Hero of the day, CPDO edition

By Felix Salmon
November 5, 2012
decision in a recent court case which was brought against ABN Amro and Standard & Poors.

" data-share-img="" data-share="twitter,facebook,linkedin,reddit,google" data-share-count="true">

I’d never heard of Australian federal judge Jayne Jagot before today, but she’s my new favorite jurist, thanks to her decision in a recent court case which was brought against ABN Amro and Standard & Poors.

The coverage of the decision (Quartz, FT, WSJ, Bloomberg, Reuters) concentrates, as it should, on the hugely important precedent being set here: that a ratings agency — in this case, S&P — is being found liable for losses that an investor suffered after trusting that agency.

S&P is appealing the decision, which runs to an astonishing 635,500 words, or almost 1,500 pages: it’s literally longer than War and Peace. At this point, it’s fair to assume that Jagot is one of the world’s foremost experts on structuring and rating CPDOs — crazy derivative instruments which had a brief moment of glory at the end of 2006 before imploding spectacularly during the financial crisis. And helpfully, her decision begins with a 56-paragraph summary of her findings, which lays out exactly how culpable and incompetent S&P really was.

I haven’t read the whole decision, of course, but I have read the summary — and as someone who’s been writing about CPDOs for six years now, I can attest that Jagot’s summary is the single best one-stop shop for understanding what happened with these things. She concedes at the beginning that “the explanation unavoidably refers to complex concepts which are likely to be unfamiliar to those without specialist expertise in structured finance” — this is not easy reading, and if you don’t enjoy nerding out with such concepts, I would never recommend it to you. But if you are reasonably familiar with credit default swaps, Monte Carlo simulations, volatility assumptions, and the like, then I highly recommend you read Jagot’s summary: it’s utterly eye-opening.

The case at heart is a simple one: 12 local councils in Australia bought a bunch of CPDOs, and they only did so because S&P had given those instruments a triple-A rating. S&P, in turn, should never have given the CPDOs that triple-A rating. So it’s S&P’s fault that the councils lost so much money — jointly with ABN Amro, which structured the things.

How does Jagot come to the conclusion that “a reasonably competent ratings agency” would never have given the CPDOs a triple-A rating? Simple: S&P used utterly bonkers assumptions in order to come to its conclusion.

The way that structured finance worked, pre-crisis, was that banks would come up with ever more ingenious ways of structuring products which just qualified for a triple-A rating; they would then try to persuade the ratings agencies that their reasoning was kosher. This case was no different: ABN Amro put together a model, plugged a bunch of numbers into the model, and — presto — the model spat out a default rate which was so low as to justify a triple-A rating. It then showed the model to S&P.

Given such a new concept and product, S&P really should have developed its own model from scratch. It didn’t, however — it used ABN Amro’s model instead. That was bad enough. But worse — much worse — was that S&P didn’t even come up with its own assumptions to plug into the model: it used ABN Amro’s assumptions! Even though those assumptions were unjustifiable.

Bloomberg’s Mark Gilbert explained the concept of model risk at the time:

Derivatives-based trading strategies rely on computer simulations to test their performance under different market scenarios. These simulations typically expect the future to be like the past; the collapse of Long-Term Capital Management LP in 1998 is proof of the danger of using inductive reasoning to extrapolate general laws from particular observable instances.

In my article about the Gaussian copula function, I explained how this played out in practice. CDOs were priced based on the assumption that the future would be like the past, and then when the future turned out to be very different, they blew up.

But as it turned out, if S&P had assumed that the future would be like the past — if it plugged the market realities of the time into ABN Amro’s model — the CPDOs would never have managed to get a triple-A rating. This is where Jagot’s summary is invaluable: it shows that in order to generate a triple-A credit rating, S&P basically had to stick its head in the sand and ignore market realities. Never mind the future, S&P couldn’t even model the present!

Just look at some of the assumptions which S&P made in order to be able to get the longed-for triple-A credit rating. The most glaring is the starting spread — the money investors put into CPDOs was then invested in a synthetic basket of corporate debt. The lower the yield on that debt, the less money the CPDO could make. And at the end of 2006, spreads on corporate debt were very low. When ABN Amro issued a CPDO called Rembrandt 2006-3, the spread in question was just 29bp. But S&P, in rating Rembrandt 2006-3, used a starting spread of 36bp.

The difference might seem small. But S&P knew that with a starting spread of 32bp, where ABN Amro had hedged the deal, the model could not spit out a triple-A rating. And ABN Amro knew that the triple-A rating could not be justified if the starting spread was lower than 35bp. Yet somehow Rembrandt 2006-3 managed to get a triple-A rating with the lower starting spread. How? S&P simply didn’t model it. Instead, they just used an earlier rating, from a deal which never got issued and which was modeled using the 36bp starting spread.

That wasn’t the only place where S&P made unjustifiable assumptions. For instance, check out the number they plugged in for volatility: ABN Amro assumed that the index had volatility of 15%. But it didn’t. In reality, the volatility of the index was somewhere between 28% and 29%. And S&P had recently rated a different product, called LSS, using a volatility assumption, for exactly the same index, of 35%. The problem was that with a volatility assumption of 35%, the CPDO would never rate triple-A. So S&P just discarded that figure entirely, and used ABN Amro’s instead.

Here’s Jagot:

S&P believed ABN Amro’s assertions that the actual average volatility of the Globoxx since inception was 15%. S&P did not calculate the volatility for itself although it could easily have done so and, in my view, was required to do so as a reasonably competent ratings agency…

This assumption as to volatility was unreasonably and unjustifiably low. It did not represent either a reasonably anticipated or expected (that is, non-stressed) input or market condition or an exceptional but plausible (that is, stressed) input or market condition. I am satisfied that but for this error about volatility the CPDO could not have been rated AAA by S&P on any rational or reasonable basis.

So S&P was seemingly incapable of looking at the market to see (a) the spread on corporate debt, or (b) the volatility of the index. Which probably explains another one of their whopping great errors: their assumption for where the index would be over the ten-year lifespan of the bonds. This was a number known as the long-term average spread, or LTAS, and once again, the higher the yield on corporate bonds, the better the CPDO would perform. Remember that when the CPDO was issued, the spread in question was around 30bp. So what did S&P assume it would be in future?

Although there was no rational or reasonable basis for doing so other than the fact that one approach enabled the CPDO to satisfy the AAA rating quantile and the other approach did not, S&P decided that its base case for modelling and rating the CPDO should be an LTAS of 40 bps for one year and 80 bps for nine years… Although the split between 40 bps and 80 bps was itself arbitrary, the most salient point is that the modelling showed that if the assumed LTAS of 40 bps was extended to two years then, all other assumptions being the same, the CPDO again did not meet the AAA rating quantile. S&P then drew a further arbitrary, irrational and unreasonable distinction between the lower LTAS of 40 bps lasting for one year as opposed to two years, the result of which was that on S&P’s approach the CPDO achieved the AAA rating quantile default rate of 0.728% (which it did not if the lower LTAS of 40 bps continued for two years).

There’s really no way of reading what S&P did, here, except that it simply massaged the assumptions it was using until it managed to find something which was consistent with the triple-A rating it wanted. When spreads are at 30bp, what makes you think they’ll average 40bp over one year and then 80bp over nine years? Especially when the index as a whole has never averaged anything like 80bp? It’s simply not a reasonable assumption, and the fact that S&P made it just goes to show how the agency was acting for its paymasters — ABN Amro — and was not putting out reliable ratings at all.

There’s more, too. S&P also plugged into its model an assumption of 7% for something called roll-down benefit, or RDB. Every six months, the CPDO would exit its existing positions, and buy new positions in the index maturing six months later. In general, bonds which mature later have higher yields, so S&P assumed that on average, the new index would yield 7bp more than the old index. And that was an utterly crucial assumption. Never mind the triple-A rating: without the RDB, the CPDO couldn’t even get an investment-grade rating. It wouldn’t even be triple-B.

That was a reasonable assumption, at the time — in terms of the general slope of the yield curve. In general, corporate bonds do tend to yield about 7bp more if they mature 6 months later. But in assuming the 7bp figure, S&P completely ignored the fact that it was comparing apples and oranges: the components of the new index would not always be the same as the components of the old index. After all, the index was an index of investment-grade corporate debt, so if a company lost its investment-grade credit rating, it wouldn’t be included in the index any more.

You’d think that a ratings agency, of all institutions, would be alive to the risk of ratings downgrades. But, it turns out, not so much. ABN Amro, in its model , simply didn’t include what’s known as “ratings migration” — and S&P, similarly, completely ignored it.

The result, in reality, was devastating. Because companies could borrow at such low rates, they were particularly vulnerable to being taken over by private-equity firms which could load them up with cheap debt, devastating their credit ratings. And that’s exactly what happened. A whole series of investment-grade companies, like Alliance Boots, Alltel, and Boston Scientific, got levered up by their new private-equity owners, and lost their investment-grade credit ratings.

When those companies dropped out of the index, the companies which were left had significantly lower yields than the index as a whole did before. As a result, rather than yielding 7bp more than the old index, the new index actually yielded 15bp less. That just devastated the CPDOs, and ultimately led them to default.

Put it all together, and you get a very shocking view of S&P. Here’s the list:

  • S&P used the wrong model input for starting spread.
  • S&P used the wrong model input for volatilty.
  • S&P used the wrong model input for average spread.
  • S&P completely ignored ratings migration.

If S&P had just got any one of these things right, the CPDO would never have gotten that triple-A rating. If it had got them all right, the CPDO would almost certainly not even have been investment grade, let alone triple-A.

S&P was not doing its job, and as a result a bunch of Australian municipalities lost a great deal of money. Jagot has found S&P liable, as she should. Good for her.

19 comments

Comments are closed.