Sağlam Koşullu Korelasyon GARCH (Sağlam DCC-GARCH)

Standard DCC-GARCH maximises a full or quasi-maximum likelihood that is sensitive to outliers and numerically intractable for many assets. Robust DCC-GARCH replaces this with an M-estimation or composite-likelihood objective that downweights extreme observations and decomposes the joint likelihood into bivariate pairs, yielding consistent, outlier-resistant parameter estimates that scale to large dimensions.

Composite likelihood sums the log-likelihoods of all bivariate marginal distributions of the asset returns rather than evaluating the full multivariate likelihood. This reduces computation from O(n^3) to O(n^2) in the number of assets and avoids numerical instabilities in large correlation matrices, while retaining consistency of the estimator.

The Huber tuning constant c (often set near 1.345 for 95% efficiency under normality) controls the threshold beyond which residuals are downweighted. A smaller c gives more robustness but lower efficiency. Sensitivity analysis over a range of c values — comparing parameter stability — is good practice.

The rmgarch R package implements composite-likelihood DCC and supports robust extensions. The ccgarch and MTS packages offer related functionality. Stata and Python (arch library) support standard DCC; fully robust variants typically require custom likelihood programming or use of rmgarch.

Robust DCC-GARCH is preferable when you want a parsimonious, interpretable dynamic correlation model and outlier resistance is the main concern. Copula-based models offer more flexibility in modelling the full dependence structure and tail dependence separately, but at substantially higher parametric complexity. For large portfolios and speed-robustness trade-off, Robust DCC-GARCH is usually the pragmatic choice.