a lot.
that's why perceptrons are interesting for removing the repainting
I do not see very enthousiasm here for making an EA out of it LOL. Anyway I am not very hot too, however I think I needed to post it because it is an important alternative for Neural Net expert trained with high quality Neural Net simulator that most of it is free of charge.