«Urban Problems and sPatial methods VolUme 17, nUmber 1 • 2015 U.S. Department of Housing and Urban Development | Office of Policy Development and ...»
In this study, we seek to extend the research on racial and ethnic heterogeneity in loan modifications to include the modification terms borrowers receive. We also follow borrowers to observe differences in borrower repayment outcomes after modifications are made. This question is crucial, because, if modifications merely delay foreclosure, they may actually make lenders and borrowers (who are making payments under the modification) worse off. If redefault rates are systematically higher for borrowers of color who have received modifications, it would suggest that additional policies may be needed if the goal is to help these borrowers resolve their delinquency and sustain homeownership.
Methodologically, we present this analysis using a national sample of subprime and Alt-A mortgages originated at the peak of the subprime lending boom that are being serviced by a wide range of bank and nonbank servicers. This study is the only one to date to use merged loan performance data to study modifications at the national level through December 2012 (covering the peak period during which modifications were made). Although our sample still covers only a segment of the mortgage market, we believe that expanding the geographic and historical coverage of the analysis adds valuable new empirical evidence to our understanding of loan modifications and their effectiveness. We provide further details on the data in the next section.
Data and Methods For this analysis, we created a unique dataset that merges loan-level data on subprime home mortgages that are managed by Corporate Trust Services (CTS) with loan-level data on
borrowers from the HMDA. This merged dataset enables us to analyze whether differences in loan modification terms are influenced by the race and ethnicity of the borrower and to assess the extent to which these modifications are successful in preventing subsequent redefault.
CTS is a subsidiary service of Wells Fargo Bank, N.A. (hereafter, Wells Fargo) that provides investment vehicles administered by the bank. The CTS data cover privately securitized mortgages for which Wells Fargo serves as the trustee, including mortgages with different interest rate structures, purposes, property types, and lien statuses (Quercia and Ding, 2009; White, 2009b).7 The database includes loans originated as early as the 1980s, tracks performance until the loan is paid off or foreclosed upon, and includes more than 4 million individual loans.
Although Wells Fargo serves as the trustee for these investor pools, the data include loans from more than 100 servicers across the country, including large bank servicers such as Bank of America Corporation and J.P. Morgan Chase & Co. and nonbank servicers such as Ocwen Financial Corporation and Nationstar Mortgage Holdings, Inc. (Goodman and Lee, 2014). The top 20 servicers in our data cover bank and nonbank servicers, and they include 7 out of the 10 largest servicers in terms of market share in 2013 (Goodman and Lee, 2014). The largest servicer in the CTS data handles more than 13 million loans, while the smallest has approximately 70,000 loans in its portfolio. The data also reflect a broad range of servicer quality as ranked by Moody’s Corporation credit rating services, including servicers who scored an SQ1, which represents strong combined servicing ability and stability, and SQ4, which represents less-than-average servicing ability and stability (Moody’s Investor Service, 2014).
Each monthly loan record contains the borrower’s FICO credit score, loan-to-value (LTV) ratio at origination, the last 12 months of delinquency history, the property ZIP Code, the type of loan, and the original and current balance of the loan. Importantly for this study, the CTS data include a modification indicator, which represents all permanent loan modifications and equals one for every period after the loan is modified. The reports also have information about the loan balance, mortgage payment, and interest rate, before and after modification, which enables us to identify whether total mortgage debt, interest rate, or mortgage payments are changed for individual homeowners.
The CTS dataset, however, does not include any information on the borrower’s race or ethnicity. For this reason, following methods used by other researchers, we merge the CTS data with loan-level HMDA data (Ding, 2013; Ding et al., 2012). HMDA data provide information on the race and ethnicity of the borrower, his or her income, and the geographic location of the property securing the loan. To match the data, we sort CTS and HMDA loans into the census tracts of the purchased property using a geographic crosswalk file.8 Within each census tract, we match loan originations on the following variables: origination date, loan amount, lien These investor report files are available at https://www.ctslink.com.
One challenge in merging these data is relating U.S. Postal Service (USPS) ZIP Codes (the scale of the CTS data) to Census Bureau geographies (the scale of the HMDA data). We used the MABLE/Geocorr12: Geographic Correspondence Engine to allocate loans in ZIP Codes to corresponding census tracts. Details about the crosswalk are available at the Missouri Census Data Center, http://mcdc.missouri.edu/websas/geocorr12.html. For robustness, the authors also tested other available crosswalks (for example, the HUD/USPS ZIP crosswalk file), but the match rate did not improve.
Cityscape 171Collins, Reid, and Urban
status, and loan purpose.9 Only loans that provide for a direct match on these variables are included in the resulting sample. We were able to match 69.2 percent of the unique loans in the servicing record to HMDA applications. We compare the sample means of CTS matched loans against those that were not matched and find no significant differences in the average loan amount, the borrower’s FICO score, or whether the loan had an adjustable interest rate.10 In addition, we compare the demographic distribution of the CTS sample against the demographic distribution of subprime loans in HMDA and find that the proportions of non-Hispanic White, Black, Hispanic, and Asian borrowers are similar across the two datasets.
The sample used in this study consists of all first-lien mortgages for owner-occupied, singlefamily residences originated in 2004, 2005, and 2006 (as the market shifted in early 2007, nonconforming subprime loans were no longer being added to the CTS database); we limit the data to loans that were active but at least 60 days delinquent as of June 2009.11 We drop observations that went into bankruptcy during the panel and loans that were prepaid in the first period of observation. We also remove loans with an original balance of more than $1 million, because they are arguably a different subset of loans.12 We observe modifications and loan performance through December 2012. Data on modifications from the OCC show that the volume of modifications peaked in early 2010 and then declined throughout 2011 and 2012, meaning that our sample captures the period during which most modifications were made (OCC, 2014).
Because our interest in this study is to understand the relationship between modification types and redefault for different types of borrowers, we focus our analysis on 42,000 modified loans and consider only permanent, not trial, modifications.
Using cross-sectional linear probability models, we examine the performance of these loans from June 2009 through December 2012, at periods 6 and 12 months after modification, controlling for a wide range of loan, borrower, and housing market characteristics.13 We create additional variables to distinguish between different types of modifications. We construct two indicator variables, “interest rate decreased” and “loan balance decreased,” that equal 1 if the rate decreased or the balance decreased, respectively.14 To assess the extent of payment relief, we calculate the percentage change in the interest rate and monthly payment (“payment change”) before and after modification. We also create an indicator variable that assesses whether The matching procedure was completed while one of the authors was at the Federal Reserve Bank of San Francisco, providing access to the nonpublic HMDA data, which include origination date. CTS loans were matched to HMDA on site, and then all identifying HMDA variables (including loan number) were deleted from the matched record, resulting in a CTS data file with race/ethnicity and income attached to each loan record, but no ability to regenerate the origination date or link the CTS records to the public HMDA file.
Other studies that have used matching to merge HMDA data with loan performance records employ a probability matching technique so data on loans with multiple matches are not lost (Bocian et al., 2011). To date, no research has compared and contrasted these methods and the strengths and weaknesses of the different approaches.
We chose to focus on delinquent loans because borrowers who receive modifications without being delinquent may differ from distressed borrowers in important and distinct ways.
Dropping loans over $1 million results in a loss of about 0.5 percent of observations.
Using a cross-sectional model design versus a panel structure did not change our substantive findings, so we present the cross-sectional results to ease interpretation.
The data do not enable us to see whether the decline in the balance is related to principal forbearance or forgiveness.
a loan was “HAMP-eligible.” Although we cannot directly see which loans were modified under HAMP, this HAMP-eligible variable includes loans that (1) were modified after the launch of HAMP, (2) had an unpaid principal balance of less than $729,750, (3) had an interest rate reduction that did not bring the interest rate to less than 2 percent (the HAMP interest rate floor), and (4) were ARMs but converted to fixed-rate mortgages after modification (in other words, ARMs that remained ARMs after modification were excluded).
The control variables in our analysis include the borrower’s race and ethnicity, the borrower’s income, the borrower’s FICO score at origination, a no-documentation indicator, a prepaymentpenalty indicator, and the combined loan-to-value (CLTV) ratio. We coded the race and ethnicity variables in the HMDA data based on the primary applicant as “Black/African-American” (Black), “Hispanic/Latino” (Hispanic), “Asian/Hawaiian/Pacific Islander” (Asian),15 and “nonHispanic White.” The variables that capture the borrower’s race and ethnicity, income, and FICO score are measured at the time of origination; one significant limitation of these data and most data that report loan performance is the inability to assess how changes in the borrower’s income or FICO score over time influence either the probability of default or the success of modification.16 We take a log transformation of income in the models because borrower income is not normally distributed. To account for changes in the housing market, we use monthly data from Zillow at the ZIP Code level and calculate relative house price changes for each loan, enabling us to see the effect of a borrower’s equity position on modification terms or the likelihood of cure. All our models also include metropolitan statistical area (MSA)-level fixed effects to account for other market-level conditions that may influence modification terms or redefault.
Exhibit 1 presents summary statistics for the CTS sample of modified loans. The descriptive means for these variables are measured at origination with the exception of “HAMP eligibility,” which is determined at the time of modification. For purposes of this study, it is noteworthy that the sample is demographically diverse. Although the plurality of borrowers is nonHispanic White (48 percent), the sample also includes 22 percent Black borrowers, 28 percent Hispanic borrowers, and 4 percent Asian borrowers. Most loans (62 percent) listed a male borrower as the primary applicant. The average credit score of borrowers in the sample was 613, which is generally considered to be subprime (consistent with the fact that these are subprime and Alt-A loans that are bundled into private-label securities). The average applicant income at origination was $85,790. Focusing next on loan characteristics, we find that the average loan balance at origination was $241,265, with a mean LTV ratio of 83.48 percent. Most loans were ARMs (69 percent), with an average interest rate of 7.37 percent. Approximately one out of four modified loans were HAMP eligible, suggesting that a fair number of loans in our sample underwent proprietary modifications.
Also includes a small percentage of Native American and other races.
Although a borrower’s race and ethnicity should be static characteristics and not change between origination and modification, researchers who have analyzed the HAMP data have found that the race and ethnicity associated with a loan can change. For example, at modification it may be the coapplicant who interfaces with the lender. Lenders may have also made assignment errors, either at origination or at time of modification. In addition, researchers have noted the high degree of nonreporting of race/ethnicity data in HMDA (Wyly and Holloway, 2002).