Markowitz mean-variance takes $\mu$ and $\Sigma$ as known; in practice both are estimated with error, and Michaud (1989) showed that optimised portfolios are wildly sensitive to that error. Robust portfolio optimisation (Ben-Tal & Nemirovski 1998; Goldfarb & Iyengar 2003) replaces point estimates with uncertainty sets and maximises the worst-case expected return inside the set. The resulting programme is a second-order cone programme (SOCP), only modestly harder than a QP, and the portfolios it produces are dramatically more stable under re-estimation.
Uncertainty set, worst-case objective, SOCP reformulation
| Symbol | Meaning |
|---|---|
| $\mu_0 \in \mathbb{R}^n$ | Nominal (point-estimate) expected-return vector |
| $\Sigma \in \mathbb{R}^{n\times n}$ | Covariance matrix (assumed known here; Goldfarb-Iyengar also allow it uncertain) |
| $\mathcal{U}_\kappa$ | Ellipsoidal uncertainty set around $\mu_0$ with radius $\kappa$ |
| $\kappa \ge 0$ | Uncertainty-set radius (0 = nominal Markowitz; large = very conservative) |
| $w \in \mathbb{R}^n$ | Portfolio weights, $\sum w_i = 1$, $w_i \ge 0$ |
| $\lambda \ge 0$ | Risk-aversion parameter on portfolio variance |
Maximise the worst-case expected return inside $\mathcal{U}_\kappa$, still subject to a variance penalty:
The inner minimisation over the ellipsoid has a closed form:
When returns are jointly Gaussian, robust worst-case and mean-CVaR portfolios coincide for appropriate choices of $(\kappa, \alpha)$: $\lVert \Sigma^{1/2} w \rVert$ is proportional to the standard deviation of $\mu^\top w$, and CVaR of a Gaussian is $\sigma \phi(\Phi^{-1}(\alpha))/(1-\alpha)$. For non-Gaussian returns the two diverge — see the CVaR sub-application.
Compare nominal Markowitz to robust as $\kappa$ varies
Gold bars: nominal Markowitz ($\kappa = 0$). Blue bars: robust ($\kappa$ as set above). Watch the robust weights flatten out as $\kappa$ grows — shrinkage toward diversification.
Risk-return plane ($\sigma, \mathbb{E}[r]$): individual assets as dots; Markowitz and robust portfolios as starred points. The robust portfolio typically sits below-left of the nominal.
Price of robustness, calibration, variants
Ben-Tal & Nemirovski proved that on convex problems, robustness induced by an ellipsoidal set is affordable: worst-case performance improves materially while nominal performance degrades only mildly, provided $\kappa$ is calibrated to actual parameter uncertainty rather than worst imaginable extremes. Choose $\kappa$ too aggressively and you underperform even the equal-weight portfolio; choose too conservatively and you recover Markowitz’s fragility.
Robust-optimisation theory and finance applications