Choosing a Lookback and Regularization for Factor Covariance Estimates

# In practice, how many days are used to estimate the covariance matrix of factor returns?


# In practice, how many days are used to estimate the covariance matrix of factor returns?












Let's say we have a factor model with $N$ factors. I understand that the unbiased estimator of the covariance matrix $\Sigma_f$ is:

$$ \Sigma_f = \frac{1}{n-1} X^T X $$

where $X$ is a matrix of daily factor returns over $n$ trading days. I'm unclear on a few details regarding the practical implementation of this:

- Is there an upper bound on the number of trading days we want to use to estimate this covariance matrix?

- Is it common to use an exponential weighting to prioritize recent factor returns?

- What are the most common ways to regularize this estimate? I've seen papers based on Random Matrix Theory that drop certain eigenvalues/eigenvectors of the sample covariance matrix (e.g. Robust Estimation of Risk Factor Covariance Matrix), but I'm unsure if this is popular in practice.

## Answer by KaiSqDist (score 1, accepted)

https://quant.stackexchange.com/a/80277

Not sure about the strand of literature on factor covariance matrix estimation, but on the historical covariance matrix estimation through log returns - this is a research topic from what I have heard and discussed.

Although you would think that to estimate the future volatility of 30 days, you would require using historical vol from the past 30 days or an option with TTM of 30 days, but interestingly, I have heard that these aren't exactly the best estimates.

Of course this can be extended to conditions of the current market sentiment as option-implied volatility tends to overestimate the true realized volatility by the volatility risk premium and historical volatility is usually below the true realized volatility and yada yada yada...

The point I am trying to get here is that I don't think there is a definitive historical lookback period to use to best estimate the true realized volatility (in the future of course), and similarly, I think that is the case for the factor covariance matrix - because the whole point about the factor covariance matrix is to decompose risk into factors - in estimating portfolio risk (I say portfolio risk and not asset risk because people less often use factors to decompose asset risk because the factors help to "shrink" the size of the covariance matrix of a portfolio) - like from an assets-based size of thousands to possibly a factor-based size of hundreds or less than hundreds.