issues.md

In vincenzocoia/copsupp: Vine Copulas made easy

Issues, Future Directions, and Ideas

1. It might be useful to have a function to check whether a vine (including the array) is valid (like in CopulaModel, but allowing for truncated vines).

2. It might be useful to somehow add functionality so that U-V asymmetric copulas can be used:

   • In addition to the 4 possible reflections, there's also the U-V flip for those -- meaning that there's 8 copula families in total. Are the latter 4 really needed?
   • Involves being careful about what variable comes first when specifying a vine edge, and getting the copula right when the vine array is moved around (such as re-leafing).

3. Package should probably be renamed "rvine". Or, perhaps notation should switch to "gvine", or "regvine".

4. The vignette needs to be improved:

   • It contains too much extraneous information that would only be useful for developers.
   • (This might follow from the previous downfall) it reads more like a pile of information, as opposed to a meaningful "story" that communicates what a user would want to use the package for.
   • Get the vignette working so that, upon typing vignette('copsupp') in R, the vignette appears.

5. fit.rvine() outputs a different vine array depending on the order of the variables input into the vars argument. It shouldn't.

6. qcondrvine() would be handy to have in some situations.

7. Problems with RVineCopSelect() in VineCopula (may have to write my own version):

   • Forces you to use a pre-defined set of bivariate copula families.
   • Doesn't let you choose the copula model matrix. It would even be nice to be able to choose parts of the copula model matrix too.
   • When you request a "bvtcop" to be fit (via familyset = 2), it can instead fit a "bvncop" (returning family = 1) even though you didn't want one to be fit.
   • (Also with BiCopSelect()) Puts negative parameters on 90- or 270-degree rotated copulas, but those models actually have positive parameters.
   • (Also with BiCopSelect()) Can't seem to fit the BB6 copula properly. See this example:

   udat <- CopulaModel::rbb6(10000, c(3,3)) lapply(VineCopula::BiCopSelect(1-udat[, 1], udat[, 2], familyset = 28), identity)

8. A fitting procedure for vines that's similar to that of the marginal should be included. So, parts of the vine might be specified, parts of the marginals might be specified, and maybe even some parameters ought to be shared amongst some edges.

   • Would it make sense to add components to an "rvine" object, such as data and corresponding measures of fit like AIC/BIC?

9. Perhaps a function edge_possibilities(i, j, G, rvine = TRUE) or something would be useful to see what variables can possibly go in G[i,j], based on the parts of G already specified. This way I can "fill-in" a vine array if need be, but more importantly so that, when building a new layer, I can choose variables so that the result is still a regular vine (so that I can finish coding the fitseq_rvine() function).

10. Multivariate integration should be sped up (in, for example, pcondrvine()). Use gauss quadrature or something (keep in mind the bounds of integration are probably always contained in [0,1]).

11. fitlayer_cnqr() should allow for partial specification of parameters.

12. If you want to partially specify parts of a vine, like in layeropts() for example, it might be useful to have a function that returns the desired matrix (or whatever) with NULLs and NAs in their appropriate places. Something like specifycop([2,3] = c("frk", "bvncop")).

13. fitrvine_basic(), and probably many others, throw an error when the data is a data frame (must be matrix). Fix this to allow for data frames.

14. Speed-up multivariate integration by using a more appropriate method than nested "integrate".

    • cubature package.
    • Or, check the suite of packages for this purpose on this page of CRAN task view.

15. Instead of making flipped copulas by appending "u", "v", and "r", perhaps make functionals like pcopu() and qcondcopu() that take a function (like pfrk() and qcondgum()) and returns the same type of function, but flipped accordingly. This way, when defining their own, people don't have to make these flipped versions. Also, it makes programming easier.

16. Make documentation rvine.object() that describes the "rvine" object and explains how to look at it. Should first look at the purpose of such documentation. For an example, see rqss.object() in {quantreg}.

17. Vectorize pnew() and others over cpar:

    • For example, pnew(u, v, cpar) should vectorize over cpar if u,v are not vectors. Vectorizing phiinv() should accomplish this -- but that involves vectorizing that algorithm for finding inverse. Maybe I should just use Vectorize() instead.

18. This code results in an error (2016-07-28): summary(rvine(matrix(1)))

19. Add bibtex citation to the package

20. When rearranging a vine (with releaf(), for example), the copulas should also be permutation-reflected where necessary (probably if an upstream variable switches places with a downstream variable, but would need to figure out exactly when to permute which copulas).

21. pcondrvine() throws an error when the inputted vine is truncated so that the last row of the copula matrix has two copulas (not 1) -- the pcondmat that is input into copreg::rVineTruncCondCDF() also is missing that last row of a single copula, which is why it errors out.

vincenzocoia/copsupp documentation built on Aug. 23, 2020, 7:37 a.m.