PaperlayerView exact revision on arXiv

Source-linked AI summary

D-vine copula based quantile regression

Daniel Kraus, Claudia Czado

arXiv:1510.04161v4stat.ME

TL;DR

Existing quantile-regression approaches can impose restrictive dependence and linearity assumptions, motivating a more flexible method for conditional-quantile prediction. The paper sequentially fits a likelihood-optimal D-vine copula with automatic variable selection, finding strong simulation performance and useful CDS applications, while identifying misspecification under non-monotonic dependence.

  • Problem

    Quantile regression needs flexible conditional-quantile models because linear approaches can be misspecified and suffer from quantile crossing, multicollinearity, and variable-transformation or interaction issues.

  • Method

    The method sequentially adds predictors and fits the required pair-copulas by maximizing conditional log-likelihood, producing a semiparametric D-vine quantile-regression model.

  • Results

    The method outperforms competing quantile-regression approaches by a great margin in simulations, while also offering faster computation than nonparametric and semiparametric competitors and supporting CDS stress testing.

  • Takeaways & Limitations

    D-vine quantile regression supports flexible conditional-quantile estimation, automatic variable selection, and multi-institution financial stress tests with quantiles extracted from one fitted model.

  • Takeaways & Limitations

    Parametric D-vines can be misspecified for non-monotonic dependence, producing persistent model bias that larger training samples do not remove.

Abstract

from arXiv · show

Quantile regression, that is the prediction of conditional quantiles, has steadily gained importance in statistical modeling and financial applications. The authors introduce a new semiparametric quantile regression method based on sequentially fitting a likelihood optimal D-vine copula to given data resulting in highly flexible models with easily extractable conditional quantiles. As a subclass of regular vine copulas, D-vines enable the modeling of multivariate copulas in terms of bivariate building blocks, a so-called pair-copula construction (PCC). The proposed algorithm works fast and accurate even in high dimensions and incorporates an automatic variable selection by maximizing the conditional log-likelihood. Further, typical issues of quantile regression such as quantile crossing or transformations, interactions and collinearity of variables are automatically taken care of. In a simulation study the improved accuracy and saved computational time of the approach in comparison with established quantile regression methods is highlighted. An extensive financial application to international credit default swap (CDS) data including stress testing and Value-at-Risk (VaR) prediction demonstrates the usefulness of the proposed method.

1 Introduction

Quantile regression is widely used for conditional prediction, particularly in finance, but conventional linear methods impose restrictive shape and dependence assumptions. The proposed methodology instead models response–covariate dependence flexibly with a D-vine copula.

  • Applications: Conditional quantiles support financial tasks including portfolio optimization, asset pricing, and systemic-risk evaluation.Applications include CoVaR, which conditions an institution’s loss distribution on other institutions being in distress.
  • Limitations of existing methods: Linear quantile regression can be misspecified when dependence deviates from a Gaussian copula, while also facing quantile crossing and linear-model pitfalls.The cited pitfalls include multicollinearity, covariate selection and significance, and handling interactions or transformations.
  • Proposed approach: The proposed method makes no precise assumptions about conditional-quantile shape and models dependence between response and covariates with a parametric D-vine copula.The D-vine is presented as a flexible alternative to the restrictive assumptions of linear quantile regression.

2 D-vine copulas

D-vine copulas decompose multivariate dependence into bivariate pair-copulas arranged along an ordered sequence, with conditional distributions represented through copula-based transformations. This construction provides the building blocks and graphical structure used by the paper’s regression method.

  • Copula foundations: A copula isolates multivariate dependence from marginal distributions by transforming each variable through its probability integral transform.For continuous variables, the joint density factors into a copula density and the marginal densities.
  • Conditional quantities: Conditional copulas describe the dependence of two variables given a conditioning set, while conditional distributions use corresponding conditional-distribution notation.The paper distinguishes conditional copulas, copula densities, conditional distributions, and conditional PIT distributions.
  • D-vine construction: A D-vine pair-copula construction represents a multivariate density through bivariate copulas arranged in the order X1–X2–. . .–Xd.When all margins are uniform, the resulting construction is called a D-vine copula.
  • Graphical representation: Each edge in the nested tree representation corresponds to a pair-copula, including conditional pair-copulas in higher trees.For a five-dimensional example, the construction lists unconditional and conditionally indexed pair-copulas across successive trees.
  • Estimation assumptions: Fitting a fixed-order D-vine estimates every pair-copula parametrically under the simplifying assumption that conditional copulas do not depend on conditioning values.The simplifying assumption sets ci,j;D(·, ·; xD) equal to ci,j;D(·, ·).
  • Conditional calculations: The conditional distribution F1|23 can be computed recursively by composing h-functions from lower-level conditional distributions and pair-copulas.The displayed expression applies h1|3;2 to F1|2 and F3|2, which themselves are obtained from marginal PITs and pair-copulas.

3 D-vine based quantile regression model

The method estimates conditional quantiles by fitting a D-vine copula to transformed response and predictor data, with an ordering selected through conditional log-likelihood. Its construction supports flexible dependence, analytic quantile evaluation, automatic variable selection, and noncrossing quantiles.

  • 3.1 Conditional quantile function: The conditional quantile is obtained by combining the response marginal inverse with a conditional copula quantile evaluated at predictor PIT values.The estimation plugs estimated marginals and the fitted conditional copula into the conditional quantile representation.
  • 3.1 Conditional quantile function: The model fits a D-vine ordered as V–U_l1–…–U_ld, allowing separate bivariate copulas and choosing the predictor order by conditional likelihood.Nested h-functions and inverse h-functions make the conditional copula quantile calculable.
  • 3.1 Conditional quantile function: The conditional copula quantile is monotonically increasing in α, so quantile functions at different levels cannot cross.This addresses a problem reported for linear and nonlinear quantile regression.
  • 3.2 Estimation process: The two-step estimator first fits marginal distributions and transforms observations into pseudo copula data, then estimates the D-vine on those transformed observations.The authors prefer nonparametric marginal estimation to avoid bias and inconsistency from misspecified fully parametric marginals.
  • 3.2 Estimation process: The sequential algorithm adds the remaining covariate that most improves conditional log-likelihood and stops when no candidate improves the model.This constructs the D-vine without comparing all d! predictor orderings.
  • 3.2 Estimation process: The procedure automatically selects influential covariates, ranks their predictive strength, excludes superfluous variables, and flexibly models dependence with the response.The paper connects these properties to handling collinearity, transformations, and covariate inclusion or exclusion without additional effort.

4 Established quantile regression methods

Established approaches include linear, additive boosting, fully nonparametric, and semiparametric copula-based quantile regression. They differ in flexibility, estimation strategy, and how conditional quantiles are computed.

  • Linear quantile regression: Linear quantile regression assumes conditional quantiles are linear in predictors, an assumption criticized as restrictive and associated with crossing regression lines.Under normal margins, the linear conditional-quantile model permits only the Gaussian copula.
  • Boosting additive quantile regression: Boosting additive quantile regression relaxes linearity by representing effects with smooth functions estimated using additive-model techniques.The cited approach uses boosting to estimate additive quantile regression functions.
  • Nonparametric quantile regression: Fully nonparametric quantile regression estimates the conditional distribution with a kernel estimator and numerically inverts it to obtain conditional quantiles.Bandwidths are selected automatically in the referenced R implementation.
  • Semiparametric quantile regression: Semiparametric quantile regression estimates marginals nonparametrically and a regular-vine copula, then obtains conditional quantiles by minimizing a weighted check function.Unlike the proposed D-vine method, the cited procedure generally cannot calculate the conditional copula quantile analytically.

5 Results of the simulation study

The simulation study evaluates D-vine quantile regression across Clayton, multivariate t, and misspecified non-monotonic dependence scenarios using out-of-sample MISE and computational-time comparisons. It performs strongly in correctly specified settings, but parametric copulas show bias under non-monotonic dependence.

  • Simulation design: The study evaluates three scenarios, including three-dimensional Clayton, five-dimensional t-copula, and non-monotonic dependence settings.The first two scenarios have known conditional quantile functions and vary dependence, margins, and training-sample settings.
  • Simulation design: Performance is assessed using estimated out-of-sample MISE over 100 replications, with relative MISE comparing competitors against D-vine quantile regression.Each replication uses simulated training and evaluation data, estimates conditional quantiles, and averages integrated squared errors.
  • Scenario C3: In Scenario C3, D-vine quantile regression fits well, with estimated MISE errors on the order of 10^-3 to 10^-2.The 50%-quantile generally fits better than the 95%-quantile, and errors decrease with larger training samples.
  • Scenario C3: In Scenario C3, linear quantile regression errors reach up to 30 times those of D-vine regression, while boosting additive models lag by factors of 2.4 to 11.9.The scenario violates linearity, and nonparametric quantile regression is described as too imprecise to be a serious competitor.
  • Computational time: D-vine quantile regression is much faster than nonparametric and semiparametric competitors, with computation time growing linearly in the training-sample size.After one model fit, different quantile levels can be extracted through inverse h-functions rather than separate optimizations.
  • Scenario t5: In Scenario t5, D-vine regression remains preferred especially for larger samples, although linear regression slightly outperforms it for the median with ntrain = 300.At α = 0.95, the linear method loses this advantage, while D-vine outperforms the semiparametric method for medians.
  • Scenario M5: In Scenario M5, parametric D-vine regression is misspecified: errors retain their order of magnitude as sample size increases, while nonparametric and boosting methods perform best.The fitted parametric copula poorly matches the non-monotonic dependence, producing comparatively high MISE values.

6 Application to the CDS data set

The CDS application uses D-vine quantile regression for stress testing and out-of-sample conditional quantile prediction across 38 financial institutions. Stress tests reveal geographically structured spillovers, while out-of-sample results favor D-vine and semiparametric methods for extreme quantiles.

  • Data and design: The dataset contains 1,371 daily CDS-spread log-return observations for 38 European, US, and Asia-Pacific financial institutions.The institutions include 18 banks and 20 insurers.
  • Stress testing: Stress testing removes serial dependence with GARCH models and evaluates conditional spillovers on the PIT-transformed scale.Stress levels are represented by κ = 0.9, 0.95, and 0.99 for moderate, severe, and extreme stress.
  • Stress testing: Extreme stress produced predicted median values up to 0.98 for European insurers Allianz and Aviva.US banks also exceeded 0.9 in the extreme stress case, but European insurance companies experienced stronger effects.
  • Out-of-sample prediction: The out-of-sample test predicts CDS-log-return quantiles conditional on institutions’ own and peer companies’ lagged values.Performance is evaluated with tick-loss across α ∈ {0.5, 0.01, 0.99, 0.995}, averaged over 38 institutions.
  • Out-of-sample prediction: D-vine and semiparametric quantile regression perform similarly and outperform competitors for extreme-quantile prediction.Linear quantile regression slightly outperforms other methods at the median, whereas its lack of tail dependence becomes apparent at extreme quantiles.

7 Conclusions and further research

The paper proposes D-vine quantile regression as a flexible method for conditional quantile prediction and applies it to financial stress analysis. The application supports geographically driven spillovers and finds that stressing major European banks has greater overall impact than stressing US banks.

  • Conclusions: D-vine quantile regression enables fast and accurate estimation of conditional quantiles using a flexible D-vine class.The method extends stress testing to situations where multiple banks and insurers are simultaneously distressed.
  • Conclusions: Geography mainly drives spillovers, with European banks exerting greater influence on European insurers than on US banks.The analysis does not support the claim that US banks have stronger influence on the international financial market than European banks.
  • Conclusions: Stress on major European banks has a greater overall impact on the financial system than stress on US banks.

A D-vine regression algorithm

The algorithm constructs a D-vine sequentially by selecting covariates that maximize conditional log-likelihood. At each step it estimates the pair-copulas needed to extend the current vine and stops according to the selection rule.

  • Sequential selection: The algorithm evaluates candidate covariates by their conditional log-likelihood contributions and selects the best one.The current vine order is V–U_l1–...–U_l{k−1}, with unselected covariates retained in the candidate set.
  • Sequential selection: Each selected covariate extends the D-vine, requiring estimation of the new conditional and higher-tree pair-copulas.
  • Stopping and output: The procedure returns the vine ordered as V–U_l1–...–U_ld after the selected number of steps.
  • Stopping and output: AIC- or BIC-corrected conditional log-likelihoods can replace the regular criterion, reversing maximization to minimization.The implementation supports multiple Gaussian, t, and Archimedean pair-copula families.

B.1 Conditional copula quantile function based on the Gaussian copula

For a Gaussian copula, conditional quantiles are obtained from the conditional normal distribution given the predictors. The quantile function follows by inverting the conditional distribution function.

  • Conditional distribution: Conditioning a multivariate normal vector yields a normal conditional distribution with shifted mean and covariance.
  • Conditional quantile: The conditional quantile of Y at level α is computed from the conditional distribution of Y given X = x.
  • Conditional quantile: The Gaussian-copula conditional quantile is obtained by inverting the conditional distribution function.

B.2 Conditional copula quantile function based on the Student’s t copula

The conditional distribution of a multivariate t random vector remains t distributed, enabling calculation of conditional quantiles. The t-copula conditional quantile follows from this conditional distribution and its parameters.

  • Conditioning a multivariate t random vector yields another t distribution with shifted mean, covariance matrix, and degrees of freedom.
  • The conditional t distribution is characterized by degrees of freedom νc, conditional mean µc(x), and conditional variance σ2.
  • The conditional quantile of the t copula is obtained from the conditional t distribution.

B.3 Conditional copula quantile function based on the two- and three-dimensional Clayton copula

The section derives conditional quantile functions for two- and three-dimensional Clayton copulas by constructing conditional distributions and inverting their h-functions. A visualization illustrates the resulting conditional median and 95% quantile functions.

  • For the three-dimensional Clayton copula, the conditional quantile is derived by inverting the h-function.
  • For a three-dimensional Clayton copula, the conditional distribution of U given V = v and W = w is derived using a version of Equation (2.2).
  • The conditional distribution of U given W = w is obtained from the three-dimensional copula by setting the omitted argument to one.
  • The inverse functions provide the transformations needed to calculate the conditional quantile functions.
  • Figure 5 visualizes the conditional median and 95% quantile function for a three-dimensional distribution with a Clayton copula.
Loading 1510.04161v4…