Nassim Nicholas Taleb, the dark knight or rather the black swan himself, said that “anything that relies on correlation is charlatanism”. I am currently reading the excellent “The signal and the noise” by Nate Silver. In Chapter 1 of the book he has a nice piece on CDOs as an example of a “catastrophic failure of prediction” where he points to certain CDO AAA tranches which were rated on an assumption of a 0.12% default rate and which eventually resulted in an actual rate of 28%, an error factor of over 200 times!.
Silver cites a simplified CDO example of 5 risks used by his friend Anil Kashyap in the University of Chicago to demonstrate the difference in default rate if the 5 risks are assumed to be totally independent and dependent. It got me thinking as to how such a simplified example could illustrate the impact of applied correlation assumptions. Correlation between core variables are critical to many financial models and are commonly used in most credit models and will be a core feature in insurance internal models (which under Solvency II will be used to calculate a firms own regulatory solvency requirements).
So I set up a simple model (all of my models are generally so) of 5 risks and looked at the impact of varying correlation from 100% to 0% (i.e. totally dependent to independent) between each risk. The model assumes a 20% probability of default for each risk and the results, based upon 250,000 simulations, are presented in the graph below. What it does show is that even at a high level of correlation (e.g. 90%) the impact is considerable.
The graph below shows the default probabilities as a percentage of the totally dependent levels (i.e 20% for each of the 5 risks). In effect it shows the level of diversification that will result from varying correlation from 0% to 100%. It underlines how misestimating correlation can confound model results.