ABS CDOs and the 2008 Correlation Failure

Take the BBB tranches of a hundred different mortgage-backed deals — the riskier, mezzanine slices that couldn't earn a AAA rating on their own — pool them together, and run them through the waterfall machinery again. That is an ABS CDO: a securitization of securitizations, whose entire economic case rested on one number, the correlation between the underlying pools' losses, and that number turned out to be catastrophically wrong.

Why the structure looked safe on paper

Recall from The Securitization Waterfall and Payment Priority that subordination lets a senior tranche survive losses well beyond the expected case, as long as losses across the pool are not too correlated. Rating agencies modeled the hundred underlying BBB tranches as having modest pairwise default correlation — loans in Ohio defaulting for reasons largely unrelated to loans in California defaulting. Under that assumption, even if individual BBB tranches had, say, a 20 percent chance of being wiped out, the chance that most of them were wiped out simultaneously was small, because their bad outcomes were assumed to be roughly independent. That let a CDO built from 100 percent BBB collateral still produce a senior tranche rated AAA, on the logic used in Default Correlation and the Asset Threshold Model: low correlation compresses the probability of a catastrophic joint outcome even when individual pieces are risky.

Worked example: what independence versus correlation actually implies

Simplify to ten underlying BBB tranches, each with a 20 percent individual probability of total loss, feeding a CDO where the senior tranche only loses money if at least seven of the ten default.

Under near-zero correlation (treat them as roughly independent for illustration), the probability of seven or more failures out of ten, each at 20 percent, computed from the binomial distribution, is under 0.1 percent — a genuinely rare, AAA-consistent tail event.

Now suppose the true correlation is high because all ten tranches are exposed to the same national housing cycle. In the limit of very high correlation, the ten outcomes move almost together: either housing holds up and close to zero default, or housing craters and close to all ten default. The probability of "seven or more defaulting" is no longer a rare tail event combined from many small, separate coin flips — it becomes close to the probability that the single shared factor (national home prices) falls hard, which historically was not a 0.1 percent event. The senior tranche's real risk was orders of magnitude higher than the independence-based model implied, for a reason that had nothing to do with any individual mortgage going bad and everything to do with a mis-specified correlation assumption.

Worked example: the domino into cash CDOs and synthetics

By late 2007, many ABS CDOs were themselves referenced by synthetic CDOs and CDS, multiplying the exposure to the same underlying correlation error many times over without adding any new mortgages — a single bad pool of loans could sit underneath billions of dollars of derivative exposure once resecuritized and referenced repeatedly. When national home prices fell roughly 30 percent peak to trough, mezzanine tranches across nearly every geography lost value together, senior ABS CDO tranches that had been rated AAA suffered losses of 80 percent or more in many deals, and the ratings that had assumed near-independence turned out to have priced a scenario that, once the shared factor moved, was nowhere near as rare as modeled.

Where you meet it in practice

Correlation assumptions sit underneath every multi-name credit product — CDOs, CLOs, basket CDS, index tranches — and the 2008 ABS CDO collapse remains the canonical case study for why those assumptions need to be stress-tested against a shared macro factor, not just estimated from a benign historical sample. Any quant pricing correlation-sensitive structured credit should be able to explain this failure from memory.