After reading Felix Salmon’s
post about the Australian judge who stood up to ratings giant Standard & Poor’s (which slapped a AAA grade on some securitizations that all too quickly crapped out), I got curious.
How badly did S&P and ABN Amro (the creator of the investment) behave? What exactly happened with these odd products (called “constant proportion debt obligations,” or CPDOs for short)?
Felix made it sound pretty bad. After digging around some, I’m convinced now that
it was even worse.
What follows are some damning things about the case that Felix -- and sometimes even the judge herself, in her
decision -- didn’t touch upon.
Judge Jayne Jagot found S&P liable for losses suffered by investors that trusted their shoddy ratings. That was a big blow to S&P, which in the U.S. is used to hiding behind the skirt of the First Amendment, protesting that its ratings are only “opinions,” as if the company’s analysts were no more than film critics for a free alternative weekly.
Now CPDOs are awfully complex (Jagot approvingly quotes a description of them as “grotesquely complicated”). The way you rate one of these Rube Goldberg-ian securitizations is through a model, into which you feed a bunch of variables, then observe how the original investment fares in a series of random trial runs.
Investors in the ill-fated “Rembrandt” 10-year CPDOs were basically selling insurance (in that credit default swap kind of way) on the members of a couple of CDS indexes, known together as the Globoxx. The insurance protects against default on the debt of the companies in the indexes (each has 125 members, I think). That put the investors in a “long” position on the underlying bonds. So they benefit when the debt becomes safer and doesn’t default.
Sort of.
PARADOX OF A CPDO
Because here’s the problem with a CPDO.
It needs higher credit spreads (which indicate higher risk, which leads to fatter “insurance premiums”), while at the same time it paradoxically needs lower credit spreads (as they indicate a lower risk of default, and thus a smaller chance of taking losses).
How the heck do you square that circle? Well, ABN Amro started by jamming bad, incomplete data into the model. And the bank succeeded because S&P stunningly accepted whatever ABN Amro spoon-fed it, like a lapdog with its eye solely on the milk bone (that would be the rating fee it was promised).
To see the problems with one bit of data in particular, you need to take a closer look at the plumbing of these things:
The Rembrandt CPDOs were actually selling insurance in a rolling series of contracts. They started out by issuing protection through 5 1/4 year default swaps on the Globoxx basket. Six months later, they exited that position and sold a fresh 5 1/4 year contract on the new Globoxx (every six months, companies that no longer meet the investment-grade criteria are replaced in the indexes). By that time, of course, the original 5 1/4 year contract has become a 4 3/4 year contract. The process of changing over from the old index to the new is called “the roll.” The old “off the run” contract gives way to the new “on the run” one.
ABN Amro won a coveted AAA rating on the CPDO partly because of its wrongheaded exploitation of “the roll.”
ROLLING DOWN THE RIVER
See, credit risk curves for a company (of investment-grade quality anyway) tend to slope slightly upwards. Not by a whole lot -- but enough. Generally the longer you commit to insuring debt, the more you want to receive each year, because the more scary unknowns may be lurking way out in the future.
Why did that matter for a Rembrandt CPDO investor? Simple: When the CPDO does “the roll,” it buys a 4.75 year Globoxx contract (in CDS land, if you’ve sold a contract, you can later buy the exact same contract to cancel out your earlier position) and starts selling a 5.25 year default swap. So, assuming nothing else has changed, you reap a neat little benefit from the six-month difference in term. You buy a 4.75 year contract at x and sell a 5.25 year at x plus a little something.
ABN Amro calculated that “little something” at 7 basis points, or 0.07%. Seems like a puny, negligible number. Yet it was anything but. The CPDO could be as much as 15 times leveraged (and 15 times 7 is 105, or more than a full percentage point).
Without that 7 basis points of roll-down benefit, occurring every six months, this AAA investment would have received a junk rating.
Yet this 7 points of roll-down benefit was a grossly flawed number -- and ABN Amro’s own model inputs showed as much!
NOW IT'S IG, NOW IT'S NOT
It was never adjusted for the danger of “ratings migration,” which Jagot describes as “the phenomenon ... by which the rating of a reference entity might decrease between rolls without default.” That’s an especially insidious problem with an investment-grade index that’s changing its composition every six months.
Here’s an example of how ratings migration could sting the CPDO: It sells a Globoxx contract at 60 basis points. The economy lurches south, and some of the companies that belong to the index (what the judge refers to as “reference entities”) slip a few notches, into junk territory. After six months, let’s say the Globoxx has climbed to 80, indicating higher levels of risk. The “junk” members are replaced by new investment-grade companies, so let's suppose the new Globoxx contract is at 60 again. Got that?
Here’s the math: +7 basis points of roll-down benefit, -20 points of ratings migration. That equals a net loss of 13 basis points.
What’s worse is ABN Amro practically assures this negative ratings migration will occur -- then apparently never adjusts the model for it!
We know that because ABN Amro had to feed another bit of data into the model called “long-term average spread” (LTAS), which we’ll call “average spread,” to keep things simple. This starts at 40 in year one, then balloons to 80 in year two. In other words: ABN Amro itself expected the average level of the Globoxx to jump 40 basis points. So it’s highly likely, if that’s true, that at least a few companies were going to migrate right out of the index. This alone should have knocked out a chunk of that roll-down benefit.
NEVER SHALL THE VARIABLES MEET
But it didn’t. One reason: a glaring weakness of the model was that various key parts apparently didn’t “talk to each other” -- which made it ripe for exploitation.
Never was this weakness more apparent than with the interaction between the model’s assumptions for average spreads and volatility and default probabilities. Quite simply, there didn't seem to be any! Each variable lived in its own walled-off silo, informing the model without disturbing variables anywhere else. That’s beyond absurd.
Here’s one illustration why.
Initially ABN Amro made certain observations about the Globoxx: there was a certain expected default rate for companies in the indexes, the historical volatility was 15% (
wrong, by the way -- it was actually almost twice that, and S&P never bothered to check either), the average spread was 40 to start out with (inexplicably, this
simple, easily confirmable fact was wrong too -- by 25 percent!).
But, in the real world, these variables don’t live in separate silos. Actually, they’re more like joined at the hip. So when ABN Amro estimated that average spread would increase from 40 to 80 after one year -- a big jump -- to be thorough and honest and reflect reality -- it should also have adjusted volatility higher and the default rate higher as well.
The CPDO should have performed worse, not better, when average spread increased. But the investment actually did better when credit risk doubled!
PLAY THE GAME
If you’re getting the impression that ABN Amro cleverly worked the S&P model like deaf, dumb, and blind Tommy would a cheap pinball machine, you’re not the only one.
From Judge Jagot’s summary: “At least one person within S&P considered that ABN Amro, whether intentionally or not, had effectively ‘gamed’ the model.” The bank would have been in a good position to figure out how to game the model, too, because two former employees of S&P were on its payroll.
One way to game a model for a CPDO, as this Federal Reserve
working paper by Michael B. Gordy and Sren Willemann shows:
If spreads widen early in the life of the CPDO and then hold steady, the higher carry on future index positions can outweigh the initial loss of NAV [net asset value].
And what scenario for spreads did ABN Amro predict? 40 basis points for the first year, then widening out to 80 basis points in the second year, and holding steady for the next nine years! Sheer coincidence?
DEEPLY FLAWED MODEL
So ABN Amro crammed bad data into the model, exploited weaknesses of the model ... but wait, there’s more. The CPDO models being used at the time, it turns out, were intrinsically flawed anyway. They assigned extremely low probabilities to credit spreads blowing out to the levels seen in late 2007. Now, this isn’t late 2008 we’re talking -- only 2007.
The Fed paper notes:
The spread levels realized in late 2007 are qualitatively comparable to the levels seen in 2002, so ought not to have been taken as extreme events.
(By the way, I haven’t even really explored the question of whether any investment using 15 times leverage, and thus susceptible to the smallest of price movements, should ever be rated AAA. Also the CPDO used a “doubling down” gambler’s strategy: whenever credit spreads moved the wrong way, leverage was increased, to a maximum of 15 times. Interestingly, the CPDO began its life at 15 times leverage, underscoring the absurdity of this strategy. How could it double down when it’s already at its limit? This is like Dumb and Dumber go to the casino with $1,000 in pocket, intending to use the “double down” approach, then put the whole thousand on the very first bet.)
THE HOUSE OF CARDS
What ABN Amro did -- and what S&P contributed to -- seems pretty much like fraud to me. But here’s the thing: at the heart of most (if not all) securitizations, I bet you’ll find similar kinds of “fraud” -- negligent and poor modeling, wrong or unrealistic data inputs, massaging of data to barely achieve desired ratings. It may not occur to this degree, but it’ll be there.
That’s because, as I’ve said before in looking at CLOs, the complexity of securitization disguises a simple truth: amid all the fee extraction and other costs, there simply isn’t enough yield available in the underlying assets -- whether they’re loans, or mortgages, or credit default swap contracts -- to justify all the high ratings, after all the slicing and dicing. This is the mathematical fraud at the heart of securitizations (liquidity and diversification arguments notwithstanding).
Someday I think someone from the world of academic finance will take a deep look at this issue, and expose securitization’s house of cards. That person could do worse than starting with such egregious instruments as these Australian CPDOs and their clearly flagrant abuses of models and ratings.