There is no doubt that those stakeholders of financial markets and a stock exchange who gets the right information before it spreads can gain advantage of them. In my thesis I am trying predict future price from historical prices with data mining and time series analysis methods. Then based on these retrieved information I am going to assemble a portfolio which provides the maximal expected profit.
For the prediction I am going to use a special kind of artificial neural networks the so called deep neural networks. These models are differs from traditional networks in the number of hidden layers: the deep networks contains at least with a magnitude more than the traditional networks. Because the train of the neural networks is computationally hard until the late 2000’s it was not feasible to train them in reasonable time. In the previous years both the hardware and the algorithmic background improved as much to became one of the hottest research areas.
Neither the best portfolio could become profitable if we cannot define the right investment ratio for each asset in the right time. In order to my calculations converge even faster to the correct results instead of the CPU I run them on the graphical card. For the code compiling to the GPU I use the Theano Python package which a powerful machine is learning tool developed by the LISA Lab at the University of Montreal.
While I was working on my thesis my goal was not only to make a working implementation but I wanted to implement an optimal solution. In order to reach optimum I based my every decision on theoretical proof when it was possible. Otherwise I tried to follow rules of thumb. In those cases when I have reached crossroads in order to made the best decisions I tried to test every possible scenario and made my choice empirically. Since I have not found such a detailed analysis of deep neural networks as I did in this paper, I hope that one day my results will be helpful for someone.