Red-Blooded Risk: The Secret History of Wall Street

Surprise!

The first surprise we had was finding out that VaR was extremely hard to compute. Theoretical approximations based either on Gaussian distributions or on nonparametric methods failed miserably. You didn't get the right number of breaks, and the breaks were clustered in time and came disproportionately on days when VaR was low. The time clustering meant the VaR didn't have the right probability every day, even if you could get the probability right on average. For example, you might have breaks on 5 percent of days, which is what you want, but have breaks 15 percent of the time on days following another VaR break, and 2 percent of time on days more than four weeks after a VaR break. That's no good; you need VaR to be right every day. Having breaks mostly when VaR was low was dangerous. It meant bad things happened often when you said VaR was low, and seldom when you said VaR was high. People lose faith quickly if you do that.

The unvarnished fact was that none of us understood our risk in the center of the distribution. People worry all the time about tail risk—extreme events. But we didn't know much about what happened on all the other days—the normal days without extreme price movements. Moreover, most of our breaks came from system or data errors. We spent more time thinking up clever approximations and defaults for missing or obviously incorrect data than about how financial prices might move. An essential aspect of VaR is that it is never recomputed. The VaR you set ...