Correlation - The Foundation of Diversification
By EC Assets Research Team, Quantitative Foundations · Published · Updated
Correlation — Correlation measures the degree to which two return series move together. It is the foundational input to portfolio construction, ranging from -1 (perfectly opposite) through 0 (independent) to +1 (identical movement).
Definition
Correlation measures the degree to which two return series move together. Mathematically, it is the covariance of two variables divided by the product of their standard deviations. The result is a dimensionless number from -1 to +1 with intuitive meaning:
- +1.0: Perfect positive correlation. The two assets always move in the same direction by proportional amounts.
- 0.0: No linear relationship. The movement of one tells you nothing about the movement of the other.
- -1.0: Perfect negative correlation. The two assets always move in opposite directions by proportional amounts.
In practice, asset correlations cluster in narrower ranges. Equity-bond correlation has averaged around 0 over multiple decades. Within-asset-class correlations are typically positive (most equities move together, with average pairwise correlation around 0.3-0.5). Cross-asset-class correlations are typically lower but variable.
Why Correlation Drives Portfolio Construction
The variance of a two-asset portfolio is:
σ²_p = w₁²σ₁² + w₂²σ₂² + 2 w₁ w₂ σ₁ σ₂ ρ₁₂
The correlation ρ₁₂ enters the equation as a multiplicative term that determines how much the two assets reinforce or offset each other. Consider two assets each with 20% volatility, combined 50/50:
| Correlation ρ | Portfolio volatility | Diversification benefit |
|---|---|---|
| +1.0 | 20% | None |
| +0.7 | 18% | 10% reduction |
| +0.3 | 16% | 20% reduction |
| 0.0 | 14% | 30% reduction |
| -0.3 | 12% | 40% reduction |
| -0.7 | 7% | 65% reduction |
| -1.0 | 0% | Complete cancellation |
The diversification benefit is non-linear in correlation. Moving from correlation 1.0 to 0.5 produces a larger benefit than moving from 0.5 to 0.
Time-Varying Correlations
[!key] Correlations are not constant. The long-run equity-bond correlation has averaged near zero over multiple decades, but the path has been highly variable: strongly negative during the 1990s-2000s deflationary regime, zero or slightly positive in the 2010s, strongly positive (+0.6) in 2022 during the rate-driven decline. Portfolio construction based on long-run averages can be substantially misled if the current regime differs materially.
Three structural reasons correlations move:
Regime shifts. Major macroeconomic regimes (inflation vs deflation, monetary easing vs tightening, growth vs recession) affect different asset classes differently. The 2022 regime of high inflation and rate hikes pushed equities and bonds down together because the same factor (rising rates) hurt both.
Stress dynamics. During severe drawdowns, correlations across risk assets converge toward +1. This is the 'all correlations go to one' phenomenon - driven by deleveraging, liquidity withdrawal, and macro shocks affecting all assets.
Structural change. Some correlation shifts persist beyond short-term regime variation. The post-1990s shift from positive equity-bond correlation (1970s-80s inflation era) to negative correlation (1990s-2010s deflation era) lasted three decades.
Implementation Considerations
Estimation window. Short windows (1-2 years) capture current regime but with high statistical noise. Long windows (10+ years) provide statistical precision but blur across regimes. Most institutional practice uses 3-5 year rolling windows with explicit awareness of regime sensitivity.
Conditional correlations. Sophisticated portfolio construction increasingly uses conditional correlations - correlations specifically during stress regimes - rather than long-run averages. The conditional correlation is typically higher and more relevant for tail-risk analysis.
Shrinkage. Statistical estimation of correlation matrices for large portfolios (50+ assets) produces noisy estimates. Bayesian shrinkage techniques produce more robust matrices by combining sample estimates with prior expectations.
Common Misconceptions
"Low correlation means low risk." False. Low correlation means low covariance contribution to portfolio risk, but the underlying assets can still be individually risky. Combining ten highly volatile but uncorrelated assets produces a portfolio that's still moderately risky.
"Zero correlation means independence." Correlation only measures linear relationships. Two variables can have zero correlation but be deterministically related through non-linear functions. Real finance examples include certain options strategies whose returns are non-linear functions of underlying prices.
"Historical correlation predicts future correlation." Sometimes; often not. The reliability of historical correlation as a forward-looking estimate depends on regime stability. During regime transitions, future correlation can differ materially from recent history.
Worked Example: Diversification Math
Consider an institutional portfolio with $100M in US equities (vol 18%) and $100M in long-duration Treasuries (vol 12%). Three scenarios:
| Scenario | Equity-Bond correlation | Portfolio volatility | Diversification benefit |
|---|---|---|---|
| 1990s-2000s deflation era | -0.4 | 9.8% | 35% reduction vs equity vol alone |
| 2010s low-rate era | -0.2 | 11.5% | 25% reduction |
| 2022 inflation regime | +0.6 | 14.8% | 5% reduction |
The same dollar allocation produced very different portfolio volatility depending on the prevailing correlation regime. An institution that maintained the 60/40 mix throughout assumed a stable correlation that actually shifted by 1.0 across these episodes.
Conditional Correlation in Stress
[!warning] Correlations rise during stress. The March 2020 COVID crash saw equity-equity correlations rise from ~0.4 (normal) to ~0.8 (stress) as everything sold off together. Credit and equity correlation rose from 0.4 to 0.85. Even traditionally diversifying assets like gold and Treasuries experienced brief positive correlation with equities during forced-deleveraging episodes. The diversification you have in normal markets is not the diversification you have when you need it most. This is the "all correlations go to one" phenomenon.
Modern portfolio construction uses conditional correlations - correlations specifically during stress regimes - for tail-risk analysis. The conditional correlations are typically 20-40 percentage points higher than long-run averages and are far more relevant for understanding portfolio behaviour during the moments that matter most.
References
- Markowitz, H. (1952). Portfolio Selection. The Journal of Finance, 7(1).
- Grinold, R. C., & Kahn, R. N. (1999). Active Portfolio Management (2nd ed.). McGraw-Hill.
- Ang, A. (2014). Asset Management: A Systematic Approach to Factor Investing. Oxford University Press.
Frequently asked questions
Why is correlation more important than individual volatility for portfolios?
Because the portfolio variance formula gives correlation a multiplicative effect on diversification benefit. Two assets each with 20% volatility combined 50/50 produce: 20% portfolio vol if correlation is +1; 14% if correlation is 0; 0% if correlation is -1. The correlation, not the individual volatilities, determines the diversification gain.
What does it mean when correlations rise during stress?
During severe market stress (2008, 2020), most risk assets decline together regardless of their normal-period correlations. The phenomenon is driven by deleveraging (forced selling across positions), liquidity withdrawal (sellers can only get out of liquid positions), and macro shocks affecting all assets simultaneously. The result: diversification benefit collapses precisely when it is most needed.
How is correlation different from covariance?
Covariance measures co-movement in raw units (depends on the volatilities of both series). Correlation is covariance normalised by the product of standard deviations, producing a unit-free number from -1 to +1. Correlation is therefore comparable across asset pairs with different volatilities.
Can correlation be misleading?
Yes. Two assets with low correlation may still move together when both are affected by the same underlying driver (interest rates, oil prices, geopolitical risk). Statistical correlation captures historical linear relationships; structural correlation through shared drivers may be higher than the statistical measure suggests.
Stay informed
Market commentary, firm news and research from EC Assets - direct to your inbox.