On Wednesday 30 May 2007 19:48, Robert Kern wrote:
I'm afraid this doesn't work, either. Correlation matrices are constrained to be positive semidefinite; that is, all of their eigenvalues must be >= 0. Calculating each of the correlation coefficients in a pairwise fashion doesn't incorporate this constraint.
But you're on the right track. My preferred approach to this problem is to find the pairwise correlation matrix as you did and then find the closest positive semidefinite matrix to it using the method of alternating projections. I can't give you the code I wrote for this since it belongs to a customer, but here is the reference I used:
Thanks your comments and for the reference. I will try to implement the algorithm sometime this month.