SVNexus · Dependence
DCC correlation
Also called: dynamic conditional correlation, DCC-GARCH
Dynamic conditional correlation (DCC) models correlations that change over time, built on GARCH volatilities for each asset.
Definition
Dynamic conditional correlation (DCC), introduced by Robert Engle in 2002, is a multivariate volatility model in which correlations between assets evolve over time. It first fits a univariate GARCH model to each asset, then models the correlation of the standardized residuals with a GARCH-like recursion.
The output is a full correlation matrix for every date.
Intuition
Rolling-window correlations react slowly and give equal weight to old and new observations. Correlations in markets often jump in stress and drift back afterwards. DCC captures that by letting today's correlation depend on yesterday's correlation and yesterday's co-movement surprise, the same way GARCH does for variance.
Separating volatility from correlation also avoids a classic trap: in high-volatility periods, raw co-movement rises even if the underlying correlation does not.
How it's calculated
Step 1: fit GARCH to each return series and form standardized residuals zi,t = εi,t / σi,t.
Step 2: model the quasi-correlation matrix Qt:
Qt = (1 − a − b) · Q̄ + a · zt−1 zt−1′ + b · Qt−1
where Q̄ is the unconditional correlation of z, and a ≥ 0, b ≥ 0, a + b < 1. Step 3: rescale to a proper correlation matrix:
Rt = diag(Qt)−1/2 · Qt · diag(Qt)−1/2
Parameters are estimated by maximum likelihood, usually in two stages. Because a and b are shared by all pairs, the model scales to many assets but assumes all correlations have the same persistence.
Example: with a = 0.05 and b = 0.93, a correlation shock decays at a rate of a + b = 0.98 per day, a half-life of ln(0.5) / ln(0.98) ≈ 34 trading days. A large joint surprise today keeps affecting the estimated correlation for weeks.
How to read it
Plot DCC correlations over time for a pair or a portfolio to see whether diversification holds in stress. A spike towards one across many pairs is a sign of a market-wide risk-off episode.
Compare with simple rolling correlation to understand how much of a move is a volatility effect.
Common mistakes
- Using raw returns instead of GARCH-standardized residuals in the correlation step.
- Assuming every pair shares the same dynamics because the model forces it.
- Over-reading day-to-day wiggles, which partly reflect estimation noise.
- Treating high correlation as causation or as a lead–lag relationship.
- Fitting DCC to assets with different trading hours without synchronizing their returns.
In Saqrid
SVNexus can build its asset network from DCC correlation links, or from return correlation, transfer entropy, tail dependence or implied-vol spillover.
Drilling into a pair shows correlation dynamics and a correlation backtest alongside tail risk, lead-lag and regime sync.
Open SVNexus.