The structure
A synthetic CDO holds no physical assets. Instead, it enters into CDS contracts on a reference portfolio of entities (bonds, MBS tranches). The CDO collects CDS premiums from the protection buyers and posts collateral. If reference entities default, the CDO pays the CDS claim (the protection buyer’s loss). The CDO then issues tranched notes to investors — again structured in senior/mezzanine/equity layers — whose returns come from the CDS premium income and whose losses arise from CDS payouts on reference entity defaults. This allowed banks to create exposures to mortgage credit risk without owning any mortgages, and to create the same exposure multiple times using the same reference mortgages.
The correlation problem
The valuation of CDO tranches — and specifically the "mezzanine" tranches that were rated investment grade but carried substantial credit risk — relied critically on the assumed correlation between defaults in the reference portfolio. If defaults were independent (low correlation), a diversified portfolio of mortgages was very safe — the law of large numbers meant only a small fraction would default simultaneously. Rating agencies used Gaussian copula models with low correlation assumptions. When the entire US housing market declined simultaneously (correlation of 1), the diversification argument collapsed entirely, and mezzanine tranches that models suggested were safe experienced near-total loss.
The Gaussian copula and David Li
David Li’s 2000 paper introducing the Gaussian copula for pricing CDO tranches was perhaps the most consequential and destructive academic paper in financial history. It provided a tractable way to price correlation between defaults — making CDO tranches priceable and therefore tradeable at scale. The model’s central parameter (correlation) was calibrated to historical CDS spread data, which proved catastrophically backward-looking. Li himself warned about the model’s limitations. The industry ignored the warnings and created trillions of dollars of structured products whose pricing rested on an oversimplified correlation assumption.
“The Gaussian copula gave the industry a number where it needed a number. That it was the wrong number only became clear when the underlying houses were underwater.”
What this means for you
The synthetic CDO episode is the canonical example of the danger of: mistaking a tractable model for reality; underestimating tail correlation in systemic events; and allowing complexity to obscure underlying risk from buyers, sellers, and regulators simultaneously. Modern structured product regulation (risk retention requirements — "skin in the game," mandatory stress testing, disclosure standards) directly targets these failures. For investors evaluating complex structured products today, the question is always: what are the correlation assumptions, what happens in a systemic scenario, and does the complexity serve a genuine purpose or obscure unfavourable risk characteristics?