Stress testing covariance

Question

Going one level beyond stressed scenarios, to parameters e.g. for a VaR measure: what are the most common approaches for stressing a covariance/correlation matrix, especially taking portfolio exposure into account (so reverse testing iirc)? How many traders want e.g. to play with eigenvectors as opposed to just a concentration plus a global variance parameters? Is it common to leave full freedom of fudging the matrix subject to a later PSD correction? I'm interested in both current practice and what the leading software risk tools are offering, less in the plethora of possibilities investigated in academia.

Here's a related question with little response.

Answer 1

From what I understand, this is how stressing the covariance matrix works (and if I take the factor regression from the technique of Rosenberg 1974):

1. Within your dataset, there should be asset returns (based on asset prices), factor exposures (specific to the assets) and the factor returns (based on factor levels). Within the Rosenberg (1974) variant – the loadings of individual security returns on the factors are determined by the observable characteristics of the firm, the cross-sectional regression produces the factor returns.
2. Estimate covariance matrix using ex-post asset returns.
3. Determine the magnitude and direction of the shocks to be applied to the relevant factors that drive the returns of your portfolio assets.
4. Using the shocks (in basis points and across multiple periods), add them into the factor returns. Along with the factor exposures, compute your asset returns for each historical point in time.
5. With the shocked asset returns, compute your stressed covariance matrix and compare it to your previous covariance matrix in (2.) to observe any differences in overall portfolio risk measures like VaR, CvaR or ES.

I would appreciate it if any risk management expert could offer their insights into the underlying mathematics, thanks!