This work is concerned with algorithmic trading by using different learning procedures implemented on different hardware platforms, such CPU and GPGPU. One of the major focus of this work is algo-trading with Hidden Markov Models when the value of an optimized portfolio is predicted by HMM. Here we take into account that HMM has a powerful modeling capability, thus it can both capture mean reverting stochastic processes, as well as other time series.
In the thesis I used slight modification of the Baum-Welch algorithm which yields faster convergence. The other major attribute of this work is the GPU based implementation of the different trading algorithms, such as HMM based trading. Besides an extensive numerical analysis revealing and comparing the performance of the methods on high frequency FOREX rates, I have also carried out a detailed speed-profiling which paves the way towards high frequency trading.
I also tried to verify the use of Independent Component Analysis as a pre-processing to improve the performance of NARX and SVM based trading.