Engle (2002) Dynamic Conditional Correlation

Summary

Engle proposes the dynamic conditional correlation (DCC) class of multivariate GARCH models. The conditional covariance is factored as $H_t = D_t R_t D_t$, where $D_t$ collects univariate GARCH conditional standard deviations and $R_t$ is a time-varying correlation matrix driven by a parsimonious GARCH-like recursion. DCC keeps the flexibility of univariate GARCH while parameterizing the correlations with very few parameters, and — crucially — can be estimated in two steps from the likelihood (univariate GARCH first, then the correlation parameters), sidestepping the dimensionality explosion of conventional VEC/BEKK multivariate GARCH.

Key Claims

• Covariance decomposition. $H_t = D_t R_t D_t$ separates volatilities (from $N$ univariate GARCH models) from correlations ($R_t$), so each piece is modeled with the tool suited to it.
• DCC recursion. On the standardized residuals $u_t = D_t^{-1}\varepsilon_t$, a pseudo-correlation process evolves as $Q_t = (1-a-b)\bar Q + a\,u_{t-1}u_{t-1}' + b\,Q_{t-1}$, then rescaled to a valid correlation matrix $R_t = \mathrm{diag}(Q_t)^{-1/2} Q_t \, \mathrm{diag}(Q_t)^{-1/2}$. Scalar $a,b$ give O(1) correlation parameters regardless of $N$.
• Two-step (quasi-)ML estimation. The log-likelihood splits into a volatility part and a correlation part; fitting $N$ univariate GARCH models and then the correlation parameters is consistent if the first two conditional moments are correctly specified — with large computational advantages over full multivariate GARCH.
• Generalizes CCC. DCC nests the constant conditional correlation model (Bollerslev 1990) as the special case $a=b=0$, and empirically dominates it because financial correlations move (spiking in crises).
• Feasible for large $N$. Because the correlation parameter count does not grow with $N$, DCC scales to dozens–hundreds of assets, where BEKK/VEC are infeasible.

Concepts Introduced or Extended

• Conditional Correlation

Entities Mentioned

• Robert Engle

Quotes

"A new class of multivariate models called dynamic conditional correlation models is proposed. These have the flexibility of univariate GARCH models coupled with parsimonious parametric models for the correlations … estimated very simply with univariate or two-step methods based on the likelihood function."

My Take

DCC is the multivariate-GARCH model that actually gets used, and this is why: it trades a small amount of flexibility (scalar correlation dynamics, no cross-asset spillovers in the base version) for the ability to fit correlation matrices among many assets at near-univariate cost. The two-step estimator is the whole game — it turns an $O(N^2)$-parameter optimization into $N$ small ones plus a two-parameter step, which is what makes portfolio- and risk-scale applications tractable. On the wiki it is the primary source behind the DCC material on the Conditional Correlation page, sitting between the simpler CCC/BEKK constructions and the large-$N$ regularization frontier. The honest caveats are already in the literature: scalar DCC imposes common correlation dynamics on all pairs, its marginalization properties are not fully understood, and rigorous QML asymptotics lag behind BEKK's.