Mining and Analyzing Financial Patterns using Parallel Genetic Programming on GPGPU

Abstract

When creating a system for automated trading of financial assets, one has to make multiple decisions under uncertainty. One significant part of such a process is related to knowledge discovery in time series data. Quantitative techniques are applied to time series data to mine informative signals and make a profit using them. Genetic optimization is an approach that can be put to use in such situations.<br/> <br/> One drawback of genetic optimization is that it requires a significant amount of computational power to be genuinely effective. I address this problem by exploiting parallelism to fully utilize modern computer hardware. Specifically, I will explore the parallelization of fitness evaluation using general purpose graphics processing units (GPGPUs). A parallel genetic search framework for knowledge discovery in time series data is proposed and is shown to be highly effective, achieving more than 1450-fold reduction in the running time of an optimization process.<br/> <br/> The proposed parallel optimization framework is applied to financial data to obtain attractive technical patterns in the Korean stock market. The latent space induced by the collected patterns by factorizing relation matrix is analyzed. The analysis shed some light on how patterns interact with each other and create a geometry.<br/> <br/> I further investigate the value of the proposed framework by studying the cryptocurrency market. Extensive experiments are performed to analyze the factors that affect the quality of the solutions found by the proposed genetic programming system.<br/> Out-of-sample performance of the patterns indicates that parallel genetic optimization process can consistently find signals that are profitable and frequent. A trading simulation with the generated patterns suggests that the captured signals are indeed useful for portfolio optimization.