Recently new research has appeared on using a co-movement measure to
construct the covariance matrix as part of the Modern Portfolio Theory (MPT)
style portfolio construction. Below is the abstract of the
Gerber, Markowith and Pujara (2015) paper whose methodology is also currently patent pending:
Markowitz’s mean-variance MPT has remained the cornerstone of portfolio selection methods after decades of research and debate. There is an extensive literature on MPT implementation, especially on estimation errors and expected return assumptions. However, covariance matrix estimation, an essential input, continues to be frequently based on historical correlations. There has been a recent new study that proposes replacing historical correlations with a robust co-movement measure called the Gerber Statistic.
In the research paper, it is stated that MPT using the Gerber Statistic outperformed portfolios using historical correlation as measured by ex-post returns under realistic investment constraints, including transaction costs and a broad range of investor types, for an investment universe of global stock indices, bonds and commodities for the period January 1994 to December 2013.
This post is to illustrate an implementation of the Gerber statistic. The focus is to compare the speed of computation for three different implementations with increasing performance
Rcpp with OpenMP for parallization
Implementation in R
Implementation in Rcpp
Implementation in Rcpp with OpenMP
Finally, let’s compare the speed gain result. The test data is based on a
return matrix of 30 securities with 2500 data points. It can be seen that the
OpenMP version of the calculation is clearly faster than the serial version
which itself is much faster than the R version.