pyscarcopula

A Python library for dynamic copula modelling: bivariate, multivariate, vine, and stochastic copula models for financial time series and risk analytics.

• About
• Install
• Features
• Mathematical background
• Examples and docs
• License

About

pyscarcopula fits bivariate and multivariate dependence models using copulas in Python. Alongside classical constant-parameter copulas, it supports stochastic copula autoregressive (SCAR) models where the copula parameter is driven by a latent Ornstein-Uhlenbeck process or Kendall's tau follows a bounded Jacobi diffusion.

The package is aimed at financial time series, risk modelling, and experiments with dynamic dependence. It provides bivariate copulas, C-vines, R-vines, conditional sampling, prediction, goodness-of-fit diagnostics, and risk metrics.

Supported estimation methods:

Install

pip install pyscarcopula

Official wheels include the compiled numerical extension and do not need a local compiler. Source and editable installs require a C++17 compiler:

• Windows: Microsoft C++ Build Tools / Visual Studio Build Tools, or MinGW-w64 GCC (see below)
• Linux: GCC or Clang with the usual Python development headers
• macOS: Xcode Command Line Tools

On Windows, a MinGW-w64 GCC toolchain (for example the MSYS2 ucrt64 GCC) can be used instead of the Microsoft compiler by opting in explicitly:

PYSCA_CPP_COMPILER=mingw32 pip install .
# or, from the source tree:
python setup.py build_ext --compiler=mingw32 --inplace

The GCC runtime is linked statically, so the resulting extension does not need MSYS2 DLLs at runtime. MSVC remains the default Windows toolchain.

For local development:

git clone https://github.com/AANovokhatskiy/pyscarcopula
cd pyscarcopula
pip install -e ".[test]"

To run the full test suite from the source tree, build the C++ extension in place first:

python setup.py build_ext --inplace
pytest --run-validation

pytest --run-validation enables optional validation tests. A source checkout without a successfully built extension is incomplete for the default bivariate GAS workflow.

Optional benchmark and large validation checks are disabled by default. Enable them explicitly:

PYSCA_RUN_BENCHMARKS=1 \
PYSCA_RUN_LARGE_BENCHMARKS=1 \
PYSCA_RUN_VINE_BENCHMARKS=1 \
pytest tests --run-validation

On Windows PowerShell:

$env:PYSCA_RUN_BENCHMARKS = "1"
$env:PYSCA_RUN_LARGE_BENCHMARKS = "1"
$env:PYSCA_RUN_VINE_BENCHMARKS = "1"
pytest tests --run-validation

Core dependencies: numpy, numba, scipy, joblib, tqdm.

Verify the compiled extension with:

python -m pyscarcopula._native_smoke

Features

VineCopula quick start

VineCopula is the primary API for regular vines. Omitting structure selects an R-vine from data; C-vines and D-vines are fixed RVineMatrix structures:

from pyscarcopula import VineCopula
from pyscarcopula.vine import RVineMatrix, cvine_structure, dvine_structure

# Data-driven regular vine
auto = VineCopula().fit(u, method="mle")

# Fixed standard structures
c_vine = VineCopula.cvine(d=u.shape[1]).fit(u, method="mle")
d_vine = VineCopula.dvine(d=u.shape[1]).fit(u, method="mle")

# Any valid regular-vine tree sequence, without writing a raw matrix
structure = RVineMatrix.from_trees(d=u.shape[1], trees=my_trees)
fixed = VineCopula(structure=structure).fit(u, method="mle")

Use vine.structure for the RVineMatrix and vine.natural_order_matrix when an integration specifically needs the natural-order runtime matrix. vine.matrix is a compatibility property. The raw pyvinecopulib matrix uses one-based labels and the opposite tree-level order above each anti-diagonal entry, so it is not obtained by simply adding one. See the matrix-layout conversion.

Copula families

• Archimedean: Gumbel, Frank, Clayton, Joe, including rotations where supported
• Elliptical: Gaussian and Student-t
• Independence copula for null models and vine pruning
• Multivariate Gaussian, Student-t, equicorrelation, and stochastic Student models
• Shared APIs for bivariate, multivariate, and vine models

Vine copulas

• One VineCopula runtime for auto-selected and fixed regular vines
• Fixed C-vine and D-vine factories backed by RVineMatrix
• Arbitrary valid structures built from decoded tree edges
• Automatic family and rotation selection per edge using AIC/BIC
• Tree-level and edge-level truncation
• Mixed MLE, SCAR, GAS, and independence edges within one vine

Sampling and prediction

• Unconditional sampling from fitted bivariate, multivariate, and vine models
• Conditional sampling for static and dynamic multivariate models and R-vines
• Exact and approximate conditional modes for R-vines
• PredictConfig for explicit prediction options
• Reproducible random generation via rng
• JSON persistence through model.save() and ModelClass.load() (include_data=False can omit stored training data)

Diagnostics and risk

• Rosenblatt-transform based goodness-of-fit tests
• Mixture Rosenblatt transform for stochastic models
• Predictive time-varying copula parameter paths
• VaR and CVaR utilities in pyscarcopula.contrib

CPU parallelism

• Explicit native threading for eligible multivariate row, emission, conditional-sampling, static-likelihood, and Monte Carlo kernels
• Process-level fit_independent and rolling risk_metrics execution
• Absolute one-thread default: omitted n_threads always means 1, regardless of environment variables
• Dependency-free C++17 linear algebra without hidden BLAS or OpenMP pools

Mathematical background

By Sklar's theorem, a joint distribution can be represented as

F(x_1, \ldots, x_d) = C(F_1(x_1), \ldots, F_d(x_d)),

where C is a copula and F_i are marginal distributions. This separates marginal modelling from dependence modelling.

For a one-parameter Archimedean copula with generator phi,

C(u_1, \ldots, u_d; \theta)
  = \phi^{-1}(\phi(u_1; \theta) + \cdots + \phi(u_d; \theta)).

In SCAR models the copula parameter is time-varying:

\theta_t = \Psi(x_t),
\qquad
dx_t = \kappa(\mu - x_t)dt + \nu dW_t,

where x_t is a latent Ornstein-Uhlenbeck process and Psi maps the latent state to the valid parameter domain. scar-tm-jacobi instead evolves Kendall's tau directly with a bounded Jacobi diffusion and maps tau back to the copula parameter for families that implement tau_to_param.

The transfer matrix method evaluates the latent-state likelihood by exploiting the Markov structure of the latent process. The path integral is computed as a sequence of matrix-vector products on a discretized grid or spectral basis, avoiding Monte Carlo variance at the cost of numerical approximation.

For SCAR-TM-OU, transition_method='auto' uses a hybrid deterministic strategy: Hermite spectral evaluation where it is reliable, matrix-based transition evaluation for regimes better handled on a grid, and local Gauss-Hermite in narrow-kernel OU cases. In broad terms, this keeps the latent path integral as repeated deterministic linear-algebra updates while choosing the most suitable transition representation automatically. See docs/guide/performance.md for the details and the available transition_method values.

result = fit(copula, u, method="scar-tm-ou")

result = fit(copula, u, method="gas")

Use the default scaling="unit" for production. scaling="fisher" remains an experimental, numerically sensitive mode.

See docs/guide/performance.md for supported families and numerical options.

Vine copulas decompose a d-dimensional dependence model into bivariate copulas arranged in a sequence of trees. VineCopula() selects a regular-vine structure from data subject to the proximity condition. VineCopula.cvine(...) and VineCopula.dvine(...) use fixed standard structures, while VineCopula(structure=RVineMatrix.from_trees(...)) accepts an arbitrary valid decoded tree sequence.

Examples and docs

Worked notebooks are available in examples/:

• 01_basic_api.ipynb
• 02_bivariate.ipynb
• 03_multivariate.ipynb
• 04_vine.ipynb
• 05_risk_metrics.ipynb
• 06_pyvinecopulib_comparison.ipynb

Additional documentation is in docs/. Estimation methods are described in docs/guide/estimation-methods.md, and performance-related details are kept in docs/guide/performance.md. CPU threading, process workers, thread safety, and scaling limits are documented in docs/guide/parallelism.md. Release history is in CHANGELOG.md.

License

MIT License. See LICENSE.txt.

Contacts

Contact me for any questions or discussion aanovokhatskiy@gmail.com