Correlation Risk Analysis

Regime-dependent correlation, diversification failure, dynamic correlation models, and correlation trading strategies.

When to Activate

• User analyzes portfolio diversification and correlation assumptions
• Correlation breakdown during market stress events
• Dynamic correlation modeling with DCC-GARCH or similar
• Correlation trading strategies (dispersion, correlation swaps)
• Risk model validation where correlation assumptions drive results
• Multi-asset portfolio construction sensitive to correlation estimates

Core Concepts

Correlation Measurement

Pearson Correlation

• Linear association between two return series: rho = cov(X,Y) / (sigma_X * sigma_Y)
• Assumes bivariate normality for inference — violated by financial returns
• Sensitive to outliers: a single extreme co-movement can dominate the estimate
• Does not capture non-linear dependence or tail dependence

Rank Correlation (Spearman, Kendall)

• Spearman: Pearson correlation on ranked data — robust to outliers
• Kendall tau: proportion of concordant vs. discordant pairs — better for small samples
• Both capture monotonic (not just linear) relationships
• Preferred for non-normal distributions and copula calibration

Rolling Correlation

• Compute correlation over a trailing window (e.g., 60, 120, 252 days)
• Reveals time-varying nature of correlation
• Window length tradeoff: short = responsive but noisy; long = stable but lagging
• Exponentially weighted correlation: more weight on recent observations, decay factor lambda

Correlation Dynamics in Crises

The most dangerous property of correlation: it increases precisely when diversification is needed most.

• During calm markets: cross-asset correlations are moderate, diversification works
• During stress: correlations spike toward 1.0 (or -1.0 for safe havens), diversification fails
• Equity correlations: pairwise stock correlations routinely jump from 0.3 to 0.7+ in sell-offs
• Cross-asset: equity-credit correlation spikes (both sell off); equity-Treasury correlation flips negative (flight to quality)
• 2022 anomaly: stocks and bonds fell together — 60/40 portfolio assumption of negative correlation failed
• Correlation asymmetry: higher correlation in down markets than up markets (documented in academic literature)

Why Correlation Increases in Stress

1. Common factor dominance: in crisis, a single "risk-on/risk-off" factor drives all assets
2. Leverage and forced selling: margin calls force liquidation across asset classes simultaneously
3. Information cascading: panic spreads, investors herd into similar trades
4. Liquidity withdrawal: market makers widen spreads on everything, co-movement increases mechanically
5. Volatility effect: correlation and volatility are positively linked (conditional correlation rises with vol)

Methodology

Dynamic Conditional Correlation (DCC-GARCH)

The DCC model of Engle (2002) captures time-varying correlation:

1. Step 1 — Univariate GARCH: fit GARCH(1,1) to each return series to obtain standardized residuals
   • h_it = omega_i + alpha_i * r_{i,t-1}^2 + beta_i * h_{i,t-1}
   • Standardized: epsilon_it = r_it / sqrt(h_it)
2. Step 2 — Dynamic correlation: model the correlation of standardized residuals
   • Q_t = (1 - a - b) * Q_bar + a * epsilon_{t-1} * epsilon_{t-1}' + b * Q_{t-1}
   • R_t = diag(Q_t)^{-1/2} * Q_t * diag(Q_t)^{-1/2} (normalize to get correlation matrix)
   • Parameters a and b control correlation dynamics (a = news impact, b = persistence)
   • Q_bar is the unconditional correlation matrix of standardized residuals
3. Estimation: two-step quasi-maximum likelihood — first step estimates univariate GARCH, second step estimates DCC parameters
4. Advantages: captures time variation, mean-reversion in correlation, parsimonious (2 parameters for any dimension)
5. Limitations: symmetric (does not distinguish up/down moves), computationally intensive for large N

Asymmetric DCC (aDCC)

Extends DCC to capture the asymmetry that correlations increase more in downturns:

• Q_t = (1-a-b)Q_bar - gN_bar + aepsilon_{t-1}epsilon_{t-1}' + bQ_{t-1} + gn_{t-1}*n_{t-1}'
• Where n_t = epsilon_t if epsilon_t < 0, else 0 (captures negative shocks)
• Parameter g captures the asymmetric correlation response to negative returns

Copula-Based Dependence Modeling

Copulas separate marginal distributions from the dependence structure:

• Gaussian copula: no tail dependence — underestimates joint extreme risk
• Student-t copula: symmetric tail dependence — better for financial data
• Clayton copula: lower (left) tail dependence — captures crash co-movement
• Time-varying copulas: fit copula parameters dynamically, similar to DCC approach
• Calibration: use Kendall's tau or maximum likelihood on pseudo-observations

Correlation Risk in Portfolio Context

Diversification Ratio

• DR = (sum of w_i * sigma_i) / sigma_portfolio
• DR = 1 means no diversification (perfect correlation); higher DR = more diversification
• Monitor DR over time; declining DR signals rising correlations

Effective Number of Bets (ENB)

• Based on PCA of the portfolio return covariance
• ENB = exp(-sum of p_i * ln(p_i)) where p_i are normalized eigenvalue shares
• Low ENB despite many positions means hidden concentration through correlation

Correlation Stress Testing

• Stressed correlation matrix: increase all pairwise correlations by a fixed amount (e.g., +0.3)
• Must ensure the resulting matrix is positive semi-definite (nearest PSD matrix if not)
• Alternatively, use shrinkage: C_stressed = (1-w)C_current + wC_crisis
• Compute portfolio risk under stressed correlations

Examples

Rolling Correlation Regime Detection

SPY-TLT rolling 60-day correlation:
  2019 (calm):     -0.35 (negative — diversification working)
  Mar 2020 (COVID): -0.55 (strongly negative — flight to quality)
  2021 (recovery): -0.20 (weakly negative)
  2022 (inflation): +0.40 (positive — stocks and bonds falling together)
  2023 (normalization): -0.25 (returning to negative)

A 60/40 portfolio assuming -0.30 correlation was severely mispriced in 2022.
Actual 2022 portfolio vol was 40% higher than model predicted due to correlation shift.

DCC-GARCH Fitted Correlation

Pair: Bank of America (BAC) vs. JPMorgan (JPM)
DCC parameters: a = 0.03, b = 0.95 (high persistence)
Unconditional correlation: 0.78
Conditional correlation range: 0.55 (calm periods) to 0.95 (crisis periods)
2008 peak conditional correlation: 0.96
2020 COVID peak: 0.93
Current: 0.75

Implication: during stress, these two positions offer almost zero diversification.
Portfolio VaR assuming constant 0.78 correlation underestimates stress losses by 15-20%.

Dispersion Trading Example

S&P 500 implied correlation (from index vol vs. component vols): 0.65
Historical realized correlation (trailing 3 months): 0.50
Trade: Sell index straddle, buy single-stock straddles (short correlation).
Profit if realized correlation stays below implied correlation.
Risk: correlation spike event (market crash) causes large loss on short index vol.
Size: limit notional so max loss at correlation = 1.0 is within risk budget.

Quality Gate

• Correlation estimates must use at least 120 observations (approximately 6 months daily data)
• DCC model must be validated: Ljung-Box test on standardized residuals, likelihood ratio test vs. constant correlation
• Stressed correlation matrix must be verified as positive semi-definite
• Portfolio risk under both current and stressed correlation must be reported
• Diversification ratio and ENB must be monitored and reported alongside standard risk metrics
• Correlation asymmetry must be tested: compare up-market vs. down-market correlations
• Copula model selection must be justified: compare Gaussian, t, and Clayton via AIC/BIC
• Tail dependence estimates must accompany linear correlation for any risk assessment
• Correlation assumptions in risk models (VaR, ES) must be stress-tested separately
• Re-estimate dynamic correlation models at least monthly; review for regime changes quarterly