I have several assets, each with different return histories.

Some of the assets have 75 days of return history, others have 40 or so days. In calculating a robust covariance matrix, should I be using a resampling with replacement technique or should I be using an expectation maximization algorithm?

I was thinking of truncating the samples so that they all share ~40 days of return history, and then resampling that distribution of returns with replacement.

But, I'm not sure that's ideal.

  • 2
    $\begingroup$ What is your primary concern: robustness (w.r.t. outliers) or missing data? Neither resampling or EM do adress the former. There are however ways to deal with both, are you interested? $\endgroup$ – Quartz Sep 2 '13 at 15:47
  • $\begingroup$ I'm concerned with both, but my primary concern is with the robustness of the covariance matrix. I actually thought resampling would address the robustness, but not the missing data. I'm thinking the two are mutually exclusive techniques, each addressing a different component. $\endgroup$ – tragen907 Sep 2 '13 at 19:40
  • $\begingroup$ Whops sure, sloppy speed writing on my part, of course I just wanted to say that neither method adresses both aspects and that they are not mutually exclusive, and having different targets they can't be compared. There's a vast literature nowadays (even just for variants of EM), no need to stick to (rough and outdated) resampling. My favourite combined method is: ecares.org/ecaresdocuments/seminars1011/frahm.pdf . However it's not extensible to asynchronous quotes. $\endgroup$ – Quartz Sep 3 '13 at 9:50

Your Answer

By clicking “Post Your Answer”, you agree to our terms of service, privacy policy and cookie policy

Browse other questions tagged or ask your own question.