Robust Mean-Variance and CVaR

Beyond shrinkage and Bayesian priors, two further families of estimators hedge against the limitations of the plug-in approach: robust optimization, which protects against worst-case parameter realizations, and CVaR optimization, which targets the tail of the loss distribution directly.

Robust mean-variance

Suppose \(\mu\) lies in an ellipsoidal uncertainty set centered at the estimate \(\hat\mu\):

\[ \mathcal{U}_\mu(\kappa) \;=\; \bigl\{\, \mu \;:\; (\mu - \hat\mu)^\top \Sigma_\mu^{-1} (\mu - \hat\mu) \le \kappa^2 \bigr\}. \tag{1} \]

The worst-case mean-variance problem is

\[ \max_w\; \min_{\mu \in \mathcal{U}_\mu(\kappa)} \;\mu^\top w - \tfrac{\gamma}{2} w^\top \Sigma w \quad \text{s.t.}\quad \mathbf{1}^\top w = 1. \tag{2} \]

The inner minimum has the closed form \(\hat\mu^\top w - \kappa \sqrt{w^\top \Sigma_\mu w}\), giving the SOCP

\[ \max_w\; \hat\mu^\top w - \kappa\,\sqrt{w^\top \Sigma_\mu w} \;-\; \tfrac{\gamma}{2} w^\top \Sigma w, \tag{3} \]

solved by any SOCP backend [@goldfarb2003]. The parameter \(\kappa\) controls the conservatism of the solution; for a \((1-\alpha)\) confidence ellipsoid under Gaussianity, \(\kappa = \sqrt{\chi^2_{n, 1-\alpha} / T}\).

CVaR (Rockafellar-Uryasev)

For loss \(L(w) = -w^\top r\) and confidence level \(\alpha \in (0, 1)\), the Value-at-Risk is

\[ \mathrm{VaR}_\alpha(w) \;=\; \inf\{\, \zeta \;:\; \mathbb{P}(L(w) \le \zeta) \ge \alpha \,\}. \tag{4} \]

The Conditional VaR is the conditional mean of losses beyond VaR:

\[ \mathrm{CVaR}_\alpha(w) \;=\; \mathbb{E}\bigl[\,L(w)\;\big|\;L(w) \ge \mathrm{VaR}_\alpha(w)\,\bigr]. \tag{5} \]

Rockafellar and Uryasev [@rockafellar2000] showed that CVaR has the variational representation

\[ \mathrm{CVaR}_\alpha(w) \;=\; \min_\zeta\; \zeta \;+\; \frac{1}{1-\alpha}\,\mathbb{E}\bigl[(L(w) - \zeta)_+\bigr]. \tag{6} \]

Given a sample \(\{r_1, \dots, r_T\}\), the empirical CVaR optimization becomes a linear program:

\[ \begin{aligned} \min_{w, \zeta, u}\;& \zeta \;+\; \frac{1}{(1-\alpha) T} \sum_{t=1}^T u_t \\ \text{s.t.}\;& u_t \ge -w^\top r_t - \zeta,\quad u_t \ge 0,\quad t=1,\dots,T,\\ & \mathbf{1}^\top w = 1, \quad w \in \mathcal{W}. \tag{7} \end{aligned} \]

This is solved directly by CVaROptimizer with any LP-capable backend.

When to use which

situation	recommended formulation
Returns approximately Gaussian, moderate \(n/T\)	Mean-variance + LW shrinkage
Strong priors or absolute views	Black-Litterman
Concern about mean estimation error specifically	Robust mean (SOCP)
Heavy-tailed returns, regulatory tail constraint	CVaR

References

[@goldfarb2003]; [@rockafellar2000]; [@ben-tal2009]. See Citations.