copula

High performance copula generators

pyscenarios.copula.gaussian_copula(cov, samples, seed=0, chunks=None, rng='Mersenne Twister')

Gaussian Copula scenario generator.

Simplified algorithm:

>>> l = numpy.linalg.cholesky(cov)
>>> y = numpy.random.standard_normal(size=(samples, cov.shape[0]))
>>> p = numpy.dot(l, y.T).T
Parameters:
  • cov (numpy.ndarray) – covariance matrix, a.k.a. correlation matrix. It must be a Hermitian, positive-definite matrix in any square array-like format. The width of cov determines the number of dimensions of the output.
  • samples (int) –

    Number of random samples to generate

    Note

    When using SOBOL, to obtain a uniform distribution one must use \(2^{n} - 1\) samples (for any n > 0).

  • chunks –

    Chunk size for the return array, which has shape (samples, dimensions). It can be anything accepted by dask (a positive integer, a tuple of two ints, or a tuple of two tuples of ints) for the output shape.

    Set to None to return a numpy array.

    Warning

    When using the Mersenne Twister random generator, the chunk size changes the random sequence. To guarantee repeatability, it must be fixed together with the seed. chunks=None also produces different results from using dask.

  • seed (int) –

    Random seed.

    With rng='SOBOL', this is the initial dimension; when generating multiple copulas with different seeds, one should never use seeds that are less than cov.shape[0] apart from each other.

    The maximum seed when using sobol is:

    pysamples.sobol.max_dimensions() - cov.shape[0] - 1
  • rng (str) – Either Mersenne Twister or SOBOL
Returns:

array of shape (samples, dimensions), with all series being normal (0, 1) distributions.
Return type:

If chunks is not None, dask.array.Array; else numpy.ndarray
pyscenarios.copula.t_copula(cov, df, samples, seed=0, chunks=None, rng='Mersenne Twister')

Student T Copula / IT Copula scenario generator.

Simplified algorithm:

>>> l = numpy.linalg.cholesky(cov)
>>> y = numpy.random.standard_normal(size=(samples, cov.shape[0]))
>>> r = numpy.random.uniform(size=(samples, 1))
>>> s = scipy.stats.chi2.ppf(r, df=df)
>>> z = numpy.sqrt(df / s) * numpy.dot(l, y.T).T
>>> u = scipy.stats.t.cdf(z, df=df)
>>> p = scipy.stats.norm.ppf(u)
Parameters:
  • df – Number of degrees of freedom. Can be either a scalar int for Student T Copula, or a one-dimensional array-like with one point per dimension for IT Copula.
  • seed (int) –

    Random seed.

    With rng='SOBOL', this is the initial dimension; when generating multiple copulas with different seeds, one should never use seeds that are less than cov.shape[0] + 1 apart from each other.

    The maximum seed when using sobol is:

    pysamples.sobol.max_dimensions() - cov.shape[0] - 2

All other parameters and the return value are the same as in gaussian_copula().