correlation and covariance matrix construction

HankScorpio · Post by **HankScorpio** » Fri Apr 04, 2008 12:12 am

IceViking, I have sent you the spreadsheet.

rakato, I shall need an email address from you, so you must either supply one in your profile (or publicly here in the forum) or just send me an email (my address is in my profile).

Cheers.

Ice Viking · Post by **Ice Viking** » Fri Apr 04, 2008 12:32 pm

Gracias hermano...Got it...also I see his 3rd edition is out

ColdBloodedOldTimes · Post by **ColdBloodedOldTimes** » Fri Apr 04, 2008 4:37 pm

I just bought this book on the recommendations of Hank and the NPers on the other thread and have found it to be most excellent.

Thanks!

rakato · Post by **rakato** » Tue May 13, 2008 5:08 pm

Is there a limit in excel to how big you can make the matrix? I am having trouble getting it to work for 85 names, but no problem with 20 names. Thanks

diogenes · Post by **diogenes** » Wed May 14, 2008 6:14 pm

I think there is a limit in older versions of excel (#cols = 250?).

rakato · Post by **rakato** » Wed May 14, 2008 6:34 pm

It worked for me up to 73x73 and then went all #VALUE! on me

MadMax · Post by **MadMax** » Wed May 14, 2008 6:45 pm

I think if you are building an 85x85 covariance matrix, that's 85x86/2=3655 parameters you are trying to estimate. I hope you have plenty of data, otherwise you have more serious problems...

athletico · Post by **athletico** » Wed May 14, 2008 7:20 pm

>> I think if you are building an 85x85 covariance matrix, that's 85x86/2=3655 parameters you are trying to estimate. I hope you have plenty of data, otherwise you have more serious problems...

Well put. Even assuming the data is 100% clean, with fewer than 3655 data points there is no guarantee the calculated correlation matrix will be positive definite.

rakato · Post by **rakato** » Wed May 14, 2008 7:44 pm

are you saying that as I get larger, my changes or crappy or corrupted data increases? What does positive "definite mean"?

hgeorgako · Post by **hgeorgako** » Wed May 14, 2008 7:47 pm

On the same subject.

I'm trying to convey to some of our traders the importance of computing a "positive definite" correlation/covariance matrix and that our data requirements scale up really fast as the number of underlyings increase. For 50 stocks, you need 1225 independent entries and you need at least 20 to 30 data-points for each one of those. ~36750 x-min returns for example.

From a mathematical point of view this makes sense. Positive definiteness, computing an inverse etc...

How do I practically convey this message? The question i get is....why not use 100 datapoints and compute all the correlations and trade based on those? Why do we need all that data? What can go wrong?