http://www.wired.com/techbiz/it/magazine/17-03/wp_quant?currentPage=1 A very interesting and well written article about the flaws in the popular model for CDO pricing.
Gaussian copula is obviously a bad tool for financial markets because one of its assumptions is low extreme correlations which is contrary to reality. When things go wrong, correlations go to extremes (1 or -1). Even T-copula doesn’t have that ridiculous assumption of extreme correlations going to zero.
a good read, thanks for sharing
be wary of geeks with formulas
maratikus Wrote: ------------------------------------------------------- > Gaussian copula is obviously a bad tool for > financial markets because one of its assumptions > is low extreme correlations which is contrary to > reality. When things go wrong, correlations go to > extremes (1 or -1). Even T-copula doesn’t have > that ridiculous assumption of extreme correlations > going to zero. The Gaussian copula does not make any assumptions about correlations since the correlation matrix is an input into the model. In fact, you could assume that all the credits have a correlation of 1. However, by assuming a normal distribution for the marginal distributions, you may not be capturing the “fat tail” risk that exists in the market. This would have the effect of underestimating the credit risk of the portfolio.
ShouldBeWorking Wrote: ------------------------------------------------------- > maratikus Wrote: > -------------------------------------------------- > ----- > > Gaussian copula is obviously a bad tool for > > financial markets because one of its > assumptions > > is low extreme correlations which is contrary > to > > reality. When things go wrong, correlations go > to > > extremes (1 or -1). Even T-copula doesn’t have > > that ridiculous assumption of extreme > correlations > > going to zero. > > > The Gaussian copula does not make any assumptions > about correlations since the correlation matrix is > an input into the model. In fact, you could > assume that all the credits have a correlation of > 1. However, by assuming a normal distribution for > the marginal distributions, you may not be > capturing the “fat tail” risk that exists in the > market. This would have the effect of > underestimating the credit risk of the portfolio. I’m not disagreeing with you. My wording was poor (I don’t even know why I said extreme correlations while I was thinking of extreme co-movements). I was talking about how Gaussian copula understates tail dependence which is an obvious flaw of the model.
What people don’t realize is that these risk management formulas are not a one-stop shop for risk control. They are just a tool that you can use to help you analyze risk but you CANNOT rely on a simple formula exclusively.
king_kong Wrote: ------------------------------------------------------- > What people don’t realize is that these risk > management formulas are not a one-stop shop for > risk control. They are just a tool that you can > use to help you analyze risk but you CANNOT rely > on a simple formula exclusively. There were a lot of smart people on Wall Street who were well aware of this. However the incentives were heavily skewed toward generating securitization fees & service charges.