In these types of situations, I normally would create a search algorithm to 'almost brute force' through the possibilities. Of course, you can't brute force it, but a smart search algo should find a good configuration.
People generally only "give away" things of value either in exchange for other items of value or else donate value with the conviction that such valuables are used for a purpose that furthers "good causes". By definition of most, an individual's purse is not part of the "good causes bucket". Hence I would not put up too much hope to find the gold nuggets without a lot of digging, a deeper understanding of this space, and the willingness to get your hands dirty. Are you willing to do such? If yes, have you meaningfully understood the basics of machine learning, neural networks, have you gone through the introductory literature and are ready to even be tested and quizzed on it? Then I think you will find people to share their value in exchange for your contributions.
Excellent. I am an FX guy too. Using pure NN seems quite interesting, what I don't like is that I am having to select which inputs are used. This means I might be missing some important, but not obvious drivers.
inputs regarding hyper parameters? If data types you trained on and later use to predict on the trained model then that is left to your imagination. If you are not interested in making choices in model setup and later training then maybe this space is not what you should invest your time in. Maybe another approach to trading then suits you perhaps better...
Hey Conduit, Perhaps I expressed myself poorly. What I was suggesting is that it would be nice/better if I could automatically select the inputs used, as opposed to selecting them myself. Might be more rigorous, and also bring in some previously (for me) unused inputs that I did not consider worthwhile...
There're some methods for it, see Input Variable Selection. There's a comprehensive review - http://cdn.intechopen.com/pdfs/14882.pdf
The neural net library I am using randomly sets the weights at the start of each "set" of iterations. If I do 10 set of iterations (all starting with different random weights). Is it "good" to average all weights before a prediction? Or should I just average the top 3?
Thanks for this Dmitry. Very interesting paper. One point that intrigues me that is NOT covered in the paper is whether inputs needs to be correlated to the desired output. The paper suggests they do, but I wonder if two inputs (each uncorrelated to output) could not somehow add value when taken together?