Robust Mean-Variance
Worst-case mean-variance optimisation over an ellipsoidal uncertainty set on the expected-return vector. It treats as an unknown element of an ellipsoid around the sample mean, rather than as a point estimate. The protection covers the specified uncertainty in expected returns only. It is not validated out of sample.
Overview
Classical MVO treats the sample mean of returns as if it were the true expected-return vector. Michaud (1989) and many subsequent authors documented that the resulting portfolios are catastrophically sensitive to small perturbations in : a 1% shift in the input mean of one asset can swing the optimal weight by tens of percent. Robust optimisation attacks the problem at its root by treating as a decision-relevant unknown.
Goldfarb and Iyengar (2003) gave the canonical formulation: replace the point estimate with an ellipsoidal confidence region around the sample mean, and optimise the worst-case Sharpe over that ellipsoid. The worst-case inner problem admits a closed-form solution, so the overall portfolio problem reduces to a second-order cone programme (SOCP). Tutuncu and Koenig (2004) extended the framework to joint uncertainty in and ; Ben-Tal and Nemirovski (2002) provide the broader theoretical infrastructure of robust optimisation. Ceria and Stubbs (2006) popularised the approach in industry.
FolioLab implements the ellipsoidal-uncertainty robust MVO. The radius of the ellipsoid is the central hyperparameter and acts as the "ambiguity budget": at the problem collapses to standard MVO; as grows the optimiser hedges against ever-larger deviations of the true from its estimate.
Mathematical Formulation
Notation
- - point estimate of the expected-return vector
- - covariance of the estimator (typically )
- - size of the ambiguity ellipsoid (radius of the -ball), fixed here as described below
- - covariance of returns (assumed known here)
Ellipsoidal uncertainty set
For multivariate-normal returns and a sample of size , the estimator covariance is , and can be calibrated to a chosen confidence level: a -confidence ellipsoid corresponds to .
Two values are fixed, and the confidence level is not 95%. The shipped radius is scaled by an uncertainty_aversion factor of 0.35, on top of a confidence level fixed at 0.75:
Neither value is a request field. The applied set is not a 75% confidence region. Its radius is 0.35 times the 75% chi-square radius. Under the method's own normal model of the mean estimator, an ellipsoid of radius holds the true mean with probability , and that probability falls fast as the number of assets grows:
| Assets | 75% chi-square radius | Applied radius | Implied coverage |
|---|---|---|---|
| 2 | 1.665 | 0.583 | 15.6% |
| 5 | 2.574 | 0.901 | 2.4% |
| 10 | 3.542 | 1.240 | 0.12% |
| 20 | 4.881 | 1.708 | 0.0003% |
So the uncertainty set is much smaller than the notation suggests, and the portfolio is close to the plain mean-variance solution. Every RobustMVO result records the settings in uncertainty_set: confidence_level, uncertainty_aversion, n_assets, chi_square_radius, the applied kappa and the implied_coverage. The PDF and Excel reports state the same values. The risk aversion on the variance term is fixed at 5.0 as well.
The point estimate is the arithmetic mean of daily returns, annualized over 252 trading days. is the Ledoit-Wolf shrunk covariance, annualized the same way, and in those annual units. The method needs at least 30 daily return observations. Per-name weight bounds, asset exclusions, and group sum limits such as sector caps enter the same program as extra rows; the worst-case mean term does not change under them.
Worst-case mean-variance
The inner minimum (worst-case mean for a given ) is solved analytically:
Substituting back, the outer problem becomes a second-order cone programme that any modern conic solver handles efficiently; FolioLab solves it with CLARABEL first, then SCS, then ECOS:
Calibrating the ambiguity radius
Three standard approaches to choosing : (1) the chi-squared confidence approach above, with equal to the quantile of a distribution; (2) cross-validation: search a grid of on in-sample data and select the one that maximises out-of-sample Sharpe; (3) practitioner heuristic: roughly corresponds to a one-standard-error ellipsoid and is a common starting point.
Advantages & Limitations
Advantages
- Mean-error aware: Takes the specified uncertainty in expected returns into the objective, instead of trusting the point estimate.
- Tractable SOCP: Solvers handle it deterministically.
- Smooth in : Sweeping the ambiguity dial gives a one-parameter family interpolating between MVO and the global min-variance portfolio.
- Sound theoretical basis: Decades of robust-optimisation literature behind it.
Limitations
- Conservative in calm markets: Worst-case lens can be too defensive.
- Symmetric ellipsoid: Doesn't encode skew in the estimator; treats up-bias and down-bias equally.
- Still needs : Robustness on the mean does not protect against covariance misspecification.
- Hyperparameter : Fixed, not calibrated. The 0.75 is a setting, not a 75% guarantee; see the coverage table above.
- Narrow scope: Protects against the specified uncertainty in expected returns only. It does not protect against covariance error, market crashes or other model error.
- Protection has a cost: The worst-case term can lower the expected return of the portfolio against plain mean-variance.
- Not validated out of sample: No claim is made that this method performs better than another.
References
- Goldfarb, D., & Iyengar, G. (2003). "Robust Portfolio Selection Problems." Mathematics of Operations Research, 28(1), 1-38.
- Tutuncu, R. H., & Koenig, M. (2004). "Robust Asset Allocation." Annals of Operations Research, 132(1-4), 157-187.
- Ben-Tal, A., & Nemirovski, A. (2002). "Robust Optimization - Methodology and Applications." Mathematical Programming, 92(3), 453-480.
- Ceria, S., & Stubbs, R. A. (2006). "Incorporating Estimation Errors into Portfolio Selection: Robust Portfolio Construction." Journal of Asset Management, 7(2), 109-127.
- Fabozzi, F. J., Kolm, P. N., Pachamanova, D. A., & Focardi, S. M. (2007). Robust Portfolio Optimization and Management. Wiley.
- Michaud, R. O. (1989). "The Markowitz Optimization Enigma: Is ‘Optimized’ Optimal?" Financial Analysts Journal, 45(1), 31-42.
- Palomar, D. P. (2025). Portfolio Optimization: Theory and Application. Cambridge University Press, Chapter 14 (Robust Portfolios), Section 14.2.
Not investment advice. Past performance is not indicative of future results.