Multi-Factor and Block Dependence Structures

Summary

Two extensions enrich the simple equidependence copula. Flexible weights (${\beta }_{i}Z$) break equidependence so pairs can differ in dependence strength. Multiple common factors ($K$-factor model) capture heterogeneous, grouped (e.g. industry) dependence. The empirically central case is block equidependence: a market-wide factor plus industry-specific factors, with common loadings within ex-ante groups — this greatly increases flexibility while keeping parameters $O(N)$ rather than $O(N^2)$.

Overview

The simple model $X_i=Z+{\epsilon }_i$ forces every pair to share one bivariate copula (equidependence). Real asset returns show heterogeneous pairwise dependence (e.g. stronger within-industry). The paper grows flexibility along two axes — heterogeneous loadings and multiple factors — while keeping the model interpretable, testable, and tractable in high dimensions. The block-equidependence structure is the workhorse of the S&P 100 application and the $N=100$ simulation.

Main Content

$K$-factor copula model

Dependence arises from $K$ common factors, for $i=1,\dots ,N$:

$$X_i=\sum _{k=1}^K{\beta }_{ik}{Z}_k+{\epsilon }_i$$ $${\epsilon }_i\sim \text{iid }{F}_{\epsilon },Z_k\perp \perp {\epsilon }_i\forall i,k$$ $$[Z_1,\dots ,Z_K{]}^{\prime }\equiv \mathbf{Z}\sim {\mathbf{F}}_z={\mathbf{C}}_{indep}(F_{z_1},\dots ,F_{z_K})$$

where ${\beta }_{ik}$ is the loading of variable $i$ on factor $k$. In full generality $\mathbf{Z}$ could have any copula ${\mathbf{C}}_Z$, but the empirically useful simplification imposes independent common factors, removing the need to specify/estimate ${\mathbf{C}}_Z$. A further simplification fixes each loading to one or zero, with weights specified in advance by grouping variables. This is a special case of the conditional independence structure of McNeil et al. (2005): variables are independent conditional on the smaller factor set $\mathbf{Z}$ (the factors are the “frailty” in credit/survival literature).

Single-factor flexible weights → heterogeneous pairs

The intermediate step $X_i={\beta }_{i}Z+{\epsilon }_i$ (see Factor Copula Construction) already breaks equidependence: pairs with larger $\min \{{\beta }_i,{\beta }_j\}$ are more dependent. Cost: $N-1$ extra parameters. To control this, assign common loadings within ex-ante groups (e.g. industry classifications), yielding a block equidependence copula — far fewer parameters than $N-1$ free loadings.

Empirical block-equidependence model (S&P 100)

The model used in the application combines one market-wide factor with seven industry factors (groups formed by first-digit SIC), for $i=1,\dots ,100$:

$$X_i={\beta }_i{Z}_0+{\gamma }_i{Z}_{S(i)}+{\epsilon }_i$$ $$Z_0\sim \text{Skew }t(\nu ,\lambda )$$ $$Z_S\sim \text{iid }t(\nu ),S=1,\dots ,7,Z_S\perp \perp Z_0\forall S$$ $${\epsilon }_i\sim \text{iid }t(\nu ),{\epsilon }_i\perp \perp Z_j\forall i,j$$

where $S(i)$ is the SIC group of stock $i$. There are 8 latent factors total, but each variable is affected by only two (its market and its own industry factor), simplifying structure and reducing free parameters. Asymmetry is allowed only in the market factor $Z_0$; industry factors and idiosyncratic shocks are symmetric (parsimony). All stocks in a group share $({\beta }_i,{\gamma }_i)$, but different groups may differ. Total: 16 parameters — more flexible than the 3-parameter equidependence model, far more parsimonious than a fully unstructured 100-dimensional copula.

Block structure of the dependence-measure matrix (App. B)

Estimation exploits the block structure. The $N\times N$ pairwise dependence matrix $D$ (entries ${\delta }_{ij}$ = rank correlation or quantile dependence) is partitioned into sub-matrices $D_{rs}$ by group. Because all pairs in groups $(r,s)$ share the same dependence, one averages within each block to form an $M\times M$ matrix $D^{\ast }$ of block-average measures:

$${\delta }_{ss}^{\ast }\equiv \frac{2}{k_s(k_s-1)}\sum \sum {\hat{\delta }}_{ij}(\text{avg of upper-triangle of diagonal block }{D}_{ss})$$ $${\delta }_{rs}^{\ast }\equiv \frac{1}{k_r{k}_s}\sum \sum {\hat{\delta }}_{ij}(\text{avg of all of off-diagonal block }{D}_{rs},r\ne s)$$

where $k_m$ is the number of variables in group $m$. Averaging the rows of $D^{\ast }$ gives an $M$-vector ${\bar{\delta }}^{\ast }$, yielding $M$ moments per dependence measure ($5M$ total for the five measures). This is what makes high-dimensional SMM feasible — see SMM Estimation of Factor Copulas.

Examples

What the industry factors reveal (S&P 100, skew $t$-$t$)

Setup: Block model estimated on 100 S&P stocks in 7 SIC groups. Result: Market-factor loadings ${\beta }_i$ range 0.88 (Food/apparel manufacturing) to 1.25 (Mining & construction), all significant at 5%. Industry-factor loadings ${\gamma }_i$ (extra intra-industry dependence beyond the market factor) range 0.17 to 1.09, all significantly $\ne 0$. Implied rank correlations span 0.39 (cross-industry pairs in SIC 1 vs 5) to 0.72 (within SIC 1). Adding industry factors makes the market factor look more fat-tailed and more left-skewed (${\nu }^{-1}\approx 1/14$, larger and more significant than the single-factor $\approx 1/25$). Interpretation: A single common factor masks both fat tails and asymmetry; controlling for intra-industry dependence reveals stronger systematic crash risk. Tests (over-identifying $J$-test, restriction tests) strongly reject removing the industry factors, removing the market factor, and reducing to equidependence.

Connections

• Factor Copula Construction — the base single-factor model these generalize.
• Tail Dependence in Factor Copulas — Proposition 3 gives tail dependence for the $K$-factor model.
• SMM Estimation of Factor Copulas — block averaging reduces $5N(N-1)/2$ pairwise moments to $5M$.
• Factor Copula Application - S&P 100 and Systemic Risk — where the block model is estimated and tested.

See Also

• Factor Copulas - Overview
• Dependence Measures for Copulas — the five pairwise moments (rank correlation, quantile dependence) used for block-averaged SMM estimation
• Factor Analysis and PPCA — parallel factor structure in the Bayesian latent-variable setting; compare continuous vs. copula factor models
• Copula Estimation — Bayesian Gaussian copula for comparison with the frequentist SMM factor copula approach
• Econometrics