Alexander (2002) Principal Component Models for Generating Large GARCH Covariance Matrices

Summary

Full multivariate GARCH is practically impossible beyond a few dimensions — with so many parameters the likelihood becomes very flat and optimization fails — yet firm-wide risk management needs very large conditional covariance matrices. Alexander describes the principal-component GARCH ("orthogonal GARCH", O-GARCH) model, which generates such matrices by fitting only univariate GARCH models to the principal components of the risk factors. The paper (consolidating Alexander-Chibumba 1996 and Alexander 2000/2001b) shows the method is accurate and efficient, works best in highly correlated systems such as term structures, extends to combined equity/FX systems with careful calibration, and derives simple conditions for the resulting covariance matrix to be positive semi-definite.

Key Claims

• Why full multivariate GARCH fails. In more than a few dimensions the parameter count makes the likelihood surface very flat, so direct estimation of the large covariance matrices needed to net the risks of a large trading book is infeasible.
• O-GARCH construction. Take the principal components of the (correlation of the) risk factors; fit a univariate GARCH to each retained component's variance; reconstruct the conditional covariance as $H_t = A\,\mathrm{diag}(h_{1t},\ldots,h_{kt})\,A'$ from the constant factor loadings $A$ and component variances $h_{it}$. Only univariate estimation is required.
• Where it works best. Highly correlated systems — e.g. interest-rate term structures — where a few principal components capture most variation; with careful initial calibration, equities and FX can be combined into one large covariance matrix.
• Positive semi-definiteness. Simple conditions on the retained components guarantee the final covariance matrix is positive semi-definite by construction.
• Applications. Directly useful for value-at-risk and other large-scale risk systems.

Concepts Introduced or Extended

• Covariance Matrix Estimation

Entities Mentioned

• Carol Alexander

Quotes

"The O-GARCH model is an accurate and efficient method for generating large covariance matrices that only requires the estimation of univariate GARCH models … It works best in highly correlated systems, such as term structures."

My Take

O-GARCH is the pragmatic answer to the curse of dimensionality in multivariate volatility: rather than fight the flat likelihood of full BEKK/VEC GARCH, collapse the system to a handful of orthogonal factors, model each with a one-line univariate GARCH, and rebuild the matrix. That it needs only univariate fits is the whole selling point — it is what makes hundreds-of-assets covariance matrices tractable for VaR. The method's Achilles heel is exactly its strength: it assumes a low-dimensional, stable factor structure, so it shines on term structures (where 2–3 PCs dominate) but demands "careful calibration" for heterogeneous equity/FX books, and the retained-component choice trades PSD-safety against fidelity. It sits naturally beside DCC as the other scalable route to large conditional covariances — factor-orthogonalization versus correlation-targeting — and is the principal-component counterpart to the static shrinkage estimators on the covariance page.