1
$\begingroup$

Have a question about SVI Zeliade Implementation (pdf, overview). The paper suggested to do 2 rounds of optimization, first for $\{a,b,\rho\}$ and 2nd for $\{m,\sigma\}$.

Does anyone know if I can invert the order of optimization. First, optimize $\{m,\sigma\}$, and then optimize $\{a,b,\rho\}$.

Benedict

||improve this question
$\endgroup$
  • 2
    $\begingroup$ It's been a while since I looked at this, but iirc you misunderstood the order: the point of the Zeliade whitepaper is that the 5 dimensional optimization can be reduced to a 2 dimensional optimization since for any given $m$ and $\sigma$ the other parameters follow as the result of an (constrained) linear program. $\endgroup$ – Bram Jan 12 '18 at 8:50
-2
$\begingroup$

I didn't find the methodologies you mentioned. My thought is you should do the five parameters optimization at the same time. Don't be scared by 5 parameters. The minimization process is actually pretty fast. We implemented in Java and to optimize SVI parameters for thousands of vol smiles take only 1 minute. Minimizing 5 parameters will definitely provide you with best fit than do them separately.

I have seen some methods that fix values on some parameters but never heard of minimizing a function separately (I could definitely be wrong). In SABR model, the backbone could be set by traders, so only three parameters could be involved in minimization although it has 4 parameters totally. It could be similar case in SVI as well to fix some parameters but not for two times minimizations

||improve this answer
$\endgroup$
  • $\begingroup$ Hui, thanks for your comment. I actually implement the calibration in a two way fashion. One was two parameters, and the other a three parameter fashion. Go check out SVI Zeliade in google and there is a paper referencing it. Thank you! $\endgroup$ – Benedict May 25 '18 at 15:46

Your Answer

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

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