DCC-GARCH Dynamic Conditional Correlation Engine

A single correlation number, fit once over a long history, hides exactly the thing a risk manager needs to know: that correlations between risky assets tend to rise sharply in stress, precisely when diversification is supposed to help most. Engle's (2002) Dynamic Conditional Correlation model lets correlation move on its own clock.

How To Use This Model

Reading This Tool

Set each asset's own GARCH(1,1) volatility dynamics (how much a shock feeds through, how persistent volatility clustering is), a long-run average correlation, and the DCC parameters that govern how quickly correlation itself reacts to joint shocks and how persistent those reactions are. The engine then simulates both return series together under this joint process.

Watch the correlation path tab: even though the long-run average correlation is fixed by you, the realized path wanders meaningfully around it, and spikes higher exactly during the volatility clusters visible in the return series above it. That single picture is the entire empirical case for modeling correlation dynamically rather than as a static number.

GARCH & DCC Parameters

Asset 1 GARCH α (Shock Sensitivity) 0.08

Asset 1 GARCH β (Persistence) 0.90

Asset 2 GARCH α (Shock Sensitivity) 0.10

Asset 2 GARCH β (Persistence) 0.85

Long-Run Correlation ρ̄ 0.30

DCC a (Correlation Shock Sensitivity) 0.03

DCC b (Correlation Persistence) 0.95

Trading Days Simulated 1500

a + b must stay below 1 for the correlation process to be stationary (mean-reverting to ρ̄); this tool caps their sum automatically. Both assets' unconditional variance is fixed so that changing α or β reshapes clustering behavior without changing average volatility.

How Far Correlation Actually Wanders

a+b: –

Realized Average Correlation

Minimum Correlation Reached

Maximum Correlation Reached

Correlation Half-Life

Time-Varying Conditional Correlation vs. The Long-Run Average

Multivariate Volatility & Correlation Modeling

The Model

h_i,t = ω_i + α_iε_i,t-1² + β_ih_i,t-1 (univariate GARCH(1,1), each asset)
Q_t = (1−a−b)Q̄ + a(z_t-1z_t-1′) + bQ_t-1 (DCC(1,1) on standardized residuals z)
R_t = diag(Q_t)^−1/2Q_tdiag(Q_t)^−1/2

Each asset's own conditional variance follows a standard GARCH(1,1) recursion. The standardized residuals from both series then drive a separate, second GARCH-like recursion (Q_t) on the correlation matrix itself, rescaled at every step to guarantee a valid correlation matrix (unit diagonal). This build was verified two ways: each univariate GARCH process's simulated long-run variance matches the theoretical unconditional variance ω/(1−α−β) to within simulation noise, and setting a=b=0 collapses the model exactly onto a constant correlation equal to ρ̄ at every single time step, both mechanical identities a correct implementation must satisfy exactly.

When To Actually Use This Model

Daily portfolio risk systems where correlation inputs materially affect reported VaR or CVaR, and where using a rolling-window historical correlation would lag genuine regime changes.
Pricing multi-asset derivatives (basket options, correlation swaps) where the correlation term itself carries meaningful risk, not just each leg's own volatility.
Explaining, quantitatively, why a well-diversified-looking portfolio can suffer outsized joint losses precisely in the periods it was built to protect against.

Key Assumptions & Limitations

DCC assumes the same a and b govern every pair in a larger universe; in practice, different asset pairs often have genuinely different correlation dynamics, which this two-asset version sidesteps but larger implementations must confront.
Standard DCC treats positive and negative shocks symmetrically; asymmetric DCC variants (allowing correlation to react more to joint bad news) often fit real equity data better.
Like any GARCH-family model, parameters here are set directly rather than estimated by maximum likelihood from real data, production use requires a genuine quasi-maximum-likelihood fit, typically in two stages: univariate GARCH first, then DCC on the resulting residuals.

Foundational References

Engle, R. F. (2002). Dynamic Conditional Correlation: A Simple Class of Multivariate Generalized Autoregressive Conditional Heteroskedasticity Models. Journal of Business & Economic Statistics, 20(3), 339-350.

Bollerslev, T. (1986). Generalized Autoregressive Conditional Heteroskedasticity. Journal of Econometrics, 31(3), 307-327.

Engle, R. F., & Sheppard, K. (2001). Theoretical and Empirical Properties of Dynamic Conditional Correlation Multivariate GARCH. NBER Working Paper 8554.

Want to build a portfolio around this kind of tail-risk behavior?

Open The CVaR Portfolio Optimization Engine →