Hi:
I ve worked with some softcomputing techniques for clustering, chaotic time series modelling. I know lot things, but there are extensive information, every day somebody create a new neural network architecture telling everybody that his neural network is the best.
Since i knew about forex i saw an interesting field for my knowledge and for some matlabs scripts made a long time, i tried predict price of EUR/USD using perceptrons but they suffer a lot of overfitting, memorize training data but predict like crap, althougth with some improvements with cross validation, local models make it better and are conceptually thousands times more simple than neural networks, but take a lot of computational time cause to create the model you need serach along all your data set nearest neighboors from your query point, in test sets it does make errors betwen 0.0018 and 0.0030 pips in data of 1h hour. (A local model is like a linear regression but you dont use all points, you only use the k-nearest to your point)
There's another class of neural networks created by russian investigators, they are polinomial neural networks, they have some few parameters, 'cause network it self is self organizative, algorithm it self decide number of nodes, polinomial function to use into the node and number of layer. An interesting work about polynomial nn in forex is here, is very good.
http://www.cs.uoi.gr/~foka/papers/foka_msc.pdf
I proved it with matlab but results are weird, i obtained a shifted version of prices like prediction, at first i thougth that i was wrong, but training an state of art neural network from neurosolutions i got the same. At last i found the very very, generalization of polinomial neural networks volterra kernels, this behaves like an infinite terms polynomial network (very complex concepts) is like a SVM(support vector machine), but i still follow getting the same shifting in predictions, this shifting make me errors in the same range of local models, if i would have to choose i would pass a local model to mq4 i still have the same results that with a very, very complex model like volterra kernel and better than perceptrons
I've read a lot about state of the art, there's two related concepts in neural networks that looks promising, but are still complex first are liquid state machines and others are echo
state networks, but i'm tired and searching more simple methods.
I thing that is better try to predict turning points using different indicator more than trying to predict price it self, trying price itself is very, very difficult (for not say nearly imposible still with state of art softcomputing techniques). Also, this strategy is very profitable like was demostrated in EA championships, my advice don't loose your time trying predict the price. I was coding differents methods in matlab in my free time during nearly three years(fortunatelly not all for forex), still the knowledge is very usefull and i've learnt sooooooooooooo much, but i insist don't loose your time trying predict the price.
See ya guys
Neural Network indicator - Any potential users?
Linear Mode
