Statistical Factor Models for Covariance Estimation

Summary

The document outlines a proposed way to estimate a covariance matrix with a statistical factor model. It begins with a sample or robust covariance estimate, decomposes that matrix into eigenvalues and eigenvectors, rotates the factors, and retains only the most significant eigenvectors. A Marčenko–Pastur distribution is suggested as one possible guide for choosing a cutoff. Returns are then regressed on the retained factors to form the factor-model covariance estimate.

The question proposes that removing weaker components may filter noise from off-diagonal covariance estimates, while more residual noise remains on the diagonal. It explicitly distinguishes this dimensionality reduction from robustness to outliers: choosing fewer factors does not itself make estimation resistant to extreme observations. The text is a question rather than a resolved analysis; it presents no empirical comparison, estimator properties, or evidence that the proposed signal-versus-noise description holds generally.

Key ideas

• The proposed workflow starts from an initial covariance estimate and extracts eigenvectors as candidate statistical factors.
• A spectral cutoff can be used to discard less significant eigenvectors before fitting returns.
• The question hypothesizes that factor reduction may suppress noise in off-diagonal covariance estimates.
• Reducing the number of factors is not the same as using an estimator robust to outliers.
• The document does not provide results that validate its proposed interpretation.

Tags

# What is special about covariance estimation from statistical factor models?


# What is special about covariance estimation from statistical factor models?












If you were to compare the usual sample covariance estimate to a robust covariance estimate (such as MCD), you can say that the robust estimate is more tolerant to outliers in the data and will not be influenced as much by the presence of these outliers. In other words, the robust estimate is not "jumping at shadows".

Is there a similar statement you can make about covariance estimation using a statistical factor model?

For the sake of completeness, the procedure for estimating covariance using a statistical factor model is as follows:

- Estimate an initial covariance matrix via some other means (sample or robust)

- Decompose into Eigen values and Eigen vectors and apply the rotation.

- Discard all but the most significant eigen vectors (you can use a Marcenko-Pastur distribution to calculate a cutoff)

- Regress the returns on the remaining factors.

- Calculate the covariance in the usual way for a factor model.

It seems to me that by discarding the non significant eigen vectors, you are ignoring a lot of the noise and the resulting covariance matrix represents more signal on the off diagonal entries, with more noise represented on the diagonal entries. This is not the same as being robust to outliers though.

Does the above seem like a fair statement? Is there anything else we can say about it?

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.