Single- and Multi-Factor Models for Return Covariance

Summary

The note compares how single-index and multi-index factor models represent the covariance matrix of stock returns. In the single-factor case, covariance is modeled as market variance times the outer product of security betas, plus a diagonal matrix of asset-specific residual variances. This structure is simple and imposes strong constraints on possible correlations.

With multiple factors, factor loadings and the factors’ own covariance matrix determine the shared component, while diagonal residual variances capture security-specific risk. Adding factors can represent a richer, higher-rank covariance structure. The model remains an approximation when there are fewer factors than securities, so it cannot express every possible covariance matrix. The note explains the structure conceptually but gives no empirical comparison or guidance for selecting factors.

Key ideas

• A single-index model represents shared return variation through one market factor and security betas.
• Idiosyncratic variances are added on the diagonal of the covariance matrix.
• A multi-index model uses factor loadings and factor covariances to represent shared risk.
• Using fewer factors than securities limits the covariance structures the model can represent.

Tags

# Answer by Alex C (score 1)


# What is the difference between the Single Index Model and Multi-Index Models in computing the variance-covariance matrix of stock returns?












Would be very grateful for some help in comparing the single index model with other multi-index models in computing the variance-covariance matrix.

## Answer by Alex C (score 1)

https://quant.stackexchange.com/a/33918

In a single factor model the covariance matrix of the returns is

$\Omega=\sigma_M^2\beta\beta^T+D$

where $\sigma_M^2$is the market variance, $\beta$ is an N vector containing the Betas of the N securities, and $D$ is a diagonal matrix containing the residual variances of the N securities. This kind of matrix is fairly simple and far from general.

In a multiple factor model with K factors this generalizes to

$\Omega= B\Omega_f B^T+D$

now $B$ is an $K \times N$ matrix of factor loadings and $\Omega_f$ is the $K \times K$ covariance of the factors. Again $D$ is a diagonal matrix of idiosyncratic variances for the N securities. This kind of model can produce matrices more complex (of higher rank) than the prior method. However, it still cannot match any arbitrary covariance matrix unless $K=N$. With $K \lt N$ it involves some approximation or simplification compared to an arbitrary covariance matrix.

Shown in full with attribution under the source's licence. Licence: CC BY-SA 4.0 (Stack Exchange)

This summary was written by Stratmill's research agent from the original; it is not a copy of the source.