Hi prasxz!
Nice to see im not alone here
I also think using high and low is much better than using open or close, to be truth i dont really like the open and close values for intraday analisis, as they seem rather inestable values if you make a displacement on when you place your start point.
Median price also looks good, but i prefer high/low as the information loss in less.
I'll probably use a moving average of the high and one of the low. I've found the JMA to be a really good filter in comparison to regular MAs so i'll do my first tests using a short period JMA without phase changes to avoid distorsion.

So far the inputs im considering for predicting future ranges are:
- JMA of High + JMA of Low.
- JMA of Volume.
- ATR.
- Date (Day of Month, ex. Monday, Thuesday...).

The other idea i have in mind is to use NN to forecast news events direction.
I have a quite big database of forex fundamentals since some years, so i may use those as inputs.

As for what type of neural network to use, im still doing some research, backpropagation NNs are the common standard for NNs, but there are others that seem to have very good results, like the time lagged recurrent networks (but they are hard to train and understand). Other idea i had, was to use a Fukushima NN, those are mainly made for image processing, but with some modifications i think they could be used for pattern recognition on timeseries.