Stanford’s Darrell
Duffie is giving a series
of lectures at Princeton on capital immobility, which will be turned into
a book to be published by Princeton University Press. The Press invited me to
dinner with Duffie last night – delicious, thanks! – and we
got to talking about ratings agencies. Duffie is a member of Moody’s Academic
Research Committee, and his take on the ratings
crisis is basically that there’s nothing there that better modelling and
clearer thinking can’t fix.
What’s the problem, he says, with Moody’s rating structured products like the
notorious CPDOs?
Not that Moody’s got the rating wrong, so much as that the product
should never have been considered a ratable instrument in the first place. Is
there a conflict, he asks, inherent in the fact that issuers, and not investors,
pay the ratings agencies? No: the ratings agencies’ credibility is much more
important to them than any benefit they might get from boosting the ratings
they give, and in any case it’s relative ratings which matter, not
absolute ones.
I wasn’t convinced. For one thing, the ratings agencies clearly and consistently
said that ratings are horizontally comparable, as it were: that a double-A rating
on a sovereign means it has the same default risk as a double-A rated municipal
bond, or corporate bond, or structured product. And it turns out that’s not
the case: structured products, especially, default much more than identically-rated
munis, and they always have – this is not news. It certainly seems as
though the ratings agencies, which made enormous profits from rating structured
products, were "nicer" to those products than they were to other issuers.
But my real reason for skepticism is that I’m in the middle of being extremely
impressed by another Princeton University Press author, Riccardo Rebonato. In
his new
book, Rebonato compellingly skewers the "frequentist" approach
to probabilities employed not only by Moody’s and the other ratings agencies
but more generally across the whole world of finance. Looking backwards at what
happened in the past gives you lots of data, which can be chopped up and examined
in any number of different ways, gives a false sense that future probabilities
can be scientifically determined to three or four significant figures.
But that whole approach is based on the idea that market moves are like coin
flips, or balls in an urn: that the markets might move, but that their
underlying structure never really changes. In fact, the defaults rates that
matter are the ones in the future, not the ones in the past. And to get a grip
on those you need to understand what Rebonato calls "subjective probability"
– which, although it might not have quite the mathematical rigor of frequentist
probability, can actually be much more useful and accurate.
Duffie does understand this. Moody’s had a habit, he says, of tweaking incredibly
complicated models until they matched the past data, and then deciding that
because the models matched the past data, they must give a good idea of what
will happen in the future. Obviously, that’s never going to work very well:
it’s the financial equivalent of shooting an arrow at a barn door, painting
a target around it, and claiming astonishing marksmanship. But I think Duffie
and I differ on the question of whether better models can fix this problem.
He thinks they can; I think they can’t, certainly so long as Bayesian statistics
remains a backwater for Wall Street’s quants.