A C++ vine copula library
C++ CMake Other
Clone or download
tvatter and tnagler Release 0.2.8 (#330)
* add mcor to tree_criterion docs (#312)

* add mcor to tree_criterion docs

* add missing $ (Vinecop wasn't displayed)

* Add getters for Kendall's tau (#319)

* add get_tau() method to Bicop class

* add get_tau() and get_all_taus() method to Vinecop class

* use MatrixXd consistently for pair copula parameters

* add inline prefix to Vinecop::get_tau()

* proper truncation of pdf values (#320)

* Minor fixes for Eigen tools (#321)

* use const reference in *_or_nan functions

* fix indentation in invert_f()

* increase search interval when initial fit is unreasonable (#322)

* lower upper bound for 2nd bb7 parameter (#324)

* ensure that boundaries are respected for Joe's hinv (#323)

* make mcor correction less agressive (#326)

* lower upper bound for joe parameter to avoid nans in hfunc (#325)

* fix travis (#328)

* try installing RcppArmadillo

* fix travis

* fix osx

* try fix ubuntu

* try fix ubuntu +1

* Loglik (#327)

* add bicop loglik

* add loglik for vinecop

* use NAN instead of 0 as default

* properly handle BicopFamily::indep

* clean-up

* Prepare 0.2.8 (#329)

* bump version

* update NEWS

* bump date

* update NEWS +1

* update NEWS +2

* update NEWS +3

* update NEWS +4

* fix loglik-related bugs (#331)

* throwing the get_loglik()/non-fitted error from bicop class (#332)

* move the error throwing to bicop class

* typo

* use boost isnan (#333)
Latest commit c53aa33 May 11, 2018

README.md

vinecopulib

Build Status Windows Build status Coverage Status License: MIT

What are vine copulas?

Vine copulas are a flexible class of dependence models consisting of bivariate building blocks (see e.g., Aas et al., 2009). You can find a comprehensive list of publications and other materials on vine-copula.org.

What is vinecopulib?

vinecopulib is a header-only C++ library for vine copula models based on Eigen. It provides high-performance implementations of the core features of the popular VineCopula R library, in particular inference algorithms for both vine copula and bivariate copula models. Advantages over VineCopula are

  • a stand-alone C++ library with interfaces to both R and Python,
  • a sleaker and more modern API,
  • shorter runtimes and lower memory consumption, especially in high dimensions,
  • nonparametric and multi-parameter families.

Status

Version 0.2.8 was released on May 4, 2018. While we did our best to design a user-friendly API, the library is still under active development and changes are to be expected. We are also working on interfaces for R and Python.

Contact

If you have any questions regarding the library, feel free to open an issue or send a mail to info@vinecopulib.org.

Documentation

Below, we give a brief overview of the most important functionality. The full set of classes and methods can be found in the API documentation.


Getting started

Requirements

To build the library, you'll need at minimum:

Optionally, you'll need:

Note that:

  • The C++11 thread support library, available along with any C++11 compiler on OSX/Windows/most-linux-distributions, is used for multithreading.
  • A findR.cmake looks for R and VineCopula in the default locations for linux and osx, but problems might occur with versions installed from R/RStudio. Therefore, prior to building the library, it is recommended to use:

sudo Rscript -e 'install.packages(c("VineCopula"), lib="/usr/lib/R/library", repos="http://cran.rstudio.com/")'

How to build the library

By default, vinecopulib is header-only. It means that we use the CMake build system, but only to build the documentation and unit-tests, and to automate installation (i.e., place headers in the usual location). If you just want to use vinecopulib, you can use the header files (located in theincludesfolder) right away.

The unix one liner (from the root folder):

mkdir build && cd build && cmake .. && make && make doc && sudo make install && bin/test_all

Alternatively, we provide an option to precompile compiled the library in order to save building time (and memory) via the CMake option VINECOPULIB_SHARED_LIB. In this case, source files are generated from header files and the CMake build system additionally allows to install the .dylib/.so/.dll object.

The unix one liner (from the root folder):

mkdir build && cd build && cmake .. -DVINECOPULIB_SHARED_LIB=ON && make && make doc && sudo make install && bin/test_all

Step Shell command
Create a build folder mkdir build
Move to the created folder cd build
Create the MakeFile via cmake cmake .. (or cmake .. -DVINECOPULIB_SHARED_LIB=ON for the compiled version)
Compile the library make or make -j n where n is the number of cores
Build the documentation (optional) make doc
Install the library on linux/OSX (optional) sudo make install
Run unit tests (optional) bin/[test_executable]

To install the library without unit tests, the MakeFile can be created via cmake .. -DBUILD_TESTING=OFF. Additionally, a Debug mode is available via cmake .. -DCMAKE_BUILD_TYPE=Debug.

On Windows, CMake will generate Visual Studio files instead of Makefiles, the following sequence of commands can be used to perform compilation using the command prompt:

md build
cd build
cmake ..
cmake --build . --config Debug
cmake --build . --config Release
cmake --build . --config Release --target install

Instead of the cmake --build commands, the generated vinecopulib.sln file can be open in the Visual Studio GUI. Furthermore, as for linux systems, the third line can be replaced by cmake .. -DVINECOPULIB_SHARED_LIB=ON to generate the source files in order to compile vinecopulib in non-header-only mode.

The following CMake flags (given with example values) will likely come handy:

-DBOOST_ROOT=c:\local\boost_1_63_0
-DEIGEN3_INCLUDE_DIR=c:\local\eigen-eigen-da9b4e14c255
-DCMAKE_INSTALL_PREFIX=c:\local\vinecopulib-install
-DCMAKE_GENERATOR_PLATFORM=x64
-DBOOST_DEBUG=1

How to include the library in other projects

Using make install, vinecopulib is installed in the usual location of the system, namely

  • <prefix>/include/ (for the headers),
  • <prefix>/lib/ (for the shared library when VINECOPULIB_SHARED_LIB=ON is used),
  • <prefix>/lib/cmake/vinecopulib (to allow cmake to find the library with find_package),

where <prefix> is e.g. /usr/ or /usr/local. Note that make install only copies vinecopulib.hpp in <prefix>/include/ and puts the other headers in a subfolder <prefix>/include/vinecopulib, but using #include <vinecopulib.hpp> is enough to load both bivariate and vine functions.

The easiest way to include vinecopulib in another project (and to avoid writing makefiles) is to use CMake. For instance, an example projet where the source code to be linked could contain

  • a CMakeLists.txt file for the project's setup,
  • a subfolder src for the source code, containing
    • the source code,
    • another CMakeLists.txt file for the project libraries and executables.

The top-level CMakeLists.txt could be:

Example

cmake_minimum_required(VERSION 3.2)

set(CMAKE_CXX_STANDARD 11)

project (Example)

# Setting default folders
set(CMAKE_RUNTIME_OUTPUT_DIRECTORY ${CMAKE_SOURCE_DIR}/bin)

# C++ compile flags
if (NOT WIN32)
  set(CMAKE_CXX_FLAGS "-std=gnu++11 -Wextra -Wall -Wno-delete-non-virtual-dtor -Werror=return-type -O2 -DNDEBUG")
endif()

# Find vinecopulib package and dependencies
find_package(vinecopulib                  REQUIRED)
find_package(Boost 1.56                   REQUIRED)
include(cmake/findEigen3.cmake            REQUIRED)
find_package(Threads                      REQUIRED)

# Set required variables for includes and libraries
# In the second line
#   * VINECOPULIB_LIBRARIES is needed if vinecopulib has been built as a
#     shared lib (does nothing otherwise).
#   * CMAKE_THREAD_LIBS_INIT is needed for some linux systems
#     (but does nothing on OSX/Windows).
set(external_includes ${VINECOPULIB_INCLUDE_DIR} ${EIGEN3_INCLUDE_DIR} ${Boost_INCLUDE_DIRS})
set(external_libs ${VINECOPULIB_LIBRARIES} ${CMAKE_THREAD_LIBS_INIT})

# Include subdirectory with project sources
add_subdirectory(src)

Assuming a single main.cpp source file (with #include <vinecopulib.hpp> at the top), the CMakeLists.txt file in /src/ could then be:

Example

# Include header files
include_directories(${external_includes})

# Add main executable
add_executable(main main.cpp)

# Link to vinecopulib if vinecopulib has been built as a shared lib
# and to pthreads on some linux systems (does nothing otherwise)
target_link_libraries(main ${VINECOPULIB_LIBRARIES} ${CMAKE_THREAD_LIBS_INIT})

Namespaces

In the examples mentioned above, it is assumed that using namespace vinecopulib; is used. While the namespace vinecopulib contains the most important functionalities described below, there are a two others that are available to the user:

  • bicop_families: convenience definitions of sets of bivariate copula families
  • tools_stats: utilities for statistical analysis

Bivariate copula models

Bivariate copula models are implemented as the Bicop class, and BicopFamily is a closely related enum class describing the type or "family" of copula.

To use bivariate copula models in your code, include the header vinecopulib/bicop/class.hpp (or simply vinecopulib.hpp) at the top of your source file.

Implemented bivariate copula families

type name BicopFamily
- Independence indep
Elliptical Gaussian gaussian
" Student t student
Archimedean Clayton clayton
" Gumbel gumbel
" Frank frank
" Joe joe
" BB1 bb1
" BB6 bb6
" BB7 bb7
" BB8 bb8
Nonparametric Transformation kernel tll

Note that several convenience vectors of families are included in the sub-namespace bicop_families:

  • all contains all the families
  • parametric contains the parametric families (all except tll)
  • nonparametric contains the nonparametric families (indep and tll)
  • one_par contains the parametric families with a single parameter (gaussian, clayton, gumbel, frank, and joe)
  • two_par contains the parametric families with two parameters (student, bb1, bb6, bb7, and bb8)
  • elliptical contains the elliptical families
  • archimedean contains the archimedean families
  • bb contains the BB families
  • itau families for which estimation by Kendall's tau inversion is available (indep,gaussian, student,clayton, gumbel, frank, joe)

Example

// print all available families
std::cout << "Available families : ";
for (auto family : vinecopulib::bicop_families::all) {
		std::cout << get_family_name(family) << " ";
}

Set up a custom bivariate copula model

There are essentially two ways of setting-up bivariate copulas:

  • with known parameters,
  • from data (i.e., with estimated parameters).

The constructor with known parameters takes 3 arguments:

  • The copula family (default to indep)
  • The rotation (default to 0)
  • The parameters (default to parameters corresponding to an independence copula)

Example

// 90 degree rotated Clayton with default parameter (corresponds to independence)
Bicop clayton(BicopFamily::clayton, 90);

// Gauss copula with parameter 0.5
Bicop gauss(BicopFamily::gaussian, 0,  Eigen::VectorXd::Constant(1, 0.5));

The constructor from data takes the same arguments as the select method and is described in the next section.

Fit and select a bivariate copula

You can either fit the parameters of a given Bicop object with fit() or select the best fitting model from a set of families with select().

Example

// create a Gauss copula with parameter 0.5 and simulate 1e3 observations
Bicop model(BicopFamily::gaussian, 0,  Eigen::VectorXd::Constant(1, 0.5));             
auto data = model.simulate(1e3);

// instantiate a gaussian copula with default parameters and fit to data
Bicop fitted(BicopFamily::gaussian);
fitted.fit(data);
std::cout <<
    "estimated parameter: " <<
    fitted.get_parameters() <<
    std::endl;

// assign another family to the same variable and fit to data
fitted = Bicop(BicopFamily::student);
fitted.fit(data);
std::cout <<
    "estimated parameter: " <<
    fitted.get_parameters() <<
    std::endl;

// alternatively, assign to a family and fit automatically
fitted.select(data);
std::cout <<
    "family: " << fitted.get_family_name() <<
    "rotation: " <<  fitted.get_rotation() <<
    std::endl;

As it's arguably the most important function of the Bicop class, it's worth understanding the second argument of select(), namely an object of the class FitControlsBicop, which contain seven data members:

  • std::vector<BicopFamily> family_set describes the set of family to select from. It can take a user specified vector of families or any of those mentioned above (default is bicop_families::all).
  • std::string parametric_method describes the estimation method. It can take "mle" (default, for maximum-likelihood estimation) and "itau" (for Kendall's tau inversion, although only available for families included in bicop_families::itau).
  • std::string nonparametric_method describes the degree of the density approximation for the transformation kernel estimator. It can take constant, linear and quadratic (default) for approximations of degree zero, one and two.
  • double nonparametric_mult a factor with which the smoothing parameters are multiplied.
  • std::string selection_criterion describes the criterion to compare the families. It can take either "loglik", "aic", or "bic"(default).
  • bool preselect_families describes a heuristic preselection method (default is true) based on symmetry properties of the data (e.g., the unrotated Clayton won't be preselected if the data displays upper-tail dependence).
  • size_t num_threads number of threads to run in parallel when fitting several families.

As mentioned above, the arguments of select() can be used as arguments to a constructor allowing to instantiate a new object directly:

Example

// instantiate an archimedean copula by selecting the "best" family according to
// the BIC and parameters corresponding to the MLE
Bicop best_archimedean(data, FitControlsBicop(bicop_families::archimedean));
std::cout <<
    "family: " << best_archimedean.get_family_name() <<
    "rotation: " <<  best_archimedean.get_rotation() <<
    best_archimedean.get_parameters() <<
    std::endl

// instantiate a bivariate copula by selecting the "best" family according to
// the AIC and parameters corresponding to Kendall's tau inversion
FitControlsBicop controls(bicop_families::itau, "itau");
controls.set_selection_criterion("aic");
Bicop best_itau(data, controls));
std::cout <<
    "family: " << best_itau.get_family_name() <<
    "rotation: " <<  best_itau.get_rotation() <<
    best_itau.get_parameters() <<
    std::endl

Work with a bivariate copula model

You can simulate from a model and evaluate the pdf, h-functions, inverse h-functions, log-likelihood, AIC, and BIC.

Example

// Gauss copula with parameter 0.5
Bicop bicop(BicopFamily::gaussian, 0,  Eigen::VectorXd::Constant(1, 0.5));

// Simulate 100 observations
auto sim_data = bicop.simulate(100);

// Evaluate the pdf
auto pdf  = bicop.pdf(sim_data);

// Evaluate the two h-functions
auto h1   = bicop.hfunc1(sim_data);
auto h2   = bicop.hfunc2(sim_data);

// Evalute the two inverse h-functions
auto hi1  = bicop.hinv1(sim_data);
auto hi2  = bicop.hinv2(sim_data);

// Evaluate the log-likelihood, AIC, and BIC
auto ll   = bicop.loglik(sim_data);
auto aic  = bicop.aic(sim_data);
auto bic  = bicop.bic(sim_data);

Bivariate copula models can also be written to and constructed from JSON files and boost::property_tree::ptree objects:

// Gauss copula with parameter 0.5
Bicop bicop(BicopFamily::gaussian, 0,  Eigen::VectorXd::Constant(1, 0.5));

// Save as a ptree object
boost::property_tree::ptree bicop_node = bicop.to_ptree();

// Write into a JSON file
boost::property_tree::write_json("myfile.JSON", bicop_node);

// Equivalently
bicop.to_json("myfile.JSON");

// Then a new Bicop can be constructed from the ptree object
Bicop bicop2(bicop_node);

// Or from the JSON file
Bicop bicop3("myfile.JSON");

Vine copula models

Vine copula models are implemented in the class Vinecop. To use this class in your code, include the header include the header vinecopulib/vinecop/class.hpp (or simply vinecopulib.hpp) at the top of your source file. This automatically enables all features for bivariate copula models.

Set up a custom vine copula model

Custom models can be created through the constructor of Vinecop. A model is represented by a std::vector<std::vector<Bicop>> containing all pair-copulas and an R-vine matrix.

Similarly to bivariate copulas, there are essentially two ways of setting-up vine copulas:

  • with known parameters,
  • from data (i.e., with estimated parameters).

The constructor with known parameters has two versions:

  • one for which the only argument is the dimension, allowing to set-up a D-vine with only independence copulas,
  • and one for which the two arguments are a matrix of integers (i.e., and R-vine matrix) and a std::vector<std::vector<Bicop>> containing all pair-copulas.

Example

// specify the dimension of the model
int d = 3;

// instantiate a three dimensional D-vine with independence copulas
Vinecop default_model(d);

// alternatively, instantiate a std::vector<std::vector<Bicop>> object
auto pair_copulas = Vinecop::make_pair_copula_store(d);  

// specify the pair copulas
auto par = Eigen::VectorXd::Constant(1, 3.0);
for (auto& tree : pair_copulas) {
    for (auto& pc : tree) {
        pc = Bicop(BicopFamily::clayton, 270, par);
    }
}

// specify a structure matrix
Eigen::Matrix<size_t, Eigen::Dynamic, Eigen::Dynamic> mat(3, 3);
mat << 1, 1, 1,
       2, 2, 0,
       3, 0, 0;

// instantiate the custom model
Vinecop custom_model(pair_copulas, mat);

The constructors from data take the same arguments as the two select methods described below.

How to read the R-vine matrix

The R-vine matrix notation in vinecopulib is different from the one in VineCopula. An example matrix is

1 1 1 1
2 2 2 0
3 3 0 0
4 0 0 0

which encodes the following pair-copulas:

tree edge pair-copulas
0 0 (4, 1)
1 (3, 1)
2 (2, 1)
1 0 (4, 2; 1)
1 (3, 2; 1)
2 0 (4, 3; 2, 1)

Denoting by M[i, j] the matrix entry in row i and column j, the pair-copula index for edge e in tree t of a d dimensional vine is (M[d - 1 - t, e], M[t, e]; M[t - 1, e], ..., M[0, e]). Less formally,

  1. Start with the counter-diagonal element of column e (first conditioned variable).
  2. Jump up to the element in row t (second conditioned variable).
  3. Gather all entries further up in column e (conditioning set).

A valid R-vine matrix must satisfy several conditions which are checked when RVineMatrix() is called:

  1. The lower right triangle must only contain zeros.
  2. The upper left triangle can only contain numbers between 1 and d.
  3. The antidiagonal must contain the numbers 1, ..., d.
  4. The antidiagonal entry of a column must not be contained in any column further to the right.
  5. The entries of a column must be contained in all columns to the left.
  6. The proximity condition must hold: For all t = 1, ..., d - 2 and e = 0, ..., d - t - 1 there must exist an index j > d, such that (M[t, e], {M[0, e], ..., M[t-1, e]}) equals either (M[d-j-1, j], {M[0, j], ..., M[t-1, j]}) or (M[t-1, j], {M[d-j-1, j], M[0, j], ..., M[t-2, j]}).

Condition 6 already implies conditions 2-5, but is more difficult to check by hand.

Fit and select a vine copula model

The method select_all() performs parameter estimation and automatic model selection when the vine structure is unknown (i.e., it modifies the structure of the object), using the sequential procedure proposed by Dissman et al. (2013). Alternatively, select_families() performs parameter estimation and automatic model selection for a known structure (i.e., using the structure of the object). In both cases, the only mandatory argument is the data (stored in a Eigen::MatrixXd), while controls argument allow for customization of the fit.

Example

// specify the dimension of the model
int d = 5;

// simulate dummy data
Eigen::MatrixXd data = tools_stats::simulate_uniform(100, d);

// instantiate a D-vine and select the families
Vinecop model(d);
model.select_families(data);

// alternatively, select the structure along with the families
model.select_all(data);

Note that the second argument to select_all() and select_families() is similar to the one of select() for Bicop objects. Objects of the class FitControlsVinecop inherit from FitControlsBicop and extend them with additional data members to control the structure selection:

  • int truncation_level describes the tree after which family_set is set to {BicopFamily::indep}. In other words, all pair copulas in trees lower than truncation_level (default to none) are "selected" as independence copulas.
  • std::string tree_criterion describes the criterion used to construct the minimum spanning tree (see Dissman et al. (2013)). It can take "tau" (default) for Kendall's tau, "rho" for Spearman's rho, or "hoeffd" for Hoeffding's D (suited for non-monotonic relationships).
  • double threshold describes a value (default is 0) of tree_criterion under which the corresponding pair-copula is set to independence.
  • bool select_truncation_level can be set to true to select the truncation level automatically (default is false).
  • bool select_threshold can be set to true to select the threshold parameter automatically (default is false).
  • size_t num_threads number of threads to run in parallel when fitting pair copulas within one tree.

As mentioned above, the arguments of select_all() and select_families() can be used as arguments to a constructor allowing to instantiate a new object directly:

Example

// specify the dimension of the model
int d = 4;

// simulate dummy data
Eigen::MatrixXd data = simulate_uniform(100, d);

// instantiate a vine from data using the default arguments
Vinecop best_vine(data);

// alternatively, instantiate a structure matrix...
Eigen::Matrix<size_t, Eigen::Dynamic, Eigen::Dynamic> M;
M << 1, 1, 1, 1,
     2, 2, 2, 0,
     3, 3, 0, 0
     4, 0, 0, 0;

// ... and instantiate a vine copula from data using the custom structure,
// Kendall's tau inversion for parameters
// estimation and a truncation after the second tree
FitControlsVinecop controls(bicop_families::itau, "itau");
controls.set_truncation_level(2);
controls.set_num_threads(4);  // parallelize with 4 threads
Vinecop custom_vine(data, M, controls);

Work with a vine copula model

You can simulate from a vine copula model, evaluate its density, distribution, log-likelihood, AIC and BIC.

Example

// 5-dimensional independence vine
Vinecop model(5);           

// simulate 100 observations
auto data = model.simulate(100)  

// evaluate the density
auto pdf = model.pdf(data)

// evaluate the distribution
auto cdf = model.cdf(data)

// evaluate the log-likelihood
auto ll = model.loglik(data)

// evaluate the AIC
auto aic = model.aic(data)

// evaluate the BIC
auto bic = model.bic(data)

Vine copula models can also be written to and constructed from JSON files and boost::property_tree::ptree objects:

// 5-dimensional vine copula
Vinecop vinecop(5);

// Save as a ptree object
boost::property_tree::ptree vinecop_node = vinecop.to_ptree();

// Write into a JSON file
boost::property_tree::write_json("myfile.JSON", vinecop_node);

// Equivalently
vinecop.to_json("myfile.JSON");

// Then a new Bicop can be constructed from the ptree object
Vinecop vinecop2(vinecop_node);

// Or from the JSON file
Vinecop vinecop2("myfile.JSON");