Since my last two posts ( currency strength indicator and preliminary tests thereof ) I have been experimenting with different ways of smoothing the indicators without introducing lag, mostly along the lines of using an oscillator leading signal plus various schemes to smooth and compensate for introduced attenuation and making heavy use of my particle swarm optimisation code. Unfortunately I haven’t found anything that really works to my satisfaction and so I have decided to forgo any further attempts at this and just use the indicator in its unsmoothed form as neural net input.
In the process of doing the above work I decided that my particle swarm routine wasn’t fast enough and I started using the BayesOpt optimisation library, which is written in C++ and has an interface to Octave. Doing this has greatly decreased the time I’ve had to spend in my various optimisation routines and the framework provided by the BayesOpt library will enable more ambitious optimisations in the future.
Another discovery for me was this Predicting Stock Market Prices with Physical Laws paper, which has some really useful ideas for neural net input features. In particular I think the idea of combining position, velocity and acceleration with the ideas contained in an earlier post of mine on Savitzky Golay filter convolution and using the currency strength indicators as proxies for the arbitrary sine and cosine waves function posited in the paper hold some promise. More in due course.