[OllieJones]

Nassim Taleb is making an argument about the structure of the banking system's risk management schemes. This is an attractive article for us to comment on because, doh, we'd like to punish somebody for the 2005-2008 Collateralized-debt-obligation and Credit-default-swap Cluster-f* (I'll call that the C3F from here on). And, surely it would be good to hold somebody accountable for what happened. Not that they'll be able to fix it. But still, it will feel like justice.

But that's not Mr. Taleb's point.

The real C3F problem is of TREMENDOUS professional interest to creative software folks. Most of us work on systems that can aggregate lots of measurements to try to get a big picture of the system. Some of us look at video and audio signals. Others of us look at web server logs. Today, I'm trying to troubleshoot slow DBMS performance. Our brothers and sisters in banking and trading look at measures of risk.

And we all know what we do to make sense of these measures. We average them. We sometimes throw out the outliers. We measure their standard deviations, or maybe their quintiles if we're sophisticated. And then we track the averages and other aggregates, assuming that it's sound to do so.

My boss asked me today, "is the average query time going up?" I responded, "wrong question! we need to look at the outliers."

This approach to averaging measurements feels like something we got from our mothers' milk as infants. But it's based on Gauss's Central Value Theorem, which shows that independent (repeat INDEPENDENT) measurements tend to have a normal bell curve distribution. (We call that a Gaussian distribution in honor of the Central Value Theorem).

So, what the heck, let's sell mortgages to poor folks, and huge mortgages to rich folks. They can't ALL fail to pay, can they? The ones who fail to pay will be the outliers, won't they? The Central Value Theorem teaches us that the average person will pay up. So we can manage the two-sigma risk by buying a credit default swap (you have AIG's phone number, call them!), and all is well.

Except for one thing. Mortgage defaults aren't independent of each other. When one property on the block goes into default, it becomes harder to sell the others or refinance them. So the Central Value Theorem's premise of INDEPENDENT measurements fails. Big time. Lo and behold, C3F.

Mr. Taleb's point is that in the real world of risk management, things aren't Gaussian. The events he calls black swans are long-tail events (that is, their probability curve falls off far slower than the Central Value Theorem predicts).

Why is this relevant to HN? Because we can easily deceive ourselves by ignoring outliers (black swans) in our fields of work. Hopefully it won't be as catastrophic as C3F, but we should beware.

Seriously, if you haven't read Mr. Taleb's book The Black Swan, it's worth your trouble.

[bwanab]

You are correct in most of your points, but it's worth pointing out that for the 25 or 30 years prior to 2007 mortgage defaults did tend to be uncorrelated. During that entire time frame, if you were a trader who chose to insure yourself against black swan events, you were not likely to have kept your job for long since the cost of that insurance would have been prohibitively high cutting deeply into your profits.

There's a great line (one of many) in Andrew Ross Sorkin's book "Too Big To Fail" in which Jamie Dimon, the CEO of JPMorganChase told Hank Paulson, at the time the Treasury Secretary, "You've got to make us do it". It's essentially the prisoner's dilemma. It would be advantageous for all the banks to institute a given change, but none of them can do it by themselves without being chewed into little pieces by the others.