Serving the Quantitative Finance Community

 
User avatar
Rafafa
Topic Author
Posts: 0
Joined: August 28th, 2012, 1:21 pm

Projecting dv01s on a few drivers - SVD

August 29th, 2012, 9:59 am

Hi all,I remember seing a few years ago that one could reduce the number of dv01 exposures by "aggregating" them. This wasn't for traders but for management. For instance we may 30 dv01 exposures but some people like to see it summed up on the 2Y, 10Y, 30Y points for instance. So the other 27 exposures are aggregated on this 3 points in a meaninful way. I remember that was linked to a SVD on the covariance matrix of the rates but that is all I remember and I cannot find anything on the internet...Would somebody have a lead or a paper on the subject?Many thanks!NB: this is linked to the PCA concept except that instead of creating new axes based on the original variables, we already know the new axes (in my examples 2,10,30Y) and want to project on them.
 
User avatar
Martinghoul
Posts: 188
Joined: July 18th, 2006, 5:49 am

Projecting dv01s on a few drivers - SVD

August 29th, 2012, 10:21 am

How about simple linear interpolation? I remember doing some work on this a long time back and concluding that sometimes simple is truly best.
 
User avatar
Rafafa
Topic Author
Posts: 0
Joined: August 28th, 2012, 1:21 pm

Projecting dv01s on a few drivers - SVD

August 29th, 2012, 10:32 am

That can indeed be a good idea. But then the points between 2 pillars will only be projected on these 2 pillars. Basically the idea is the same except I'd like to be able to extract all the covariance information.And then to be honest, I would also like to apply it to a volatility surface, where I could also do linear interpolations in 2D but I guess the result would be less satisfactory because of the expiry/tenors overlaps..
 
User avatar
Blazes
Posts: 0
Joined: September 6th, 2004, 7:36 am

Projecting dv01s on a few drivers - SVD

August 29th, 2012, 11:07 am

Run your PCA analysis and see what the the first 3 PCA coefficients are for each YC point. Then for each non reference point solve a simultaneous system to map its PCA exposure to the reference points and aggregate across all points. My memory of running PCA on single yield curves is that generally 95%++ of the variability is captured by the first three components. However I don't think you can reduce the number of dimensions and preserve all of the covariance information. Not sure if this achieves what you want?
 
User avatar
Rafafa
Topic Author
Posts: 0
Joined: August 28th, 2012, 1:21 pm

Projecting dv01s on a few drivers - SVD

August 29th, 2012, 11:19 am

Hi Blazes, it is indeed the idea but it is not clear to me yet how I can map the PCA weights of the different points on each other. I will look further. As far as I can remember, the code was less than 10 lines so it is not something too exotic. Tks!