Mean-Variance Portfolio Optimization

• Mean-variance portfolio optimization (MVO) is a quantitative framework that balances expected return and risk through quadratic optimization, defining efficient frontiers.
• It uses formulations like minimum variance at fixed return and risk-adjusted return to provide explicit closed-form solutions under Gaussian assumptions.
• Extensions to non-Gaussian returns using NMVM models allow for adjusted mean inputs, enabling robust portfolio construction with SSD-consistent risk measures.

Mean-variance portfolio optimization (MVO) is the foundational quantitative framework for the trade-off between expected return and risk in portfolio allocation. The Markowitz mean–variance paradigm defines efficient frontiers, yields explicit closed-form solutions in the Gaussian setting, and remains the central benchmark for portfolio selection, robust to significant generalizations including heavy-tailed distributions, law-invariant risk measures, and various practical constraints.

1. Fundamental Formulation and Solution Structure

Given $d$ risky assets with joint return vector $X \in \mathbb{R}^d$, mean $\mu = E[X] \in \mathbb{R}^d$, and covariance matrix $$\Sigma = \mathrm{Cov}(X) \in \mathbb{R}^{d \times d}$$, MVO seeks portfolio weights $w \in \mathbb{R}^d$ that balance expected return and variance. Two canonical formulations are:

• Minimum variance at fixed return:

$$\min_{w} \; w^T \Sigma w \quad \text{s.t.} \quad w^T \mu = r, \quad e^T w = 1, \quad w \ge 0$$

where $r$ is the target expected return, $e = (1,\ldots,1)^T$, and $w \ge 0$ enforces a long-only constraint.

• Risk-adjusted return (Lagrangian):

$$\min_{w} \left\{ \frac{1}{2}w^T \Sigma w - \lambda w^T \mu \right\} \quad \text{s.t.} \quad e^T w = 1$$

for risk-aversion parameter $X \in \mathbb{R}^d$0.

In the unconstrained case (no short-sale ban), the solution exhibits explicit "two-fund separation":

$X \in \mathbb{R}^d$1

The mean–variance efficient frontier defined by these solutions is a quadratic (parabolic) curve in the $X \in \mathbb{R}^d$2 plane (Sayit, 2022).

2. Generalized Return Distributions and Risk Criteria

When asset returns are modeled beyond Gaussian—specifically as normal mean–variance mixtures (NMVM), $X \in \mathbb{R}^d$3 with mixing variable $X \in \mathbb{R}^d$4—law-invariant, convex, SSD-consistent risk measures admit an explicit solution for the mean–risk efficient frontier. The central result is:

• For any finite, law-invariant, SSD-consistent convex risk measure (e.g., CVaR$X \in \mathbb{R}^d$5, spectral measures), the mean–risk optimal portfolio under $X \in \mathbb{R}^d$6 NMVM$X \in \mathbb{R}^d$7 is the solution to a classical mean–variance problem with adjusted mean $X \in \mathbb{R}^d$8 and unchanged covariance $X \in \mathbb{R}^d$9:

$\mu = E[X] \in \mathbb{R}^d$0

This equivalence collapses the general mean–risk optimization (for an entire class of risk measures) to an MVO problem with transformed inputs (Sayit, 2022).

3. Stochastic Dominance, SSD, and Efficient Frontier Characterization

A sufficient condition for second-order stochastic dominance (SSD) in NMVM models is derived: if $\mu = E[X] \in \mathbb{R}^d$1 and $\mu = E[X] \in \mathbb{R}^d$2 for $\mu = E[X] \in \mathbb{R}^d$3, then $\mu = E[X] \in \mathbb{R}^d$4 if $\mu = E[X] \in \mathbb{R}^d$5 and $\mu = E[X] \in \mathbb{R}^d$6, i.e., higher mean and smaller variance imply dominance. This suffices to order portfolios for any concave utility and applies directly in the optimization of portfolios under general law-invariant risk measures, reinforcing the reduction to MVO with adjusted mean (Sayit, 2022).

4. Practical Implementation and Numerical Illustration

A practical guideline is established: when empirical returns are well-modeled by an NMVM process and the risk measure is law-invariant, convex, and SSD-consistent, one computes the adjusted mean, retains the original $\mu = E[X] \in \mathbb{R}^d$7, and solves a classical MVO. This significantly simplifies portfolio construction even under distributional heterogeneity (skewness, fat tails, volatility clustering). Short-sale or further convex constraints are handled numerically after this adjustment. Necessary conditions for this reduction include integrability $\mu = E[X] \in \mathbb{R}^d$8 and SSD-consistency of the risk measure (note: VaR is not compatible).

Numerical examples in the literature confirm that incorporating the $\mu = E[X] \in \mathbb{R}^d$9 adjustment to $$\Sigma = \mathrm{Cov}(X) \in \mathbb{R}^{d \times d}$$0 leads to efficient frontiers that strictly improve the variance-return trade-off for the same target return, compared to the naive (unadjusted) MVO (Sayit, 2022).

5. Limitations and Scope of the Theory

The reduction holds provided that:

• $$\Sigma = \mathrm{Cov}(X) \in \mathbb{R}^{d \times d}$$1 is integrable, i.e., $$\Sigma = \mathrm{Cov}(X) \in \mathbb{R}^{d \times d}$$2 for finite mean adjustment,
• The risk measure $$\Sigma = \mathrm{Cov}(X) \in \mathbb{R}^{d \times d}$$3 is law-invariant and SSD-consistent,
• The portfolio constraints are convex (more general constraints may require iterative methods post-adjustment).

Value-at-Risk (VaR) is a notable exception, as it lacks SSD-consistency. In such cases, the closed-form reduction is inapplicable (Sayit, 2022).

This framework unifies the analysis for a wide spectrum of risk-averse utility functions and convex risk measures, subsuming traditional Markowitz, mean–semideviation, and spectral-risk-based optimizations, as long as the conditions above are satisfied.

6. Connection to Broader Portfolio Optimization Theory

The NMVM-based reduction extends the classical Markowitz paradigm to account for non-Gaussian features in asset returns without losing tractability. For nonlaw-invariant or nonconvex measures, or for returns with path dependence or time variation in the mean/covariance, more general stochastic control techniques are required; otherwise, the mean–variance structure provides an analytically tractable and interpretable solution (Sayit, 2022).

Further, models where returns exhibit time series dependence or are estimated empirically may require regularization or Bayesian approaches to address estimation error—a distinct issue but compatible with the conclusion that for law-invariant, SSD-consistent risk, the optimal portfolio can always be characterized as a solution to a (possibly transformed) MVO problem.

7. Summary Table: Equivalence Principle for Mean–Variance Optimization under NMVM

Return Model	Risk Measure $$\Sigma = \mathrm{Cov}(X) \in \mathbb{R}^{d \times d}$$4	Closed-form MVO Reduction	Condition on $$\Sigma = \mathrm{Cov}(X) \in \mathbb{R}^{d \times d}$$5
Multivariate Normal	Any quadratic or convex law-invariant	Yes, $$\Sigma = \mathrm{Cov}(X) \in \mathbb{R}^{d \times d}$$6, $$\Sigma = \mathrm{Cov}(X) \in \mathbb{R}^{d \times d}$$7 unchanged	Always (Gaussian)
NMVM	Law-invariant, SSD-consistent	Yes, replace $$\Sigma = \mathrm{Cov}(X) \in \mathbb{R}^{d \times d}$$8, use $$\Sigma = \mathrm{Cov}(X) \in \mathbb{R}^{d \times d}$$9	SSD-consistency required
NMVM	VaR	No, reduction fails	VaR not SSD-consistent

This equivalence principle represents a comprehensive generalization of Markowitz mean–variance theory, enabling its closed-form frontier characterization under rich, real-world return dynamics and for a large class of risk preferences (Sayit, 2022).