…the big problem was that it turned out that VaR could be gamed. That is what happened when banks began reporting their VaRs. To motivate managers, the banks began to compensate them not just for making big profits but also for making profits with low risks. That sounds good in principle, but managers began to manipulate the VaR by loading up on what Guldimann calls “asymmetric risk positions.” These are products or contracts that, in general, generate small gains and very rarely have losses. But when they do have losses, they are huge.
I find this interesting: reporting and acting on VaRs is not different from reporting and acting on the results of any other probabilistic model. Engineers and operations researchers do it all the time when modeling the failure rates of process units and finished products or services. The same applies to variations in process inputs (crude oil composition for a refinery, particle size distribution for a powdered pharmaceutical drug and so on). Even something as mundane as the residence time distribution in a reactor is a probabilistic model that important decisions are based on. Yet the failure modes are usually not catastrophic system meltdown. Why?
There are two questions I cannot fully answer yet: It is relatively easy to cross-check an engineering model with theory using back-of-the-envelope estimates from first principles. Is the same possible for financial models? It is also relatively easy to cross-check a model with data from controlled experiments, not just observations seen in the wild with confounding factors. Are controlled experiments feasible and practical for financial systems?
Of course, no model would be complete without a car analogy:
David Einhorn, who founded Greenlight Capital, a prominent hedge fund, wrote not long ago that VaR was “like an air bag that works all the time, except when you have a car accident.”