Questions: Correlation and Covariance Matrices in Portfolio Optimization

Question 1

An investor builds a portfolio combining equities and bonds, observing a historical correlation of -0.2. During the 2008 financial crisis, what typically happened to such correlations across risky assets?

• A. They stayed near historical levels, confirming that historical correlation reliably predicts crisis behavior
• B. They dropped below -0.5, providing even better diversification when most needed
• C. They rose sharply toward 1.0, causing assets to fall together and eliminating diversification benefits
• D. They became undefined because market volatility makes correlation incalculable in crises

Answer: (C)

This is the 'cruel irony of diversification': correlations between most risky assets spike toward 1.0 during financial crises as panic selling, margin calls, and forced liquidation sweep across all asset classes simultaneously. The portfolio that appeared well-diversified using historical data provides far less protection than predicted precisely when markets are most stressed. A portfolio optimizer using calm-period correlations would dramatically underestimate crisis risk — this is the central practical limitation of mean-variance optimization.

Question 2

For a two-asset portfolio with equal weights, portfolio variance is best described as:

• A. The simple average of the two individual asset variances
• B. The weighted sum of individual variances plus a covariance term that can reduce or increase total risk
• C. The product of the two assets' standard deviations
• D. The larger of the two individual variances

Answer: (B)

Portfolio variance = w₁²σ₁² + w₂²σ₂² + 2w₁w₂Cov(R₁,R₂). With equal weights (w=0.5 each): 0.25σ₁² + 0.25σ₂² + 0.5Cov(R₁,R₂). The covariance term is what captures diversification: negative covariance reduces portfolio variance below the weighted average of individual variances, positive covariance adds to it. The covariance matrix Σ generalizes this to any number of assets through the compact expression σ²_p = w'Σw — all pairwise interactions are encoded in Σ.

Question 3

The off-diagonal entries of the covariance matrix capture pairwise co-movement between assets, and negative off-diagonal values indicate potential diversification benefits.

• T. True
• F. False

Answer: True

The covariance matrix Σ has asset variances on its diagonal (σᵢ²) and pairwise covariances off-diagonal (Cov(Rᵢ,Rⱼ)). Negative covariance means the assets tend to move in opposite directions — when one falls, the other tends to rise — which reduces overall portfolio variance. The correlation matrix standardizes these entries to [-1, +1], making the magnitude of co-movement easier to interpret across assets with different return scales.

Question 4

A well-diversified portfolio constructed using historical correlations will perform as predicted during a market crash because diversification reduces portfolio risk in most market conditions.

• T. True
• F. False

Answer: False

This is the central practical failure of naive mean-variance optimization. Historical correlations are estimated in one market regime (normal times) and can be wildly inaccurate during another (crisis). The correlation instability phenomenon — correlations spiking toward 1 during crashes — causes diversification benefits to evaporate exactly when they are most needed. A portfolio that looks well-hedged in calm markets can suffer simultaneous losses across all positions in a crisis. Robust portfolio construction must account for regime-dependent correlations and stress testing, not just historical averages.

Question 5

Why does the covariance matrix need to be positive semi-definite, and what problem arises when estimating it from historical data with many assets?

Model answer: Positive semi-definiteness ensures that no portfolio weight vector produces a negative portfolio variance (w'Σw ≥ 0 for all w), which would be mathematically meaningless. When estimating Σ from historical data with N assets and T observations, if T < N the sample covariance matrix is singular and not invertible. Even when T > N, estimating N(N-1)/2 pairwise covariances from limited data introduces substantial noise — a 100-asset portfolio requires estimating 4,950 covariance pairs. These estimation errors compound in optimization, causing the optimizer to take extreme positions in poorly-estimated assets. Shrinkage estimators (blending the sample Σ toward a structured target) and factor models (expressing covariances through a small number of common drivers) are standard practical remedies.

The dimensionality problem is why textbook portfolio optimization often fails in practice. A clean mathematical framework requires clean inputs, but real-world covariance estimation is noisy, regime-dependent, and high-dimensional. Understanding this gap is essential for applying the theory responsibly.