Pragmatic question about use of Stress Tests/metrics

Question

I have been looking at the onboarding of some derivative products, and the subject of our internal stress framework. I suspect like similar businesses, we have a set of stress scenarios, mostly based on historical events but given ad hoc tweaks, augmentations etc. at the discretion of the risk group. These are circulated and the outcome of the stress calculations are used for internal reporting, limits etc. in conjunction with other standard methods, VaR etc.

My question is, without trying to be too delicate, given the scenarios are circulated and calculated on a fixed schedule (they are not calculated daily) how might one stop the front office from optimizing their portfolio to the schedule and/or choice of scenarios? It is of distinct value to for the desks to get low scores on these tests.

Given the stress scenarios are large, deterministic moves, on fixed dates, it is relatively straightforward to design option portfolios that pay off on those particular dates for those particular moves. e.g. short dated ratio call/put spreads where the strikes and maturities are adjusted to give maximum benefit for the particular scenarios.

The most obvious first step would be to randomize the date of the stress, if one has a good enough handle on the term structures involved.

Given the philosophy of stress, it seems better not randomize the magnitudes, but one could randomize the selection from a given set of scenarios?

Is anyone aware of anything published anything on the matter? Or had thoughts on the subject?

Many thanks!

Answer 1

Your process of calculating the impact of market stress scenarios sounds more manual / less automated than best industry practices. The disadvantages of having manual processes incclude:

• it's expensive, so you're reluctant to add new scenarios to your portfolio of scenarios. (You can't just click a button and add "March 2020", like some lucky people can.)

• there's a good chance that the humans doing it will sometimes screw something up accidentally

• there's a tiny chance that the humans will screw something up on purpose in cahoots with your traders.

I think, if you fully automate it, then you'll be able to run more market stress scenarios daily, rather than monthly, and even intraday, and worry much less about the cost of adding scenarios.

But full automation is a big ask if you don't have it. Assuming that you work within this limitation:

Your concern that the traders will flatten their risks on days when they know you will check, and, conversely, knowing that you won't check right after for another month, go to town on risk-taking sounds very valid.

• You could randomize the days on which you run your monthly tests. I.e., the probability of picking "today" should be about the same approximately for the first 30 days after running the test, and then rapidly reach 1.

• you could select random scenarios from a large portfolio of (redundandish) scenarios.

• you coul perform reverse analysis (related question Stress testing by Banks ) - given the book, what are the classes of plausible market scenarios (not necessarily stressed) that would hurt it? If you calculate (daily) VaR (historical or Monte Carlo), you can see the adverse scenarios right there, but there may be other plausible adverse scenarios.

As for the idea of randomizing the market moves in the scenarios, I'm unsure how useful that would be for market risk management. A growing number of firms do generate (as often as daily) random (extreme) stress market scenarios and run their pricing models and related risk infrastructure through that. It's a regression test on a predefined test-case trades and sometimes on the actual trades as well. But its goal is not market risk management, but rather model risk management - the ongoing performance monitoring. They verify that the pricing models don't fail, and all the associated plumbing (P&L Explain etc) works as expected.

But for market risk management, suppose you know that the P&L from a scenario where S&P 500 moved 1000+random*10000 is USD -1 billion. How do you know whether this is a lot? For a non-random stress scenario, figuring out the limit is non-trivial. for randomized one, I just don't see a good way of judging whether the P&L is within your appetite.