Challenges in Estimating Stable Covariance Matrices for Portfolios

Summary

The document describes a practitioner’s difficulty estimating a variance–covariance matrix that reduces portfolio volatility out of sample for a Russell 3000 universe. Approaches tried include direct covariance estimates from daily, weekly, or monthly returns over lookback windows ranging from three months to five years, as well as factor models with 10 to 100 factors. The author also tried linear regression and other regularization methods to improve factor beta estimates, but reports no out-of-sample volatility reduction.

The author is exploring techniques from López de Prado’s 2016 and 2019 papers and asks whether others have achieved positive out-of-sample results or have alternative methods for robust covariance estimation. The document presents a research problem and methods attempted, but no successful solution, comparative evidence, or response. Its observations are specific to the author’s data and implementation; they do not establish that the listed methods generally fail or that any particular alternative will improve portfolio volatility.

Key ideas

• The author reports difficulty finding covariance estimates that reduce realized portfolio volatility out of sample.
• The tested approaches include direct return-based estimates over multiple frequencies and lookback periods.
• The author also tried factor models, regression, and regularization without a reported volatility improvement.
• The post asks about López de Prado’s methods and other ways to obtain stable covariance estimates.
• No solution or evidence comparing alternative methods is provided.

Tags

# Robust estimates of variance covariance matrix


# Robust estimates of variance covariance matrix












I am looking for help from other people with experience creating variance covariance matrix that have enough predictive power to actually lower portfolio volatility out of sample.

Using real world data I haven't been able to estimate a variance covariance matrix that has enough stability out of sample to lower portfolio volatility. My universe is the Russell 3000.

I've tried using a direct calculation of the covariance matrix from daily / weekly / monthly returns over several lookback periods ranging from 3 months to 5 years. I have also tried estimating the variance covariance matrix using a factor model with 10 - 100 factors. I've used linear regressionand other regularization methods to estimate factor betas to try to improve out of sample performance but nothing has lead to an actual decrease in out of sample volatility.

Initial research has lead me to look at Lopez de Prado's techniques from his 2016 and 2019 papers. Has anyone achieved positive out of sample results using his methods or have any other ideas about how to estimate robust / stable covariance matrix?

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.