Skip to content
All library documents

How Covariance Shrinkage Can Make Risk Parity Nearly Equal Weighted

Article Quant Q&A · Author: Riskay
Read the original

Summary

This discussion examines why a risk-parity optimizer may return nearly equal asset weights when its inputs use a Ledoit–Wolf shrunk covariance matrix. The proposed explanation is that strong shrinkage reduces differences in the covariance estimates, leaving the optimizer with little basis for assigning unequal risk contributions. The accepted answer clarifies the relevant limiting case: equal variances and uniform correlations produce an equal-weight risk-parity portfolio.

The answer notes that Ledoit–Wolf shrinkage toward that structure can have this effect when the fitted shrinkage coefficient is at or near its maximum. It suggests checking the fitted coefficient and trying a lower, explicitly chosen shrinkage level. Even then, equal weights can persist if the original sample covariance already has equal variances and uniform correlations. The exchange explains a diagnostic relationship between the covariance matrix and risk-parity weights, but it does not compare the out-of-sample performance of shrinkage choices or establish that covariance shrinkage is generally appropriate for mean-variance, Black–Litterman, or risk-parity portfolios.

Key ideas

  • Risk parity gives equal weights when assets have equal variances and uniform correlations.
  • Strong covariance shrinkage can move estimates toward that structure.
  • A shrinkage coefficient near one can therefore produce nearly equal risk-parity weights.
  • Inspect the fitted shrinkage coefficient when diagnosing unexpectedly equal allocations.
  • A lower shrinkage level may restore differences unless the sample covariance is already uniform.

Tags

Portfolio constructionRisk managementStatistics
Full text
# Ledoit/Wolf covariance shrinkage in risk-parity optimisation


# Ledoit/Wolf covariance shrinkage in risk-parity optimisation












This is more of a theoretical question.

I have been working on some mean-variance / Black-Litterman models and played around with Ledoit/Wolf's covariance shrinkage method (sklearn function in Python). If I shrink the covariance matrix for mean variance portfolios, I get really nice portfolios along the efficient frontier (except for corner solutions), and the same goes for Black-Litterman.

Since I am comparing mean-variance and Black-Litterman portfolios to 1/N and Risk Parity allocations, I thought let's just feed the shrunk covariance matrix into the risk-parity optimizer. What happens is that I now get equal weights for all assets (in Risk Parity). I guess that is because the covariance matrix was shrunk to the extent that the differences in covariances are now too insignificant, and hence I get equal weights?

Did anybody run into this issue, or can anybody confirm my simple assumption is the reason for this? I am just playing a bit around but found this quite interesting. Made me question a bit whether it makes much sense to apply shrinkage to mean-variance and Black-Litterman in the first place.

Cheers, have a good week guys. RSK

## Answer by MGL (score 6, accepted)

https://quant.stackexchange.com/a/55350

The Risk Parity portfolio will be equal weighted if the assets have uniform correlation and equal variance. This would be the case for the shrunk covariance matrix if the shrinkage coefficient used equals unity.

In sklearn, you can check the shrinkage coefficient for the Ledoit-Wolf shrinkage after fitting it from the instance's .shrinkage_ attribute. If the shrinkage coefficient is (sufficiently close to) 1, then the Risk Parity portfolio will have (very close to) equal weights.

You could try running a shrinkage with the sklearn.covariance.ShrunkCovariance -class and explicitly set the shrinkage parameter to be well under 1. Using the resulting shrunk covariance, the resulting Risk Parity portfolio should not have equal weights (unless your sample covariance matrix does indeed have uniform correlations and equal variances).

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.