Mean-variance optimization is the method, introduced by Harry Markowitz in 1952, of choosing portfolio weights to get the highest expected return for a given level of risk, or the lowest risk for a given return. Risk is measured by variance, and diversification comes from how assets move together.
Inputs and idea
- Expected returns for each asset
- Volatilities and correlations, together forming the covariance matrix
- An objective such as maximising return minus a penalty for variance
A worked two-asset example
| Asset A | Asset B | |
|---|---|---|
| Expected return | 10% | 6% |
| Volatility | 20% | 10% |
| Correlation | 0.3 | |
A 50/50 mix has an expected return of 8%. Its variance is 0.25 x 0.04 + 0.25 x 0.01 + 2 x 0.25 x 0.3 x 0.20 x 0.10 = 0.0155, so volatility is about 12.4%, lower than the 15% average of the two volatilities. That gap is the benefit of diversification. The numbers are made up. See efficient frontier.
The practical problem
The optimiser treats the inputs as exact, so it loads up on assets whose estimated return is high or estimated risk is low, which are often the assets with the biggest estimation errors. The result is concentrated, unstable weights that change a lot when the inputs change slightly. This is sometimes called error maximisation.
Common remedies
- Constraints on weights, sectors and turnover. See optimization constraints.
- Better covariance estimates, such as shrinkage.
- Blending views with market equilibrium. See Black-Litterman model.
- Methods that avoid return forecasts, such as minimum variance, risk parity and hierarchical risk parity.
See also portfolio optimization.
Frequently asked questions
What is mean-variance optimization?
It is a method for choosing portfolio weights that maximises expected return for a given variance, or minimises variance for a given return.
Who developed mean-variance optimization?
Harry Markowitz, in a 1952 paper that founded modern portfolio theory.
Why is mean-variance optimization unstable?
Because small errors in expected returns and covariances produce large changes in the optimal weights, often leading to concentrated portfolios.
What are alternatives to mean-variance optimization?
Minimum variance, risk parity, hierarchical risk parity, Black-Litterman and constrained or shrinkage-based approaches.