It seems that I have arrived late at the party. The thread seems to be dead or dying, but it is of great interest to me, so I post anyway in hope someone is still out there.
It seems that one difficulty with this approach to a net for FX is that one must choose an indicator as a target. I have been experimenting with NN's for FX for several years, and have been able to train many nets to predict an EMA for example, with very small mse. The problem always seems to be that all the error occurs at the pivot points, just when you want accuracy. I've never been sure that the net was any better than a simple linear prediction, which is also very good except at the turns.
There is another approach to be found in a recent article on the MQL site "Using Neural Networks In MetaTrader". The author has managed to integrate FANN into MT4, so that one can train a net based on how well the associated EA performs. This avoids several problems discussed above: 1. One does not train for mse, but instead for signals that give positive return. 2. One does not have to set arbitrary thresholds for the output signal, but can find the best threshold by genetic optimization over the training set.
This latter approach has it's own drawbacks. First and worst, the training must be done manually. This takes a great deal of time. It can be overcome by writing a windows script (in autoit e.g.) that does the training, but it will still take lots of time. Second, there is no possibility to use a validation set to decide when to stop.. this must be done manually. Third, the available networks are restricted to the set in FANN, so that other nets like recursive Laguerre, that are better suited to time series, cannot be used.
Finally, if anyone is still interested, it would be great if a competent programmer ( I am incompetent) would integrate a more powerful Net training program with MT4 and then somehow include a script, or other mechanism to allow training and validation.