- Jonathan81
**Posts:**122**Joined:**

Mr quant i don't think that your formula is good look at the formula p16 between local correlation (marginal compound correlation ) and base correlation

- Jonathan81
**Posts:**122**Joined:**

last question aconze"and then mapping these strikes into strikes in the X domain using the large pool model"I don't see what are you talking about (N(xk)=L(K) ????)thank you

Finally, does anyone try to implement Societe Generale's model ??

One of the thing that I observe is that when you compute Cumm Loss Distribution function from the continuous local correlation then the CDF is humped at the base but on the other hand a spiky local correlation gives a continuos local correlation.How it can be explained bcoz local correlation should give the right CDF ...?

Please ignore the previous reply............One of the thing that I observe is that when you compute Cumm Loss Distribution function from the continuous local correlation then the CDF is humped at the base but on the other hand a spiky local correlation gives a continuos CDF.How it can be explained bcoz local correlation should give the right CDF ...?

This thread faded away fast... Is it because of lack of public interest? I want to try to revive it with one request and one question:1. Someone on this forum mentioned another technical paper of SG guys with more details on their local correlation model. Could anyone please upload it here if you have it. Would be much appreciated.2. I am confused about their calculation of d K/d (Spread) on p. 19. (here Spread = index spread) in the computation of delta for a tranche with strike K. They define subordination K as a function of total probability to be in-the-money, which is obtained with their "market law of losses". If the index spread changes, thenthe market-implied cumulative loss distribution P(L>K) for the strike K will change as well, partly in a calculable way (due to the first copula-based term in the formula on p. 7), and partly in an uknown way, as we do not know how the correlation smile (that enters the second term in the same formula on p.7) will change if the index spread will move tomorrow. Therefore, it does not appear possible to calculate the change of P(L>K) for a given level K when the index spread S moves. As a consequence, as they assume that subordination is a function of P(L>K), I don't understand how they manage to calculate the change of subordinationdK/dS, and eventually to come up with the values of delta they report. Am I missing anything?

Hi all,has anyone tried to implement this model?Regards,SurferD

It is an interesting idea. However the paper uploaded earlier contains several errors. This ends up with the formulas on page 9 as well most of appendix of the uploaded paper being incorrect.The main error is that individual A_j's are no londer normal, when rho(X) is not degenerate, but have a quite complicated distribution, which in fact depends on the function rho(.) itself (call it F(x;rho(.))). Therefore it is incorrect to write inverse normal cdf, for example, in the second equation on page 9 (for epsilon*) . They repeat the same mistake through out the appendix as well.Second of all, it is a mistake to say "Local correlation can be derived from this relationship by finding the roots of a second order polynomial equation" at the bottom of page 9. It is not a second-order equation. They just didn't notice that there is a term in it, namely, N^(-1)(p), which is in fact should be F^(-1)(p;rho(.)). Therefore the dependence on rho in this equation is much more complicated. I don't think it can be solved analitically. There could be some recursive numerical methods, but it requires an analysis of existence/uniqueness.Hope they find and correct the mistakes...Cheers...

Observer, you're right but I did some tests on this model and actually, this error (that we can assume as an approximation) has a small impact.and any way, we can find a anlalytic solution for this problem without using the approximation above. (using some numerical integration tough)

Finally, has someone verified the claimed robustness of such a model in the pricing of bespoke tranches?

Last edited by Zub on September 14th, 2006, 10:00 pm, edited 1 time in total.

- Jonathan81
**Posts:**122**Joined:**

Observer : During "Les petits dejeuners de la Finance" Mr Turc presented a new version of local correlation with new slides and he said that it is a good approximation to take normal inverse of p even the it is not a normal.

Last edited by Jonathan81 on September 14th, 2006, 10:00 pm, edited 1 time in total.

GZIP: On