Using GARCH Volatility in Mean-Variance Portfolio Optimization

Summary

The document raises the question of how to incorporate GARCH volatility forecasts into a Markowitz mean-variance portfolio optimization. The investor’s existing process estimates expected returns and a variance-covariance matrix, then seeks a portfolio with an attractive Sharpe ratio. GARCH is considered as a way to represent volatility clustering, where volatility changes over time rather than staying constant.

The central issue is whether asset-level conditional variance estimates can simply replace the usual variances in the covariance matrix, or whether covariances also need to be modeled dynamically. The question points to multivariate GARCH models as one way to estimate time-varying covariance relationships. However, the document contains no answer, model specification, empirical comparison, or optimization results. It therefore frames a methodological choice rather than providing a validated procedure; any implementation would still need coherent expected-return and covariance forecasts aligned to the portfolio’s horizon.

Key ideas

• GARCH models can represent volatility that changes over time and clusters.
• Mean-variance optimization requires both expected returns and a covariance matrix.
• Univariate variance forecasts alone do not specify how correlations or covariances should change.
• Multivariate GARCH is raised as a possible approach to conditional covariance estimation.
• The document poses the modeling question but supplies no recommendation or empirical evidence.

Tags

# GARCH for Mean Variance Optimization


# GARCH for Mean Variance Optimization












I am currently trying to carry out a mean variance optimisation, with the implementation of GARCH. I'm not sure if this is going to make complete sense as my understanding of GARCH is limited.

In the past whenever I have carried out mean variance optimisation (under Markowitz) I have calculated the expected returns, created a var-covar matrix and maximised the Sharpe Ratio [(Er-rf)/St.dev^2].

Currently, instead of carrying this out as usual, a friend who works in risk management suggested that I look at GARCH models to more accurate model volatility to account for clustering.

My question is what changes do I have to make to my mean variance optimisation for this to work? Can I take the variance outputs for each asset and run the var-covar matrix as normal and then maximise sharpe or do I have to continue to make further changes for the optimisation to actually make sense? I have read in some places that to create covariance matrix's for GARCH its best to run multivariate models.

Any help would be hugely appreciated:)

Shown in full with attribution under the source's licence. Licence: CC BY-SA 4.0 (Stack Exchange)

This summary was written by Stratmill's research agent from the original; it is not a copy of the source.