User specified covariance matrix about riskfolio-lib

[dcajasn]

Hi, not by now, when I implement that feature I will release and example.

I think is good to have this feature asap if you want to further increase the visibility and outreach of the package because industry people usually work with their own covariance matrix. De Prado's research was very influential on this.

[dcajasn]

Hi, well by now is not my priority to include that feature. Now I'm focus on other features like additional hierarchical clustering techniques, NCO model and relaxed risk parity.

Up to you. Sound interesting additions. But my advice is that people would care more about a feature having a user-specified covariance because this is at the heart of all portfolio max problems. In practice, people first try to remove as much of the noise in the covariance matrix as possible by applying their own customised techniques before they use any optimisation routine.

Anyways, looking forward to these new features then.

[dcajasn]

Covariance is only the heart of Markowitz related models. HRP and HERC use naive risk parity and only need covariance when risk measure is variance. For covariance case, Riskfolio-Lib has options to reduce noise, for example we can use shrinkage methods like ledoit-wolf, oas and oracle; also you can use ewma methods to increase the weights of last observations in estimated covariance.

Correct. But read De Prado's Asset Management on the covariance matrix. I am only telling you my opinion that between those priorities my sense is that the industry would rather prefer to have the feature with a user-specified covariance matrix, rather than other features you mentioned.

But again, up to you. You did an excellent job so far and thanks for the library.

[dcajasn]

I know that some people prefer to build their own covariance matrix, in Portfolio object there is that option but not in HCPortfolio object. I will try to add this feature when I implement NCO (something like factors_stats, blacklitterman_stats, blfactors_stats and custom cov) but it's not my priority for now. As Riskfolio-Lib is my hobby, in most cases the features that I add are models that seems interesting to me from the mathematical perspective.

[dcajasn]

Hi @msh855, I implement the option to use custom covariance in hierarchical clustering portfolios. Update your Riskfolio-Lib version to the latest to try the new features.

Thanks for letting me know. Though covariance matrices may not change anything to a hierarchical clustering portfolio because such methods construct clusters based on the relative distance between assets which different covariances may not affect much. So, clustering and hence optimal shares are likely to be the same across different covariance matrices. I have tried this in the past and indeed the portfolio was robust to covariance matrices. But happy to investigate again based on your RiskFolio capabilities.

[dcajasn]

Hi @msh855, custom covariance could affect hierarchical portfolios in two ways:

• If the distance metric is covariance-based, the custom covariance change the distance metric and the hierarchical structure.
• In the optimization process, custom covariance change the inter and intra cluster variance calculations.