DatagenCopulaBased.jl

Data generator based on copulas
Author iitis
Popularity
21 Stars
Updated Last
12 Months Ago
Started In
September 2017

DatagenCopulaBased.jl

Coverage Status DOI

Copula based data generator. Returns data in the form of the t x n matrix U wheret numerates the number of realizations, and n numerates the number of marginals. By the copula definition each marginal uᵢ is uniformly distributed on the segment [0,1]. Realizations of such marginal would be U[:,i].

Interdependence between marginals is modeled by the n-variate copula, see e.g.: R. B. Nelsen, 'An introduction to copulas', Springer Science & Business Media (2007). See also K. Domino: 'Selected Methods for non-Gaussian Data Analysis', Gliwice, IITiS PAN, 2019, [arXiv:1811.10486] (https://arxiv.org/abs/1811.10486).

This module support the following copula families:

  • Elliptical copulas (Gaussian, t-Student),
  • Frechet copulas (maximal, minimal, independent),
  • Marshall-Olkin copulas,
  • Archimedean copulas (Clayton, Frank, Gumbel, Ali-Mikhail-Haq),
  • Archimedean nested copulas.

Installation

Within Julia, run

pkg> add DatagenCopulaBased

To install the files Julia 1.0 or higher is required.

Sampling data

To sample t realisations of data from copula::TypeOfCopula use

julia> simulate_copula(t::Int, copula::TypeOfCopula; rng::AbstractRNG = Random.GLOBAL_RNG)

where rng is the random number genrator that can be selected.

julia> Random.seed!(43);

julia> simulate_copula(3, GaussianCopula([1. 0.5; 0.5 1.]))
3×2 Array{Float64,2}:
 0.589188  0.815308
 0.708285  0.924962
 0.747341  0.156994

For simulate_copula all mentioned below copulas are supported.

Given U the preallocated matrix of Float64 it can be filled by size(U,1) sample of the copula by running:

julia> simulate_copula!(U::Matrix{Float64}, copula::TypeOfCopula; rng::AbstractRNG = Random.GLOBAL_RNG)

For simulate_copula! all mentioned Archimedean copulas (including nested and the chain) as well as the Frechet and the Marshal-Olkin copulas are supported. Number of marginals size(U,2) in the preallocated matrix must equal to these in the copula model, else the Assertionerror will be raised.

julia> u = zeros(6,3)
6×3 Array{Float64,2}:
 0.0  0.0  0.0
 0.0  0.0  0.0
 0.0  0.0  0.0
 0.0  0.0  0.0
 0.0  0.0  0.0
 0.0  0.0  0.0

julia> Random.seed!(43);

julia> c = ClaytonCopula(3, 3.)
ClaytonCopula{Float64}(3, 3.0)

julia> simulate_copula!(u, c)

julia> u
6×3 Array{Float64,2}:
 0.740919   0.936613   0.968594
 0.369025   0.698884   0.586236
 0.0701388  0.185901   0.0890538
 0.535579   0.516761   0.538476
 0.487668   0.549494   0.804122
 0.653199   0.0923366  0.387304

Elliptical copulas

Elliptical copula is derived form the multivariate elliptical distribution (such as the Gaussian or the t-Student). Suppose F(x₁, ..., xₙ) is the Cumulative Density Function (CDF) of such multivariate distribution, and Fᵢ(xᵢ) is the univariate CDF of its ith marginal (we assume it is continuous). Hence uᵢ = Fᵢ(xᵢ) is modeled by the uniform distribution on [0,1]. Given the elliptical multivariate distribution, the elliptical copula is: C(u₁, ..., uₙ) = F(F₁⁻¹(u₁), ..., Fₙ⁻¹(uₙ)).

The Gaussian copula

julia> GaussianCopula::Matrix{Float64})

The Gaussian copula is parameterized by the correlation matrix Σ that needs to be symmetric, positively defined and with ones on the diagonal. The number of marginals is given by the size of Σ.

julia> GaussianCopula([1. 0.5; 0.5 1.])
GaussianCopula{Float64}([1.0 0.5; 0.5 1.0], 2)

The t-Student copula

julia> StudentCopula::Matrix{Float64}, ν::Int)

The t-Student copula is parameterized by the Σ matrix a in the Gaussian copula case, and by the integer parameter ν > 0 interpreted as the number of degrees of freedom. The number of marginals is given by the size of Σ.

julia> StudentCopula([1. 0.5; 0.5 1.], 1)
StudentCopula{Float64}([1.0 0.5; 0.5 1.0], 1, 2)

The Marshall-Olkin copula

The Marshall-Olkin copula is derived form the Marshall-Olkin exponential distribution with positively valued parameters. The Marshall-Olkin copula models the dependency between the random variables subjected to external shocks. The shock connected with the single variable is modeled there by λₖ, while the shock connected with two variables by λₖₗ, etc...

julia> MarshallOlkinCopula::Vector{Float64})

Parameters are ordered as follow in the argument vector λ = [λ₁, λ₂, ..., λₙ, λ₁₂, λ₁₃, ..., λ₁ₙ, λ₂₃, ..., λₙ₋₁ₙ, λ₁₂₃, ..., λ₁₂...ₙ], all must be non-negative. The number of marginals of such implemented Marshal-Olkin copula is n = ceil(Int, log(2, length(λ)-1)).

To generate data from the Marshall-Olkin copula we use algorithm presented in P. Embrechts, F. Lindskog, A McNeil 'modeling Dependence with Copulas and Applications to Risk Management', 2001.

julia> c = MarshallOlkinCopula([1., 2., 3.])
MarshallOlkinCopula{Float64}(2, [1.0, 2.0, 3.0])

julia> Random.seed!(43);

julia> 5×2 Matrix{Float64}:
 0.854724   0.821831
 0.885202   0.858624
 0.471677   0.244436
 0.834864   0.356275
 0.0661758  0.033564

The Frechet copula

The two parameters Frechet copula is C(u₁, u₂) = α C_{max}(u₁, u₂) + β C_{min}(u₁, u₂) + (1- α - β) C_{⟂}(u₁, u₂). Here C_{max}(u₁, u₂) yields maximal 1 cross-correlation, while C_{min}(u₁, u₂) minimal -1 cross correlation. The C_{min}(u₁, u₂) is the copula only in the bivariate case. Obviously we require 0 ≤ α ≤ 1 , where 0 ≤ β ≤ 1 and 0 ≤ 1-α - β ≤ 1.

julia> FrechetCopula(n::Int, α::Float64, β::Float64)

is supported only for n = 2.

julia> c = FrechetCopula(2, 0.4, 0.4)
FrechetCopula{Float64}(2, 0.4, 0.4)

julia> Random.seed!(43);

julia> simulate_copula(5, c)
5×2 Matrix{Float64}:
 0.180975   0.775377
 0.924876   0.408278
 0.599463   0.400537
 0.661781   0.661781
 0.0950087  0.0950087

The one parameter Frechet copula C(u₁, ..., uₙ) = α C_{max}(u₁, ..., uₙ) + (1-α) C_{⟂}(u₁, ..., uₙ), where 0 ≤ α ≤ 1 is supported for any n ≥ 2.

julia> FrechetCopula(n::Int, α::Float64)
julia> c = FrechetCopula(3, 0.4)
FrechetCopula{Float64}(3, 0.4, 0.0)

julia> Random.seed!(43);

julia> simulate_copula(5, c)
5×3 Matrix{Float64}:
 0.180975    0.775377   0.888934
 0.408278    0.912603   0.828727
 0.661781    0.661781   0.661781
 0.125437    0.0950087  0.130474
 0.00463791  0.0288987  0.521601

The Archimedean copulas

The bivariate Archimedean copula C(u₁,u₂) = φ⁻¹(φ(u₁)+φ(u₂)) is defined by the continuous strictly decreasing generator function φ(t) parameterized by θ. Such generator must fulfill φ(t): [0,1] →[0, ∞). The n-variate Archimedean copula can be defined analogically: C(u₁,..., uₙ) = φ⁻¹(φ(u₁)+...+φ(uₙ)). Here the constrains for the θ parameter are more strict, see: M. Hofert, 'Sampling Archimedean copulas', Computational Statistics & Data Analysis, 52 (2008), 5163-5174.

Following Archimedean copulas are supported in the module:

  • Clayton copula - parameter domain: θ ∈ (0, ∞) for n > 2 and θ ∈ [-1, 0) ∪ (0, ∞) for n = 2,
julia> ClaytonCopula(n::Int, θ::Float64)
  • Frank copula - parameter domain: θ ∈ (0, ∞) for n > 2 and θ ∈ (-∞, 0) ∪ (0, ∞) for n = 2,
julia> FrankCopula(n::Int, θ::Float64)
  • Gumbel copula - parameter domain: θ ∈ [1, ∞),
julia> GumbelCopula(n::Int, θ::Float64)
  • Ali-Mikhail-Haq copula - parameter domain: θ ∈ (0, 1) for n > 2 and θ ∈ [-1, 1] for n = 2
julia> AmhCopula(n::Int, θ::Float64)

For implemented sampling algorithms see as well P. Kumar, 'Probability Distributions and Estimation of Ali-Mikhail-Haq Copula', Applied Mathematical Sciences, Vol. 4, 2010, no. 14, 657 - 666; and R. B. Nelsen, 'An introduction to copulas', Springer Science & Business Media (2007).

julia> c = ClaytonCopula(3, 3.)
ClaytonCopula{Float64}(3, 3.0)

julia> Random.seed!(43);

julia> simulate_copula(5, c)
5×3 Array{Float64,2}:
 0.740919   0.936613  0.968594
 0.369025   0.698884  0.586236
 0.0701388  0.185901  0.0890538
 0.535579   0.516761  0.538476
 0.487668   0.549494  0.804122  

The optional third empty type <: CorrelationType parameter is used to compute θ from the expected Kendall - KendallCorrelation or Speraman SpearmanCorrelation cross-correlation. Here only positive correlations are supported, and there are some limitations are for the Ali-Mikhail-Haq copula due to limitations on θ there.

julia> c = ClaytonCopula(3, 0.5, KendallCorrelation)
ClaytonCopula{Float64}(3, 2.0)

julia> x = simulate_copula(500_000, c);

julia> corkendall(x)
3×3 Array{Float64,2}:
 1.0       0.500576  0.499986
 0.500576  1.0       0.501574
 0.499986  0.501574  1.0      
julia> c = ClaytonCopula(3, 0.5, SpearmanCorrelation)
ClaytonCopula{Float64}(3, 1.0760904048732394)

julia> x = simulate_copula(500_000, c);

julia> corspearman(x)
3×3 Array{Float64,2}:
 1.0       0.499662  0.499637
 0.499662  1.0       0.500228
 0.499637  0.500228  1.0   

The reversed Gumbel, Clayton and Ali-Mikhail-Haq copulas are supported as well:

julia> GumbelCopulaRev(n::Int, θ::Float64)
julia> ClaytonCopulaRev(n::Int, θ::Float64)
julia> AmhCopulaRev(n::Int, θ::Float64)

The reversed copula is introduced by the following transformation ∀ᵢ uᵢ → 1-uᵢ. For modeling justification see: K. Domino, T. Błachowicz, M. Ciupak, 'The use of copula functions for predictive analysis of correlations between extreme storm tides', Physica A: Statistical Mechanics and its Applications 413, 489-497, (2014); and K. Domino, T. Błachowicz, 'The use of copula functions for modeling the risk of investment in shares traded on the Warsaw Stock Exchange', Physica A: Statistical Mechanics and its Applications 413, 77-85, (2014).

julia> c = ClaytonCopulaRev(2, 5.)
ClaytonCopulaRev{Float64}(2, 5.0)

julia> Random.seed!(43);

julia> simulate_copula(10, c)
10×2 Array{Float64,2}:
 0.246822   0.0735546
 0.0214448  0.154414
 0.453721   0.598829
 0.87328    0.91861  
 0.896485   0.899053
 0.966261   0.981044
 0.0372783  0.0100412
 0.899013   0.758491
 0.473352   0.334147
 0.0438898  0.256301

The nested Archimedean copulas

The Nested Archimedean copula is C_θ(C_ϕ₁(u₁₁, ..., u₁,ₙ₁), ..., C_ϕₖ(uₖ₁, ..., uₖ,ₙₖ), u₁ , ... uₘ). Here θ is the parameter of the parent copula while ϕᵢ is the parameter of the child copula. If m > 0, some random variables will be modeled by the parent copula only. The example is:

julia> NestedClaytonCopula(childred::Vector{ClaytonCopula}, m::Int, θ::Float64)
julia> a = ClaytonCopula(2, 3.)
ClaytonCopula{Float64}(2, 3.0)

julia> b = ClaytonCopula(2, 4.)
ClaytonCopula{Float64}(2, 4.0)

julia> NestedClaytonCopula([a,b], 0, 1.)
NestedClaytonCopula{Float64}(ClaytonCopula{Float64}[ClaytonCopula{Float64}(2, 3.0), ClaytonCopula{Float64}(2, 4.0)], 0, 1.0, 4)

Only the nesting within the same family is supported. The sufficient nesting condition requires parameters of the children copulas to be larger than the parameter of the parent copula. For sampling one uses the algorithm form McNeil, A.J., 'Sampling nested Archimedean copulas', Journal of Statistical Computation and Simulation 78, 567–581 (2008).

julia> a = ClaytonCopula(2, 0.9, KendallCorrelation)
ClaytonCopula{Float64}(2, 18.000000000000004)

julia> b = NestedClaytonCopula([a], 1, .2, KendallCorrelation)
NestedClaytonCopula{Float64}(ClaytonCopula{Float64}[ClaytonCopula{Float64}(2, 18.000000000000004)], 1, 0.5, 3)

julia> x = simulate_copula(500000, b);

julia> corkendall(x)
3×3 Array{Float64,2}:
 1.0       0.899927  0.201108
 0.899927  1.0       0.201252
 0.201108  0.201252  1.0    

For the Gumbel copula the double nesting is supported. Double Nested copula is: C_θ(C_ϕ₁(C_Ψ₁₁(u,...), ..., C_C_Ψ₁,ₗ₁(u...)), ..., C_ϕₖ(C_Ψₖ₁(u,...), ..., C_Ψₖ,ₗₖ(u,...))). These are in the following form.

julia> DoubleNestedGumbelCopula(children::Vector{NestedGumbelCopula}, θ)
julia> a = GumbelCopula(2, 2.)
GumbelCopula{Float64}(2, 2.0)

julia> b = GumbelCopula(2, 3.)
GumbelCopula{Float64}(2, 3.0)

julia> c = GumbelCopula(2, 4.)
GumbelCopula{Float64}(2, 4.0)

julia> p = NestedGumbelCopula([a,b], 0, 1.75)
NestedGumbelCopula{Float64}(GumbelCopula{Float64}[GumbelCopula{Float64}(2, 2.0), GumbelCopula{Float64}(2, 3.0)], 0, 1.75, 4)

julia> p1 = NestedGumbelCopula([c], 1, 1.5)
NestedGumbelCopula{Float64}(GumbelCopula{Float64}[GumbelCopula{Float64}(2, 4.0)], 1, 1.5, 3)

julia> gp = DoubleNestedGumbelCopula([p, p1], 1.2)
DoubleNestedGumbelCopula{Float64}(NestedGumbelCopula{Float64}[NestedGumbelCopula{Float64}(GumbelCopula{Float64}[GumbelCopula{Float64}(2, 2.0), GumbelCopula{Float64}(2, 3.0)], 0, 1.75, 4), NestedGumbelCopula{Float64}(GumbelCopula{Float64}[GumbelCopula{Float64}(2, 4.0)], 1, 1.5, 3)], 1.2, 7)


julia> Random.seed!(43);

julia> simulate_copula(2, gp)
2×7 Array{Float64,2}:
 0.103462  0.358534  0.068492  0.0914353  0.90365   0.861869  0.0716466
 0.755824  0.946489  0.745881  0.916382   0.448706  0.354352  0.676657

Another example

julia> a = GumbelCopula(2, .9, KendallCorrelation)
GumbelCopula{Float64}(2, 10.000000000000002)

julia> b = GumbelCopula(2, 0.8, KendallCorrelation)
GumbelCopula{Float64}(2, 5.000000000000001)

julia> p = NestedGumbelCopula([a], 1, 0.6, KendallCorrelation)
NestedGumbelCopula(GumbelCopula[GumbelCopula(2, 10.000000000000002)], 1, 2.5)

julia> p1 = NestedGumbelCopula([b], 1, 0.5, KendallCorrelation)
NestedGumbelCopula{Float64}(GumbelCopula{Float64}[GumbelCopula{Float64}(2, 10.000000000000002)], 1, 2.5, 3)

julia> pp = DoubleNestedGumbelCopula([p,p1], 0.1, KendallCorrelation)
DoubleNestedGumbelCopula{Float64}(NestedGumbelCopula{Float64}[NestedGumbelCopula{Float64}(GumbelCopula{Float64}[GumbelCopula{Float64}(2, 10.000000000000002)], 1, 2.5, 3), NestedGumbelCopula{Float64}(GumbelCopula{Float64}[GumbelCopula{Float64}(2, 4.0)], 1, 1.5, 3)], 1.1111111111111112, 6)


julia> x = simulate_copula(750_000, pp);

julia> corkendall(x)
6×6 Array{Float64,2}:
 1.0        0.90994    0.599545   0.0981701  0.0982193  0.099066
 0.90994    1.0        0.599713   0.0981766  0.0982359  0.0990366
 0.599545   0.599713   1.0        0.099002   0.0989463  0.0991464
 0.0981701  0.0981766  0.099002   1.0        0.819611   0.499707
 0.0982193  0.0982359  0.0989463  0.819611   1.0        0.499743
 0.099066   0.0990366  0.0991464  0.499707   0.499743   1.0    

Hierarchical nested Gumbel copula is supported as well C_θₙ₋₁(C_θₙ₋₂( ... C_θ₂(C_θ₁(u₁, u₂), u₃) ,..., uₙ₋₁)uₙ). Here bivariate Gumbel copulas are nested one in the another. The most inner is the ground ... ground child one and the most outer is the ground ... ground parent one. Numbel of marginals is n = length(θ)+1.

julia> HierarchicalGumbelCopula::Vector{Float64})

Here θ is the parameter vector, starting form the ground ... ground child one and ending on the ground ... ground parent one. Hence elements of θ must be sorted in the descending order.

julia> c = HierarchicalGumbelCopula([5., 4., 3.])
HierarchicalGumbelCopula{Float64}(4, [5.0, 4.0, 3.0])

julia> Random.seed!(43);

julia> simulate_copula(3, c)
3×4 Array{Float64,2}:
 0.100353  0.207903  0.0988337  0.0431565
 0.347417  0.217052  0.223734   0.042903
 0.73617   0.347349  0.168348   0.410963
julia> c = HierarchicalGumbelCopula([.9, 0.7, 0.5, 0.1], KendallCorrelation)
HierarchicalGumbelCopula{Float64}(5, [10.000000000000002, 3.333333333333333, 2.0, 1.1111111111111112])

julia> x = simulate_copula(750_000, c);

julia> corkendall(x)
5×5 Array{Float64,2}:
 1.0       0.900078  0.700316   0.499483   0.100431
 0.900078  1.0       0.700184   0.499376   0.100567
 0.700316  0.700184  1.0        0.499635   0.0999502
 0.499483  0.499376  0.499635   1.0        0.0999432
 0.100431  0.100567  0.0999502  0.0999432  1.0  

The chain of bivariate copulas

The chain of the bivariate copulas is determined by the sequence of bivariate copulas C₁, C₂, ..., Cₙ₋₁ where each model the subsequent pair of marginals Cₖ(uₖ, uₖ₊₁). Hence the cross-correlation is introduced locally and decreases as the distance between marginals grows.

The chain of bivariate Archimedean copulas

In this case, each element of the copula chain is the Archimedean copula (Clayton, Frank and Ali-Mikhail-Haq families are supported). Hence the chain is parameterized by the parameters vector θ (with parameters domains as in the case of bivariate copulas) and the vector of string determining the copula family. Following families are supported: Clayton - key: "clayton", Frank - key: "frank" and Ali-Mikhail-Haq - key: "amh". The ith element of the vector θ (and the ith element of the string with copulas names) determine the cross-correlation between the ith and the i+1th marginal. Number of marginals is n = length(θ)+1

julia> ChainArchimedeanCopulas::Vector{Float64}, copulas::Vector{String})

if one want to use one copula type, use

julia> ChainArchimedeanCopulas::Vector{Float64}, copulas::String)
julia> ChainArchimedeanCopulas([2., 3.], ["clayton", "frank"])
ChainArchimedeanCopulas(3, [2.0, 3.0], ["clayton", "frank"])
julia> ChainArchimedeanCopulas([2., 3.], "frank")
ChainArchimedeanCopulas(3, [2.0, 3.0], ["frank", "frank"])
julia> c = ChainArchimedeanCopulas([0.7, 0.5, 0.7], "clayton", KendallCorrelation)
ChainArchimedeanCopulas{Float64}(4, [4.666666666666666, 2.0, 4.666666666666666], ["clayton", "clayton", "clayton"])

julia> x = simulate_copula(750_000, c);

julia> corkendall(x)
4×4 Array{Float64,2}:
 1.0       0.699611  0.443936  0.399355
 0.699611  1.0       0.499728  0.443053
 0.443936  0.499728  1.0       0.699855
 0.399355  0.443053  0.699855  1.0       

Negative correlations are supported here as well:

julia> c = ChainArchimedeanCopulas([0.7, 0.5, -0.7], "clayton", KendallCorrelation)
ChainArchimedeanCopulas{Float64}(4, [4.666666666666666, 2.0, -0.8235294117647058], ["clayton", "clayton", "clayton"])

julia> x = simulate_copula(750_000, c);

julia> corkendall(x)
4×4 Array{Float64,2}:
  1.0        0.69992    0.445151  -0.372628
  0.69992    1.0        0.500511  -0.413707
  0.445151   0.500511   1.0       -0.699821
 -0.372628  -0.413707  -0.699821   1.0    

The chain of bivariate Frechet copulas

Here, each bivariate copula is the two parameters Frechet one Cₖ = C_{αₖ,βₖ}(uₖ, uₖ₊₁), where αₖ and βₖ are elements of parameter vectors α and β that must be of the equal size. Number of marginals in n = length(α) = length(β).

julia> c = ChainFrechetCopulas(α, β)
julia> c = ChainFrechetCopulas([0.2, 0.3], [0.5, 0.1])
ChainFrechetCopulas{Float64}(3, [0.2, 0.3], [0.5, 0.1])

julia> Random.seed!(43);

julia> simulate_copula(3, c)
3×3 Array{Float64,2}:
 0.828727  0.171273  0.180975
 0.400537  0.408278  0.775377
 0.429437  0.912603  0.888934

Correlation matrix generation

We supply a few methods to generate a n x n correlation matrix Σ.

Fully random cases

to generate randomly a correlation matrix run

julia> cormatgen(n::Int)

or

julia> cormatgen_rand(n::Int)

For different methods we have different outputs

julia> Random.seed!(43);

julia> cormatgen(4)
4×4 Array{Float64,2}:
 1.0       0.396865  0.339354  0.193335
 0.396865  1.0       0.887028  0.51934
 0.339354  0.887028  1.0       0.551519
 0.193335  0.51934   0.551519  1.0     

julia> cormatgen_rand(4)
4×4 Array{Float64,2}:
 1.0       0.659183  0.916879  0.486979
 0.659183  1.0       0.676167  0.808264
 0.916879  0.676167  1.0       0.731206
 0.486979  0.808264  0.731206  1.0  

In general the second case gives higher values of correlations.

Deterministic cases

To generate a correlation matrix with constant elements run

julia> cormatgen_constant(n::Int, α::Float64)

parameter α should satisfy 0 <= α <= 1

julia> cormatgen_constant(4, 0.4)
4×4 Array{Float64,2}:
 1.0  0.4  0.4  0.4
 0.4  1.0  0.4  0.4
 0.4  0.4  1.0  0.4
 0.4  0.4  0.4  1.0

the generalisation is with two parameters 0 <= α <= 1 and α > β

julia> cormatgen_two_constant(n::Int, α::Float64, β::Float64)
julia> cormatgen_two_constant(4, 0.5, 0.2)
4×4 Array{Float64,2}:
 1.0  0.5  0.2  0.2
 0.5  1.0  0.2  0.2
 0.2  0.2  1.0  0.2
 0.2  0.2  0.2  1.0

Here the first constant refer to the "nesting" of the higher correlation for the first half of marginals. To generate the Toeplitz matrix with parameter 0 <= ρ <= 1 run:

julia> cormatgen_toeplitz(n::Int, ρ::Float64)

julia> cormatgen_toeplitz(4, 0.5)
4×4 Array{Float64,2}:
 1.0    0.5   0.25  0.125
 0.5    1.0   0.5   0.25
 0.25   0.5   1.0   0.5  
 0.125  0.25  0.5   1.0  

Partially random and partially deterministic cases

To generate constant matrix with the noise run

julia> cormatgen_constant_noised(n::Int, α::Float64; ϵ::Float64 = (1.-α)/2.)

where the parameter ϵ must satisfy 0 <= ϵ <= 1-α

julia> Random.seed!(43);

julia> cormatgen_constant_noised(4, 0.5)
4×4 Array{Float64,2}:
 1.0       0.314724  0.590368  0.346992
 0.314724  1.0       0.314256  0.512183
 0.590368  0.314256  1.0       0.538089
 0.346992  0.512183  0.538089  1.0   

Analogically to generate noised two constants matrix run

julia> Random.seed!(43);

julia> cormatgen_two_constant_noised(4, 0.8, 0.2)
4×4 Array{Float64,2}:
 1.0       0.754384  0.194805  0.171162
 0.754384  1.0       0.117009  0.27321
 0.194805  0.117009  1.0       0.139476
 0.171162  0.27321   0.139476  1.0    

Here the first constant refer to the "nesting" of the higher correlation for the first half of marginals. Finally to generate noised Toeplitz matrix run:

julia> cormatgen_toeplitz_noised(n::Int, ρ::Float64; ϵ=(1-ρ)/(1+ρ)/2)

where the parameter ϵ must satisfy 0 <= ϵ <= (1-ρ)/(1+ρ)

julia> Random.seed!(43);

julia> cormatgen_toeplitz_noised(4, 0.5)
4×4 Array{Float64,2}:
 1.0        0.376483  0.310246  0.0229948
 0.376483   1.0       0.376171  0.258122
 0.310246   0.376171  1.0       0.525393
 0.0229948  0.258122  0.525393  1.0  

Changes the subset of marginals of multivariate Gaussian distributed data

To change the chosen marginals subset, determined by the vector of indicesind, of the multivariate Gaussian distributed data x by means of t-Student sub-copula with a parameter ν run:

julia> gcop2tstudent(x::Matrix{Float64}, ind::Vector{Int}, ν::Int; naive::Bool = false, rng = Random.GLOBAL_RNG)

all univariate marginal distributions will be Gaussian, hence unaffected by the transformation. The keyword naive means the naive resampling if true. Custom random number generator is supported.

julia> Σ = [1. 0.5 0.5; 0.5 1. 0.5; 0.5 0.5 1];

julia> Random.seed!(42);

julia> x = Array(rand(MvNormal(Σ), 6)')
6×3 Array{Float64,2}:
 -0.556027  -0.662861   -0.384124
 -0.299484   1.38993    -0.571326
 -0.468606  -0.0990787  -2.3464  
  1.00331    1.43902     0.966819
  0.518149   1.55065     0.989712
 -0.886205   0.149748   -1.54419

julia> gcop2tstudent(x, [1,2], 6)
6×3 Array{Float64,2}:
 -0.519458  -0.498377   -0.384124
 -0.37937    1.66806    -0.571326
 -0.432902  -0.0178933  -2.3464  
  1.01216    1.50814     0.966819
  0.226484   1.12436     0.989712
 -0.727203   0.238701   -1.54419

To change the chosen marginals subset of the multivariate Gaussian distributed data x by means of the Archimedean copula run:

julia> gcop2arch(x::Matrix{Float64}, inds::Vector{Pair{String,Vector{Int64}}}; naive::Bool = false, notnested::Bool = false, rng = Random.GLOBAL_RNG)

Marginals to be changed are list in inds[i][2], while the corresponding Archimedean copula is determined in inds[i][1]. Many disjoint subsets of marginals with different Archimedean copulas can be transformed. All univariate marginal distributions are Gaussian hence unaffected by the transformation. The keyword naive indicates the use of the naive data resampling if true. The keyword notnested if true indicates the use of one parameter Archimedean copula instead of a nested one. Custom random number generator is supported.

julia> Σ = [1. 0.5 0.5; 0.5 1. 0.5; 0.5 0.5 1];

julia> Random.seed!(42)

julia> x = Array(rand(MvNormal(Σ), 6)')
6×3 Array{Float64,2}:
 -0.556027  -0.662861   -0.384124
 -0.299484   1.38993    -0.571326
 -0.468606  -0.0990787  -2.3464
  1.00331    1.43902     0.966819
  0.518149   1.55065     0.989712
 -0.886205   0.149748   -1.54419

julia> gcop2arch(x, ["clayton" => [1,2]])
6×3 Array{Float64,2}:
 -0.742443   0.424851  -0.384124
  0.211894   0.195774  -0.571326
 -0.989417  -0.299369  -2.3464
  0.157683   1.47768    0.966819
  0.154893   0.893253   0.989712
 -0.657297  -0.339814  -1.54419


julia> Random.seed!(42);

julia> x = Array(rand(MvNormal(Σ), 6)');

julia> gcop2arch(x, ["gumbel" => [1,2]])
6×3 Array{Float64,2}:
0.178913   1.60797    -0.384124
0.579476   0.880272   -0.571326
-0.986662  -0.0180474  -2.3464  
1.20299    2.55397     0.966819
0.857086   1.86212     0.989712
-0.548206  -0.439289   -1.54419

To change the chosen marginals subset list in ind of the multivariate Gaussian distributed data x by means of the Frechet maximal copula run:

julia> gcop2frechet(x::Matrix{Float64}, ind::Vector{Int}; naive::Bool = false)

all univariate marginal distributions are Gaussian as they are unaffected by the transformation. The keyword naive means naive resampling if true.

julia> Σ = [1. 0.5 0.5; 0.5 1. 0.5; 0.5 0.5 1];

julia> Random.seed!(42)

julia> x = Array(rand(MvNormal(Σ), 6)');

julia> gcop2frechet(x, [1,2])
6×3 Array{Float64,2}:
 -0.875777   -0.374723   -0.384124
  0.0960334   0.905703   -0.571326
 -0.599792   -0.0110945  -2.3464
  0.813717    1.8513      0.966819
  0.599255    1.56873     0.989712
 -0.7223     -0.172507   -1.54419

To change the chosen marginals subset list in ind of themultivariate Gaussian distributed data x by means of the bivariate Marshall-Olkin copula run:

julia> gcop2marshallolkin(x::Matrix{Float64}, ind::Vector{Int}, λ1::Float64 = 1., λ2::Float64 = 1.5; naive::Bool = false, rng = Random.GLOBAL_RNG)

all univariate marginal distributions are Gaussian and unaffected by the transformation. In the keyword naive is true uses the naive resampling. The algorithm requires length(ind) = 2 λ₁ ≥ 0 and λ₂ ≥ 0. The parameter λ₁₂ is computed from expected correlation between both changed marginals. Custom random number generator is supported.

julia> Σ = [1. 0.5 0.5; 0.5 1. 0.5; 0.5 0.5 1];

julia> Random.seed!(42);

julia> x = Array(rand(MvNormal(Σ), 6)');

julia> gcop2marshallolkin(x, [1,2])
6×3 Array{Float64,2}:
 -0.790756   0.784371  -0.384124
 -0.28088    0.338086  -0.571326
 -0.90688   -0.509684  -2.3464  
  0.738628   1.71026    0.966819
  0.353654   1.19357    0.989712
 -0.867606  -0.589929  -1.54419

Helpers

Converting marginals

Takes matrix X: size(X) = (t, n) ie t realisations of n-dimensional random variable, with all uniform marginal univariate distributions ∀ᵢ X[:,i] ∼ Uniform(0,1), and convert those marginals to the common distribution d with parameters p[i]

julia> convertmarg!(x::Matrix{T}, d::UnionAll, p::Union{Vector{Vector{Int64}}, Vector{Vector{Float64}}}; testunif::Bool = true)

If testunif = true each marginal is tested for uniformity.

julia> c = GaussianCopula([1. 0.5; 0.5 1.])

julia> Random.seed!(43);

julia> x = simulate_copula(10, c);


julia> convertmarg!(x, Normal, [[0, 1],[0, 10]])

julia> x
10×2 Array{Float64,2}:
10×2 Matrix{Float64}:
  0.225457      8.97627
  0.548381     14.3926
  0.666147    -10.0689
 -0.746662     -9.03553
 -0.746857     17.2101
 -0.608109     -3.45649
 -0.136555      0.700419
  0.215631     -7.34409
 -0.00352701   -0.434793
 -0.876853      2.39009

To convert i th marginal to univariate distribution d with parameters array p run

julia> using Distributions

julia> quantile.(d(p...), x[:,i])
julia> c = GaussianCopula([1. 0.5; 0.5 1.])

julia> Random.seed!(43);

julia> x = simulate_copula(10, c);

julia> quantile.(Levy(0, 1), x[:,2])
10-element Array{Float64,1}:
  18.327904335047272
 112.72788160148863  
   0.4992650891811052
   0.5642861403809334
 350.0676959136128   
   1.2175971128674394
   2.510078079677982
   0.6980591543550244
   2.0290242635860944
   3.527994542141473

To convert all marginals to the same d with the same parameters p run

julia> using Distributions

julia> quantile.(d(p...), x)
julia> quantile.(Levy(0, 1), x)
10×2 Array{Float64,2}:
 3.42919    18.3279  
 7.14305   112.728   
 9.6359      0.499265
 0.687009    0.564286
 0.686835  350.068   
 0.827224    1.2176  
 1.71944     2.51008
 3.3597      0.698059
 2.18374     2.02902
 0.582946    3.52799

BigFloat implementation, developement version.

For some copulas: Marshall-Olkin, Frechet, Gumbel, Frank, Ali-Mikhail-Haq,and Nested Gumbel the BigFloat implementation is supported. However it is a development version that requires enhancement of other Julia packages.

julia> θ = BigFloat(2.);

julia> Random.seed!(43)
MersenneTwister(43)

julia> simulate_copula(3, c)
3×3 Matrix{BigFloat}:
 0.201305  0.277557  0.387158
 0.395095  0.536081  0.152753
 0.612178  0.39509   0.426144

Citing this work

This project was partially financed by the National Science Centre, Poland – project number 2014/15/B/ST6/05204;

  • while reffering to gcop2arch(), gcop2frechet(), and gcop2marshallolkin() - please cite K. Domino, A. Glos: 'Introducing higher order correlations to marginals' subset of multivariate data by means of Archimedean copulas', [arXiv:1803.07813] (https://arxiv.org/abs/1803.07813);

  • while reffering to gcop2tstudent() - please cite K. Domino: 'Multivariate cumulants in outlier detection for financial data analysis', Physica A: Statistical Mechanics and its Applications Volume 558, 15 November 2020, 124995 (https://doi.org/10.1016/j.physa.2020.124995);

  • you may also cite K. Domino: 'Selected Methods for non-Gaussian Data Analysis', Gliwice, IITiS PAN, 2019, ISBN: 978-83-926054-3-0, [arXiv:1811.10486] (https://arxiv.org/abs/1811.10486)

Used By Packages

No packages found.