5.1.5 Failure Modes & Real-World Practice

Basic Prerequisites: 04 · Min-Variance & Constraints.

---

1. Intuition & Practical Objective

Mean-variance optimization is mathematically beautiful and empirically fragile in one decisive way: it turns estimation error in the inputs - above all the expected returns $\mu$ - into violent swings in the weights. This page makes that precise and gives a practitioner the numbers to distrust, and the tests to run before trusting an optimizer. The objective is not cynicism: it is to know exactly which input errors matter (the means, ten-fold over variances - Chopra & Ziemba 1993), how they amplify (${\Sigma }^{-1}$), and why the headline symptom is extreme long/short weights that are useless to an investor.

The three failures, in one line each:

1. The optimizer is an “estimation-error maximizer” (Michaud’s phrase): it places its biggest bets on the assets whose $\mu$ estimates are worst (Best & Grauer 1991 quantify this).
2. Covariance inversion instability: near-collinear assets make ${\Sigma }^{-1}$ enormous, so tiny eigenvalue noise becomes huge weights.
3. Means dominate everything: errors in $\mu$ are ~11× more damaging than errors in variances (and ~21× more than covariances) in the MV objective (Chopra & Ziemba 1993) - and $\mu$ is exactly what we know least.

---

2. Mathematical Ground Truth & Derivations

The sensitivity formula. With the budget-only (unconstrained, shortable) tangency/frontier solution $w={\Sigma }^{-1}(\lambda 1+\gamma \mu )$, derivative with respect to one mean:

$$\frac{\partial w}{\partial {\mu }_j}={\Sigma }^{-1}[\frac{\partial \lambda }{\partial {\mu }_j}1+\frac{\partial \gamma }{\partial {\mu }_j}\mu +\gamma e_j],$$

the multiplier derivatives $\partial \lambda /\partial {\mu }_j,\partial \gamma /\partial {\mu }_j$ come from re-imposing the two budget/return constraints; they are what makes the response amplified (the Best–Grauer effect) rather than the bare ${\Sigma }^{-1}{e}_j$. i.e. a change in the mean of asset $j$ is pushed through the full leverage matrix ${\Sigma }^{-1}$. When $\Sigma$ has small eigenvalues (assets nearly collinear), ${\Sigma }^{-1}$ has large eigenvalues, and a tiny change in one ${\mu }_j$ can move many weights by multiples of 100%. This is the mathematical core of “error maximizer.”

Ranking of input damage (Chopra & Ziemba 1993). In terms of lost value in the MV objective, as sample size grows the cost of an error in means stays roughly constant while the cost of errors in variances/covariances falls, so for realistic sample sizes:

$$\text{cost}(\mu \text{ error})\approx 11\times \text{cost}({\sigma }^2\text{ error}),\text{cost}({\sigma }^2\text{ error})\approx 2\times \text{cost}(\text{covariance error}),$$

(Chopra & Ziemba: errors in means are ~11$\times$ as damaging as variance errors and ~21$\times$ as damaging as covariance errors; the ranking is $\mu >{\sigma }^2>{\sigma }_{ij}$.) Why: the objective rewards $w^T\mu$ linearly but penalizes $w^T\Sigma w$ quadratically in weight - noise in $\mu$ selects extreme weights whose variance cost is then borne fully.

Best & Grauer’s headline (1991). In a 100-asset equally-weighted universe, the mean increase needed to drive the most sensitive asset out of the portfolio is just 0.08%; five assets 0.18%; ten 0.30%; and about 11.6% drives half the universe out - yet the portfolio’s own return and standard deviation change almost not at all (the frontier is deep; the corner weights are shallow).

---

3. Computational Implementation - watching the optimizer blow up

Stdlib only. Five nearly-collinear assets (pairwise $\rho \approx 0.97$, small asymmetry) → an ill-conditioned $\Sigma$; compute the tangency portfolio, nudge one mean by 0.2%, and watch the weights swing.

Read it twice: at a perfectly ordinary near-collinear covariance ($\kappa =169$), the tangency portfolio already asks you to hold leverage of −103% to +165% - and a +0.2% nudge to one asset’s mean changes asset 3’s weight by +0.727 (a 73% swing) while asset 2 drops 0.179. That is the estimation-error maximizer in action: the smallest input error, hugely amplified through ${\Sigma }^{-1}$.

---

4. Failure Modes & First-Principles Breakdowns (numbered)

1. Mean-error amplification (the #1 killer). Because $w\propto {\Sigma }^{-1}(\mu -r_{f}1)$, and $\mu$ is the least reliably estimated input, the optimizer bets hardest precisely where the evidence is thinnest. Best & Grauer: a 0.08% mean shift drops the most-sensitive asset. Test: perturb $\mu$ by sampling error and watch $\Vert w\Vert$; real deployments must shrink or regularize (06 · Advanced Extensions).
2. Small-eigenvalue explosion. As soon as two assets are near-duplicates ($\rho \to 1$), ${\Sigma }^{-1}$ has a huge eigenvalue and the weights become garbage - the structural root of “why not just optimize over 500 correlated names.”
3. Means > variances > covariances, always. Chopra–Ziemba’s ~11× factor means that pouring effort into a better $\Sigma$ while feeding naive sample means still leaves you dominated by mean noise. Priorities: fix $\mu$ (Black–Litterman, shrinkage) before polishing $\Sigma$.
4. The “portrait of the frontier is stable” illusion. The frontier curve is stable under input noise (Best–Grauer: return/SD barely move) while the corresponding weights are not. An investor who plots the frontier and trusts the point on it is trusting a knife-edge. Resampling (Michaud) is one fix.
5. Constraint leak / nonpositive reality. Once you add $w\ge 0$, turnover caps, and impact costs, the raw optimizer’s extreme weights are partially tamed but the signal is also lost - the transfer-coefficient idea behind Transaction Costs.

---

5. References

• Best, Michael J. & Grauer, Robert R.: On the Sensitivity of Mean–Variance-Efficient Portfolios to Changes in Asset Means, RFS 4(2):315–342 (1991)
• Chopra, Vijay K. & Ziemba, William T.: The Effect of Errors in Means, Variances, and Covariances on Optimal Portfolio Choice, Journal of Portfolio Management 19(2):6–11 (1993)
• Michaud, Richard O. (with Robert O. Michaud): Efficient Asset Management, 2nd ed., OUP (2008) - “Markowitz optimization en masse”: resampling and “estimation-error maximizers”.
• DeMiguel, Garlappi & Uppal: Optimal Versus Naive Diversification, RFS 22(5) (2009)
• Kan & Zhou: Optimal Portfolio Choice with Parameter Uncertainty, JFQA 42(3) (2007)

---

6. Connected Graph Bridges

• Back: 04 · Min-Variance & Constraints · Index Hub
• Forward: 06 · Advanced Extensions
• Sibling fixes: Covariance Shrinkage & RMT · Black–Litterman · HRP