Stylized Facts for Evaluating Generated Correlation Matrices

Summary

The document lists statistical and structural features proposed as checks on whether generated financial correlation matrices resemble empirical ones. These include a pairwise-correlation distribution shifted toward positive values, an eigenvalue spectrum broadly following Marchenko–Pastur behavior with several unusually large eigenvalues, and a leading eigenvector with positive entries. It also points to hierarchical clustering and scale-free structure in the minimum spanning tree as possible diagnostics.

The post frames these features as a question for portfolio managers: should they routinely verify such realism properties, or prioritize whether a matrix works for portfolio optimization and asset allocation? It does not answer that question, provide thresholds, or present tests on generated matrices. The listed properties are therefore a proposed diagnostic checklist, not evidence that any particular GAN output is realistic or suitable for investment decisions. Practical validation would need to consider the matrix’s intended portfolio use as well as its statistical structure.

Key ideas

• Empirical correlation matrices may show pairwise correlations shifted toward positive values.
• Their eigenvalue spectra can have a broad Marchenko–Pastur pattern alongside a few large eigenvalues.
• A positive leading eigenvector, hierarchical clusters, and scale-free minimum spanning tree structure are proposed realism checks.
• The document raises, but does not resolve, whether these diagnostics matter beyond suitability for portfolio optimization.
• The listed properties are not accompanied by thresholds or validation results for any specific generated matrix.

Tags

# Verification of the Realism of Generated Correlation Matrices


# Verification of the Realism of Generated Correlation Matrices












I was reading up on the generation of correlation matrices with generative adversarial networks (GANs) on the blog (https://gmarti.gitlab.io/ml/2019/07/15/financial-correlations-stylized-facts.html) with all credits to Gautier Marti and his amazing stuffs (do check it out), and he mentioned that empirical/realistic correlation matrices have several mathematical properties:

- Distribution of pairwise correlations is significantly shifted to the positive,

- Eigenvalues follow the Marchenko–Pastur distribution, but for a. a very large first eigenvalue, b. a couple of other large eigenvalues,

- Perron-Frobenius property (first eigenvector has positive entries),

- Hierarchical structure of clusters,

- Scale-free property of the corresponding minimum spanning tree (MST).

And these correlation matrices live on a particular $N$-dimensional elliptope depending on the number of assets $N$. Also see (https://gmarti.gitlab.io/ml/2019/06/23/CorrGan-3D.html) on his explanation of how the $N$-dimensional elliptope came about.

My question is - for those of you working in portfolio management, how often do you verify if your correlation matrices meet these properties, or is it only important for you to check if the matrices are suitable for portfolio optimization/asset allocation purposes?

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.