I have started reading Ed Gately’s book about neural networks in finances. I have also spent some time on writing my own Expert Advisors.
I have a doubt about correct choice of output parameter for neural network. The examples are: price in x days in the future; some value from 0 to 1, where 0.0-0.3 means enter long, 0.7-1.0 enter short, 0.3-0.7 no entry; percentage change of current price and so on. Some out output values are better, some are worse.
Let’s assume that neural network works in ‘open prices only’ mode. New bar comes, network says 0.2 what means enter long, next bar comes, network says again I should enter long. After several bars I have five opened trades, then new bar says between 0.3-0.7 so I do not open new trades.
I have many trades opened (what potentially may increase risk as I put much more in simultaneous trades). Then network says 0.8 - and what should I do? Should I close all already opened trades? Or should the network somehow close all the opened trades at the same time? (How?) And what about opening short trades (on reverse pair)? How to distinguish if 0.7-1.0 is about closing long trades and opening short trades? I guess EA should not be present on the market all the time.
Yes, you should close all open trades because the network expects that the price will go down.
Aside from that, it is known that neural networks are normally a poor method for price prediction. Unless you have some sophisticated filter for input and output to the network, their results extremely suffer from overfitting. So they tend to fail with out of sample data and live trading.
I agree with jcl, neural networks – for the most part – plays out much differently in OOS/live trading versus the learning sample.
But if you do continue to pursue neural nets, I would consider placing more emphasis in closing prices of each bar, as well as strategies aside from back propagation (if that is what you are currently using) as it’s been overused and quite frankly with very little results. Some other techniques is to look at derivatives of price rather than the price series itself.
Take a look at MINE from Harvard University, it might be worth your time to play around with.