neural network designs and experiences

[boskop]

greetings,

i'm trying out a kind of genetically trained recurrent elman network to predict the price of assets in the future.
what are your experiences in neural network designs? how did you train it? what where the questions that the output neurons got asked? and, most importantly, how many computers of what configuration were you able to utilize

 

[boskop]

in my research i stumbled upon the Nguyen-Widrow weight initialization method:
http://www.heatonresearch.com/encog/articles/nguyen-widrow-neural-network-weight.html

which seems to speed up evolution by about 10 %. yay!

the original paper: http://www.stanford.edu/class/ee373b/nninitialization.pdf

 

[the hare]

I use PNN's but only for trend identification. I honestly dont think prediction of future price is the way to go at all. I wish you luck with it. There used to be some decent discussions behind the scenes at the wealth lab forums which might be worth checking out.

 

[Pat494]

The neural networks of the past could do ok on trends but were hopeless at oscillations. Much the same as moving averages really, so unless some wunderkind has come up with a better system then don't waste good money on them imho

 

[alexander]

The neural networks of the past could do ok on trends but were hopeless at oscillations. Much the same as moving averages really, so unless some wunderkind has come up with a better system then don't waste good money on them imho

+1

NNs are basically fitting tools.

 

[FxGreenTrade]

I use PNN's but only for trend identification. I honestly dont think prediction of future price is the way to go at all. I wish you luck with it. There used to be some decent discussions behind the scenes at the wealth lab forums which might be worth checking out.

hare: I agree this is the best approach as NN's are not going to be the holy grail by predicting actual contract values. May I ask you for tips on using NN's or similar in trend identification. I would like to work on classifying trend vs. range days.

 

[trader3cnd]

NNs are too old to have any predictive value at this point. They have lost their edge, if any at all.

 

[FxGreenTrade]

NNs are too old to have any predictive value at this point. They have lost their edge, if any at all.

I guess I am not specifically speaking of NN's but of any sort of predictive analytics. NN's are an example and another one could be support vector machines. I am looking for a way to detect next period/day trending or sideways market.