Who this is for — Readers who want to understand what mean–variance optimisation actually solves, how the efficient frontier is formed and why a mathematically optimal portfolio may be fragile out of sample.
Mean–variance optimisation combines expected returns with a covariance matrix to compare portfolios in a single-period problem. A portfolio is mean–variance efficient if no other feasible portfolio offers greater expected return for the same variance, or lower variance for the same expected return. The set of such solutions is the efficient frontier.
The framework associated with Harry Markowitz organises a trade-off under stated assumptions. It does not know future returns, prove that variance exhausts risk, or choose objectives, constraints and costs. Estimated inputs and the feasible set are part of the answer.
The two core equations
For a weight vector w, expected-return vector μ and
covariance matrix Σ:
E[Rₚ] = wᵀμVar(Rₚ) = wᵀΣwFor two components:
σₚ² = w₁²σ₁² + w₂²σ₂² + 2w₁w₂Cov(R₁,R₂)Covariance explains why portfolio risk is not an average of stand-alone risks. Correlation standardises dependence, but the matrix used by the optimiser is expressed in covariance units.
A common formulation minimises wᵀΣw for a target return subject to
constraints such as ∑wᵢ = 1. Another maximises expected return less
a variance penalty. Results depend on risk-aversion parameters and constraints
on shorting, leverage, turnover, concentration and liquidity.
What the frontier shows
In expected-return–volatility space, portfolios below the frontier are dominated within the model; the leftmost point is the global minimum-variance portfolio; moving along the upper branch changes the estimated trade-off; adding constraints changes the feasible set; and changing inputs moves the entire frontier.
“Efficient” is therefore conditional, not an absolute verdict. A model- efficient solution can be unsuitable for liabilities, drawdown, tail risk, liquidity or capacity. Investors using the same universe may choose different points because their objectives and constraints differ.
Estimation error
Expected returns, volatilities and covariances are estimated. Small input changes can create large weight changes, especially when similar assets compete on small expected-return differences. Common criticisms include input sensitivity, extreme concentration, missing skew and tails, diversification by label rather than factor, weak linkage to liabilities, and a single-period structure that can omit tax, cost and rebalancing.
DeMiguel, Garlappi and Uppal show how difficult it was for many optimised rules
in their datasets to beat an equal-weight baseline out of sample. Their result
does not make 1/N universally optimal. It makes comparison with
simple baselines, realistic costs and unused test data essential.
Robustness tools
| Tool | Purpose | Limitation |
|---|---|---|
| Weight constraints | avoid extreme solutions | may hide weak inputs without correcting them |
| Shrinkage | stabilise means or covariance | requires a target and intensity |
| Resampling or scenarios | expose sensitivity to alternative inputs | depends on the simulation model |
| Black–Litterman | combine equilibrium and views with uncertainty | views and parameters still require choices |
| Minimum variance | reduce reliance on expected returns | remains covariance-dependent |
| Risk budgeting | control contributions under a chosen measure | does not optimise return automatically |
| Turnover penalties | represent transition frictions | needs a realistic cost model and starting portfolio |
A credible robustness exercise varies windows, frequencies, universes, constraints and costs while preserving an out-of-sample test. A single solution displayed to many decimal places communicates numerical precision, not economic certainty.
Two-asset example
Suppose estimated annual volatilities are 10% and 20%, correlation is 0.25,
and weights are 60% and 40%. Covariance is
0.25 × 0.10 × 0.20 = 0.005. Estimated variance is:
0.60²×0.10² + 0.40²×0.20² + 2×0.60×0.40×0.005 = 0.0124Volatility is approximately √0.0124 = 11.14%. This illustrates
covariance; it neither forecasts future volatility nor proves that 60/40 is
efficient. Expected returns, constraints and all feasible combinations are
still required.
Publishing an optimisation
A reproducible result states the universe and survivorship rule; currency, frequency and date range; estimators for returns and covariance; objective function; constraints and cash treatment; starting positions, turnover and cost model; missing-data and outlier rules; sensitivity analysis; out-of-sample validation; and the date and version of inputs.
Common error — Presenting the highest point of an estimated frontier as the “best portfolio”. The optimiser solves the supplied objective; it does not decide whether inputs, objective and constraints represent the real problem.
Sources
- Harry Markowitz, Portfolio Selection, The Journal of Finance (1952) — original mean–variance formulation.
- CFA Institute, Principles of Asset Allocation — 2026 refresher reading — objectives, frontiers, criticisms, liabilities, factors and practical use.
- Victor DeMiguel, Lorenzo Garlappi and Raman Uppal, Optimal Versus Naive Diversification, The Review of Financial Studies — out-of-sample comparison with
1/N. - Fischer Black and Robert Litterman, Global Portfolio Optimization, Financial Analysts Journal (1992) — combining equilibrium and investor views.
- Richard O. Michaud and Robert Michaud, Estimation Error and Portfolio Optimization: A Resampling Solution — estimation error, resampling and solution stability.