MoneyTec 2.0 - Online Forex Traders & Brokers Social Networking Community » Trading Talk » Beginners Forum

[szczupak]

Many traders are said to unsuccessfull... this may be due to a number of different factors... the financial literature has shown that you can't apply tools for linear regression if markets are non-linear... what's the solution...? The solution are neural networks... if some of you don't know what it is all about visit a blog i started some time ago... i will keep you up to date with the literature and i will also build models to make predictions even more accurate than those i'm makingnow... the address is: http://neuralmarkets.blogspot.com
I hope you find it interesting and useful...
take care

Szczupak

[Aleph]

Nice video of the hr-2 robot gotta love this little dude

aleph

[initialsBB]

I've been using NN for some time now, with promising results, so i wonder which software do you use? i use TradingSolutions, but besides using the built-in NN i also try to build more complex NN in NeuroSolutions (although by now unsuccessfully).

[szczupak]

Initials could you please provide some results and the architecture of your NN with all the layers and so on... what type of NN are you using? what markets have you applied it for?
thanks

p.s. i use a number of different programs depending on the analysis i have to make and the control i want to have on the training and testing process but i don't accept black boxes that do not allow you to set parameters manually..

[initialsBB]

Thus far good results returned Basic Multi-layer Perceptron with 1 hidden layer and Time-lag Recurrent with 1 or 2 hidden layers on AUDUSD, EURJPY and NZDUSD (on those pairs i focused on). TS does allow you to manipulate the parametars of the networks, but it does not have networks with more than 2 hidden layers-for that you'd have to buy NeuroSolutions. I wonder how much time you wait (or allow) for out-of-sample performance before you decide that model is worth trading? and how often do you retrain the network? some do it every 3 month, while some every 6 months?

[hanzahar]

i dont get it...what does it have to do with forex?

[wkh]

i think neural networks are so old that much of the potential they have to discover inefficiencies and the like will already have been exploited by market participants. multi-layer perceptrons are so 80s, dude.

if someone could respond to my thread, i just have a simple question. thanks.

[szczupak]

The literature from 2005 and 2006 suggests that NN can outperform all kinds of econometric analysis... and regression in particular... so their value is not to be underestimated... as i said in my first post ii will write more about it in my blog...

[initialsBB]

Originally Posted by wkh

i think neural networks are so old that much of the potential they have to discover inefficiencies and the like will already have been exploited by market participants. multi-layer perceptrons are so 80s, dude.

if someone could respond to my thread, i just have a simple question. thanks.

This is so funny that i really don't have to prove its absurdity. I guess you're some newbie on a "new-indicator" hunt

[szczupak]

i agree with you Initials...

[szczupak]

look at the results please...

[initialsBB]

But is this data that was used to train the network or is it out-of-sample? Neurals tend to learn the rules of the training data well, but usually that performance deteriorates on out-of-sample. Of course some deteriorate only slightly, but i had examples with 95% accuracy on training and 35% on out-of-sample. Therefore i usually take into consideration only performance on data that the network hasn't seen before.

[szczupak]

out of sample buddy

[initialsBB]

Have you done any real trading with those predictions? How do you compare that style of trading with the one that actually learns profitable patterns from the past and generates buy,sell and stop signals when they appear in the future?

[szczupak]

old patterns do not work any longer... especially TA classic patterns known from the 80s... NN can adapt themselves... the results are 72% trend accuracy on s&p500 and 84% on eur/usd