The minimum variance portfolio is the combination of assets with the lowest possible volatility. It depends only on the volatilities and correlations of the assets, not on their expected returns, which makes it more robust than portfolios that need return forecasts.
A worked two-asset example
Asset A has a volatility of 20%, Asset B 10%, and the correlation is 0.3. The weight in A that minimises variance is:
wA = (σB2 - ρσAσB) / (σA2 + σB2 - 2ρσAσB)
| Step | Value |
|---|---|
| Numerator | 0.01 - 0.3 x 0.2 x 0.1 = 0.004 |
| Denominator | 0.04 + 0.01 - 2 x 0.3 x 0.2 x 0.1 = 0.038 |
| Weight in A | 0.004 / 0.038, about 10.5% |
| Weight in B | about 89.5% |
| Portfolio volatility | about 9.8%, lower than either asset alone |
The numbers are made up. With many assets, the weights come from the inverse of the covariance matrix.
Why people use it
- It avoids forecasting returns, the least reliable input in mean-variance optimization.
- Empirically, minimum variance stock portfolios have had lower risk than the market and returns not much lower, in line with the low-volatility finding.
Cautions
- Concentration. Unconstrained, it can put most weight in a few low-risk assets or sectors, so constraints are usual. See optimization constraints.
- Hidden tilts. It tends to favour low-beta, defensive stocks and so underperforms in strong rallies.
- Estimation error in the covariance matrix still matters, so shrinkage is common.
Frequently asked questions
What is a minimum variance portfolio?
It is the mix of assets with the lowest volatility, calculated from the assets' volatilities and correlations alone.
Does the minimum variance portfolio need expected returns?
No. It depends only on the covariance matrix, which is why it is more robust than methods that need return forecasts.
Does minimum variance always hold the lowest-volatility stocks?
Not exactly. It also uses correlations, so it can include a more volatile asset if it diversifies the others.
What are its risks?
Concentration, defensive and low-beta tilts that lag in strong rallies, and sensitivity to estimation errors in the covariance matrix.