This article makes me laugh inside: Credit Swaps Thwart Fed's Ease as Debt Costs Surge
The punch line of the article, similar to the signs we have seen at quant hedge funds is that we are not as smart as we thought we were.
In other words, the same over-confidence that led people to believe that they could construct complex derivatives at a ridiculously breakneck pace without changing the nature of the financial system is now leading to models that just don't work. To me this is another symptom of a society that has become to easily convinced by "proof" through statistical regression. Statistics are useful and powerful, especially in the context of scientific and medical research. However, once one leaves the sciences and moves into areas that involve human beings and even more complex systems like international financial markets, the idea that we can build a model to predict the future becomes less believable.
At a very simple level, if it was so easy to predict what was going to happen tomorrow, no one would make money by investing because all of this information would be priced into today's security values. The so-called efficient market hypothesis reflects this basic intuition.
The problem is: predicting the future is ridiculously hard in one person's life, not to mention trying to do it for a group of people, a society of people, or the world markets as a whole. Add in some leverage, a multi-trillion dollar market of obscure and complex derivatives, some people falling on rough times underneath the whole thing, and you are left with a system fluctuating out of control.
What will slow down the volatility and bring us back down to reality? Hard to know, but one thing is clear: don't trust the model.
Thursday, March 6, 2008
Subscribe to:
Post Comments (Atom)
2 comments:
In some ways, predicting over groups of things, whether they are people or not, can make the predictor's job easier.
Consider the example of something which is completely random, such as a coin toss. It is easy enough to characterize a single coin toss as having a 0.5 probability of landing "heads". Any particular coin toss, though, will yield an outcome which is either "heads" or "tails", guaranteeing an error of 0.5. Predicting the aggregate outcome of 100 independent coin tosses, however, is much easier. There is a very very good chance that the proportion of "heads" outcomes will be near 0.5.
Naturally, things are more complicated than that in real life since the grouped outcomes in financial markets are not completely independent, and the proportion of the time which sees "heads" changes over time.
Does this mean that empirical modeling "does not work"? No. It means that:
1. Empirical modeling, like all human endeavors involves some imperfection. How much is "too much" depends on one's situation.
2. Empirical models need to be checked with some frequency. If they are not, then empirical models will lose track of reality. Given my observations of other modelers in the financial industry, I suspect this is part of the problem in today's markets.
-Will Dwinnell
Data Mining in MATLAB
Thanks Will. I think that is right. The problem for me arises in the context of derivatives of pools of assets. While I agree that generalization across large sets of observations allows for predictable outcomes in well-ordered distribution fields like coin flips, I think modeling something like a CDO of subprime RMBS is fundamentally more difficult. There is no "normal distribution" underlying the performance of borrowers taking out mortgages on houses for the first time at the peak of a real estate boom, not to mention a pool of a pool of tranches of claims on a wide variety of said mortgages. The problem is one of predicting how a novel set of financial instruments will perform based on a novel set of empirical circumstances on the backs of an unpredictable set of actors (unpredictable because there is no "history" to draw from).
It is not impossible, just very very hard. And the overconfidence with which people approached this problem is having dire consequences for all of us.
Post a Comment