I know, I know, I already used that title before...
Anyway, in the wake of Bill Gross claiming the ratings agencies were "fooled" into giving certain ratings on ABS CDO deals, I thought I'd give a little explanation of how the ratings process works in CDOs and my take on why it is breaking down now.
First let me say that just because bonds aren't performing well doesn't mean the rating agencies were wrong. For example, according to Moody's, historically there is a 2.5% chance a Aa-rated bond will default over 10-years. So the fact that any given Aa bond winds up defaulting doesn't mean that the rating was wrong to begin with. By extension, if the ratings agency believed a certain event was a low probability, just because that event comes to fruition, doesn't mean they weren't right about it being a low probability. If you are asked to predict the roll of two dice, what single number will you pick? 7 right? If you are asked to pick a 3 number range, you'd say 6-8 right? If the number rolled turns out to be 10, were you wrong in your prediction? If a priori, you correctly predict the odds, its still possible something unusual happens. That doesn't make your initial analysis wrong.
Now back to CDOs. The CDO rating process is very quantitative. Basically the process works like this. Each agency has their own Monte Carlo simulation software. They take a CDO portfolio, make certain assumptions about correlation of defaults, then use the Monte Carlo to figure out what the probability of suffering various levels of defaults.
The correlation of defaults is absolutely critical. If you have a portfolio of similar credits, the odds are good that several of them will default at once. Think of default probability as a two part function: a issue specific variable and a common variable among a group of issues. For example, if its GM and Ford, the common variable would be the health of the US auto market. We could extend this into a more complex function, where you might have GM, Ford, and Dollar General. GM and Ford are both impacted by the auto industry, but all three are impacted by general consumer spending.
The Monte Carlo simulation takes this correlation into account. The consequences of high correlation is that the results become very bimodal. There becomes a high probability of very low defaults or very high defaults. Keep this in mind for later.
Once we've established the expected level of defaults, the rating agencies then run a cash flow simulation of the CDO structure at various default levels. They used "stressed" default patterns, basically trying to bunch the defaults together to find the most stressful situation. They each have slightly different methods of interpreting the results. S&P and Fitch set a level of defaults at which a given tranche must survive (not default) to earn a given rating. Moody's sets an "expected loss" which means that they take the range of portfolio defaults, the level of loss at each default level, and do a probability weighting.
Where can this process break down? What's happening in the sub-prime market that is causing CDOs to perform so poorly? The obvious answer is "defaults are high, stupid!" Granted. Obviously high default levels can break a CDO. But if we were merely going through a high default period, I'd argue that doesn't amount to a "mistake" by the ratings agencies. They acknowledge in their models that there is a chance defaults come in high. Basically, they said its possible that you roll a 2, just not likely. Just because you actually do roll a 2, doesn't make the rating agencies wrong.
I argue that misinterpreting the correlation of defaults is the bigger problem. Two things happened in the sub-prime market to alter historical correlations. First, lending standards declined significantly. Second, interest rates got extremely low, then rose sharply. So questionable loans were being made, and since most sub-prime loans are 2/28, these weaker borrowers are (or will be soon) getting hit with huge payment shocks all at the same time. Therefore the correlation of defaults in this market is substantially higher than was assumed.
Thinking back to our multi-part default function, every bond in a portfolio of residential MBS has multiple common variables: interest rates, home prices, and employment trends. Even if we assume the rating agencies didn't know that sub-prime lending standards were weakening, anyone with a basic understanding of economics, real estate, or mortgage lending knew that loan performance was going to be, in part, a function of those three elements. While the inner workings of rating agencies Monte Carlo simulators are proprietary, I feel confident to say that the power of these common variables in creating high correlation was underestimated.
Commenter Chris among others has pointed out there should have been some caveat emptor here. He's 100% right. In fact, I talked to a Bear Stearns CDO trader about 2 years ago about ABS CDOs, and he said he liked the CLO market better specifically because he was worried about correlation of defaults in RMBS. Baa/BBB rated tranches of ABS CDOs have been trading cheaper than the same tranches in CLOs for several years. So its not like the correlation issue was a big secret.
I think this is another example of investors assuming that one Baa/BBB bond is the same as another just because the rating agency assigned the same rating. But the market traded Baa/BBB rated ABS CDO tranches far far cheaper than similarly rated CDOs with other collateral, or for that matter, plain vanilla corporate bonds. As an investor, you have to realize that stuff trades cheap for a reason. The market isn't stupid. Bill Gross may think the rating agencies are fools, but in this case, anyone who blindly follows them is the more foolish.