Neural Network Trading, Serious people only! | Page 5 | Forex Factory

Options

Go to
Permalinks
Print Thread

Similar Threads

Neural network trading 34 replies

Neural Network Trading with Simbrain? 0 replies

Neural Network trading using Matlab and Metatrader 8 replies

Interested in Neural Network Assisted Trading 4 replies

neural network based systems in forex 22 replies

Attachments: Neural Network Trading, Serious people only!
Exit Attachments

Neural Network Trading, Serious people only!

  • #81
  • Jan 15, 2009 3:10am Jan 15, 2009 3:10am
I was able to predict something like 60 % prediction using NN ( at best 70 % ). But 60 % is not that great with NN when you know that it works sometimes and it does not works at the other time.

You can achieve better prediction consistantly with using some decent academic method such ARIMA or GARCH modelling integrated with regressor.

In fact GARCH does very good job even without regressor if you are only predicted for Volatility!!!!!!!

Hope this helps.
  • #82
  • Jan 16, 2009 3:10am Jan 16, 2009 3:10am
Quoting auto9817
I was able to predict something like 60 % prediction using NN ( at best 70 % ). But 60 % is not that great with NN when you know that it works sometimes and it does not works at the other time.

You can achieve better prediction consistantly with using some decent academic method such ARIMA or GARCH modelling integrated with regressor.

In fact GARCH does very good job even without regressor if you are only predicted for Volatility!!!!!!!

Hope this helps.
I use a ANN to confirm a trigger, I get much better results in backtesting, but still there is slippage and with a scalping system that can eat up profits fast.

My AI interests of the moment is Reinforcement learning and it's many flavors. There is a lot of academic research down these lines. Here are a few links.

http://www.cis.upenn.edu/~mkearns/papers/rlexec.pdf
http://www.icsi.berkeley.edu/~moody/JForecastMoodyWu.pdf
http://papers.klab.caltech.edu/273/1/425.pdf
http://www.stanford.edu/class/cs229/proj2006/Molina-StockTradingWithRecurrentReinforcementLearning.pdf
http://www.cs.bham.ac.uk/~wbl/biblio/cache/http___mahd-pc.jbs.cam.ac.uk_archive_PAPERS_2004_WP18.pdf

There are many more and several to do with Forex.

Dan
.
  • #83
  • Jan 30, 2009 2:03pm Jan 30, 2009 2:03pm
Hi,

Here is a strategy with 16 JMA NNs which have predictions with extremly low Rs error. I'm not an author of this strategy and i think it has a major fault in it
see this link, this is a reason that it gives unconsistent results.

http://www.investireoggi.it/forum/ne...32426-315.html

post 3145 down

If anybody in interested in improving this strategy, please post here.

Krzysztof
Attached File(s)
File Type: zip 16 Networks for JMA.zip   1.7 MB | 21,514 downloads
  • #84
  • Jan 30, 2009 5:04pm Jan 30, 2009 5:04pm
NSDT Pro R 5.0 A 4

Pair : GBPUSD
TF : H1
Attached File(s)
File Type: zip GBPUSD60.zip   48 KB | 1,321 downloads
  • #85
  • Feb 1, 2009 6:29am Feb 1, 2009 6:29am
Hi,

Here is an advanced strategy for NS, it uses JMAs crossovers, HP filter and Ehlers trend line. It uses also LEAD function, the usage of this function in real time trading is unclear for me.

Included chart has 3 Strategies, JMACROSOVER, FIlter1 and FIlter2, it you have a problem with accessing existing data calculations or adding an alarm, it is necessary to extract all strategies separatelly and put to the new chart.

Krzysztof
Attached File(s)
File Type: zip JMACROSSOVER.zip   1.1 MB | 2,144 downloads
  • #86
  • Feb 3, 2009 4:12am Feb 3, 2009 4:12am
Another advanced strategy. Needs to be converted to NS 5.x

Krzysztof
Attached File(s)
File Type: zip TS#1of2.zip   1.2 MB | 1,733 downloads
  • #87
  • Feb 3, 2009 4:26am Feb 3, 2009 4:26am
This error is due to loosing of optimization data. Just enough to write optimization data manually from original strategy.

Krzysztof
Attached Image(s) (click to enlarge)
Click to Enlarge Name: 16n1.JPG Size: 185 KB
Click to Enlarge Name: 16n2.JPG Size: 76 KB
  • #88
  • Feb 4, 2009 5:00am Feb 4, 2009 5:00am
Here is an example of real time repaint in Neuroshell. Reason - non causual data in Neural Net prediction.

Krzysztof
Attached Image(s) (click to enlarge)
Click to Enlarge Name: 2.JPG Size: 99 KB
Click to Enlarge Name: 3.JPG Size: 99 KB
Click to Enlarge Name: 4.JPG Size: 100 KB
Click to Enlarge Name: 5.JPG Size: 99 KB
  • #89
  • Feb 5, 2009 9:17am Feb 5, 2009 9:17am
One of those three indicators causes this. See below - sell order canceled and buy moved.

Krzysztof
Attached Image(s) (click to enlarge)
Click to Enlarge Name: Ich1.JPG Size: 117 KB
Click to Enlarge Name: Ich2.JPG Size: 118 KB
Click to Enlarge Name: Ich3.JPG Size: 69 KB
  • #90
  • Feb 5, 2009 12:20pm Feb 5, 2009 12:20pm
Has anyone been able to update the cortex indicator to a 4h or 15m indicator and retrained it for current volatile market conditions?If so, I really like to know.
  • #91
  • Feb 5, 2009 12:53pm Feb 5, 2009 12:53pm
From where i can get this Cortex indicator ?? Is it for NS ?? How you are measure it's quality ??

I believe the most important is the training data rather than NN type.

Krzysztof
  • #92
  • Feb 5, 2009 1:43pm Feb 5, 2009 1:43pm
Quoting fajst_k
From where i can get this Cortex indicator ?? Is it for NS ?? How you are measure it's quality ??

I believe the most important is the training data rather than NN type.

Krzysztof
http://cortex.snowcron.com/forex_nn.htm

They made one for 1 hour, but trained at slow markets. I like to see it modified. I have not been successful after hours of plotting...

But it used to work pretty good, just these markets are very different now.
  • #93
  • Feb 6, 2009 7:34am Feb 6, 2009 7:34am
Hi Ilovepippin, General question. I see you have Neuroshell. Than you use CORTEX. Can you tell me what is a logic behind using CORTEX than ?? Does it gives better results than NS neural nets ??? Did you make a comparision test for the same pair, TF, period and the same training data set ?? What was the criterium of choosing Cortex, can you post comparision results ?? I believe using external application like a CORTEX will be difficult to train NN properly simply because you can not feed NNs with proper data like .e.g in case when you use NS detrended data or certain specific data like e.g. wavelet transform outputs or outputs from other indicators. Krzysztof
  • #94
  • Feb 6, 2009 7:47am Feb 6, 2009 7:47am
Quoting fajst_k
Hi Ilovepippin,...
Hi fajst_k, I don't have neuroshell. From what I understand it is not really a good nn.I certainly don't want to use it for trading.I used the cortex indicator to predict a significant turn of the market and used it in combination with technical analysis. It worked really good until these volatile markets. Now, it will tell you when the markets turns, but you might still see a 200pip drop or gain on the market until it turns.But in general I like to see the big turns on the 4h chart and I read the 15m chart MACD and like to use the cortex_nn for confirmation. However, I have been sitting hours upon end putting new weights in, but I cant get it to work for 15m or 4h.
  • #95
  • Feb 6, 2009 8:22am Feb 6, 2009 8:22am
Perhaps this may give fresh ideas, the forum is in MQL4

http://forum.mql4.com/19239

It's modified version of previously available trainer + trading EA that's now combined to enable self training automatically
Adaptive autotrading is the future...
  • #96
  • Feb 6, 2009 10:43am Feb 6, 2009 10:43am
Can you post some statistics of buy/sell signals from CORTEX indicators together with info about training data etc..I can do the same using NS for the same pair and we can compare.

I don't see any reason why CORTEX should produce better signals than WARD PNN nets used in NS.

Krzysztof.
  • #97
  • Feb 7, 2009 7:50am Feb 7, 2009 7:50am
Here is a comparision of performance recurrent net against standard Ward net for the same training period and data set, calibration and out of sample period. Input for both NN was stochastic indicator.
Attached Image(s) (click to enlarge)
Click to Enlarge Name: r1.JPG Size: 174 KB
Click to Enlarge Name: r2.JPG Size: 114 KB
Click to Enlarge Name: r3.JPG Size: 76 KB
Click to Enlarge Name: r4.JPG Size: 57 KB
Click to Enlarge Name: r5.JPG Size: 52 KB
  • #98
  • Feb 7, 2009 7:57am Feb 7, 2009 7:57am
So for the same period (1 day out of sample) standard net earned 39 pips and recurrent 112 pips (both comission 3 pips). Paper trading results were also much better for recurrent net. Input data was low quality, so possible space for improvement is big here by feeding the net with high quality training and optimalization data to obtain stronger non linearity.

Krzysztof
  • #99
  • Feb 7, 2009 9:13pm Feb 7, 2009 9:13pm
Quoting charlinks
Thank you all who participated.

Neural networks do work!

Let's see how closely and if this prediction will work.
the result of this time pridiction is not good enough , the trend went up at last
Attached Image(s) (click to enlarge)
Click to Enlarge Name: 2009-02-08_111222.png Size: 25 KB
Click to Enlarge Name: 2009-02-08_111246.png Size: 26 KB
  • #100
  • Feb 8, 2009 12:11pm Feb 8, 2009 12:11pm
is here

http://www.trade2win.com/boards/soft...roshell-9.html

Krzysztof
0 traders viewing now