 |
 |
Like the macro models, the dependent variable generally was the proportion of the pool that terminated each month. However, pool level models often utilized more sophisticated statistical techniques, such as logistic regression, that constrained the outcome to between 0 and 100%.
Given the continuing importance of the agency market, pool level models remain in heavy use throughout the mortgage market. Because the available information is relatively sparse (compared to the large number of attributes of the underlying mortgages in a pool), researchers have devoted significant resources to using this data as effectively as possible.2
4. Non-Agency Loan Level Models: 1990 and beyond
The next advance in data availability came about during the resolution of the S&L crisis in the early 1990’s. The Resolution Trust Corporation (RTC), in its role disposing of the assets of failed thrifts, set new standards of data disclosure, providing very detailed information on the collateral underlying its securities. As the non-agency market began to develop during this same period, non-agency issuers followed the RTC’s lead in terms of data availability, disclosing loan-level data in order to induce investors to hold their securities.
Hence, "loan level" analyses were the next mortgage modeling innovation resulting from changes in data availability. These loan level behavioral models attempted to explain why, despite good performance by pool level models, apparently identical pools still showed a significant level of unexplained variation in termination behavior.
2 See Davidson, A., et. al. (2003), Securitization: Structuring and Investment Analysis, John Wiley & Sons.
