Andersen-Bollerslev-Christoffersen-Diebold (2004) Practical Volatility and Correlation Modeling for Financial Market Risk Management

Summary

Andersen, Bollerslev, Christoffersen, and Diebold (2004) survey and unify recent developments in time-varying volatility, translated into practical guidance for financial risk management. The paper argues that the two dominant industry approaches — historical simulation and RiskMetrics — are both fundamentally flawed for different reasons, and proposes a hierarchy of generalized autoregressive conditional heteroscedasticity (GARCH)-based and high-frequency-data-based alternatives that range from simple univariate models to large-scale multivariate covariance systems. A central theme is that conditionality is paramount: value at risk (VaR) must reflect today's information about tomorrow's distribution, not an unconditional average.

Key Claims

• Historical simulation (HS-VaR) has no conditionality. The 1% HS-VaR from a 250-day rolling window can have a true conditional coverage probability as high as 10% on any given day. HS-VaR cannot produce a multi-day term structure beyond the trivial square-root-of-h scaling (invalid under non-i.i.d. returns).
• RiskMetrics has a flat volatility term structure. Exponential smoothing (EWMA, $\lambda=0.94$) is optimal if and only if squared returns follow a local-level (random-walk-plus-noise) model; that flatness is counterfactual and produces implausible k-day VaR forecasts.
• GARCH(1,1) is the minimum viable conditionality model. Mean-reversion parameter $(\alpha+\beta) < 1$ produces a realistic volatility term structure; k-step forecast: $E[\sigma^2_{t+k}] = \bar\sigma^2 + (\alpha+\beta)^{k-1}(\sigma^2_t - \bar\sigma^2)$.
• Component GARCH (Engle-Lee 1999): GARCH(2,2) with a time-varying long-run variance component $q_t$ approximates long memory and separates short- and long-run volatility dynamics.
• Dimensionality kills vech-GARCH: Full vech-GARCH(1,1) for $N$ assets has $O(N^4)$ parameters — 51 million for $N=100$. Practical solutions: Scalar GARCH (common $\alpha,\beta$), Flex-GARCH ($N(N+1)/2$ bivariate models, positive semi-definite [psd] by construction), dynamic conditional correlation (DCC; $N+2$ parameters total).
• DCC (Engle 2002) decouples variance and correlation: Step 1 — $N$ univariate GARCH models; Step 2 — GARCH(1,1) on standardized return outer products $Q_t = \bar Q(1-a-b) + a u_{t-1}u_{t-1}' + bQ_{t-1}$; normalize $R_t = \text{diag}(Q_t)^{-1/2}Q_t\text{diag}(Q_t)^{-1/2}$. Only 2 extra scalar parameters.
• Realized variance (RV) from intraday returns: $RV_t = \sum_{j=1}^D r^2_{t,j/D}$ approaches integrated volatility as $D \to \infty$; at 15–30 min intervals, measurement error is small. Log(RV) is approximately lognormal; RV autocorrelations are significant for 100+ lags (long memory). autoregressive moving average (ARMA) / autoregressive fractionally integrated moving average (ARFIMA) models on log(RV) outperform GARCH in out-of-sample volatility forecasting.
• Near-normality of RV-standardized returns (Andersen-Bollerslev-Diebold-Labys [ABDL] 2001, Zhou 1996): Daily returns standardized by realized volatility are close to standard normal; GARCH-standardized returns still show fat left tails. This enables the log-normal/normal mixture model: $\log(\sigma_t)|F_{t-1} \sim N(\mu,\tau^2)$, $r_t|\sigma_t \sim N(0,\sigma^2_t)$.
• Filtered Historical Simulation (FHS): Estimate a parametric GARCH model; standardize returns; bootstrap from standardized residuals to construct the conditional return distribution; simulate multi-day paths from the GARCH recursion using resampled innovations.
• Extreme equity correlations (Longin-Solnik 2001): Correlations between extreme negative equity returns are much larger than the bivariate normal distribution implies; bivariate normal predicts correlations → 0 in both tails, but data show strongly elevated negative-extreme correlations.
• Base-asset factor approach for large $N$: Model $N_F \ll N$ liquid base assets with realized covariance; map illiquid assets via $r_t = \Lambda f_t + \varepsilon_t$; FHS using factor innovations $f_t$ and idiosyncratic $\varepsilon_t$ resampled separately.

Concepts Introduced or Extended

• GARCH and BEKK-GARCH — component GARCH, Flex-GARCH, DCC, realized variance, FHS
• Value at Risk — HS-VaR pathology, FHS, log-normal/normal mixture, extreme tail correlations
• Realized Volatility — realized variance from intraday returns; near-normality of RV-standardized returns; ARFIMA forecasting

Entities Mentioned

• Torben G. Andersen — Northwestern/Kellogg
• Tim Bollerslev — Duke
• Peter F. Christoffersen — McGill
• Francis X. Diebold — UPenn

Quotes

"For most financial risk management purposes, the conditional perspective is exclusively relevant, notwithstanding, for example, the fact that popular approaches based on historical simulation and extreme-value theory typically adopt an unconditional perspective."

"Although the daily realized variance is just an estimate of the underlying integrated variance and likely measured with some error, it presents an intriguing opportunity: it is potentially highly accurate, and indeed accurate enough such that we might take the realized daily variance to be the true daily variance."

My Take

The paper is a well-organized survey that earns its place in a risk manager's reading list. The clearest contribution is the HS-VaR critique — the simulation in Figure 2 makes vivid what "lack of conditionality" means quantitatively, not just rhetorically. The realized variance discussion is the most academically forward-looking section: the near-normality of RV-standardized returns is a surprising empirical regularity that genuinely simplifies distributional modeling. The main limitation is that the paper was written at the boundary between theory and practice — many ideas (non-synchronous trading corrections for realized covariances, microstructure noise-robust estimators) were still "future research." A decade later, these are active and well-developed literatures. The practical ranking is roughly: GARCH > RiskMetrics > HS-VaR for univariate risk measurement; DCC > scalar GARCH > diagonal GARCH for multivariate; and realized-variance-based approaches dominate all GARCH-based ones when high-frequency data are available.