Why Mean-Variance Portfolio Weights Can Oppose Expected-Return Signs

Summary

The document examines whether an asset with a positive expected return must receive a positive weight in an unconstrained mean-variance portfolio. For the stated objective, the solution depends on the inverse covariance matrix multiplied by the full expected-return vector. As a result, an individual weight reflects how that asset’s covariance relationships interact with the expected returns and risks of all other candidates; the sign of its own expected return alone does not determine the sign of its weight.

The response also cautions that mean-variance optimization can amplify errors in expected-return estimates, which are often less reliable than volatility estimates. A positive expected return may encourage a long position, particularly when the asset compares favorably with the rest of the universe, but it does not guarantee one. The explanation is conceptual and gives no numerical example, constraints, or remedies such as shrinkage. Its main lesson is to interpret portfolio weights jointly rather than reading them directly from each asset’s return estimate.

Key ideas

• An asset’s mean-variance weight depends on the full expected-return vector and covariance matrix.
• A positive expected return does not guarantee a positive portfolio weight.
• Covariances with other assets can alter the sign and magnitude of an individual weight.
• Mean-variance optimization can magnify errors in expected-return estimates.
• Portfolio weights should be interpreted in the context of the entire candidate universe.

Tags

# Sign retention in mean variance optimization


# Sign retention in mean variance optimization












The mean variance optimization to the objective: $h^T\alpha - \lambda h^T V h$ results in the solution:

$h = \frac{V^{-1} \alpha}{2 \lambda}$

Would a positive value for an asset in $\alpha$ result in a positive value in the weights vector $h$ ?

## Answer by develarist (score 2)

https://quant.stackexchange.com/a/49912

The sign of the portfolio weight $h_n$ assigned to one asset cannot be solely determined by the sign of that asset's expected return, $\alpha_n$, due to the model having to also take as an input that asset's dependence structure with the other $N-1$ assets being considered, captured within the covariance matrix $V$. Those other assets are also vying for inclusion in the portfolio.

Since the mean-variance model was recognized by Michaud to be an "error maximization" model in that it favors (overweights) assets that have high expected returns and low variance, however, even though these same assets are likely to be the most prone to misestimation due to the nature of expected return estimates being much more unreliable than asset volatility estimates, the sign of an asset's expected return does, at least, suggest that the model will "pick it" for a long position, especially if that asset has the highest expected return (plus lowest risk) within the candidate investment pool. But again, it all depends on the overall data for all assets being considered.

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.