5.9.5 Failure Modes & Practice

Basic Prerequisites: 04 · Carry & Styles and Covariance Shrinkage & RMT.

---

1. Intuition & Practical Objective

Every result on the previous pages rests on an estimated covariance matrix and an assumed set of expected returns. This page names precisely how multi-asset and factor allocation fails - so a practitioner knows which inputs to distrust and how the failure shows up in money terms. There is no cynicism here: the discipline is knowing where the model is an approximation so the residual risk can be measured.

The four failures, in one line each:

1. Correlations rise in a crisis - diversification is cheapest exactly when you need it most, and the covariance matrix you estimated on calm data understates joint losses.
2. Factor crowding - factors are traded; as capital crowds a factor its premium decays and its correlation to everything else rises.
3. Estimation error is the dominant term - the mean-variance optimizer loads hardest on the least estimable input (the mean) and loses out-of-sample to naive $1/N$.
4. Carry crashes - carry strategies have negative skew: steady gains, rare large losses.

The one-sentence essence. “Diversification is a state-dependent quantity, not a constant: the same portfolio that has a 1.46 diversification ratio in calm markets can approach a single bet when correlations all move to 1 at once.”

---

2. Mathematical Ground Truth & Derivations

Correlation is not constant. Model the joint return as a mixture of a calm regime ${\Sigma }_{\text{calm}}$ and a crisis regime ${\Sigma }_{\text{crisis}}$, with crisis covariance exhibiting a single dominant factor (everything loads on the stress factor). Then

$${\sigma }_p^2=w^{\top }\Sigma w\text{with}\Sigma \in \{{\Sigma }_{\text{calm}},{\Sigma }_{\text{crisis}}\},$$

and the effective number of bets - a measure of how many independent risks the portfolio truly holds - is the participation ratio of the correlation matrix’s eigenvalues,

$$N_{\text{eff}}=\frac{({\sum }_k{\lambda }_k{)}^2}{{\sum }_k{\lambda }_k^2},{\lambda }_k=\text{eigenvalues of the correlation matrix}.$$

When one factor dominates the correlation structure, one eigenvalue captures most of the total $N={\sum }_k{\lambda }_k$ and $N_{\text{eff}}\to 1$. In the numbers below, $N_{\text{eff}}$ drops from $3.17$ to $2.20$ and the equal-weight portfolio’s volatility rises $1.19\times$.

Estimation error compounds through ${\Sigma }^{-1}$. The tangency portfolio $w^{\top }\propto {\Sigma }^{-1}(\mu -r_{f}1)$ is non-linear in the inputs; small errors in $\mu$ and $\Sigma$ are amplified by matrix inversion. Chopra & Ziemba (1993): in the MV objective, errors in means dominate variances $\approx 10.5\times$ and covariances $\approx 21\times$ (variances dominate covariances $\approx 2\times$). Best & Grauer (1991): a 1% shift in a single mean can move weights by 50%+. The theoretical reason $1/N$ is so hard to beat (DeMiguel–Garlappi–Uppal 2009) is exactly this amplification: the estimation-error penalty of optimising cancels the benefit of the better in-sample frontier.

Factor crowding as a correlation increase. If a factor is crowded, its flow-driven component loads on a common “deleveraging” factor; empirically this shows up as the factor’s correlation to the market and to other crowded factors rising during drawdowns - the same correlation-crisis mechanism, applied to factors rather than asset classes.

Carry crash as negative skew. Carry’s return distribution has small positive mean, negative skew, and heavy left tail: $\mathrm{Skew}(r^{\text{carry}})<0$. A mean-variance investor ignores the third moment and therefore over-sizes carry; a defensive (low-vol / anti-beta) overlay is the structural hedge.

---

3. Computational Implementation - the failures in numbers

Two experiments in one runnable block (numpy). Experiment A (correlation crisis): the same equal-weight portfolio under a calm and a crisis covariance. Experiment B (estimation error): a mean-variance optimizer fitted on 24 months of noisy data, evaluated out-of-sample against naive $1/N$.

Experiment A: in the crisis regime the same portfolio’s volatility jumps from $7.9\%$ to $9.4\%$, and the effective number of bets falls from $3.17$ to $2.20$ - diversification quietly disappears. Experiment B: the optimizer fitted on real-world-length samples posts an out-of-sample Sharpe of $0.14$ with $66\%$ volatility, against naive $1/N$‘s $0.30$ Sharpe at $7.9\%$ volatility. MVO is not wrong; it is an estimation-error maximiser, exactly as the theory says.

---

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

1. Correlation crisis (diversification illusion). Calm-data covariances understate joint losses. The equal-weight portfolio’s volatility rises $1.19\times$ and its effective bets collapse $3.17\to 2.20$ when correlations move to crisis levels. Fix: stress covariance, regime conditioning (page 06), tail-aware allocation.
2. Factor crowding. A crowded factor’s premium decays and its correlation rises - in drawdowns, crowding becomes correlation. Detect via factor crowding/capacity diagnostics and de-size (see Factor Crowding & Capacity).
3. Estimation error dominates. Out-of-sample, unconstrained MVO lost to $1/N$ here by a wide margin. Fix: shrink means (Black–Litterman), shrink $\Sigma$ (shrinkage & RMT), constrain weights, or resample.
4. Carry crash (negative skew). Carry’s steady gains hide a heavy left tail; mean-variance ignores skew and over-sizes. Fix: explicit tail hedge, leverage caps, defensive overlay.
5. Strategic/tactical confusion. Repeated tactical re-optimisation re-writes the strategic plan with noise; the tracking-error budget is what keeps the two separate.

---

5. References

• Chopra & Ziemba, “The Effect of Errors in Means, Variances, and Covariances…,” JPM 19(2):6–11, 1993
• Best & Grauer, “On the Sensitivity of Mean–Variance-Efficient Portfolios to Changes in Asset Means,” RFS 4(2):315–342, 1991
• DeMiguel, Garlappi & Uppal, “Optimal Versus Naive Diversification,” RFS 22(5):1915–1953, 2009
• Laloux, Cizeau, Bouchaud & Potters, “Noise Dressing of Financial Correlation Matrices,” PRL 83(7):1467–1470, 1999
• Koijen et al., “Carry,” JFE 127(2):197–225, 2018
• Ang, Asset Management (2014)

---

6. Connected Graph Bridges

• Back: 04 · Carry & Styles · Index Hub
• Forward: 06 · Advanced Extensions
• Sibling: Covariance Shrinkage & RMT · Black–Litterman · Robust Portfolio Optimization
• Cross-pillar: Factor Crowding & Capacity · Quantitative Risk