Multivariate GARCH Model Comparison: DCC, CGARCH, and GOGARCH • tsbs

Overview

The tsbs package supports three multivariate GARCH correlation specifications for modeling time-varying dependence in financial time series:

DCC (Dynamic Conditional Correlation) - Engle (2002)
CGARCH (Copula GARCH) - Patton (2006), Jondeau & Rockinger (2006)
GOGARCH (Generalized Orthogonal GARCH) - van der Weide (2002)

This vignette compares these three approaches, highlighting their:

Theoretical foundations and model structure
Key assumptions and flexibility
Computational considerations
Appropriate use cases
Integration with the MS-VARMA-GARCH bootstrap framework

Model Taxonomy

Multivariate GARCH Correlation Models
├── Correlation-Based Approaches
│   ├── DCC (dcc_modelspec)
│   │   └── Models correlation dynamics directly
│   └── CGARCH (cgarch_modelspec)
│       └── Separates marginals from dependence via copulas
└── Factor-Based Approach
    └── GOGARCH (gogarch_modelspec)
        └── Derives correlation from independent components

Part I: DCC (Dynamic Conditional Correlation)

Model Structure

The DCC model of Engle (2002) specifies time-varying correlations through a two-stage process:

Stage 1: Univariate GARCH

For each series $i = 1, \ldots, k$ : $r_{i,t} = \mu_{i,t} + \varepsilon_{i,t}, \quad \varepsilon_{i,t} = \sigma_{i,t} z_{i,t}$ $\sigma_{i,t}^2 = \omega_i + \alpha_i \varepsilon_{i,t-1}^2 + \beta_i \sigma_{i,t-1}^2$

Stage 2: Correlation Dynamics

The standardized residuals $z_t = (z_{1,t}, \ldots, z_{k,t})'$ follow DCC dynamics: $Q_t = (1 - \alpha - \beta) \bar{Q} + \alpha z_{t-1} z_{t-1}' + \beta Q_{t-1}$ $R_t = \text{diag}(Q_t)^{-1/2} \, Q_t \, \text{diag}(Q_t)^{-1/2}$

where $\bar{Q}$ is the unconditional covariance of $z_t$ , and $R_t$ is the time-varying correlation matrix.

Key Parameters

Parameter	Description	Typical Range
$\alpha$	News impact (ARCH effect)	0.01 - 0.10
$\beta$	Persistence (GARCH effect)	0.85 - 0.98
$\alpha + \beta$	Total persistence	< 1 (stationarity)

Assumptions

Normality or Student-t: Standardized residuals follow MVN or MVT
Scalar dynamics: Same $\alpha, \beta$ for all correlation pairs
Symmetric response: No distinction between positive/negative shocks

Usage in tsbs

spec <- list(
  list(
    var_order = 1,
    garch_spec_fun = "dcc_modelspec",
    distribution = "mvn",  # or "mvt"
    garch_spec_args = list(
      dcc_order = c(1, 1),
      garch_model = list(...)
    )
  )
)

Strengths

Simple, parsimonious (only 2 correlation parameters)
Computationally efficient
Well-established asymptotic theory
Good for capturing smooth correlation dynamics

Limitations

Assumes common dynamics across all pairs
Limited flexibility for marginal distributions
May miss asymmetric dependence patterns
“Flat beta problem” in inference (see inference guide)

Part II: CGARCH (Copula GARCH)

Model Structure

Copula GARCH separates marginal distributions from the dependence structure using copula theory.

Stage 1: Univariate GARCH + PIT

Same univariate GARCH as DCC, but with Probability Integral Transform: $u_{i,t} = F_i(z_{i,t} | \theta_i)$

where $F_i$ is the marginal CDF. This transforms residuals to uniform $[0,1]$ .

Stage 2: Copula Transformation

Transform uniform margins to copula residuals: $\tilde{z}_{i,t} = \Phi^{-1}(u_{i,t}) \quad \text{(Gaussian copula)}$ $\tilde{z}_{i,t} = t_\nu^{-1}(u_{i,t}) \quad \text{(Student-t copula)}$

Stage 3: Correlation Dynamics

Apply DCC-style dynamics to copula residuals: $Q_t = (1 - \alpha - \beta) \bar{Q} + \alpha \tilde{z}_{t-1} \tilde{z}_{t-1}' + \beta Q_{t-1}$

PIT Transformation Methods

Method	Description	Use Case
`parametric`	Uses fitted distribution CDF	Standard approach
`empirical`	Uses empirical CDF	Non-parametric, robust
`spd`	Semi-parametric distribution	Flexible tails

Copula Distributions

Copula	Properties	Parameters
`mvn`	Gaussian	None (symmetric)
`mvt`	Student-t	Shape $\nu$ (tail dependence)

ADCC Extension

CGARCH supports Asymmetric DCC (ADCC) for leverage effects: $Q_t = (1 - \alpha - \beta - \delta\bar{\gamma}) \bar{Q} + \alpha \tilde{z}_{t-1} \tilde{z}_{t-1}' + \beta Q_{t-1} + \gamma n_{t-1} n_{t-1}'$

where $n_t = \tilde{z}_t \cdot \mathbf{1}(\tilde{z}_t < 0)$ captures negative shocks.

Usage in tsbs

spec <- list(
  list(
    var_order = 1,
    garch_spec_fun = "cgarch_modelspec",
    distribution = "mvn",  # copula distribution
    garch_spec_args = list(
      dcc_order = c(1, 1),
      dynamics = "dcc",           # or "adcc", "constant"
      transformation = "parametric",  # or "empirical", "spd"
      copula = "mvn",             # or "mvt"
      garch_model = list(univariate = list(...))
    )
  )
)

Comparison: DCC vs CGARCH

Aspect	DCC	CGARCH
Marginal distributions	Implicitly normal	Flexible via PIT
Tail dependence	Limited (MVT only)	Student-t copula
Transformation step	None	PIT to uniform
Asymmetric dynamics	Not standard	ADCC available
Complexity	Lower	Higher
Parameters	2 (α, β)	2-4 (α, β, [γ], [ν])

Strengths

Flexible marginal distributions
Captures tail dependence with MVT copula
ADCC for asymmetric correlation response
SPD transformation for robust tail modeling

Limitations

Higher computational cost
More parameters to estimate
Same “flat beta problem” as DCC
PIT step introduces additional estimation uncertainty

Part III: GOGARCH (Generalized Orthogonal GARCH)

Model Structure

GOGARCH takes a fundamentally different approach: instead of modeling correlations directly, it models multivariate volatility through independent components.

Stage 1: ICA Decomposition

Decompose observed residuals into independent components: $\varepsilon_t = A \cdot s_t$

where $A$ is the mixing matrix and $s_t$ are independent components. Equivalently: $s_t = W \cdot \varepsilon_t$ where $W = A^{-1}$ is the unmixing matrix.

Stage 2: Component GARCH

Each independent component follows univariate GARCH: $s_{i,t} = \sqrt{h_{i,t}} \cdot \eta_{i,t}$ $h_{i,t} = \omega_i + \alpha_{1,i} s_{i,t-1}^2 + \beta_{1,i} h_{i,t-1}$

where $\eta_{i,t} \stackrel{iid}{\sim} (0, 1)$ .

Stage 3: Covariance Reconstruction

Time-varying covariance is reconstructed via: $\Sigma_t = A \cdot H_t \cdot A'$

where $H_t = \text{diag}(h_{1,t}, \ldots, h_{k,t})$ .

The correlation matrix is: $R_t = D_t^{-1} \Sigma_t D_t^{-1}$

where $D_t = \text{diag}(\sqrt{\Sigma_{11,t}}, \ldots, \sqrt{\Sigma_{kk,t}})$ .

ICA Methods

Method	Algorithm	Properties
`radical`	RADICAL	Robust, outlier-resistant
`fastica`	FastICA	Fast, widely used
`jade`	JADE	Joint diagonalization

Key Insight: No Direct Correlation Parameters

Unlike DCC/CGARCH, GOGARCH has no α, β parameters for correlation dynamics. Instead, time-varying correlations emerge from:

The fixed mixing matrix $A$ (estimated via ICA)
The time-varying component variances $h_{i,t}$ (via GARCH)

Usage in tsbs

spec <- list(
  list(
    var_order = 1,
    garch_spec_fun = "gogarch_modelspec",
    distribution = "norm",  # or "std"
    garch_spec_args = list(
      ica_method = "radical",  # or "fastica"
      garch_model = list(
        model = "garch",
        garch_order = c(1, 1)
      )
    )
  )
)

Comparison: DCC/CGARCH vs GOGARCH

Aspect	DCC/CGARCH	GOGARCH
Correlation source	Explicit dynamics	ICA + component volatility
Main parameters	α, β (correlation)	ω, α₁, β₁ per component
Number of params	2-4 total	3k (k = series)
Estimation	Two-stage MLE	ICA + GARCH MLE
Correlation dynamics	Smooth, autoregressive	Can be more complex
Cross-sectional	Scalar (common)	Component-specific

Strengths

No assumption of common dynamics across pairs
Can capture complex correlation patterns
Natural for factor-based modeling
Higher-order moments (coskewness, cokurtosis) available

Limitations

ICA estimation uncertainty
Requires more series for reliable ICA
Mixing matrix interpretation can be challenging
Higher computational cost for many series

Part IV: Model Selection Guide

When to Use DCC

Few series (2-10): DCC is parsimonious
Smooth correlation dynamics: DCC captures gradual changes
Quick estimation: DCC is fastest
Established benchmarks: DCC is standard in finance

When to Use CGARCH

Non-normal marginals: When univariate distributions are clearly non-Gaussian
Tail dependence matters: For risk management with extreme events
Asymmetric correlation response: Use ADCC for leverage effects
Robust tail modeling: SPD transformation for heavy tails

When to Use GOGARCH

Factor structure suspected: When underlying independent factors drive returns
Higher-order moments needed: Coskewness/cokurtosis for portfolio optimization
Complex correlation patterns: When DCC dynamics are too restrictive
Many series with latent structure: Natural for factor models

Decision Flowchart

Start
  │
  ├─ Are marginals clearly non-normal?
  │   ├─ Yes → Consider CGARCH
  │   └─ No  → Continue
  │
  ├─ Is there a suspected factor structure?
  │   ├─ Yes → Consider GOGARCH
  │   └─ No  → Continue
  │
  ├─ Need asymmetric correlation response?
  │   ├─ Yes → Use CGARCH with ADCC
  │   └─ No  → Continue
  │
  ├─ Is parsimony important?
  │   ├─ Yes → Use DCC
  │   └─ No  → Consider all three
  │
  └─ Default: DCC (simplest, well-understood)

Part V: Inference Considerations

The Flat Beta Problem (DCC & CGARCH)

Both DCC and CGARCH share the same correlation dynamics, leading to:

Likelihood surface ~8x flatter in β direction than α
Hessian-based SEs underestimate β uncertainty by ~5-6x
Suggestion: Use bootstrap for β inference

GOGARCH Inference

GOGARCH has different challenges:

ICA estimation uncertainty not captured by Hessian
Two-stage estimation (ICA then GARCH)
Suggestion: Always use bootstrap for all parameters

Summary Table

Model	α SE Method	β SE Method	Other Parameters
DCC	Hessian OK	Bootstrap	Shape: Hessian OK
CGARCH	Hessian OK	Bootstrap	γ, ν: Bootstrap
GOGARCH	Bootstrap	Bootstrap	All: Bootstrap

Part VI: Computational Considerations

Estimation Time (Typical)

Model	Small (k=2)	Medium (k=5)	Large (k=10)
DCC	Fast	Fast	Moderate
CGARCH	Moderate	Moderate	Slow
GOGARCH	Moderate	Slow	Very Slow

Memory Usage

DCC/CGARCH: Stores k×k correlation matrices per time point
GOGARCH: Stores ICA decomposition + k volatility series

Parallelization

All three models support parallelization in the tsbs bootstrap:

library(future)
plan(multisession, workers = 4)

result <- tsbs(
  data = data,
  spec = spec,
  n_boot = 1000,
  parallel = TRUE
)

Part VII: Integration with MS-VARMA-GARCH

All three models integrate with the Markov-Switching framework:

# Two-regime model with DCC
spec_ms <- list(
  # Regime 1: Low volatility
  list(
    var_order = 1,
    garch_spec_fun = "dcc_modelspec",
    distribution = "mvn"
  ),
  # Regime 2: High volatility
  list(
    var_order = 1,
    garch_spec_fun = "dcc_modelspec",
    distribution = "mvt"
  )
)

# Can mix model types across regimes
spec_mixed <- list(
  # Regime 1: Simple DCC
  list(garch_spec_fun = "dcc_modelspec", ...),
  # Regime 2: CGARCH for crises (better tail modeling)
  list(garch_spec_fun = "cgarch_modelspec", ...)
)

References

DCC: - Engle, R. (2002). Dynamic conditional correlation: A simple class of multivariate GARCH models. Journal of Business & Economic Statistics, 20(3), 339-350. - Aielli, G. P. (2013). Dynamic conditional correlation: On properties and estimation. JBES, 31(3), 282-299.

Copula GARCH: - Patton, A. J. (2006). Modelling asymmetric exchange rate dependence. International Economic Review, 47(2), 527-556. - Jondeau, E. & Rockinger, M. (2006). The Copula-GARCH model of conditional dependencies. Journal of International Money and Finance, 25(5), 827-853.

GOGARCH: - van der Weide, R. (2002). GO-GARCH: A multivariate generalized orthogonal GARCH model. Journal of Applied Econometrics, 17(5), 549-564. - Boswijk, H. P. & van der Weide, R. (2011). Method of moments estimation of GO-GARCH models. Journal of Econometrics, 163(1), 118-126.

ICA: - Learned-Miller, E. G. (2003). ICA using spacings estimates of entropy. JMLR, 4, 1271-1295. - Hyvärinen, A. & Oja, E. (2000). Independent component analysis: algorithms and applications. Neural Networks, 13(4-5), 411-430.