Skip to contents

Vine copula models and structures

Source: vignettes/vine-copula-models.Rmd
vine-copula-models.Rmd

A vine copula decomposes a multivariate dependence model into a sequence of bivariate copulas. The structure determines which conditioned and conditioning variables belong to each edge; one bicop_dist object supplies the model for that edge.

Read an R-vine structure

Consider the triangular array

4 4 4 4
3 3 3
2 2
1

It represents these pair copulas:

Tree Edge Conditioned pair Conditioning set
1 1 1, 4 —
1 2 2, 4 —
1 3 3, 4 —
2 1 1, 3 4
2 2 2, 3 4
3 1 1, 2 3, 4

In R, the structure can be constructed from its order and off-diagonal rows:

structure <- rvine_structure(
  order = 1:4,
  struct_array = list(c(4, 4, 4), c(3, 3), 2)
)
structure
#> 4-dimensional R-vine structure ('rvine_structure')
#> 4 4 4 4
#> 3 3 3  
#> 2 2    
#> 1
as_rvine_matrix(structure)
#> 4-dimensional R-vine matrix ('rvine_matrix')
#> 4 4 4 4
#> 3 3 3  
#> 2 2    
#> 1

The square R-vine matrix contains the same information with zeros filling the unused triangle. as_rvine_structure() converts it back to the compressed representation.

cvine_structure() and dvine_structure() construct the two common order-determined special cases. Random valid structures are available through rvine_structure_sim().

cvine_structure(c(4, 1, 3, 2))
#> 4-dimensional R-vine structure ('rvine_structure')
#> 2 2 2 2
#> 3 3 3  
#> 1 1    
#> 4
dvine_structure(c(4, 1, 3, 2))
#> 4-dimensional R-vine structure ('rvine_structure')
#> 1 3 2 2
#> 3 2 3  
#> 2 1    
#> 4
rvine_structure_sim(4)
#> 4-dimensional R-vine structure ('rvine_structure')
#> 3 3 3 3
#> 1 1 1  
#> 4 4    
#> 2

See ?rvine_structure for the complete validity and proximity conditions.

Construct a vine copula from components

For a three-dimensional model, the first tree has two pair copulas and the second has one. The nested list follows pair_copulas[[tree]][[edge]].

pair_copulas <- list(
  list(
    bicop_dist("gaussian", parameters = 0.6),
    bicop_dist("clayton", parameters = 1.5)
  ),
  list(bicop_dist("frank", parameters = 2))
)
model <- vinecop_dist(
  pair_copulas,
  structure = dvine_structure(1:3)
)
model
#> 3-dimensional vine copula model ('vinecop_dist')
summary(model)
#> # A data.frame: 3 x 10 
#>  tree edge conditioned conditioning var_types   family rotation parameters df
#>     1    1        1, 2                    c,c gaussian        0        0.6  1
#>     1    2        2, 3                    c,c  clayton        0        1.5  1
#>     2    1        1, 3            2       c,c    frank        0          2  1
#>   tau
#>  0.41
#>  0.43
#>  0.21

The model can immediately be evaluated and simulated:

u <- rvinecop(5, model)
dvinecop(u, model)
#> [1] 0.3040095 0.3516670 1.8689282 1.3118066 1.7333509
pvinecop(u[1:2, ], model, n_mc = 1000)
#> [1] 0.143 0.204

Fit and select a model

With no structure supplied, vinecop() selects trees sequentially using the Dissmann algorithm and fits a pair copula on every selected edge.

n <- 150
z <- rnorm(n)
x <- cbind(
  z + rnorm(n, sd = 0.7),
  0.6 * z + rnorm(n),
  -0.4 * z + rnorm(n),
  rnorm(n)
)
u <- pseudo_obs(x)

fit <- vinecop(u, family_set = "onepar", selcrit = "bic")
fit
#> 4-dimensional vine copula fit ('vinecop')
#> nobs = 150   logLik = 17.21   npars = 6   AIC = -22.42   BIC = -4.36
summary(fit)
#> # A data.frame: 6 x 11 
#>  tree edge conditioned conditioning var_types   family rotation parameters df
#>     1    1        4, 2                    c,c      joe        0        1.1  1
#>     1    2        3, 1                    c,c    frank        0       -1.3  1
#>     1    3        2, 1                    c,c gaussian        0       0.33  1
#>     2    1        4, 1            2       c,c      joe       90        1.1  1
#>     2    2        3, 2            1       c,c  clayton      270       0.15  1
#>     3    1        4, 3         1, 2       c,c      joe       90        1.2  1
#>     tau loglik
#>   0.060    1.4
#>  -0.143    3.3
#>   0.217    8.1
#>  -0.043    1.3
#>  -0.068    1.2
#>  -0.087    2.0

The most important pair-copula controls (family_set, par_method, nonpar_method, mult, selcrit, weights, presel, and allow_rotations) have the same meaning as in bicop(). The bivariate copula guide describes the available families, rotations, and estimation methods.

Inspect the selected trees

Getters work on both manually constructed and fitted models. Tree and edge indices are one-based in the R interface.

get_structure(fit)
#> 4-dimensional R-vine structure ('rvine_structure')
#> 2 1 1 1
#> 1 2 2  
#> 3 3    
#> 4
get_matrix(fit)
#> 4-dimensional R-vine matrix ('rvine_matrix')
#> 2 1 1 1
#> 1 2 2  
#> 3 3    
#> 4
get_pair_copula(fit, tree = 1, edge = 1)
#> Bivariate copula fit ('bicop'): family = joe, rotation = 0, parameters = 1.11, var_types = c,c
get_family(fit, tree = 1, edge = 1)
#> [1] "joe"
get_parameters(fit, tree = 1, edge = 1)
#>          [,1]
#> [1,] 1.111375
get_ktau(fit, tree = 1, edge = 1)
#> [1] 0.06016693
get_all_families(fit)
#> Nested list of lists for the copula families of a 4 dimensional vine with all trees: 
#> - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - 
#> x[[1]] -> a list with the 3 copula families of tree 1. 
#> x[[2]] -> a list with the 2 copula families of tree 2. 
#> x[[3]] -> a list with the 1 copula families of tree 3.

summary(fit) returns the edge table invisibly, which is convenient for programmatic inspection.

Control structure selection

The edge weight used to build each tree is selected by tree_crit:

  • "tau" for absolute Kendall’s tau;
  • "rho" for absolute Spearman’s rho;
  • "hoeffd" for Hoeffding’s D;
  • "mcor" for maximum correlation;
  • "cxi" for symmetrized Chatterjee’s xi;
  • "joe" for a Gaussian-copula mutual-information criterion;
  • a custom function of data and weights.

tree_algorithm = "mst_prim" and "mst_kruskal" select a maximum spanning tree. "random_weighted" and "random_unweighted" draw random spanning trees, which can be useful in ensemble or exploratory workflows.

custom_strength <- function(data, weights) {
  if (length(weights)) {
    abs(weighted.mean(data[, 1] * data[, 2], weights))
  } else {
    abs(mean(data[, 1] * data[, 2]))
  }
}

custom_fit <- vinecop(u, tree_crit = custom_strength)
random_fit <- vinecop(u, tree_algorithm = "random_weighted")

Custom criteria run on the calling thread. Pair-copula fitting may still use multiple cores.

Fix all or part of a structure

Supply a full structure to keep it fixed while selecting pair-copula families. A truncated structure fixes its existing trees and asks vinecop() to select the remaining trees up to trunc_lvl.

fixed_structure_fit <- vinecop(
  u,
  structure = dvine_structure(1:4),
  family_set = c("gaussian", "clayton")
)

first_tree <- rvine_structure(order = 1:4, struct_array = list(rep(4, 3)))
partial_fit <- vinecop(
  u,
  structure = first_tree,
  family_set = "onepar"
)

Truncation and thresholding

trunc_lvl limits the number of fitted trees. Pair copulas above that level are independence copulas. threshold replaces edges whose selection strength is below the threshold by independence copulas.

truncated <- vinecop(u, family_set = "onepar", trunc_lvl = 1)
thresholded <- vinecop(u, family_set = "onepar", threshold = 0.1)
truncate_model(fit, trunc_lvl = 1)
#> 4-dimensional vine copula fit ('vinecop'), 1-truncated
#> nobs = 150   logLik = 12.74   npars = 3   AIC = -19.49   BIC = -10.46

Set trunc_lvl = NA or threshold = NA to select the value automatically with mBICV. The fitted threshold and truncation level are visible in the model summary.

Refit an existing model

Pass a fitted model through vinecop_object to update parameters while keeping its structure and families fixed. This is useful for rolling windows or bootstrap samples.

refitted <- vinecop(u, vinecop_object = fit)
stopifnot(identical(get_all_families(refitted), get_all_families(fit)))

For a manually constructed distribution, use vinecop() to select from data or construct a new vinecop_dist() after changing its components.