Sign up ×
Quantitative Finance Stack Exchange is a question and answer site for finance professionals and academics. It's 100% free, no registration required.

I'm currently working on my Masters project related to accelerating Greeks computations for CVA on mixed interest rate portfolios. I would like to know about the status of technology for CVA and its Greeks computations in the industry (mainly related to speed of computation).

Example situation:

  1. Portfolio of 100 000 instruments
  2. Mixture of IR Swaps, Swaptions on multiple currencies
  3. Consider case with credit/IR correlations AND without them

Question: How long (approximately, or simply mention the order) would it take on your system (or system you know) to compute total CVA (including all the netting agreements, collaterization stuff...) and sensitivities of it to every yield curve used, vol surface?

If it is not too confidential, mention the underlying technology (cpu cluster, gpus) and maybe also methods used (like Longstaff-Schwartz); you can skip the name of institution.

Why I need this? I do have a few numbers from local smaller banks, but I'd like to get a broader picture for the need of accelerated methods for these computations.

(Basel III is coming soon, so this will be mandatory for every single serious bank.)

I hope it is clear what I'm seeking.

share|improve this question
Community wiki? –  quant_dev Mar 25 '11 at 12:40
A link ? If i google about it - i get a few papers on CVA on SWAPS (mainly D.Brigo) & lots of forum posts on how people have not clue how it works. –  Vytautas Mar 25 '11 at 12:57
Can you share the papers? –  Student T Mar 29 '11 at 0:40
start here: See "Counterparty Risk" –  Vytautas Mar 29 '11 at 9:43
What are the numbers you see from local banks? –  Phil H Oct 19 '12 at 9:45

2 Answers 2

Claudio Albanese has a paper on the topic of GPUs and CVA computations. Here is one of his papers: link to paper

share|improve this answer
:) I've read the paper and quite a few others by Luca Capriotti. I'd like to mark this question as closed as I dont think I'll get a straight answer to initial question - which was how time-consuming these computations were on existing systems. Yes with infinite parallelization - comp time goes to zero, but nobody has infinite resources... –  Vytautas May 6 '11 at 8:10
Then the question is even more vague :) You either speed up by adding hardware or find more of analytical solution to a very multi-factor problem. Haven't seen Capriotti papers, will take a look. GPUs is an easy add-on to most desktop/server systems and with just 2 Tesla cards you can get 1000 cores on one server, inexpensive speedup(especially for MC problems) for just reprogramming the software. –  Ari May 6 '11 at 19:50
lol at "just reprogramming the software." –  Tim Barrass Sep 21 '12 at 12:38

This book is quite good as a starting point:

share|improve this answer
The book & the theory is not what i'm looking for. <br> I'm trying to get a picture on the computational complexity people face in practical implementations, runtime. –  Vytautas Mar 25 '11 at 13:31
Gregory actually discusses this in his book. A generic answer people can give you is "there's lots of complexity" (why? because you need to evalute lots of future scenarios). If you want a detailed answer, you need to be more precise in your question. –  quant_dev Mar 25 '11 at 14:31
Some details for the question: 3000 scenarios, 10y horizon, 100 observation dates for the PFE. I managed to get my hands on the mentioned book, but i still don't see any example timings, which is mainly what i'm searching for... –  Vytautas Mar 25 '11 at 15:07
Can you evaluate scenarios in parallel? –  quant_dev Mar 25 '11 at 20:13

Your Answer


By posting your answer, you agree to the privacy policy and terms of service.

Not the answer you're looking for? Browse other questions tagged or ask your own question.