Lewis (1992) provides an excellent history of consumer scoring for retail lending as a whole. Scoring models tend to use single-period classification (logit and variations) techniques to assess the probability of default for a loan. These techniques have been commercially developed as Fair Isaac's FICO score, Fannie Mae's Desktop Underwriter, etc. Historically, such "origination" or "behavior" scores predict default likelihoods over 12-18 months.

Dubitsky (2003) is a good example of so-called "roll rate" or "rating migration"-type models. These tools historically have looked at delinquency status and default, or alternatively bond ratings, as states in a multi-state time-constant Markov model. Banks and thrifts have applied roll-rate models to mortgages, credit cards, personal loans, etc. Ratings migration models are popular with secondary market participants in the MBS and ABS sectors, as well.

For a description of the life-of-loan loss or actuarial approach, Mays (2003) is an authoritative reference. This methodology generates estimates of lifetime defaults based on origination characteristics. Often, ratings agencies and large banks use this methodology, along with an assumed timing curve, to project loss behavior for pools or portfolios of loans over time.

The final method used to measure credit risk is applying option-based structural models. Kau et. al. (1985) and Capozza (1998) are good examples from the academic literature on mortgages. Such models look upon default as an American option on underlying value of the collateral. While this approach has proven successful in the corporate sector (i.e., Moody's KMV), consumer lending tends to be characterized by "inefficient exercise" of the default option.

Default Models for Mortgages
For consumer lending, in general, academic and practitioner research confirms that data-driven approaches outperform pure option-based models such as Kau (Merton-like option model). Most banks utilize internally developed models, rather than externally developed vendor models. In particular, "data-driven" approaches tend to dominate mortgage default modeling (i.e., S&P Levels).

However, the traditional consumer default methods outlined above all tend to have significant drawbacks. Scoring and life-of-loan loss models ignore timing of default and the effect of prepayment, as well as fail to account for covariates changing over time. Roll rate models do not incorporate predictive variables, and assume that "history repeats itself" with respect to macroeconomic risk factors. Finally, option-based structural models insist that default option is ruthlessly exercised.

Mortgages pose a particular challenge for default modelers. As with other types of consumer assets, >>>