Those who advised about various aspects(through the forum as well as PM), Thank you very much.
Now it is the most important thing.
What should be the outputs?
At first I thought one day forecast would be more than enough(and it can be done with a greater accuracy) but thinking further in to this, 12hr advance notice is not enough to make an investment decision. Even though one day forecast is fairly accurate, it will be difficult to detect the change in market trend with a single forecast.
I think at least five day forecast is required in order to get some idea about the market.
Did some testing on a simple feed forward(back propagation) NN.
Will make another post about network training next time, for the time being results are as follows.
Training Data : 1500++ data records from Jan/2004~Dec/2010
Test Data : 73 data records from (A)Jan/2004, (B)Dec/2010&Jan/2011, (C)April/2011
Test Symbol : NTB(Normal)
Network Design: Feed forward back propagation NN. 12 Inputs, 6 Hidden neurons, 5 Outputs.
Performance based on Mean Absolute Percentage Error,
Bellow image is the training map showing how the predicted value converge in to actual value during training.
Bellow two graphs compare "One Day" & "Five Day" predicted values with actual price.
One day forecast error
Five day forecast error
Please note that, x% error does not mean that forecast is (100-x)% accurate.
It is the deviation of predicted value from the actual.
Also above network is far away form perfect and need a lot of training and fine tuning.
Thinking about trying a couple of different NN designs later.
Feels like, things related to field of computing is out of the scope of this forum
and will not go in to details here.
Any comments about network outputs are highly appreciated.
After all daily forecast may not be the ideal output
All that matters to us is a simple buy/sell signal