An unsupervised parallel genetic cluster algorithm for graphics processing units
School of Computational and Applied Mathematics, University of the Witwatersrand, South Africa
During times of stock market turbulence, monitoring the intraday clustering behaviour of financial instruments allows one to better understand market characteristics and systemic risks. While genetic algorithms provide a versatile methodology for identifying such clusters, serial implementations are computationally intensive and can take a long time to converge to the global optimum. We implement a Master-Slave parallel genetic algorithm (PGA) with a Marsili and Giada log-likelihood fitness function to identify clusters within stock correlation matrices. We utilise the Nvidia Compute Unified Device Architecture (CUDA) programming model to implement a PGA and visualise the results using minimal spanning trees (MSTs). We demonstrate that the CUDA PGA implementation runs significantly faster than the test case implementation of a comparable serial genetic algorithm. This, combined with fast online intraday correlation matrix estimation from high frequency data for cluster identification, may enhance near-real-time risk assessment for financial practitioners.
March 23, 2014 by hgpu