Factor Copulas - Overview

Summary

Oh & Patton (2012) propose a class of copula (dependence) models for economic variables built from a simple latent factor structure, designed specifically for high dimensions (50+ variables). The models lack a closed-form density, but admit analytical tail-dependence results via extreme value theory and are estimable by a fast simulated method of moments (SMM) using rank statistics. Applied to all 100 S&P 100 constituents, they find a fat-tailed, asymmetric common factor: dependence is stronger in crashes than in booms, with implications for systemic risk.

Overview

The financial crisis of 2007-2008 exposed a failure of dependence models: assets that had previously moved mostly independently suddenly crashed together. Correlation-based models built on multivariate Gaussianity capture dependence through the second moment but impose zero tail dependence and symmetric treatment of booms and crashes — they neglect the possibility that large crashes are correlated across assets. At the same time, the econometric toolkit for modelling risk across many variables was thin: most copula methods in the literature did not scale beyond ~20 dimensions.

A copula decomposes a joint distribution into marginal distributions and a dependence structure (Sklar’s theorem). For a vector $\mathbf{Y}=[Y_1,\dots ,Y_N{]}^{\prime }$ with joint distribution $\mathbf{F}$, marginals $F_i$, and copula $\mathbf{C}$:

$$\mathbf{Y}\sim \mathbf{F}=\mathbf{C}(F_1,\dots ,F_N)$$

This separation lets the researcher (i) estimate marginals using the large univariate-modelling literature, and (ii) model the copula separately — useful in high dimensions where data sparseness and parameter proliferation cause problems, enabling multi-stage estimation.

Main Content

The two primary contributions

1. A new class of factor copulas. The dependence structure is generated by a latent linear factor model (see Factor Copula Construction). By allowing the common factor to be fat-tailed (e.g. Student’s $t$), the model captures correlated crashes (tail dependence). By allowing the common factor to be asymmetrically distributed (skew $t$), dependence can be stronger in downturns than upturns. Multiple common factors allow heterogeneous pairwise dependence. The flexibility/parsimony trade-off is chosen by the researcher according to dimension and data; simplifying assumptions (e.g. equidependence) are interpretable and testable.

2. An empirical study of the S&P 100. All 100 constituents (April 2008-Dec 2010, $T=696$ days) — one of the highest-dimension copula applications in econometrics. They find significant evidence of a fat-tailed common factor (non-zero tail dependence; the Normal copula is rejected) that is asymmetrically distributed (crashes more correlated than booms), and that the factor copula yields superior estimates of systemic-risk measures.

A third contribution is a detailed simulation study confirming the SMM estimator and its asymptotic theory have good finite-sample properties up to $N=100$.

Position relative to the literature

The models extend Hull & White (2004): they keep a simple linear, additive factor structure but allow the latent variables to have flexibly specified distributions. Related factor copulas appear in Andersen & Sidenius (2004) and van der Voort (2005) (non-linear structures) and McNeil et al. (2005) (times-to-default). Prior work largely focused on calibration/pricing, not estimation of unknown parameters. Alternatives that struggle in high dimensions: the Normal copula (Li 2000; zero tail dependence, symmetric); the $t$ / grouped-$t$ copula (Demarta & McNeil 2005; Daul et al. 2003 — usable up to 100 variables but forces equal upper/lower tail dependence, strongly rejected for equities); Archimedean copulas (Clayton, Gumbel — too few parameters for many variables); and vine copulas (Aas et al. 2009; hard-to-interpret/test assumptions). The formal SMM estimation of high-dimension copulas is new to the literature.

Why a factor structure for the copula

The latent variable $\mathbf{X}$ generated by the factor model is used only for its copula $\mathbf{C}(\mathbf{\theta })$; its implied marginals are discarded and replaced by separately-estimated marginals $F_i$. This is analogous to mixture-model constructions (Normal-inverse Gaussian, generalized hyperbolic). The leading closed-form case is when both factor and idiosyncratic distributions are Gaussian, giving a Gaussian copula with equicorrelation ${\sigma }_z^2/({\sigma }_z^2+{\sigma }_{\epsilon }^2)$; for any other distributions the copula has no closed form, motivating simulation-based estimation.

Examples

Four illustrative bivariate copulas (Figure 1)

Setup: Marginals fixed at $N(0,1)$; latent variances ${\sigma }_z^2={\sigma }_{\epsilon }^2=1$ (common factor explains half the variance of each $X_i$). Vary the factor and idiosyncratic distributions:

1. $F_z=F_{\epsilon }=N(0,1)$ → Normal copula (no tail dependence, symmetric).
2. $F_z=F_{\epsilon }=t(4)$ → symmetric copula with positive tail dependence.
3. $F_{\epsilon }=N(0,1)$, $F_z=\text{skew }t(\infty ,-0.25)$ → asymmetric dependence (crashes more correlated) but zero tail dependence.
4. $F_{\epsilon }=t(4)$, $F_z=\text{skew }t(4,-0.25)$ → both asymmetric dependence and positive tail dependence.

Result: Fat tails (low DoF) produce clustering in the joint upper and lower tails; negative skew produces stronger clustering in the joint negative quadrant than the positive one. Interpretation: A single parsimonious class spans Normality, symmetric tail dependence, and asymmetric tail dependence — the researcher dials in the features the data demand.

Connections

• Factor Copula Construction — the formal latent-variable model that generates this copula class.
• Tail Dependence in Factor Copulas — the EVT results that quantify correlated crashes/booms.
• Multi-Factor and Block Dependence Structures — how multiple factors yield heterogeneous dependence.
• SMM Estimation of Factor Copulas — the estimation method (no closed-form likelihood).
• Factor Copula Application - S&P 100 and Systemic Risk — the high-dimensional empirical study.
• SMM Estimator for Copulas — the companion Oh & Patton (2011) paper providing the estimator this paper applies.
• Bayesian copula estimation Describing correlated joint distributions — a PyMC Gaussian-copula tutorial; contrast the Bayesian Gaussian-copula approach with the frequentist, fat-tailed, factor-based approach here.

See Also

• 19. Simulated Method of Moments Estimation — Computational Methods for Economists using Python — general SMM method; this paper is a flagship application.
• Vine Copulas - Overview — the alternative high-dimensional copula architecture
• Econometrics