5.6.4 Constraint-Based Robustness & Resampling (Michaud)

Basic Prerequisites: 03 · Robust Formulations and Transaction Costs & Turnover.

---

1. Intuition & Practical Objective

Robustness does not require uncertainty sets. Two older, blunter, and remarkably effective techniques sit between “naive MVO” and “formal robust optimization”:

• Constraints as robustness. Long-only, weight caps, sector bounds, and turnover limits are the practitioner’s default. They are not cosmetic - they change the geometry of the feasible set so the optimizer can no longer express its extreme opinions. A $35\%$ cap is a hard statement that “no model output, however noisy, justifies a $500\%$ position.”
• Resampling (Michaud). Instead of optimizing one estimated $(\hat{\mu },\hat{\Sigma })$, optimize many resampled versions, then average the resulting weights. The idea (Michaud & Michaud 1998) is that a portfolio averaged over the estimation-error distribution is less brittle than the single “optimal” point.

The practical objective is to know what each actually buys you - and, crucially, what it does not. This page shows, on the shared data, that constraints are the real stabilizer, while naive resampling fails to cure a biased estimate and can even amplify it. That negative result is the honest bridge to the formal robust formulations of 03.

---

2. Mathematical Ground Truth & Derivations

2.1 The constrained model

The practically-used MVO is a quadratic program over a box (plus simplex):

$$\underset{w}{\max }{\hat{\mu }}^{\top }w-\frac{\delta }{2}{w}^{\top }\hat{\Sigma }w\text{s.t.}{1}^{\top }w=1,w_{\min }\le w\le w_{\max }.$$

With $w_{\min }=0$ it is long-only (no short sales); with a common $w_{\max }$ it caps concentration. This is a convex QP solved by active-set/interior-point methods - every real optimizer since Markowitz uses some version.

Turnover constraint. To control trading cost, add $\Vert w-w_{\text{prev}}{\Vert }_1\le \tau$ (an ${\ell }_1$ ball) - convex, and itself a robustness statement about the path, not just the endpoint (see Transaction Costs & Turnover).

Why constraints help (first principles). Unconstrained MVO’s optimum is $w^{\star }=\frac{1}{\delta }{\hat{\Sigma }}^{-1}\hat{\mu }$. The problem is that this point sits outside any sane region of portfolio space, in the short-saturated tail where estimation error dominates. Clipping the feasible set to the simplex/box removes exactly those directions. The cost is real - Clarke, de Silva & Thorley (2002) quantify it as a transfer coefficient $\mathrm{TC}<1$ that discounts achievable information ratio ($\mathrm{IR}\approx \mathrm{IC}\times \sqrt{\text{breadth}}\times \mathrm{TC}$) - but the benefit (survival) usually dominates.

2.2 Resampling (Michaud)

Michaud’s procedure:

1. Estimate $(\hat{\mu },\hat{\Sigma })$ from the sample.
2. Draw $B$ bootstrap samples of the return history (non-parametric resample, or parametric draws from $\mathcal{N}(\hat{\mu },\hat{\Sigma })$).
3. Optimize MVO on each; collect the $B$ weight vectors $\{w^{(b)}\}$.
4. Report the average $\bar{w}=\frac{1}{B}{\sum }_b{w}^{(b)}$ (and/or a resampled efficient frontier in which portfolios are ranked by their resampled risk/reward).

The claim is that $\bar{w}$ is more stable than the single $w^{\star }$. The subtlety - and where the method is often over-sold - is that step 2 resamples around the same point estimate that is itself biased by the overfit. Averaging over resamples reduces the variance contribution of the frontier, but it does not remove the bias in $\hat{\mu }$; if the center is wrong, all $B$ optimizations inherit the wrong center. Goldfarb & Iyengar (2003, §1) make this critique precisely: sampling-based and scenario approaches “do not provide any hard guarantees on the portfolio performance.”

---

3. Computational Implementation - constraints vs resampling

Same universe. We compare the unconstrained optimum, the long-only box, and the Michaud average, and we measure the dispersion of resampled weights - the quantity resampling is supposed to tame.

Two verified, and deliberately honest, results:

• The constraint is the workhorse. Clipping to the long-only box collapses the $-2.26/+4.96$ short-saturated optimum to a diversified $[0.078,0.158,0.205,0.214,0.171,0.173]$ with gross exposure $1.00$. (Here the $35\%$ cap is not even binding - long-only alone already does the job on this $N=6$ universe.) The robust max-min portfolio of 03 lands in essentially the same place: box-uncertainty and hard caps both act as penalties on large positions.
• Naive resampling does not stabilize the unconstrained problem. The Michaud average has gross exposure $13.92$ - more extreme than the single naive solution ($11.91$), and the per-draw gross exposure has standard deviation $5.49$, roaming around a mean of $16.45$. Resampling from a model centered on an overfit $\hat{\mu }$ reproduces the overfit in every draw. This is the failure the formal robust formulation is designed to fix, not a bug in the code.

The design lesson. If you are starting from an unreliable center, resampling propagates the unreliability. What actually helps is (a) imposing constraints (drop the extreme feasible directions) and/or (b) shrinking the center toward something stable (shrinkage, Black–Litterman prior, or a robust worst-case set). Resampling is best used on top of those, to build a distribution of portfolios rather than to replace them.

---

4. Failure Modes & First-Principles Breakdowns

1. Constraints are a blunt instrument with a measurable cost. Long-only and caps improve stability but discard information - the transfer coefficient $\mathrm{TC}<1$ (Clarke, de Silva & Thorley 2002) is exactly how much of your signal the constraints “leak.” Cap too tight and you have simply built $1/N$ with extra steps.
2. Resampling is not a cure for a biased estimate. §3 shows it; Goldfarb & Iyengar (2003) state it. Resampling reduces frontier-sampling variance, not input bias. Marketing it as a guarantee is a category error.
3. Turnover constraints can hide regime change. A tight ${\ell }_1$ turnover budget keeps you glued to a stale portfolio when the world has moved; robustness about inputs is not robustness about regimes.
4. Cap asymmetry / hidden concentration. Capping each asset at $0.35$ while leaving correlations free lets the portfolio concentrate in a single risk factor. Caps on weights are not caps on risk - the residual factor exposure must be checked separately (bridge to factor risk models).

---

5. References

• Michaud, R. & Michaud, R. Efficient Asset Management, 2nd ed., OUP, 2008
• Lobo, Fazel & Boyd (2007), Portfolio Optimization with Linear and Fixed Transaction Costs, Annals of OR 152:341–365
• Clarke, de Silva & Thorley (2002), Portfolio Constraints and the Fundamental Law of Active Management, FAJ 58(5):48–66
• DeMiguel, Garlappi & Uppal (2009), RFS 22(5)
• Goldfarb & Iyengar (2003), Math. of OR 28(1):1–38
• Frost & Savarino (1988), For better performance: constrain portfolio weights, JPM

---

6. Connected Graph Bridges

• Back: 03 · Robust Formulations
• Method siblings: Transaction Costs & Turnover · Risk Parity & ERC (a constraint-free stabilizer) · HRP (inversion-free)
• Forward: 05 · Failure Modes & Practice · Index Hub