A response to Danielsson and Macrae.
In “The appropriate use of risk models” I presented a synopsis of the Danielsson and Macrae document by the same name, and urged you to read it (it’s not very long).
I highlighted the sentence:
This suggests that models used to constrain risk should be substantially simpler than models used to understand risk.
What does that mean in practice? Perhaps that Ledoit-Wolf type variance estimates should be used for optimization. They are simple and seem to have good properties.
The first two (of five) reasons that Danielsson and Macrae give for the large uncertainty of risk models are:
- The model estimation period is too short
- There are structural breaks during the estimation period
I don’t think that is quite right. I see the process as continually dynamic. The term “structural break” implies periods of stability. I would replace the two items above with:
- There is an uncertainty principle at work much like that in quantum mechanics
We can’t see both location and momentum. With a short sample period we can get a vague idea of recent values. With a long sample period we get a better idea of average values. What we want is a good idea of recent values.
GARCH type models would seem to give us that. They do, but only for short time horizons — much shorter horizons than needed for most applications.
Risk models fail during crises. One way of approaching the problem is to think of two states: equilibrium and disequilibrium.
When we are in equilibrium, then risk models will work okay (though they may still have confidence intervals that are much wider than we would hope).
In disequilibrium we know that the models are untrustworthy. I’m not convinced that it is very feasible to create disequilibrium risk models. For one thing, there is very little data — not many crisis periods. Furthermore, we are likely to have the Anna Karenina problem: every crisis is a crisis in its own way. But perhaps there are ways of taking advantage of the observation that the worst case in equilibrium is likely to be the best case in crisis.
A key problem with this line of thinking is that we don’t know when a crisis will come. A current worry is that Europe will become a black hole. Just as with a physical black hole, we aren’t going to know when we soar across the event horizon into inevitable destruction. Also a crisis may appear due to something that flew entirely under the radar.
The best preparation may be to know that we are going to be unprepared. How did explorers prepare when they went into Here-be-dragons?
Much of the aim of Danielsson and Macrae seems to be on systemic risk. They point to work on banks creating living wills as encouraging.
What’s the worst that can happen?
Reducing the severity and probability of “seriously bad” seems to me to be sort of, kind of a good idea.
Same birds that followed me
To school when I was young
Were they trying to tell me something?
Were they tellin’ me to run?
from “Things That Scare Me” by Neko Case and Tom Ray