Related Topics

I was involved in writing some of the code for those tests.

Lots of monte carlo simulations. What you do calculate a lot of simulated interest rate curves and stock market curves and assumed defaults. Then you throw them against the products that the bank holds and see how the valuations change in response to the simulated inputs. Repeat several thousand times with more simulated data, and look at the results. You go through the banks mortgages, and pretend that X% go bad, what happens?

This takes a ton of compute power because you are throwing simulated inputs against the banks books and then bank can have several tens of thousands of different instruments. One problem is that if you have the models in different computer systems then you will have to combine those systems to do a bank wide stress test. The other problem is just involves optimizing the code so that you can run the tests at all.

Also, I strongly disagree with the idea that these tests are useless. Yes, they don't take into account feedback loops, but if your tests break in response to a situations without positive feedback loops, you have a *big* problem. If you pass static tests, it doesn't mean that you are out of the woods or that you are safe, but if the rocket blows up on the launch pad, then you aren't making it to the moon.

Also, I also disagree that you can fit the model to get the test result you want. The model that you use is the one that tells the trader how much the bank will buy or sell an asset if someone calls the bank, so you really can't fiddle with them too much. The pricing models that you use aren't predictive ones. They are the ones that the bank uses to charge for things.

The other thing is that the models are audited, so you just can't put in a new model if it turns out that it's causing bad test results. If you have to change a model, then the regulators are notified and they are going to be asking a lot of questions about why you are changing the model. Before the financial crisis the regulators would sometimes listen to you if you argued that model X, which was more favourable to the bank was better, but after the financial crisis, forget about it.

Also, it's not obvious before the test how a model will effect the output, and if it turns out that whether a bank will survive or not depends critically on how you model an instrument, then people are going to be looking very closely at those models. The problem was that before the financial crisis, banks didn't run stress tests so that just didn't know if something would cause the bank to blow up in a crisis. The reason that only a few banks ran these tests was that they take up a huge amount of supercomputing power, and so a bank is not going to spend them money on supercomputers and numerical programmers unless the Fed tells them to do it or else.

As far as lack of intellectual ability, building these models (and auditing them) is why banks hire physicists and mathematicians for this sort of thing.

The other thing is that I wish some of this technology could be made available to individual users, because this sort of thing would be extremely useful for personal finance. The way this would work is that you'd type in your age, and then your investment portfolio. The computer simulation would then simulate some cash flows, and then would see what you look like financially after X years. Unlike the tools on this market, what the simulator would do is to simulate life events. For example, you can see what happens if you have 15% unemployment and you get unemployed for a year. In the course of writing these stress tests, the banks came up with a lot of useful code and computational techniques, and I would hope that some of it could be open sourced so that people could use it for personal finance.