Stability-speed in bicop estimation and vine selection

[oezgesahin]

Hello,

The current lower/upper bounds for bivariate copula families' parameters are too low/high in the mle approach. While doing/using successive optimization to find bicop parameters, they are creating numerical instabilities:

• The implementation uses the same upper bounds as the package rvinecopulib except for the Gumbel copula. Its upper bound can be 17 rather than 50.

RVineSeqEst loops over all trees without considering a truncation level or fit of independence copula in high tree levels:

• The implementation loop over all trees and pair-copulas up to a truncation level or fit of independence copulas, which appears to be a faster approach.

[tnagler]

Thanks!

I notice that you're changing the parameter bounds also for other copula families. The bounds in rvinecopulib are expected to be a bit different simply because the implementation is a bit different. I don't mind changing the bounds in VineCopula if there's a good reason. So if you experienced problems with all these bounds, all god. If not, please only change the bounds that caused proiblems for your.

In any case, the bounds need also to be changed in BicopCheck().

[oezgesahin]

Thanks for all explanations!

Yes, I needed to implement all these bounds not to have any numerical problems in some analyses. These include changes in bounds of Clayton, Gumbel, Frank, Joe and their possible survival extensions.

If you agree, I can try to change the bounds in BiCopCheck() accordingly or?

[tnagler]

OK great, just want to make sure we're not introducing unnecessary restrictions. Just change the bounds in BiCopCheck() then and we're good to go.