Covariance
Cov(X,Y) = E[(X - mu_X)(Y - mu_Y)]Covariance preserves the units and scale of the joint movement.
Foundations
Covariance Estimation and Correlation
Covariance measures joint variability in units; correlation rescales it to a dimensionless sensitivity between −1 and 1.
Mental model
Covariance measures joint movement in original units, while correlation expresses that movement on a comparable scale.
Covariance captures how two variables move together but depends on their units and scales.
Correlation normalises covariance by the product of standard deviations, so it is unit-free and lies between −1 and 1.
Explore the mechanism
How does rescaling joint movement change what can be compared across risk factors?
This scatter shows simulated returns for two assets X and Y. The blue cloud is the sample; the amber ellipse represents the one-standard-deviation contour implied by the target covariance.
Target ρ: 0.70 · Sample ρ̂: 0.66
Numbers
Sample var(X) ≈ 3.08e-4 · var(Y) ≈ 7.52e-4
Sample cov(X, Y) ≈ 3.16e-4
Target ρ = 0.700 · sample ρ̂ ≈ 0.656
Implied population cov ≈ 4.20e-4
Interpretation
Scaling σ_X or σ_Y stretches the cloud and the covariance, but correlation stays the same: it measures shape (tilt) rather than absolute size. That is why covariance depends on units and correlation does not.
With small n the sample covariance and ρ̂ jump around their target values; as n grows, they stabilise and the cloud matches the amber ellipse more closely. The mental model: covariance is “joint variance in units”, correlation is the same object viewed through a scale-free lens, and estimation error comes from limited data.
UNDERSTAND
What to notice
Sample covariances and correlations are noisy estimates; their stability depends on sample size and the underlying structure.
The quantitative structure
Notation
X, Y- two random variables or return series
mu_X, mu_Y- their means
sigma_X, sigma_Y- their standard deviations
Correlation
rho_XY = Cov(X,Y) / (sigma_X sigma_Y)Correlation rescales covariance to a dimensionless number between -1 and 1.
This model assumes
- The observations used by the estimator are representative of the dependence being measured.
- Moments exist and are estimated consistently for the chosen sampling process.
Where it appears
- portfolio risk aggregation
- factor-model estimation
- hedge-ratio design
- scenario dependency modeling
Where the intuition breaks
What this view leaves out
- Sample covariance and correlation can be unstable with short histories, regime changes, outliers, or many variables.
Common mistakes
“Zero correlation means two variables are independent.”
Not quite.Zero correlation rules out linear association, not general statistical dependence, except under stronger distributional assumptions.
Where this concept sits
Follow the strongest prerequisites, applications, extensions, and contrasts.
Connections for Covariance Estimation and Correlation
Covariance Estimation and CorrelationYOU ARE HERE
Built onReturns and Log-ReturnsCovariance and correlation are estimated from joint movements in return or change series.
Extended byHigh-Dimensional Covariance ShrinkageShrinkage improves unstable sample covariance estimates by pulling them toward a structured target.
Used byPower–Gas–CO₂ Correlation StructureCross-commodity dependence is quantified with covariance and correlation estimates.
Used byLinear Regression (OLS) in Factor ModelsFactor-model regression relates returns through estimated betas and residual covariance structure.
Used byValue at Risk (VaR)Parametric portfolio VaR commonly uses covariance and correlation to aggregate risk across positions.
SupportsCredit Intensity Models and Default CorrelationDefault correlation is a dependence problem, so the general ideas of covariance and correlation provide useful statistical grounding.
Built on · Returns and Log-Returns
Covariance and correlation are estimated from joint movements in return or change series.
Explore concept →