(prepayment or foreclosure) or through December 31, 2012 if still performing. Loans originated in 2009
and after were not used in the study due to the lack of time to observe their performance.
For the FHA segment of the market, we used data from FHA’s single‐family data warehouse (SFDW).
Similar to the HLP data, the SFDW data include loan‐level information on the origination and
performance of mortgages insured by FHA. Compared to the HLP data, the FHA database contains a
less detailed loan history, but it includes a richer set of information on the borrowers. The file contains
the main underwriting variables, such as borrower credit score, initial loan balance, and sale price of the
house, along with indicators for the source of the down payment (borrower, seller, government
program, etc.). All FHA loans have full borrower documentation. We also have indicators for claim
termination (FHA pays insurance for a credit loss, usually as a result of a foreclosure), non‐claim
termination (a prepaymen t), and the date that a 90‐day delinquency episode starts, along with
information on how the delinquency spell was resolved. However, for the FHA data we do not have
detailed monthly paymen t history. Because FHA only began the routine collection of borrower credit
scores in 2004, we have limited the evaluation of FHA mortgages to loans originated from 2004 to 2008.
As with the GSE mortgages, FHA loan performance is observed through December 2012. Originations
from 2009 and 2010 were not used due to the lack of sufficient loan performance history.
For estimation purposes, we merged the GSE and FHA data with quarterly data from FHFA on MSA‐ and
state‐level house price indexes, data from Bureau of Labor Statistics (BLS) on state‐level unemployment
rates, data from the Federal Reserve on Treasury yields for 2‐ and 10‐year maturities, and data from
Freddie Mac on prevailing mortgage interest rates (Primary Mortgage Market Survey). For simulating
delinquency and foreclosure rates, we used forecasts of house prices, interest rates, and unemployment
rates from Moody’s Analytics. We primarily relied on Moody’s “base case” projection scenario, but for
sensitivity analysis we have also incorporated alternative Moody’s forecasts, namely a “pessimistic case”
that assumes a slow recovery, and falling house prices.
In terms of mortgage product type, our analysis focused on traditional 30‐year fixed‐rate mortgages, as
this is the most common product type and represents the largest share of the total origination volume
each year. We further limited the analysis universe to home purchase mortgages. In other words,
refinances were excluded. In addition, we excluded from the analysis “investor loans” and loans
classified as Alt‐A by the GSEs. These loan types were excluded because they had underwriting
requirements and performance history that were different from those of the owner‐occupied home
purchases. The underlying mortgage performance model in terms of the relevant variables and model
parameters would most likely differ significantly for these loan types. Therefore, excluding them from
the analysis would allow us to avoid confounding factors and arrive at a more precise estimate of the
effect of the down paym ent on loan outcomes.
5