Source code for skfolio.distribution.copula._gumbel

"""Bivariate Gumbel Copula Estimation."""

# Copyright (c) 2025
# Author: Hugo Delatte <delatte.hugo@gmail.com>
# Credits: Matteo Manzi, Vincent Maladière, Carlo Nicolini
# SPDX-License-Identifier: BSD-3-Clause

import numpy as np
import numpy.typing as npt
import scipy.stats as st
import sklearn.utils.validation as skv

from skfolio.distribution.copula._base import BaseBivariateCopula
from skfolio.distribution.copula._utils import (
    CopulaRotation,
    _apply_copula_rotation,
    _apply_margin_swap,
    _apply_rotation_cdf,
    _apply_rotation_partial_derivatives,
    _select_rotation_itau,
    _select_theta_and_rotation_mle,
)

# Gumbel copula with a theta of 1.0 is just the independence copula, so we chose a lower
# bound of 1.0001. After 80, the copula is already imposing extremely strong upper tail
# dependence and increasing it will make it impractical.
_THETA_BOUNDS = (1.0001, 80.0)


[docs] class GumbelCopula(BaseBivariateCopula): r"""Bivariate Gumbel Copula Estimation. The Gumbel copula is an Archimedean copula characterized by strong upper tail dependence and little to no lower tail dependence. In its unrotated form, it is used for modeling extreme co-movements in the upper tail (i.e. simultaneous extreme gains). Rotations allow the copula to be adapted for different types of tail dependence: - A 180° rotation captures extreme co-movements in the lower tail (i.e. simultaneous extreme losses). - A 90° rotation captures scenarios where one variable exhibits extreme losses while the other shows extreme gains. - A 270° rotation captures the opposite scenario, where one variable experiences extreme gains while the other suffers extreme losses. Gumbel copula generally exhibits weaker upper tail dependence than the Joe copula. It is defined by: .. math:: C_{\theta}(u, v) = \exp\Bigl(-\Bigl[(-\ln u)^{\theta}+(-\ln v)^{\theta}\Bigr]^{1/\theta}\Bigr) where :math:`\theta \ge 1` is the dependence parameter. When :math:`\theta = 1`, the Gumbel copula reduces to the independence copula. Larger values of :math:`\theta` result in stronger upper-tail dependence. .. note:: Rotations are needed for Archimedean copulas (e.g., Joe, Gumbel, Gumbel) because their parameters only model positive dependence, and they exhibit asymmetric tail behavior. To model negative dependence, one uses rotations to "flip" the copula's tail dependence. Parameters ---------- itau : bool, default=True If True, :math:`\theta` is estimated using the Kendall's tau inversion method; otherwise, the Maximum Likelihood Estimation (MLE) method is used. The MLE is slower but more accurate. kendall_tau : float, optional If `itau` is True and `kendall_tau` is provided, this value is used; otherwise, it is computed. tolerance : float, default=1e-4 Convergence tolerance for the MLE optimization. random_state : int, RandomState instance or None, default=None Seed or random state to ensure reproducibility. Attributes ---------- theta_ : float Fitted theta coefficient :math:`\theta` > 1. rotation_ : CopulaRotation Fitted rotation of the copula. Examples -------- >>> from skfolio.datasets import load_sp500_dataset >>> from skfolio.preprocessing import prices_to_returns >>> from skfolio.distribution import GumbelCopula, compute_pseudo_observations >>> >>> # Load historical prices and convert them to returns >>> prices = load_sp500_dataset() >>> X = prices_to_returns(prices) >>> X = X[["AAPL", "JPM"]] >>> >>> # Convert returns to pseudo observation in the interval [0,1] >>> X = compute_pseudo_observations(X) >>> >>> # Initialize the Copula estimator >>> model = GumbelCopula() >>> >>> # Fit the model to the data. >>> model.fit(X) >>> >>> # Display the fitted parameter and tail dependence coefficients >>> print(model.fitted_repr) GumbelCopula(theta=1.27, rot=180°) >>> print(model.lower_tail_dependence) 0.2735 >>> print(model.upper_tail_dependence) 0.0 >>> >>> # Compute the log-likelihood, total log-likelihood, CDF, Partial Derivative, >>> # Inverse Partial Derivative, AIC, and BIC >>> log_likelihood = model.score_samples(X) >>> score = model.score(X) >>> cdf = model.cdf(X) >>> p = model.partial_derivative(X) >>> u = model.inverse_partial_derivative(X) >>> aic = model.aic(X) >>> bic = model.bic(X) >>> >>> # Generate 5 new samples >>> samples = model.sample(n_samples=5) >>> >>> # Plot the tail concentration function. >>> fig = model.plot_tail_concentration() >>> fig.show() >>> >>> # Plot a 2D contour of the estimated PDF. >>> fig = model.plot_pdf_2d() >>> fig.show() >>> >>> # Plot a 3D surface of the estimated PDF. >>> fig = model.plot_pdf_3d() >>> fig.show() References ---------- .. [1] "An Introduction to Copulas (2nd ed.)", Nelsen (2006) .. [2] "Multivariate Models and Dependence Concepts", Joe, Chapman & Hall (1997) .. [3] "Quantitative Risk Management: Concepts, Techniques and Tools", McNeil, Frey & Embrechts (2005) .. [4] "The t Copula and Related Copulas", Demarta & McNeil (2005) .. [5] "Copula Methods in Finance", Cherubini, Luciano & Vecchiato (2004) """ theta_: float rotation_: CopulaRotation _n_params = 1 def __init__( self, itau: bool = True, kendall_tau: float | None = None, tolerance: float = 1e-4, random_state: int | None = None, ): super().__init__(random_state=random_state) self.itau = itau self.kendall_tau = kendall_tau self.tolerance = tolerance
[docs] def fit(self, X: npt.ArrayLike, y=None) -> "GumbelCopula": r"""Fit the Bivariate Gumbel Copula. If `itau` is True, estimates :math:`\theta` using Kendall's tau inversion. Otherwise, uses MLE by maximizing the log-likelihood. Parameters ---------- X : array-like of shape (n_observations, 2) An array of bivariate inputs `(u, v)` where each row represents a bivariate observation. Both `u` and `v` must be in the interval [0, 1], having been transformed to uniform marginals. y : None Ignored. Provided for compatibility with scikit-learn's API. Returns ------- self : object Returns the instance itself. """ X = self._validate_X(X, reset=True) if self.itau: if self.kendall_tau is None: kendall_tau = st.kendalltau(X[:, 0], X[:, 1]).statistic else: kendall_tau = self.kendall_tau # For Gumbel, the theoretical relationship is: tau = 1 - 1/theta. abs_kendall_tau = min(abs(kendall_tau), 0.9999) self.theta_ = np.clip( 1.0 / (1.0 - abs_kendall_tau), a_min=_THETA_BOUNDS[0], a_max=_THETA_BOUNDS[1], ) self.rotation_ = _select_rotation_itau( func=_neg_log_likelihood, X=X, theta=self.theta_ ) else: self.theta_, self.rotation_ = _select_theta_and_rotation_mle( _neg_log_likelihood, X=X, bounds=_THETA_BOUNDS, tolerance=self.tolerance ) return self
[docs] def cdf(self, X: npt.ArrayLike) -> np.ndarray: r"""Compute the CDF of the bivariate Gumbel copula. .. math:: C(u,v) = \exp\Bigl(-\Bigl[(-\ln u)^{\theta}+(-\ln v)^{\theta}\Bigr]^{1/\theta}\Bigr). Parameters ---------- X : array-like of shape (n_observations, 2) An array of bivariate inputs `(u, v)` where each row represents a bivariate observation. Both `u` and `v` must be in the interval [0, 1], having been transformed to uniform marginals. Returns ------- cdf : ndarray of shape (n_observations,) CDF values for each observation in X. """ skv.check_is_fitted(self) X = self._validate_X(X, reset=False) cdf = _apply_rotation_cdf( func=_base_cdf, X=X, rotation=self.rotation_, theta=self.theta_ ) return cdf
[docs] def partial_derivative( self, X: npt.ArrayLike, first_margin: bool = False ) -> np.ndarray: r"""Compute the h-function (partial derivative) for the bivariate Gumbel copula with respect to a specified margin. The h-function with respect to the second margin represents the conditional distribution function of :math:`u` given :math:`v`: .. math:: \begin{aligned} h(u \mid v) &= \frac{\partial C(u,v)}{\partial v}\\[6pt] &= C(u,v)\,\Bigl[(-\ln u)^{\theta}+(-\ln v)^{\theta}\Bigr]^{\frac{1}{\theta}-1} \,(-\ln v)^{\theta-1}\,\frac{1}{v}. \end{aligned} Parameters ---------- X : array-like of shape (n_observations, 2) An array of bivariate inputs `(u, v)` where each row represents a bivariate observation. Both `u` and `v` must be in the interval [0, 1], having been transformed to uniform marginals. first_margin : bool, default=False If True, compute the partial derivative with respect to the first margin `u`; otherwise, compute the partial derivative with respect to the second margin `v`. Returns ------- p : ndarray of shape (n_observations,) h-function values :math:`h(u \mid v)` for each observation in X. """ skv.check_is_fitted(self) X = self._validate_X(X, reset=False) p = _apply_rotation_partial_derivatives( func=_base_partial_derivative, # Computation remains unchanged X=X, rotation=self.rotation_, first_margin=first_margin, theta=self.theta_, ) return p
[docs] def inverse_partial_derivative( self, X: npt.ArrayLike, first_margin: bool = False ) -> np.ndarray: r"""Compute the inverse of the bivariate copula's partial derivative, commonly known as the inverse h-function. Let :math:`C(u, v)` be a bivariate copula. The h-function with respect to the second margin is defined by .. math:: h(u \mid v) \;=\; \frac{\partial\,C(u, v)}{\partial\,v}, which is the conditional distribution of :math:`U` given :math:`V = v`. The **inverse h-function**, denoted :math:`h^{-1}(p \mid v)`, is the unique value :math:`u \in [0,1]` such that .. math:: h(u \mid v) \;=\; p, \quad \text{where } p \in [0,1]. In practical terms, given :math:`(p, v)` in :math:`[0, 1]^2`, :math:`h^{-1}(p \mid v)` solves for the :math:`u` satisfying :math:`p = \partial C(u, v)/\partial v`. Parameters ---------- X : array-like of shape (n_observations, 2) An array of bivariate inputs `(p, v)`, each in the interval [0, 1]. - The first column `p` corresponds to the value of the h-function. - The second column `v` is the conditioning variable. first_margin : bool, default=False If True, compute the inverse partial derivative with respect to the first margin `u`; otherwise, compute the inverse partial derivative with respect to the second margin `v`. Returns ------- u : ndarray of shape (n_observations,) A 1D-array where each element is the computed :math:`u = h^{-1}(p \mid v)` for the corresponding pair in `X`. """ skv.check_is_fitted(self) X = self._validate_X(X, reset=False) u = _apply_rotation_partial_derivatives( func=_base_inverse_partial_derivative, # Computation remains unchanged X=X, rotation=self.rotation_, first_margin=first_margin, theta=self.theta_, ) return u
[docs] def score_samples(self, X: npt.ArrayLike) -> np.ndarray: r"""Compute the log-likelihood of each sample (log-pdf) under the model. For Gumbel, the PDF is given by: .. math:: c(u,v) = \exp\Bigl(-\Bigl((-\ln u)^{\theta}+(-\ln v)^{\theta}\Bigr)^{1/\theta}\Bigr) \cdot \left[\Bigl((-\ln u)^{\theta}+(-\ln v)^{\theta}\Bigr)^{1/\theta-1} \left\{ \frac{(-\ln u)^{\theta}}{u}+\frac{(-\ln v)^{\theta}}{v}\right\}\right]. Parameters ---------- X : array-like of shape (n_observations, 2) An array of bivariate inputs `(u, v)` where each row represents a bivariate observation. Both `u` and `v` must be in the interval [0, 1], having been transformed to uniform marginals. Returns ------- density : ndarray of shape (n_observations,) The log-likelihood of each sample under the fitted copula. """ skv.check_is_fitted(self) X = self._validate_X(X, reset=False) X = _apply_copula_rotation(X, rotation=self.rotation_) log_density = _base_sample_scores(X=X, theta=self.theta_) return log_density
@property def lower_tail_dependence(self) -> float: """Theoretical lower tail dependence coefficient.""" skv.check_is_fitted(self) if self.rotation_ == CopulaRotation.R180: return 2.0 - np.power(2.0, 1.0 / self.theta_) return 0 @property def upper_tail_dependence(self) -> float: """Theoretical upper tail dependence coefficient.""" skv.check_is_fitted(self) if self.rotation_ == CopulaRotation.R0: return 2.0 - np.power(2.0, 1.0 / self.theta_) return 0 @property def fitted_repr(self) -> str: """String representation of the fitted copula.""" return ( f"{self.__class__.__name__}(theta={self.theta_:0.2f}, rot={self.rotation_})" )
def _neg_log_likelihood(theta: float, X: np.ndarray) -> float: """Negative log-likelihood function for the Gumbel copula. Parameters ---------- X : array-like of shape (n_observations, 2) An array of bivariate inputs `(u, v)` where each row represents a bivariate observation. Both `u` and `v` must be in the interval [0, 1], having been transformed to uniform marginals. theta : float The dependence parameter (must be greater than 1). Returns ------- value : float The negative log-likelihood value. """ return -np.sum(_base_sample_scores(X=X, theta=theta)) def _base_sample_scores(X: np.ndarray, theta: float) -> np.ndarray: r"""Compute the log-likelihood of each sample (log-pdf) under the bivariate Gumbel copula. Parameters ---------- X : array-like of shape (n_observations, 2) Bivariate samples `(u, v)`, with each component in [0,1]. theta : float The dependence parameter (must be greater than 1). Returns ------- logpdf : ndarray of shape (n_observations,) Log-likelihood values for each observation. """ if theta <= 1: raise ValueError("Theta must be greater than 1 for the Gumbel copula.") Z = -np.log(X) s = np.power(np.power(Z, theta).sum(axis=1), 1 / theta) s = np.clip(s, a_min=1e-10, a_max=None) log_density = ( -s + np.log(s + theta - 1) + (1 - 2 * theta) * np.log(s) + (theta - 1) * np.log(Z.prod(axis=1)) + Z.sum(axis=1) ) return log_density def _base_cdf(X: np.ndarray, theta: float) -> np.ndarray: r"""Bivariate Gumbel CDF (unrotated). .. math:: C(u,v) = \exp\Bigl(-\Bigl[(-\ln u)^{\theta}+(-\ln v)^{\theta}\Bigr]^{1/\theta}\Bigr). """ cdf = np.exp(-np.power(np.power(-np.log(X), theta).sum(axis=1), 1.0 / theta)) return cdf def _base_partial_derivative( X: np.ndarray, first_margin: bool, theta: float ) -> np.ndarray: r""" Compute the partial derivative (h-function) for the unrotated Gumbel copula. For Gumbel, the copula is defined as: .. math:: C(u,v)=\exp\Bigl(-\Bigl[(-\ln u)^{\theta}+(-\ln v)^{\theta}\Bigr]^{1/\theta}\Bigr). The partial derivative with respect to v is: .. math:: \frac{\partial C(u,v)}{\partial v} = C(u,v)\,\Bigl[(-\ln u)^{\theta}+(-\ln v)^{\theta}\Bigr]^{\frac{1}{\theta}-1} \,(-\ln v)^{\theta-1}\,\frac{1}{v}. Parameters ---------- X : array-like of shape (n_observations, 2) An array of bivariate inputs `(u, v)` with values in [0, 1]. first_margin : bool, default=False If True, compute with respect to u (by swapping margins); otherwise, compute with respect to v. theta : float The dependence parameter (must be > 1). Returns ------- p : ndarray of shape (n_observations,) The computed h-function values. """ X = _apply_margin_swap(X, first_margin=first_margin) _, v = X.T x, y = -np.log(X).T p = ( np.exp(-np.power(np.power(x, theta) + np.power(y, theta), 1.0 / theta)) * np.power(np.power(x / y, theta) + 1.0, 1.0 / theta - 1.0) / v ) return p def _base_inverse_partial_derivative( X: np.ndarray, first_margin: bool, theta: float ) -> np.ndarray: r""" Compute the inverse partial derivative for the unrotated Gumbel copula, i.e. solve for u in h(u|v)=p. In other words, given - p, the value of the h-function, and - v, the conditioning variable, solve: .. math:: p = C(u,v)\,\Bigl[(-\ln u)^{\theta}+(-\ln v)^{\theta}\Bigr]^{\frac{1}{\theta}-1}\, (-\ln v)^{\theta-1}\,\frac{1}{v}, for u ∈ [0,1]. Since no closed-form solution exists, we use a numerical method. Parameters ---------- X : array-like of shape (n_observations, 2) An array with first column p (h-function values) and second column v (conditioning variable). first_margin : bool, default=False If True, treat the first margin as the conditioning variable. theta : float The dependence parameter (must be > 1). Returns ------- u : ndarray of shape (n_observations,) A 1D-array where each element is the solution u ∈ [0,1] such that h(u|v)=p. """ X = _apply_margin_swap(X, first_margin=first_margin) p, v = -np.log(X).T s = v + p + np.log(v) * (theta - 1.0) # Initial guess x = v.copy() max_iters = 50 tol = 1e-8 for _ in range(max_iters): x_new = x * (s - (theta - 1.0) * (np.log(x) - 1.0)) / (theta + x - 1.0) x_new = np.clip(x_new, x, None) diff = np.max(np.abs(x_new - x)) x = x_new if diff < tol: break u = np.exp(-np.power(np.power(x, theta) - np.power(v, theta), 1.0 / theta)) return u