high performance computing on graphics processing units: hgpu.org

The optimisation of Technical Trading parameters is a computationally intensive exercise. Models comprising a modest number of Technical Indicators require many thousands of simulations to be executed over a sample period of data, with the best performing sets of parameters employed to generate future trading signals. The purpose of this research is to investigate the suitability of GPU Computing for running the simulations in parallel and to develop a working Prototype optimiser based on the CUDA architecture. The cumulative nature of Profit and Loss over a sample period is a restricting factor in the design of a data-parallel trading simulator. Thus, different approaches to the distribution of the parallel workload are researched and an appropriate design for the Prototype is derived. Past studies are examined, including parallel Genetic Programming implementations. The remarkable speedups enjoyed by the Prototype are discussed in detail and a number of key design strategies are proposed. These include a per-thread solution identification methodology, a modification to Welford’s Standard Deviation algorithm which results in the avoidance of divergent threads, and a suitable parameter distribution policy.