Re: [Quantlib-users] copulas library proposal

Posted by Marek2 on
URL: http://quantlib.414.s1.nabble.com/Re-Quantlib-users-copulas-library-proposal-tp11877p11878.html

Hi Ferdinando


>>thanks for the contribution, I've just added it in the trunk code base.
>>I've pruned redundant inclusions, move inclusion to cpp file when
>>possible, expanded error messages to be more informative, and avoided
>>instantiating objects in the GaussianCopula::operator() method since
>>they could be instantiated once for all in the constructor.

Thanks for improvements, now code looks and works better.

>>It would be nice if you would contribute a unit test which reproduces
>>known tabulated values. This way I could have checked that I didn't
>>introduce any error ;-)

Unfortunately, I don't have any source for copulas values, neither in printed
form nor by commercal software packages, I only find free library for R.
I check code and I don't find any error. I think accuracy is as good as accuracy
of exp and other standard functions ;)
 

>>One question: you contributed bidimensional copulas. Is there an
>>efficient standard approach how to generalize to arbitrary dimensions?

As far I know in general there is no easy way to generalize arbitrary copula
to n-dimension, but some special families like elliptical it's quite easy.
For example for gaussian copula in place of bivariate_normal place
multivariate_normal, but for the time being in quantlib there is no multivariate
normal except bivariate of course ;) Additional problem is interface,
for more than n-dimenisional copulas I suggest to use vector of n-variables.
It could be interesting to develop, but I think multivariate distributions
should be done first.

Best Regards
Marek