Joshua D. Coval, Jakob W. Jurek, and Erik Stafford write,

First, the securities’ credit ratings provided a downward biased view of their actual default risks, since they were based on the credit ratings agencies’ naïve extrapolation of the favorable economic conditions. Second, the yields failed to account for the extreme exposure of structured products to declines in aggregate economic conditions (i.e. systematic risk). The spuriously low yields on senior claims, in turn, allowed the holders of remaining claims to be overcompensated, incentivizing market participants to hold the “toxic” junior tranches. As a result of this mispricing, demand for structured claims of all seniorities grew explosively. The banks were eager to play along, collecting handsome fees for origination and structuring. Ultimately, the growing demand for the underlying collateral assets lead to an unprecedented reduction in the borrowing costs for homeowners and corporations alike, fueling the real estate bubble…

In March 2007, First Pacific Advisors discovered that Fitch used a model that assumed constantly appreciating home prices, ignoring the possibility that they could fall.

The authors then quote a transcript of a phone conversation between First Pacific and Fitch that makes the latter (one of the credit rating agencies) appear to be idiots. The authors continue,

It certainly appears that rating agencies did not fully appreciate the fragility of their estimates or the possible effects of modest errors in assumptions about default correlations and probabilities in their credit ratings. But this lack of understanding was apparently shared by the regulators that tied bank capital requirements to ratings, as well as by the investors who outsourced their due diligence to rating agencies without sufficient consideration of whether credit ratings meant the same thing for structured finance as they had for single-name securities.
[emphasis added]

So we have the suits vs. geeks divided, and the problem of regulatory capital requirements, both of which I have been emphasizing.

Thanks to Tyler Cowen for the pointer. I strongly recommend reading the whole paper.