Thanks OnTheRoad
Yes you are right, increasing the number of neurons can lead to better accuracy, but just slightly. If you have bad results with 20 hidden neurons, it won't be perfect forecast with 200 neurons for sure. You also have to be careful about not overfitting the NN : if you use a lot of neurons in the NN, you have to feed the NN with a lot of data. Matlab NN toolbox doesn't calculate an ajusted Rē to control over-fitting issues.
aegis : I used a few MA slopes (Ma value(t)-Ma value(t-1)) to predict the MA slope a few bars ahead. I'm not really sure about the trading power of this tool as I didn't backtested the system. But it looks it can take trades before the MA cross happen wich is slightly better compared to a basic MA cross system.
MrM: I didn't implement a dll solution. I coded some NN's under MQL4 as I found no real advantage from linking Matlab results with MQL4 as the NN doesn't need to be retrained every seconds. I just get the weights from matlab and copy-paste in MQL4
I was off for a few days, so I have no fancy graphics to post today!