Build Neural Network Indicator in MT4 using Neuroshell

1. What were you using it for?
2. Time frames you developed.
3. Ease of use
4. Increased effectiveness in terms of prediction?
Bottom line, does it perform as advertised?

Their actual website isn't working currently. Any idea why? Out of business?

Regards,
Merlin
 
1. What were you using it for?
2. Time frames you developed.
3. Ease of use
4. Increased effectiveness in terms of prediction?
Bottom line, does it perform as advertised?

Their actual website isn't working currently. Any idea why? Out of business?

Regards,
Merlin

I was using them to find optimal TF for Goertzel indicator, TFs used were betwene 1-10 min. Yes they are easy to use. For predictions i didn't check if it gives any improvement.

I think those indis are quite old. Have a look to russian forums, there is more info about them.

Krzysztof
 
Thanks. Found the indi, but it's circa 2006. Doesn't seem happy to install.

Perhaps they'll get back in business someday ;)
 
Arryex..
Have you considered using your next bar high and low predictions on longer TF's to predict when a trend is happening?
e.g. Predict next bar high and low for TF H4, or maybe H1, and then when the prediction shows both high and low significantly up or down, trade based on predictions for TF M15. That might help avoid ranging markets, but it might not be easy to write the EA.
MadCow
 
Hi Arryex..
Where did you get 15 minute data for non FX markets (e.g. Nasdaq 100 index or Gold) to train your nets? I can only get 2K data points at any time frame from UTC or ODL. 2K at 15 minutes don't go back very far, and you have used data from 1999. Maybe your 1999 nets use only FX data?

Also did you use raw data (OHLCV) or did you preprocess?

MadCow
 
Hi Fralo,

Please do not get confuse, I have made several NN based indicators:
- Trend indicator based on EMA, non multimarket analysis. The inputs are indicators in EURUSD itself.
- predicting High, low based on multimarket.

The one that I presented lastly is the first one. I did explained previously as well where I can get the non forex data for multi market analysis (use MT4 from SIG, FXpro) but not for 1999 data.

Here attached excel files for my own trade analysis and trade data, I generate directly the trade data from my indicator and edit manually to get profit/loss, trade signal, etc. I do not know yet to use MSA to make my equity smoothly or to maximize the profit with some facility inside MSA.

Arryex
 

Attachments

  • msa1.png
    msa1.png
    98.2 KB · Views: 774
  • NNM17iM030210a Trade Result.xls
    188 KB · Views: 913
Hi Fralo,

Please do not get confuse, I have made several NN based indicators:
- Trend indicator based on EMA, non multimarket analysis. The inputs are indicators in EURUSD itself.
- predicting High, low based on multimarket.

The one that I presented lastly is the first one. I did explained previously as well where I can get the non forex data for multi market analysis (use MT4 from SIG, FXpro) but not for 1999 data.

Here attached excel files for my own trade analysis and trade data, I generate directly the trade data from my indicator and edit manually to get profit/loss, trade signal, etc. I do not know yet to use MSA to make my equity smoothly or to maximize the profit with some facility inside MSA.

Arryex

Thanks for the answer. I understand.

I thought that you had an EA that worked with strategy tester. If so, why use a spreadsheet? I am afraid I am still confused, but that's just the way I am.:LOL:

How did you get the entry and exit prices in the spreadsheet? I see commission; did you account for the spread?
BTW did you notice that about 2/3 of the profit that you show on the spreadsheet could be attributed to the long term trend in EURUSD? (Buy/Hold strategy results in ~42000, but does not account for swap (difference in interest rates) while holding.)

MadCow
 
Ideal Hardware

As you know, NNet testing can be computationally intensive.

Check out the offerings at: http://www.colfax-intl.com/nvidiaGPU.html

Asssuming there's something left in your budget after the Holidays, and you can program with Visual Studio 2010 and CUDA, one of these systems could be the answer to your prayers. :)

If you put: colfax tesla in the youtube search engine, there are a fair number of good videos. CUDA, also.

I hope to have one down here by June.

Regards,
Merlin
 
Thanks for the answer. I understand.

I thought that you had an EA that worked with strategy tester. If so, why use a spreadsheet? I am afraid I am still confused, but that's just the way I am.:LOL:

How did you get the entry and exit prices in the spreadsheet? I see commission; did you account for the spread?
BTW did you notice that about 2/3 of the profit that you show on the spreadsheet could be attributed to the long term trend in EURUSD? (Buy/Hold strategy results in ~42000, but does not account for swap (difference in interest rates) while holding.)

MadCow

I use the spread sheet to fed it into MSA... it just a draft result, commission is included but not complete with spread or swap...you may included into excel spreadsheet.

Arryex
 
Re: Ideal Hardware

As you know, NNet testing can be computationally intensive.

Check out the offerings at: http://www.colfax-intl.com/nvidiaGPU.html

Asssuming there's something left in your budget after the Holidays, and you can program with Visual Studio 2010 and CUDA, one of these systems could be the answer to your prayers. :)

If you put: colfax tesla in the youtube search engine, there are a fair number of good videos. CUDA, also.

I hope to have one down here by June.

Regards,
Merlin

i think you dont need such advanced hardware. Its just enough ordinary NVIDIA graphic
card. Free NN software with CUDA support you have here.

http://dit.ipg.pt/MBP/

Krzysztof
 
Hi Krys,

Thanks for your opinion & the good link.

As it happens, I have a generous budget and can afford the best. Ultimately, one good trade can pay for the system.

You'll also notice on http://dit.ipg.pt/MBP/bugs.aspx that this system is significantly buggy and *far* from being ready for serious production work.

Regards,
Merlin
 
Last edited:
Hi,

I dont know what is your knowledge about CUDA and CUDA programming but each application must be basically programmed from scratch so i think better is to verify if the idea works and test on non CUDA enviroment and than when you come to too long optimization time reprogram in CUDA.

Otherwise it can be just lose of time for programming or money for programmer.


Krzysztof
 
Arry..
In post 14 you posted images showing a multi-market net that predicted the 10-day close for EU. The images show you used 1353 training patterns and 346 test patterns, and they show about a 2.5 pip average error on the test set. Your error factor range on the test set show that you very seldom get an error over 10 pips.

I find it hard to believe that you can predict the daily close so far out and so well, but you seem to have done it.
If the training set and test set do not overlap, that is very good prediction for the daily close, and the test set is longer than 1 year. It looks like you could make an excellent system using this net.:clap:

So, why have you gone off chasing the trend indicator that does not appear to be as readily useful?:LOL:

MadCow
 
Madcow

Ok, let me verify again the strategy on post 14, all made in NSDT. I do not have NSDT here in offshore, I will revert to you when I am in home.
For the moment I do not care yet about the result of prediction, I am still digging out the idea how to make NN indicator and implement them. It seems most of prediction work well on daily chart, I am not sure for lower time frame.

Thank you
Arryex
 
I found the previous multimarket data and try to retrain it using several Neural Network Architecture and training methods.

Using daily market data since 1 Jan 2005 up to 26 Jan 2010 (1406 bard data, 80% Training patterns and 20% test patterns).

With 281 bars test patterns (, i.e. 24-Dec-2008 up to 26 jan 2010) the best minimum average error is 0.0000585 (= 0.6 pip on daily chart), trained only within 5 minutes.

See attached Training short report. I am sure that this empirical result may be varied upon the data and prediction methods.

Please Use "alif-syifa" to open the document.

Next step is to create indicator to shown it on MT4 and create an EA, any advice is highly appreciated.

Arryex
 

Attachments

  • Multimarket Training short report.pdf
    636.4 KB · Views: 721
I found the previous multimarket data and try to retrain it using several Neural Network Architecture and training methods.

Using daily market data since 1 Jan 2005 up to 26 Jan 2010 (1406 bard data, 80% Training patterns and 20% test patterns).

With 281 bars test patterns (, i.e. 24-Dec-2008 up to 26 jan 2010) the best minimum average error is 0.0000585 (= 0.6 pip on daily chart), trained only within 5 minutes.

See attached Training short report. I am sure that this empirical result may be varied upon the data and prediction methods.

Please Use "alif-syifa" to open the document.

Next step is to create indicator to shown it on MT4 and create an EA, any advice is highly appreciated.

Arryex
Thanks for the pdf. (y)

I don't have NS2, so can you tell me what a standard network is in that program? Is it a simple multi-layer perceptron, or is it more complicated? Also are all the nodes tanh functions, or did you use something else?

I have a suggestion about the indicator. I think that the indicator that you used for Post4 must repaint. That is why the prediction is so perfect when you look back at it, but when you look at what actually happened for the next 5 bars, the prediction is more realistic, making errors...

If your indicator just paints the endpoint that the net outputs for each bar, you can avoid the repaint problem. It is a common problem when an indicator predicts into the future. Avoiding repaint is not always easy.:(

If you get all that done, I have some suggestions for an EA, but they will depend on how well your indicator works, so I won't make them just yet. If it works as well as your pdf indicates, you won't need any suggestions for the EA, just buy low and sell high, your indicator will know where those points are.

MadCow
 
Madcow,

In version 4,NN design availables are:
1. Backpropagation: standards (3, 4 & 5 layers includes 1 input and 1 output layer), jumps(3, 4 & 5 layers includes 1 input and 1 output layer), recurrent (dampened feedback, hiddens layer feedback, output feedback), wards (2 hidden layer, 3 hidden layer, 2 hidden with jump layer).
2 Kohonen (unsupervised)
3. Probabilistic (PNN)
4. General regretion (GRNN)
5. Polynomial (GMDH)

Activation functions available (can be choosen freely):
linear [0,1], linear<<0,1>>, linear [-1,1], linear<<-1,1>>, logistic, tanh,gaussian, sin, tanh15, symetric logistic, Gaussian composite
logistic -- f(x)=1/(1+exp(-x))
linear -- f(x)=x
tanh -- f(x)=tanh(x), the hyperbolic tangent function
tanh15 -- tanh(1.5x)
sine -- sin(x)
symmetric_logistic -- 2/(1+exp(-x))-1
Gaussian -- exp(-x^2)
Gaussian-complement -- 1 - exp(-x^2)
I have the detail NN configuration in my Training long report

I will have a look to the indicator in #4 after.

Arryex
 
Calibration ??

Hi Arry,

I had a quick look to NS2 tutorial. Did you use calibration option during creation of your net
maybe ?? Because the test period is automatically set to 20% of total set and in this case its not allowed to use this 20% as an out of sample period due to future leak - this period is used for data snooping. I think it is the same as 'paper trading' in NSDT. See screens.

Krzysztof
 

Attachments

  • c1.JPG
    c1.JPG
    110.5 KB · Views: 557
  • c2.JPG
    c2.JPG
    96.6 KB · Views: 564
  • c3.JPG
    c3.JPG
    112.2 KB · Views: 564
  • c4.JPG
    c4.JPG
    142 KB · Views: 600
Top