Dynamic Conditional Correlation (DCC-GARCH)
Allows volatilities and conditional correlations to change over time in multivariate returns.
How it works
DCC typically fits each return series with a univariate conditional volatility model and then estimates correlations using their standardised residuals. A matrix recursion makes correlations respond to new observations and decay toward their baseline. Normalising that matrix gives a correlation matrix, which combines with the individual volatilities to produce conditional covariance.
Worked example
If the two daily volatilities are 2% and 1%, a conditional correlation of 0.25 gives covariance 0.00005 in squared decimal-return units. Raising the correlation to 0.75 raises covariance to 0.00015, even though the volatility forecasts stay unchanged. A spread hedge can therefore change risk with no move in either standalone volatility.
Validation and limitations
Check stationarity conditions and positive definiteness, investigate instability across regimes and test covariance forecasts against realised portfolio risk. The baseline DCC specification typically constrains correlation persistence similarly across series. Test whether that is plausible for the chosen assets.