1.9.6 Advanced Extensions

Basic Prerequisites: Linear Algebra (matrices, positive definiteness, Cholesky) and Econometrics & Time Series (forecasting).

---

1. Intuition & Practical Objective

So far every model handled one series. Portfolio risk, however, lives in the dependence structure: two assets each at 20% vol is a very different risk if their correlation is 0 or 0.9. Multivariate GARCH extends volatility modeling to a conditional covariance matrix ${\Sigma }_t$, and the central tension is that the naive extension explodes - a $k$-asset covariance matrix has $k(k+1)/2$ entries, and a full BEKK parameterises them all.

The resolution is two-step / correlation decomposition: model each asset’s volatility univariately (a problem already solved), then model only the correlation matrix dynamically. This is the CCC (constant conditional correlation, Bollerslev 1990) and DCC (dynamic conditional correlation, Engle 2002) family. DCC keeps the parameter count fixed (${\theta }_1,{\theta }_2$ regardless of $k$) and guarantees positive definiteness by construction.

The payoff: portfolio VaR, risk-parity weights, and hedging ratios that move with the market’s correlation regime - because correlations, like volatilities, cluster and rise together in crises (“correlations go to one in a crash”). This page also covers volatility forecasting and term structure, the bridge to risk applications.

---

2. Mathematical Ground Truth & Derivations

The multivariate setup. With $a_t$ the $k$-vector of shocks, ${\Sigma }_t=\mathrm{Var}(a_t\mid {\mathcal{F}}_{t-1})=\mathbb{E}[a_t{a}_t^{\prime }\mid {\mathcal{F}}_{t-1}]$, and ${\Sigma }_t^{1/2}$ any matrix square root, $a_t={\Sigma }_t^{1/2}{z}_t$, $z_t\stackrel{iid}{\sim }(0,I_k)$.

VEC / DVEC and BEKK (the parameter explosion). The general VEC has each covariance element following its own GARCH-like recursion (via the Hadamard product $\odot$, i.e. DVEC); it does not guarantee positive definiteness and allows no clean cross-dependence. BEKK (Engle–Kroner 1995),

$${\Sigma }_t=A{A}^{\prime }+\sum _{i=1}^m{A}_i(a_{t-i}{a}_{t-i}^{\prime })A_i^{\prime }+\sum _{j=1}^s{B}_j{\Sigma }_{t-j}{B}_j^{\prime },$$

is positive definite almost surely if $A{A}^{\prime }$ is, but has $k^2(m+s)+k(k+1)/2$ parameters - for $k=10$ that is hundreds, none individually interpretable. The shocks in the BEKK quadratic term are raw innovations $a_{t-i}{a}_{t-i}^{\prime }$, not standardized.

Correlation decomposition (the practical route). Write

$${\Sigma }_t=D_t{R}_t{D}_t,D_t=\mathrm{diag}\{\sqrt{{\sigma }_{11,t}},\dots ,\sqrt{{\sigma }_{kk,t}}\},R_t=({\rho }_{ij,t}).$$

Here each ${\sigma }_{ii,t}$ is a univariate GARCH(1,1) (fast, well-understood) and $R_t$ is the dynamic correlation matrix.

• CCC (Bollerslev 1990): $R_t=\bar{R}$ constant. The log-likelihood separates into $k$ univariate pieces plus a correlation piece - estimation is trivial and consistent, but it ignores that correlations move.
• DCC (Engle 2002): let ${\epsilon }_{it}=a_{it}/\sqrt{{\sigma }_{ii,t}}$ be the standardized shocks, and evolve a pseudo-correlation matrix

$$Q_t=(1-{\theta }_1-{\theta }_2)\bar{Q}+{\theta }_1{\epsilon }_{t-1}{\epsilon }_{t-1}^{\prime }+{\theta }_2{Q}_{t-1},R_t=J_t{Q}_t{J}_t,J_t=\mathrm{diag}\{q_{ii,t}^{-1/2}\}.$$

The rescaling $J_t{Q}_t{J}_t$ forces unit diagonal, turning $Q_t$ into the true correlation matrix. ${\theta }_1+{\theta }_2<1$ is the stationarity condition, and the whole model adds just two parameters regardless of dimension. Tse–Tsui (2002) is the direct analogue ${\rho }_t=(1-{\theta }_1-{\theta }_2)\bar{\rho }+{\theta }_1{\rho }_{t-1}+{\theta }_2{\psi }_{t-1}$.

Estimation (two-step). (1) Fit $k$ univariate GARCH(1,1)s → ${\hat{D}}_t$, standardized residuals ${\hat{\epsilon }}_t$. (2) Fit $({\theta }_1,{\theta }_2)$ by Gaussian likelihood on ${\hat{\epsilon }}_t$. Consistent under the two-step (Engle 2002); full ML is possible but expensive.

Portfolio VaR with dynamic correlation. For a two-asset portfolio,

$$Va{R}_{1+2}=\sqrt{Va{R}_1^2+Va{R}_2^2+2\rho Va{R}_{1}Va{R}_2}.$$

Tsay §10.7 (Cisco+Intel, $1M each, 5%) reports $57,117 (univariate) < $57,648 (time-varying corr) < $58,180 (constant corr) - the ordering showing that ignoring dynamic correlation understates joint risk when correlation rises.

Forecasting & term structure. From a fitted GARCH(1,1), the $h$-step variance forecast is ${\sigma }_h^2(\ell )={\alpha }_0+({\alpha }_1+{\beta }_1){\sigma }_h^2(\ell -1)$, decaying geometrically to the unconditional variance. The term structure (the shape of ${\sigma }_h^2(\ell )$ vs $\ell$) is upward-sloping when today’s vol is below the long-run mean and downward-sloping when above - the “volatility cone”. For $h$-day VaR, integrate the term structure: $Va{R}^{(h)}=z_p\sqrt{{\sum }_{\ell =1}^h{\sigma }_h^2(\ell )}$.

---

3. Computational Implementation - DCC recursion and the vol term structure

Standard library only. (a) Simulates two GARCH(1,1) assets with an average 0.60 correlation, runs the DCC(1,1) correlation recursion on the standardized residuals, and confirms the recovered mean correlation. (b) Takes a fitted GARCH(1,1) state after a $-3\%$ shock and reads off the volatility term structure and the corresponding 99% VaR at 1d/1w/2w/1m.

The DCC recursion recovers a mean correlation of 0.5916 against a true 0.60 - the small downward bias comes from the $Q_t$ targeting and finite sample - and it moves over 1949 days (range $[0.384,0.749]$): correlation is dynamic, exactly as the model claims. Panel (b) shows the term structure after a $-3\%$ day: 20.3% annualized today decaying toward the 10% unconditional vol - the forecast is downward-sloping because the current shock has pushed vol above its long-run level. The 99% VaR shrinks monotonically with horizon per day ($2.98\%\to 2.77\%$), because a single shock’s influence fades - the $h$-day VaR is sub-linear in $h$, not $\sqrt{h}$ (a genuine, exploitable improvement over the naive rule).

---

4. Failure Modes & First-Principles Breakdowns

1. Dimensionality is the enemy. BEKK/VEC parameter counts grow as $k^2$ and estimates are uninterpretable past $k\approx 5$; the DCC/CCC decomposition exists precisely to keep estimation feasible. Prefer factor or Cholesky structures for large $k$ (Tsay §10.5–10.6).
2. DCC scalar dynamics assume common persistence. The single $({\theta }_1,{\theta }_2)$ forces all pairwise correlations to share the same persistence - implausible when some pairs (e.g. within-sector) are far more tightly coupled than others.
3. DCC has no leverage. The basic DCC correlation recursion is symmetric; the multivariate-$t$ + leverage extension (Tsay 2006) is needed for the crisis behaviour where correlations spike.
4. Two-step estimation error propagates. Fitting the univariate GARCHs first and the correlations second ignores the uncertainty in step 1; standard errors are understated.
5. Correlations spike in crises. Because $R_t$ rises toward 1 in a sell-off, a CCC or low-correlation assumption systematically understates portfolio VaR exactly in the tail - the verified Tsay ordering ($57.1k<57.6k<58.2k$) is a mild illustration; the effect is severe for concentrated equity books.
6. Forecast horizon vs re-estimation. Term-structure forecasts assume parameters are constant over the horizon; over weeks–months the structural-break failure of 05 dominates, and the “decay to unconditional variance” is anticipated by nothing.

---

5. References

• Engle, Robert F. (2002): Dynamic Conditional Correlation: A Simple Class of Multivariate Generalized Autoregressive Conditional Heteroskedasticity Models, J. Business & Economic Statistics 20(3), 339–350
• Bollerslev, Tim (1990): Modelling the Coherence in Short-Run Nominal Exchange Rates: A Multivariate Generalized ARCH Model, Review of Economics and Statistics 72(3)
• Engle, Robert F. & Kroner, Kenneth F. (1995): Multivariate Simultaneous Generalized ARCH, Econometric Theory 11(1)
• Tse, Yiu K. & Tsui, Albert K. (2002): A Multivariate GARCH Model with Time-Varying Correlations, J. Business & Economic Statistics 20(3).
• Tsay, Ruey S.: Analysis of Financial Time Series (3rd ed., 2010)

---

6. Connected Graph Bridges

• Base: Linear Algebra (PD matrices, Cholesky) · Econometrics & Time Series (forecasting)
• Prior: 02 · ARCH & GARCH · 05 · Failure Modes · Hub: Index
• Applied: VaR & Expected Shortfall (DCC covariance → portfolio VaR) · Covariance Shrinkage & Denoising · Risk Parity · Signal Processing & Kalman (SV vs GARCH, state-space covariance)
• Home: GARCH & Volatility Modeling - Index Hub