Partial least squares

danielbogod · May 9th, 2006, 6:04 pm

I have to model a fund with a set of indexes. However, all the indexes are correlated, and I have to figure out the weights of each index in the fund. Its probably not a good idea to use OLS, because the results are meaningless, so I was thinking of PLSR or some sort of ICA. any suggestions? Also what would be a good measure for model fitting(i.e. its prediction power)?Thank you!

Frashe · May 9th, 2006, 11:14 pm

What are you going to do with the model when you have it? That determines the method used. But I suspect that if the OLS results are meaningless then so will be any interpretation of the results from any analysis. This is an overhyped area.

danielbogod · May 10th, 2006, 1:30 pm

I need the model to figure out the weights of the different indexes in the fund. later I will use this to hedge

Frashe · May 10th, 2006, 9:38 pm

You're going to hedge on the results of a model where the parameter uncertainties are huge? Good luck!

danielbogod · May 12th, 2006, 12:29 pm

any suggestions?

Frashe · May 12th, 2006, 11:37 pm

Suck it and see! Perhaps the covariance between the fund and heding instruments isn't too bad, in which case the actual choice of hedge won't really matter. Historical simulation augmented by some fat tails scenarios might show you can get away with a simple robust regression.

PaperCut · May 13th, 2006, 1:57 am

QuoteOriginally posted by: Frashe Suck it and see! ? ? ? ? ? ? ? ?

Frashe · May 13th, 2006, 9:00 am

Try it out.