![Sample our Built Environment journals, sign in here to start your access, Latest two full volumes FREE to you for 14 days]({"type":"image" , "src" : "/sda/5926/TOC-Subject-Sample-2021-Built-Environment.jpg"})
Abstract
This paper examines change in the redefault rate of delinquent mortgage borrowers granted a loan modification during the 2008–2011 period, in comparison to similarly situated self-cured borrowers. We document a larger decline in the redefault rate of modified relative to self-cured loans, controlling for differences in their observable characteristics using a propensity score matching process. We attribute the relatively rapid improvement in performance of the modified loans in part to an increasing share of principal-reduction modifications and to increasingly generous interest rate reductions. Even after accounting for these and other factors, we still observe a larger decline in the redefault rate for modifications. This higher success rate for later modification cohorts likely reflects servicer “learning-by-doing.” Overall, our findings are supportive of public policy that encourages mortgage modification, which has proved to be successful in allowing distressed homeowners to keep their homes and reducing losses to lending institutions.
Notes
1 They attribute the relative sparsity of loan modifications before 2009 to a high likelihood that a loan would self-cure in the absence of a loan modification. In addition, they find evidence that servicers increased their reliance on loan modification as borrower self-cure rates declined between 2006 and 2010.
2 This study examines borrower response to the introduction of a mass modification program by Countrywide Financial Corporation at the end of 2008 pursuant to a legal settlement, whereby interest rate modifications were offered to all subprime mortgages that were at least 60 days past due.
3 Piskorski et al. (Citation2010) emphasize three main reasons that securitized loans may be serviced differently from portfolio loans. First, servicer financial incentives may differ between the investor and portfolio contexts. Second, private service agreements (PSAs) may legally bar servicers from offering some types of modifications. Third, coordinating an agreement among the bondholders on a modification program is more difficult than agreeing to a simple foreclosure.
4 Haughwout et al. (Citation2009) analyze a sample of privately securitized, nonprime mortgages that were modified between December 2005 and March 2009 and find a “distressingly high” 56% average redefault rate within one year of modification. Similarly, Quercia et al. (2009) examine a sample of privately securitized, mostly nonprime mortgages originated in 2005 and 2006 that were modified in the second quarter of 2008. By the end of 2008, 45% of the modified loans had returned to delinquent status.
5 In addition, Voicu et al. (2011) find that HAMP modifications were less likely to redefault compared to non-HAMP modifications. Schmeiser and Gross (Citation2015) find that term extension modifications that increase the amount of principal due are most likely to redefault.
6 Richter et al. (2010) also note decline in redefault rates on modifications after 2008 and cite larger payment reductions as the primary underlying factor.
7 Scharlemann and Shore (2015) apply a regression discontinuity approach to identify and mitigate the selection effect associated with modification type in evaluating the impact of principal reduction on redefault, in the context of HAMP modifications. On the margin, consistent with a selection effect, the study finds that principal-reduction modifications have a more beneficial impact on the redefault rate after accounting for the impact of reduced LTV.
8 The algorithm used by Equifax CRISM to merge the two data sets, which is proprietary, uses information common to both component databases, including mortgage origination amount, mortgage origination date, ZIP code of the property (mortgage servicing data), ZIP code of the borrower (credit report data), current balance on the mortgage (at the end of each quarter), and the borrower’s payment history.
9 The mortgage servicing data are collected from the ten largest U.S. mortgage servicers and account for approximately 75% of all mortgages in the U.S. as of year-end 2010 (Black Knight McDash estimate). However, the data are provided without a servicer key that would allow for distinguishing between servicers.
10 We apply a procedure to identify mortgage modification and/or self-cure that is adapted from Adelino et al. (Citation2013), as described in what follows.
11 Modifications (as identified by changes in loan terms) that do not return a loan to nondelinquent status appear to be relatively uncommon.
12 As explained in Adelino et al. (Citation2013), loss mitigation flags included in servicing data are not comprehensive, necessitating construction of modification indicators using reported changes in loan terms.
13 Specifically, a rate modification is indicated by an interest rate reduction that is at least one percentage point and at least 50 basis points in excess of the decline in the index rate since the previous reset (or since origination, in the absence of a prior reset). The index rate, or market rate to which the mortgage note rate is tied, is represented by the three-month Treasury bill rate for this calculation.
14 Since 2008 Black Knight McDash has provided detailed servicer reported information on modifications for an unspecified subset of servicers, comprising roughly three-fourths of the loans included in the servicing data set. Since this is not a random subset of the servicing data, we do not use this data set for the analysis in the paper. However, it does provide a useful benchmark for cross-validation of our modification detection algorithm, against which our algorithm correctly classifies 97% of loans transitioning from 90 or more days past due to current. This suggests that our algorithm accurately distinguishes self-cures and modifications while allowing us to use the full data set.
15 Refreshed LTV is calculated as the principal balance of the mortgage divided by the current property value. The latter is based on the original appraised value updated using the county-level house price index.
16 Propensity score matching, introduced by Rosenbaum and Rubin (1983) is commonly used to account for observable heterogeneity across “treated” and “nontreated” entities, based on the conditional probability of treatment given observable characteristics to reduce selection bias in treatment. In our study, the treated loans are modified loans and the nontreated loans are self-cured loans. For further discussion of the propensity score matching concepts and algorithms, see Guo et al. (2006).
17 Inherent in the logistic representation of the choice between modification and self-cure is the assumption that this choice is independent of the possible alternative of continuing through to the end of a foreclosure process. That assumption is consistent with, but does not require, simultaneous choice among the three alternatives. It is also consistent with a hierarchical decision context (nested logit), such that the decision whether to continue through foreclosure precedes the choice between self-cure and modification (with the latter being conditional on opting to not continue through foreclosure).
18 We group this way because servicers had broader discretion over modification decisions for loans in private label securities or held in portfolio, compared to FHA, VA, and GSE loans that were subject to guidelines of those agencies. Since there are relatively few portfolio loans, we pool them with privately securitized loans.
19 FICO® Score at origination is collected from the Black Knight McDash database. The mortgage servicing data also include a servicer-provided classification of the mortgage as subprime. We use that to identify a mortgage as subprime if the FICO® Score at origination is missing from the Black Knight McDash database.
20 The ARM categories distinguished are pay-option loans allowing negative amortization (option ARM), mortgages that allow monthly rate adjustment (variable ARM), mortgages with one-year initial fixed-rate periods, those with two- or three-year initial fixed-rate periods, and those with an initial fixed-rate period longer than three years (other ARM).
21 The specified ranges of refreshed LTV are > 80% and ≤ 90%, > 90% and ≤ 100%, and > 100%. The specified delinquency status ranges are ≥ 120 and < 180, ≥ 180 and < 270, ≥ 270 and < 360, and ≥ 360 days past due.
22 This method involves selecting pairs of modified and self-cured loans within the same characteristic segment, with absolute difference in propensity scores less than a specified caliper value (α). A larger caliper interval would increase the matching rate but would be more likely to result in less accurate matching.
23 In cases in which this limit would be exceeded (an individual self-cured loan matches to more than 25 modified loans), we prioritize use in one-to-one matches, such that no other matching self-cures are available for the individual modified loan. Beyond that, we prioritize use among fewer total matches for an individual modified loan, and beyond that, the selection is random.
24 The larger decline in redefault rates of modifications cannot be attributed to differential rates of prepayment between modifications and self-cures. The likelihood of loans prepaying after modification or self-cure is small—only 11% of loans in our sample prepay within three years of their modification or self-cure date. Modified loans in our sample exhibit slower prepayment speeds, and a smaller increase in prepayment speeds over time, compared to self-cured loans.
25 We limit local economic shocks to 18 months because roughly three-fourths of redefaults happen during this period and because inclusion of additional variables measuring economic conditions over the remainder of the observation period (18 to 36 months) was found to have no important impact on the estimation results.
26 Specification testing indicated that these exclusions did not materially affect the estimation results. Origination year dummy variables are typically included when modeling the initial transition to delinquency, in order to capture unobserved variation in underwriting standards. Criteria applied at the time of origination are of secondary importance for modeling post-default performance, and origination year dummies are more difficult to interpret in this context.
27 The potential impact of selection on sustainability of self-cures relative modifications is ambiguous. For example, borrowers who self-cure may be inherently more resilient due to unobservable improvements in their financial condition that enable cure. Alternatively, decisions made under duress to enable cure may cause them to be less resilient, such as opting for an unstable employment opportunity.
28 Also, close to one would suggest that any selection effect between impacting modified versus self-cured loans would be confined to the intercept term and vintage indicators in Equation (3); it would be neutral with respect to redefault behavior in relation to the explanatory variables in
29 For payment reductions of less than 10% (which are uncommon) the estimated coefficient of the linear spline is essentially zero, so that in the final specification we exclude this portion of the linear spline.
30 A simpler approach would have been to estimate a single equation using the combined sample of matched modified and self-cured loans. That equation would append to Equation (2) an indicator for modification versus self-cure and interactions of this indicator with the vintage dummy variables and with the variables capturing the financial benefit of modification. This simpler approach assumes that borrowers who receive a loan modification have the same redefault behavior as similarly situated, self-cured borrowers after accounting for the factors represented by these additional terms. The three-step approach is advantageous in that it enables us to assess the validity of that assumption by observing whether the estimated coefficient of
is close to one.
31 Among the local economic variables appended to Equation (4), the only statistically significant relationship is for house price appreciation, and only for the privately securitized or portfolio category, such that likelihood of redefault is more responsive to rising home values for loan modifications compared to self-cures.
32 We are more agnostic regarding the further improvement in performance indicated by the estimated vintage effect for 2011, during which the economy and housing prices had begun to stabilize. On the one hand, these later defaulters, who had survived the worst of the downturn, might be systematically different along unobservable dimensions than earlier cohorts. On the other hand, additional learning-by-doing might have occurred as the economic context evolved.
33 This interpretation is consistent with the smaller improvement in modifications relative to self-cures that we note for the FHA/VA sample. The FHA has enhanced documentation requirements and traditionally has conducted relatively strong monitoring of delinquent borrowers, limiting the scope for learning-by-doing. We are grateful to an anonymous referee for suggesting this interpretation.
34 One caveat is that modifications may appear increasingly successful because servicers are more frequently selecting for loan modification borrowers with greatest likelihood of self-curing. This explanation is counterintuitive, as it suggests that servicers became less efficient. In contrast, learning-by-doing is consistent with servicers seeking to minimize the losses associated with defaulted mortgages.
35 The FDIC guide emphasizes reducing monthly payments to sustainable levels (characterized as a 31% to 38% ratio of debt payment to income).
36 This program also offered a monetary incentive for servicers to provide modifications.
37 The results also are robust to excluding from the sample borrowers who have multiple first-lien mortgages.
38 We do, however, continue to observe a mitigating impact of principal forgiveness on the likelihood of redefault for the ARM segment in this category. The estimated equations for the FRM segment in this category resemble those for the full (ARM and FRM) sample. Inclusion of the product-type indicators have no material impact on the estimated coefficients of other variables in the redefault equations for GSE loans.
39 The specific Countrywide Mortgage program studied by Mayer et al. (Citation2014) likely encouraged strategic default because it became widely publicized that borrowers were eligible for modification as early as 60 days or more past due.
40 Specifically, we randomly order the loan modifications, and then proceed to select the nearest neighbor self-cure for each modification in that order. If a self-cured loan has previously been selected as a nearest neighbor, it cannot be reused, and the closest remaining nearest neighbor is selected. We repeat this random selection process 25 times, and then weight each observation in the resulting sample by 1/25th.
41 The postmatch count of self-cured loans drops from 416,938 to 126,762 in the Agency sample and from 205,878 to 131,364 in the private securitized and portfolio sample. The decline in the postmatch number of modified loans is negligible.