Build Neural Network Indicator in MT4 using Neuroshell

Hi,

I found only up to version 4.5 you can get it free..search using google to get it. Last version is 10.x.x.

We can only generate any C code with enterprise version cost $3000 (x32) and $6000 (x64). Not sure if it is worth-ed :(


Hi,

Anyone has DTReg to share? ;-)

Thanks
 

Attachments

  • DTREG4.5.png
    DTREG4.5.png
    59.1 KB · Views: 709
  • dtreg version.png
    dtreg version.png
    91.9 KB · Views: 1,099
I've tried with Google but I had no chances...just old torrents links which didn’t work.

Maybe you could send a private email, with the web address...that would be super nice.

Thanks
Claudio
 
Hi Claudio,

I will find first the address, it has been downloaded few months a go, I will let you know if I found where it has been downloaded.

Here implementation sample of CH model in forex (just trying and may not profitable).
 

Attachments

  • EURUSD CH Model.pdf
    145.3 KB · Views: 779
different day - different data

Long time not to post here, too busy with training my daughter in tennis.
But maybe this comment can be useful for somebody.

The major mistake in my opinion what forum people are doing using NN for trading is to trying it on different data sets (different TFs, pairs and time ranges)
In this case is impossible to compare performance of different algos. In my case I applying like 10 algos to the same data set than trying to train on different number of bars, changing SL/TP and comparing results.

See example here trading date 2011.04.18 EURUSD from Dukas 1 min

Confusion_EURUSD1dukas_2011.04.18_ALL_PegasosLR__IS_bars=20000_OOS_bars=1440_PP=40.8377_SL=15_TP=30_AC=0.40903_Profit=86400_epoch=5_instantPip170

looks great - 86k earned

than another day 2011.09.15 - 54k loss

Confusion_EURUSD1dukas_2011.09.15_ALL_PegasosLR__IS_bars=20000_OOS_bars=1440_PP=29.7662_SL=15_TP=30_AC=0.36458_Profit=-54194_epoch=5_instantPip170

in both cases methodology was exactly the same. In this case algo used was very fast SVM like solver Pegasos.

In both cases behaviour of other algos is similar - either all (with some exceptions) they perform good or bad depends of the data. So in order to be able to compare apples to apples data set must be fixed. This is the way how AI people are working i.e have certain data sets(e.g MNIST) and trying to develop better algo.

So I believe data from 2011.09.15 is very algo unfriendly.....

Krzysztof
 

Attachments

  • pegasos.pdf
    360.8 KB · Views: 1,305
This is my personal conclusion how to implement neural network in MT4.....

Time to make complete ebook .... I think :)
 

Attachments

  • Implement Neural Network in MT4.pdf
    57.1 KB · Views: 3,699
I've tried with Google but I had no chances...just old torrents links which didn’t work.

Maybe you could send a private email, with the web address...that would be super nice.

Thanks
Claudio

Why do you like this programme ?
 
Hi and thanks...

There is no special reason why, I just liked the idea of having different “self-adjusting” algorithms.

I've seen for realtime application that you can make predictions with a simple cmd line, that you can call from MT4 (pag 28 of the manual)
 
Hi and thanks...

There is no special reason why, I just liked the idea of having different “self-adjusting” algorithms.

I've seen for realtime application that you can make predictions with a simple cmd line, that you can call from MT4 (pag 28 of the manual)

Hi and thanks for the reply.
Have you backtested the programme ?
I found a NN prog called Tiberius that I have been analysing with the S&P 500. There's a free trial for 30 days.
It correctly predicted the direction for the 15th & 16th but not the 17th of Feb 2012.
 
Hi Pat,

Someone can break the tiberius.xls code, then customize it...:cool:

Cheers
 

Attachments

  • TiberiusXL VB code break.png
    TiberiusXL VB code break.png
    95.4 KB · Views: 688
Yes working with excel is limited, I found other software that look better than tiberius, even only limited net configuration seems good as a starting point. Have a look to this link
Neural Network And Genetic Algorithm Software

I tried it can handle wider amount of data, three training methods (BP with momentum, Conjugate gradient or GA). As the result we can get the network data (weight) for deployment purpose. Here is the screenshot.

But I think you can not expect more for a cheaper software..:(
 

Attachments

  • Sample 4CastXL.pdf
    1.4 MB · Views: 746
...a nice reading on the subject. :confused:
 

Attachments

  • The Million Model Test.pdf
    679.5 KB · Views: 3,594
Development update..try to implement neural network indicators on Forex pair ...:)
 

Attachments

  • Neural Network Indicators Implementation on Forex Pairs.pdf
    232 KB · Views: 1,073
Here is a related paper.
http://math.nyu.edu/faculty/avellane/AvellanedaLeeStatArb071108.pdf

They propose a portfolio allocation model based on the Ornstein-Ulenbek process. They do a cross-sectional PCA on the instrument of the portfolio. What I do is different; I use CSSA on a single instrument. It is equivalent to doing an incremental cross-sectional PCA but on lagged versions of itself. Anyway, don’t pay too much attention to this stuff. My system is very simple:

The Ornstein-Ulenbek is like that: dX(t) = k (m – X(t))dt + sdW(t) (see p16 of the pdf)

My system is:
1. Do dX(t) = CSSA(price) – detrended price
2. Use a neural net to predict price and do k(m–X(t))dt = prediction

Use the method described p16 and 17 to calculate the parameters k m and s of the model. When done, calculate the s-score as described p20; s-score = (X – m) / sigma. You are almost done.

Use the s-score to generate signals:
s-score > threshold1 go short
s-score < threshold2 go long

that’s it!!

Isn´t this process the normal calculation of the Z-score, just with different inputs?
Like
z-Score = ("actual value" - "mean of value") / "Standard deviation of value"?

I think your approach is really interesting. Do you detrend the price by calculating the logarithm of the price?
 
This is the comparison result the NS2 neural network configuration made in Matlab, with some consideration:
- all transfer function in NS2 may not available yet in matlab, such as tanh15, symmetric logistic, gaussian, etc.
- Any single link in NS2 contains weight and bias, while in matlab we will have only a single bias in a layer. Means if you have two link into a slab then you will have 2 weight and 2 bias in NS2 but in matlab you will have 2 weight and 1 bias.
- Slab 1 in NS2 shall be considered as input normalization, then Slab N in NS2 equal to Layer N-1 in Matlab.

Please consider for educational purpose only, no intention in advising to use any software, technology, or trading ..you are responsible to your decision:D

(y) "this file is password protected" he he he thx
 
Thank you arryex and others for very interesting thread.

How do you use chaoshunter model together with neural net? do you use a committee or one as the filter for the other or something else?

I do not get how you get such great results with ch. The training period equity curve is always very smooth but after that it is usually down or choppy.
 
and another question, what advantage is there to predicting indicators (macd and rsx) instead of %change a few bars ahead?
 
Hi Nsvv,

You have a free choice to insert any predicted nets into a new net, and you can put also them into a committee in order can select which are the best.

My result is an empirical result, a method may not work in other instruments, for example one method that used for gold may not work in eurusd. Most of time, on CH, I change the input formula and also the search setting (evolution or swarm) .

Basically, Neuroshell trader advised to predict percent change, you may have benefit to predict MACD since MACD is less whipsaw compare to other indicator, means you will have a less false signal as well.

Sorry for the late reply.
 
Last edited:
Top