Covariance
Covariance is a statistical measure of the joint variability of two assets' returns, indicating whether and to what degree the returns of two assets move together, with positive covariance signalling co-directional movement and negative covariance indicating inverse movement.
Covariance was the raw, unscaled measure of how two assets' returns moved in relation to each other, calculated as the average product of the deviations of each asset's return from its mean. Mathematically, Cov(A, B) = E[(R_A – μ_A)(R_B – μ_B)], where R represented the return in a period and μ the mean return. A large positive covariance meant the two assets tended to both exceed their means (or both fall below their means) in the same periods. A large negative covariance meant that when one exceeded its mean, the other tended to fall below it.
Covariance was a foundational input in Markowitz's portfolio variance formula. For a two-asset portfolio with weights w_A and w_B, the portfolio variance was: σ²_p = w_A² × σ²_A + w_B² × σ²_B + 2 × w_A × w_B × Cov(A,B). The last term — which could be positive, zero, or negative depending on the sign of covariance — captured the diversification effect. Negative covariance reduced portfolio variance below the weighted sum of individual variances, while positive covariance increased it.
The challenge with interpreting covariance directly was its dependence on the scale of returns. A covariance of 50 between two stocks could signify a strong or weak relationship depending on the magnitudes of each stock's individual variance. This is why correlation — covariance standardised by dividing by the product of the two standard deviations — was more commonly communicated and interpreted by portfolio managers and analysts, as it provided the same directional information in an intuitive -1 to +1 scale.
For a portfolio of N assets, the covariance matrix contained N² elements, of which N were individual variances and N×(N-1) were pairwise covariances. For a 50-stock portfolio, this meant 1,225 unique covariance estimates. The challenge of estimating so many parameters accurately from historical data — particularly given that correlations shifted through time — made large-scale mean-variance optimisation prone to instability, motivating the use of factor models that reduced the dimensionality of the covariance estimation problem by expressing individual stock covariances through a smaller number of common factors.
In India, fund houses offering quantitative or systematic strategies computed covariance matrices using rolling windows of historical return data, with typical window lengths ranging from one to five years depending on the strategy's investment horizon. Stress testing the portfolio under crisis-period covariance estimates (using data from the 2008 crash, 2020 COVID sell-off, or other acute episodes) was considered a best practice to assess portfolio robustness when normal-period diversification benefits evaporated.