Impact of Information Asymmetry and Servicer Incentives on Foreclosure of Securitized Mortgages

Transcription

1 Impact of Information Asymmetry and Servicer Incentives on Foreclosure of Securitized Mortgages Dimuthu Ratnadiwakara March 2016 ABSTRACT In this paper I examine how servicer characteristics affect foreclosure rates of delinquent securitized mortgages. Using a sample of mortgages originated by IndyMac Bank and Countrywide Home Loan Servicing that underwent a change in servicer during the financial crisis period, I first show using a difference-in-difference methodology that the servicer has a significant effect on foreclosure rates. Consistent with the idea that information asymmetry affect foreclosure rates, I find a significantly lower foreclosure rate among mortgages where the originator also acts as servicer. Finally, foreclosure rates are ceteris paribus higher for servicers that experienced high loan growth in the past, which suggests that servicer capacity constraints affected foreclosure rates. Keywords: Securitization, Foreclosure, Information Asymmetry, Servicer JEL Codes: G21, G28, G33 University of Houston, C.T. Bauer College of Business, Houston, TX , phone: , dnratnadiwakara@uh.edu. For helpful comments and suggestions, I thank Kevin Roshak, Rauli Susmel and Vijay Yerramilli.

2 I. Introduction It is widely acknowledged that excessive foreclosures of mortgages are costly to almost all market participants (Campbell and Giglio (2009), Mian, Su, and Trebbi (2008) and Posner and Zingales (2009)). At the height of the housing crisis, it was predicted that one in nine residential borrowers would go into foreclosure by , and academics, as well as policymakers, advocated reforms to reduce the foreclosure rates of residential mortgages. Programs such as Home Affordable Modification Program (HAMP) were introduced in early 2009 in an attempt reduce the foreclosures. However, these programs did not have the desired effect and foreclosure rates remained high, particularly for private-label securitized mortgages (Figure 1). Private-label securitized mortgages accounted for more than 50% of the foreclosures during the housing crisis 2, and empirical evidence shows that private-label securitized mortgages are more likely to be foreclosed than similar loans held on banks balance sheets (Piskorski, Seru, and Vig (2010),Agarwal, Amromin, Ben-David, Chomsisengphet, and Evanoff (2011a) and Kruger (2014)). Even though securitization is often blamed for the high foreclosure rates seen in private-label securitized residential mortgages, there is no agreement on the channel through which this effect takes place. Servicers muted response to HAMP, which provided financial incentives to servicers to encourage modifications, suggests that current understanding of the servicer incentives related to foreclosure decision is not complete. It is essential to understand the drivers of high foreclosure rates of securitized mortgages to successfully implement the necessary reforms to reduce the foreclosure rates. In this paper, I use a sample of private-label securitized mortgages to explore the factors that may have contributed to high foreclosure rates. There are few reasons why securitized loans are more likely to be foreclosed. First, it may be the case that most of the securitized loans are of lower quality (Keys, Mukherjee, Seru, and Vig (2010)) and the foreclosure rate is inversely related to the loan quality. Second, the modification restrictions may have resulted in higher foreclosure rates (Levitin (2009)). Next, servicers may 1 Credit Suisse Fixed Income Research, Foreclosure Update: Over 8 Million Foreclosures Expected (Dec.4, 2008) 2 Total RMBS outstanding in 2009 was larger than 20% of the total U.S. bond market (Inside Mortgage Finance, 2010) and this consists of agency and private-label RMBS. Private-label RMBS data accounted for most of the foreclosures during the recent housing crisis. Further agency RMBS securities do not bear any credit risk, as was evident during the financial crisis. Agency RMBS are backed by the government or a government-sponsored enterprise (GSE) (Ginnie Mae, Fannie Mae, or Freddie Mac) and therefore from the perspective of the investors the issue of high foreclosures is not a concern. Private-label RMBS on the other hand expose investors directly to credit risk and level of foreclosures can have a significant impact on the returns of investors. 2

3 increase the foreclosure rate to discourage strategic defaulters who try to gain from modifications by exploiting the information asymmetries between the borrowers and the servicers as predicted by theoretical models such as Riddiough and Wyatt (1994) and Wang, Young, and Zhou (2002). Finally, misalignment of the servicer incentives may have contributed to high foreclosure rate in securitized mortgages. While some (e.g.: Thompson (2011)) have argued that servicers deliberately increase the foreclosures and extend the delinquent cycles to extract as much income from the delinquent loans as fees, others such as Bernake (2008) and Maturana (2014) have argued that capacity constraints of the servicers may have lead to higher foreclosure rates. The argument for the latter is that default loan administration is highly labor intensive and during the housing boom servicers did not have incentives to add additional staff. This may have inhibited their ability to handle growing number of defaults. Capacity constrained servicers may thus have had the strongest incentive to use mortgage servicing software to foreclose the defaulted loans without careful analysis. Since empirical studies have shown foreclosures are higher for securitized privatelabel mortgages even after accounting for borrower quality(piskorski et al. (2010),Agarwal et al. (2011a) and Kruger (2014)) and modification restrictions (Kruger (2014)) my focus in this paper is on the last two factors, both of which are related to the servicer characteristics such as information asymmetry and capacity constraints, and examine the impact of servicer on the foreclosure rate. More specifically I examine whether the information asymmetry and capacity constraints of the servicers contribute to the high foreclosure rate of securitized private-label mortgages. Identifying the impact of servicer on foreclosure rate is difficult. The securitized loans serviced by different servicers may differ on unobservable loan characteristics 3, which means that comparing foreclosure rates across different servicers can be misleading due to unobservable differences in loan characteristics. To alleviate this concern I exploit the servicer changes of mortgages initially serviced by IndyMac Bank and Countrywide Financial Corporation. There are few notable instances where the initial servicer of a pool of securitized mortgages changed and these are two of the most notable servicer changes during the housing crisis. For each case, I employ a matched difference-in-difference estimate on the basis of the predicted probability of mortgage being serviced by these servicers using the propensity score method. 3 For example some servicers specialize in certain types of mortgages such as Alt-A mortgages for which credit decision is mostly based on the borrower s credit score 3

4 Testing the information asymmetry explanation entails testing two hypotheses. The first hypothesis is that borrowers are more likely to strategically default when the probability of their loan being modified is high. A simple regression of default on modification rate may be biased since modification decision is endogenous. To address this concern I use the case of IndyMac Bank, where FDIC implemented a systematic modification program which resulted in an exogenous shock to modification rate, in a matched difference-in-difference setup. The second hypothesis is that when information asymmetry is more serious, the servicer does not have the necessary information to identify strategic defaulters. Therefore, the servicers increase foreclosure rate when information asymmetry is more serious to discourage strategic defaulters. However, empirical investigation of this hypothesis is challenging since the level of information asymmetry is not observable. I argue that when the originator continues to service the mortgages it originated, the information asymmetry is less serious and they are better able to identify strategic defaulters. I use a dummy variable indicating whether the originator is same as the servicer in my regression to test this hypothesis. I use the 1 year lagged annual loan growth for each servicer as the main measure of how capacity constrained servicer s are. Descriptive evidence shows that for most servicers the number of loans serviced increased by more than 100% annually before the housing crisis. I argue that servicers experienced a higher loan growth are more likely to be capacity constrained. I also use servicer size, measured in terms of the number of loans serviced by each servicer, as a proxy for servicer capacity. When the servicer is larger they have stronger incentives to automate due to economies of scale and, therefore, are more likely to be capacity constrained when handling defaulted mortgages leading to more foreclosures. My analysis is conducted on monthly loan performance data from Bloomberg for more than 1 million non-prime mortgages securitized between 2000 and 2008 and my main findings are as follows. I find significant cross-sectional variation in foreclosure rate across servicers which is consistent with the findings of Agarwal, Amromin, Ben-David, Chomsisengphet, Piskorski, and Seru (2012a). The time series variation of foreclosure rate is significantly less than the cross-sectional variation. The difference-in-difference regressions show that there are significant changes in foreclosure rate when the servicer changes for mortgages initially serviced by IndyMac Bank and Countrywide Financial Corporation. Both this evidence implies that factors related to servicer are important determinants of foreclosure rates. 4

5 I find that implementing the modification program by FDIC for IndyMac mortgages induced 6.5% more defaults by IndyMac borrowers. This is more than 50% increase in default rate compared prior to FDIC implementing modification program. This evidence is stronger than Mayer, Morrison, Piskorski, and Gupta (2011) and is further confirmation that borrowers strategically default when a modification is more likely. The effect of modification program is even stronger-2.6 times more likely to default-when I subset with higher FICO score and lower loan-to-value (LTV) ratio, the subset of borrowers who are less likely to default. This evidence supports the conjecture that the modification program induced defaults among borrowers who were less likely to default otherwise. Next, I find that servicers are 30% less likely to foreclose when they are also the originator of the mortgage. Information asymmetry is more important for low documentation loans as there is limited information on applicant s income and assets. Results are similar when I restrict the sample to only low documentation loans. This set of evidence supports the hypothesis that servicers increase foreclosure rate to discourage strategic defaults. In my final set of results, I show that a one standard deviation increase in lagged loan growth of servicer increases the probability of foreclosure by 8%. This is more than 13% increase relative to the average foreclosure rate of about 60%. One potential concern with this analysis is that results are driven by the rapid depreciation of house values in which case the foreclosure may be the better alternative. To address this concern I re-estimate the models using the loans that were likely to have positive equity balance to the borrower on the default date and results do not change. The effect of lagged loan growth is still significant when it is interacted with the dummy variable indicating whether originator remains as the servicer which implies that information asymmetry and capacity constraints are both important determinants of foreclosure rate. Results are similar when number of loans serviced by servicer is used instead of loan growth. Finally, I do not find any evidence to support the claim that servicers deliberately extended delinquent cycles to extract income from the borrowers. I find that time to foreclosure since the delinquency, in fact, decreased from an average of about 6 months before the crisis to about 3 months during the crisis. Time to complete the foreclosure remains uniform during the entire period at about 16 months. Therefore the capacity constraints of the servicer, proxied by loan growth, is a more plausible explanation for the higher foreclosure rate of private-label securitized mortgages. My paper is related to the empirical literature that shows securitized mortgages are more likely 5

6 to be foreclosed. While Piskorski et al. (2010), Agarwal, Amromin, Ben-David, Chomsisengphet, and Evanoff (2011b) and Kruger (2014) argue that securitized mortgages are more likely to be foreclosed, there is very little work on the drivers of this high foreclosure rates of securitized mortgages. Agarwal et al. (2012a) show that there is significant variation of modification rates across the servicers. In this paper, I provide empirical evidence supporting the claim that information asymmetry and capacity constraints of the servicers lead to higher foreclosure rates. Even though the capacity constraints of the servicer was discussed at length during the recent housing crisis (e.g: Cordell, Dynan, Lehnert, Liang, and Mauskopf (2009),Cong. Oversight Panel (2009) and Bernake (2008)), to my knowledge, this paper is the first to provide empirical evidence that servicer capacity constraints lead to higher foreclosure rates. This paper also relates to the studies that look at the internal workings of the securitization. There is a large body or literature that looks at the adverse selection and moral hazard between the originator and the investors (for e.g: Agarwal, Chang, and Yavas (2012b),Keys et al. (2010) and Elul (2009)). However, there is not much research available on the frictions between other entities of securitization. This paper looks at the information asymmetry between servicer and the borrowers and moral hazard problem between the servicer and investors. This paper is organized as follows: Section II describes the theoretical background and develops the hypotheses. The data sources, key variables and descriptive statistics are presented in Section III. Section IV discusses empirical methodology and presents the results. Section V concludes. II. Theoretical Background and Hypothesis Development Piskorski et al. (2010), Agarwal et al. (2011b) and Kruger (2014) show empirically that securitized sub-prime mortgages are more likely to be foreclosed. While controlling for observable loan characteristics, these papers use instruments and identification strategies to alleviate the concern that originators endogenously securitized the loans that are of unobservably lower quality. This implies that the foreclosure bias in securitized loans is more likely to be driven by the factors associated with the frictions introduced by securitization. These could be issue-specific factors such as modification restrictions imposed by PSAs or servicer related frictions. However, Kruger (2014) 6

7 controls for issue specific modification restrictions and still finds economically significant foreclosure bias in securitized mortgages. Hunt (2006) looked at a number of subprime MBS contracts and observed that these restrictions are rare and when present were never binding. That leaves the servicer related frictions. The servicer makes the decision to foreclose a delinquent mortgage and therefore factors associated with the servicer might be important in understanding the variation of the foreclosure rate of securitized mortgages. Agarwal et al. (2012a) show that there is significant variation in the rate of modifications across the servicers and they argue that this is due to servicer-specific factors such as preexisting organizations capabilities. In this paper I focus on two key servicer related frictions, information asymmetry between the servicer and the borrowers and capacity constraints of the servicer. A. Information Asymmetry Hypothesis Past research has argued that information asymmetry inherent in securitization may have lead to higher foreclosure rates (Levitin (2009)). Among others, models such as Riddiough and Wyatt (1994) and Wang et al. (2002) predict higher foreclosure rates for securitized mortgages due to information asymmetry between the intermediaries and borrowers. In these theoretical models, there are two types of borrowers-distressed borrowers and Non-distressed borrowers. Distressed borrowers default regardless of whether the bank is willing to renegotiate. They either do not have the ability to make monthly payments or have the ability to make monthly payments but costs of default is less than the benefits of defaulting. Non-distressed borrowers (who are also referred to as strategic defaulters) are the borrowers who gain from default only if the loan is modified 4. Therefore non-distressed borrowers, whose cost of default is high, default when the cost of default is less than the expected benefits of modifications. When the probability of modification is greater, expected benefit of modification is greater for the non-distressed borrowers, which makes them more likely to default. Distressed borrowers default irrespective of the likelihood the loan will be modified. Riddiough and Wyatt (1994) assume that the borrower type (whether distressed or non-distressed) is observable by the lender but lender s foreclosure cost is not observable by the borrower. Wang et al. (2002) assume that both borrower type and lender s foreclosure costs are 4 Strategic default is commonly referred to as the decision by a borrower to stop making payments on mortgage due to negative equity, despite having the financial ability to make the payments (Guiso, Sapienza, and Zingales (2013)). My definition of strategic defaulters is similar to Mayer et al. (2011) 7

8 not observable. When the lenders foreclosure cost is not observable, non-distressed borrowers use lender s previous choice of modification versus foreclosure as a signal of lender s foreclosure costs. Non-distressed borrowers interpret the high level of modifications and generous concessions in the past as a signal of higher foreclosure costs to the lender. They believe this would increase their chances of receiving a modification if they default. Therefore, this leads to an equilibrium where lenders trying to disguise their foreclosure costs by increasing the foreclosure rate to discourage strategic defaults when the servicer doesn t have sufficient information to identify the strategic defaulters. Moreover Adelino, Gerardi, and Willen (2013) show that foreclosure can dominate modification in the presence of asymmetric information related to extent of self-cure risk and redefault risk, which are important in determining expected benefit of loan renegotiations. Typically in securitizations servicers usually handle a large number of loans and this increases the distance between the borrowers and the servicers. This means that the servicer generally has very limited information about the borrowers default costs, the probability of self-cure and the probability of re-default compared to the information a bank would have on the loans held on their balance sheets. In most cases servicers only possess hard information such as FICO and the loan-to-value ratio. For example Keys et al. (2010) observed that originators maintained information on whether the borrower is self-employed or steadily employed by a firm but were not transferred when securitized. Given the large number of loans serviced by each intermediary and loss of soft information, cost of screening borrowers in default to understand their default costs will be prohibitive and, therefore, they are more likely to foreclose a higher fraction of loans in default. B. Capacity Constraints Hypothesis Misalignment of servicer incentives with those of other stakeholders in securitization has also been widely cited as a potential reason for high foreclosures in securitized mortgages (Piskorski et al. (2010),Agarwal et al. (2012a) and Thompson (2011)). In a typical securitization, potential homeowners borrow from originators by mortgaging their house. The originator, in turn, sells these loans to a Special Purpose Vehicle (SPV). SPV offers investment products to investors backed by the future receivables of this loan pool. In this structure, the servicer has two key responsibilities-transaction processing and Default loan administration. Transaction processing is managing the day-to-day operations and this involves collecting pay- 8

9 ments from the borrowers and transferring them to investors through the trustee. Under default loan administration the servicer is required by the Pooling and Servicing Agreement (PSA) to take appropriate actions such as modification or foreclosure for the defaulted loans. In doing so the servicer is expected to manage the defaults in such a way that they would do it for their own account. This means that the servicer should take the action that maximizes the net present value of future payoffs of the defaulted loans. This is generally implied in the PSA and there are certain examples where PSAs state this requirement explicitly. Following are two such extracts from PSAs that explicitly mention that servicer is required to administer the defaulted mortgage loans as if they do it for their own portfolio. The Servicer shall service and administer the Mortgage Loans [...] in the same manner in which it services and administers similar mortgage loans for its own portfolio... (ABFC Asset-Backed Certificates/Series 2005-OPT1) The Servicer shall service and administer the Mortgage Loans through the exercise of the same care that it customarily employs for its own account (Goldman Sachs Mortg. Co. & Bank One, N.A.) Advancing is another important responsibility of the servicer which is related to default loan administration. To the extent there is a shortfall in monthly collections from a particular borrower, the servicer is required, if it reasonably believes that such advance ultimately will be recoverable from the individual loan, to make an advance for missed payments, and to advance costs necessary to protect and foreclose on the underlying mortgaged property. Such advances have the highest seniority above investors, which means that these advances rank before all the other payments in the structure and therefore, the servicers are not concerned about the recoverability of such advances. However, servicers cannot charge interest on these advances and servicers have to bear the interest cost of advances. The longer the servicer advances more interest the servicer has to pay. More than 70% of the servicer compensation is in the form servicer fees. The servicer fee is about 25 to 50 basis points on total capital outstanding depending on the type of the mortgage. 9

10 Servicers may get additional income in the form of float income and ancillary fees. Float income is earned by investing the funds they receive from mortgagors for a short period before remitting them to the trust. For servicers default loan administration is expensive compared to transaction processing activities. Under default loan administration, servicers have to make the foreclosure versus modification decision for the delinquent mortgages. The modification requires hands-on involvement of employees since a good understanding of each borrower circumstances is necessary to recommend and implement the appropriate modifications. On the other hand transaction processing activities are repetitive and routine activities and can be easily automated. Servicers could achieve greater economies of scale by automating transaction processing activities. Given the competitive nature of the servicing business, the servicers have little control over fees and, as a result, servicers attempted to cut costs by turning to automation. Most of the servicers were using the MSP software (Levitin and Twomey, 2011) which is mainly used to automate payment processing activities. These types of software are also capable of handling foreclosure process without much involvement of people. The systems are capable of creating automatic referrals to attorneys with specific work orders and supporting loan documentation. On the other hand, it is difficult for any software to implement a successful modification program. There are different modification actions available to the servicer such as principal reduction and interest rate changes. One needs to carefully analyze the borrower s circumstances to figure out the best modification strategy in which human judgment plays a critical role. For most of the securitization transactions, fees were agreed on during the housing boom and were based on the notion that default management would be a relatively small part of servicer responsibility. Given the agreed fees servicers had no incentives to retain excess modification capacity more than what was required under normal circumstances. According to Berry (2007) the fees should have been at least five times the fees agreed if the servicers had anticipated the default rates to go up to the actual levels that were seen in Different investors in securitization are exposed to different parts of the loss distribution therefore, investors in senior tranches would not be willing to pay higher servicing fees to retain additional staff to handle defaults ex-ante before the crisis. Around , when the default rates were increasing, many of the servicers also struggled financially and it may not have been financially feasible for them to hire new employees. 10

11 Even if it was financially feasible, it would take time to hire new employees and train them. This implies that when the defaults were increasing rapidly, servicers didn t have enough people and/or couldn t hire more people to handle that many defaults. Therefore, when defaults were increasing capacity constrained servicer would have had stronger incentives to foreclose the defaulted loans using the automated systems. The robo-signing incident in 2010 gives us some evidence of this servicer bias. 13 banks were found guilty of robo-signing where servicers rapidly approve numerous foreclosures with only cursory glances at the glut of paperwork to determine if all the documents were in order. Maturana (2014) provides indirect evidence that supports the hypothesis that the servicers were capacity constrained during the 2008 crisis period. The paper shows that when an incentive fee was introduced in the agency RMBS market to increase the modifications, the modification rate of non-agency loans serviced by the servicers who also service agency loans decreased. He argues that this is because the servicers have a limited capacity to undertake modifications and those servicers concentrated modification efforts in the agency markets since it was more profitable to them. B.1. Simplified Illustration This simplified illustration would help us to better understand why the servicers would prefer to foreclose more when the default rates are increasing. For a particular servicer i let N it be the number of loans serviced at time t. Let α it be the fraction of loans that are delinquent and should be modified to maximize payoff to investors and p it is the present value of the future servicer income per loan. Let C(x it N it ) be the cost of modifying x it N it loans for the servicer i. Assume there are no re-defaults. It is reasonable to assume C(x itn it ) x it > 0 and 2 C(x it N it ) > 0 since the servicer is constrained in terms of human capital and modification x 2 it costs are likely to increase with the number of modifications. Lets assume that the foreclosure cost is zero and no advancing requirement. Zero foreclosure cost is a reasonable assumption since it is relatively easier for the servicer to recover the foreclosure related expenditure (Levitin (2009), Cordell et al. (2009)). Assuming N it is fixed, Panel A of figure 2 shows how the servicer net income (p it x it N it ) changes with x it. The servicer s income initially increases and then is maximized at a point x it and then starts to decrease. Therefore if α it > x it the servicer earns less income by modifying α itn it loans 11

12 compared to modifying only x it N it loans. As shown in Panel B of figure 2, the servicer s net income is maximized if the servicer limits the modifications to x it N it and foreclose (α it x it )N it loans. This kind of a setup can also explain why in certain circumstances the servicers would increase the number of modification more than the optimum level. For example in 2013 the Ocwen Financial Corporation was blamed for offering too many modifications to borrowers of the loans that were serviced by Ocwen. It is clear from Panel A of figure 2 that if a it < x it the servicer can increase its profits by modifying more loans than the optimum level α it. This is primarily applicable in a scenario where the default rates are low. Under this scenario, expected marginal future income after modification is greater than marginal cost of modifying. C. Summary of Hypothesis In sum, the main hypothesis investigated in this paper are as follows. 1. Does the servicer impact the foreclosure rate of securitized mortgages? 2. Do borrowers exhibit strategic behavior when the probability of modification is high? and do servicers increase foreclosure rate when information asymmetry is high to discourage strategic defaults? 3. Are capacity constrained servicers more likely to foreclose? III. Data In order to test my hypotheses, I require data on loan histories of securitized mortgages, which enables me to identify the time at which each loan defaulted and time at which the foreclosure action was initiated. I also need the servicer identity for each loan for each month as the servicer changes were fairly common after the financial crisis in For example Countrywide Financial s loans were acquired by Bank of America in 2008 and IndyMac Bank s loans were acquired by OneWest Bank in

13 A. Data Sources The main source of data for this study is Bloomberg which contains data for 4,235 private label RMBS issues. Apart from issue level data, Bloomberg provides a snapshot of all the loans of a particular issue for each month. For example, if we select January 2010 snapshot for a particular issue, for each loan we get information such as the monthly status of the loan up to January 2010, the interest rate on January 2010, the loan balance as at January 2010 and the loan servicer on that date. Correctly identifying the servicer of each mortgage is crucial to this paper and therefore, loan level data for each issue was obtained at three different points in time (2006, 2009 and 2012) to capture the changes in the servicer. By comparing the servicer at these three different points we can identify the loans or issues that had any changes to the servicer. Even though we can identify the servicer changes this way, this doesn t identify the exact month on which the servicer change happened 6. After identifying the unique combinations of servicer changes based on these three snapshots, for each combination a small sample of issues (10 issues or less depending on availability) were selected. For each sample, all the snapshots were checked to identify the exact month in which the servicer change happened. Due to the time-consuming nature collecting and compiling loan histories from the snapshots, a random sample of issues originated between 2000 and 2007 was selected for this analysis. This sample was used to construct a loan-month panel with loan status and servicer for each loan for each month. Metropolitan Statistical Area (MSA) level unemployment data was obtained from U.S. Bureau of Labor Statistics website and three-digit zip code level House Price Indices (HPI) were obtained from Federal Housing Finance Agency website. These HPIs are broad measures of the movement of single-family house prices. Each HPI is a weighted, repeat-sales index, that measures average price changes in repeat sales or refinancings on the same properties in a given region. B. Key Variables I identify the default and foreclosure initiation from the loan histories of each loan. Loan history for each loan is a string variable with each character representing the loan status in each month 6 For example some loans will have IndyMac as the servicer in 2006 and OneWest as the servicer in 2009 and 2012, resulting in the servicer combination IndyMac, OneWest, OneWest. 7 Loan level data was collected for 700 issues but loan level data was available for only 681 issues 13

14 since loan origination 8. Following Piskorski et al. (2010) and Agarwal et al. (2011a), I use the Mortgage Bankers Association (MBA) definition of 60+ day delinquency, which is represented by character 6 in the loan history, as the definition of default in this paper. Foreclosure initiation was identified as the first F after the default. Both default and foreclosure initiation were coded as binary variables in the loan-month panel. My key independent variable of interest is 1 year lagged annual loan growth for each servicer which serves as a measure of degree of servicer capacity constraints. This measure was estimated only for the top 20 servicers from the loans included in the sample. Smaller servicers were ignored to minimize the effect of any sample selection biases. Reduction of the sample size due to this restriction is not significant since the top 20 servicers service 88% of the loans in the sample. I also use the log of number of loans serviced by each top servicer as an alternative measure of the degree of capacity constraints. This was also calculated from the loans included in the sample. The loan-month panel has detailed information for loans at the time of origination such as borrower s FICO score, Loan-to-Value(LTV) ratio, loan amount, initial interest rate, interest rate basis and documentation level. FICO score, a machine generated score from individual credit reports, ranges from 400 to 900 and a higher score implies a lower probability of default. LTV is calculated by dividing the original loan amount by appraised value of the property and high LTV ratios are generally seen as higher risk. Loan documentation generally takes the values full, limited or no documentation. Full documentation loan refers to a loan where all income and assets are documented. The dataset also provides information such as the state and the ZIP code of each property. Whether a borrower has negative equity position on the mortgage is an important factor in determining whether the mortgage should be foreclosed. I approximate the borrower s equity in the home for each loan-month observation using the three-digit zip code level HPIs together with original loan amount, the original LTV and the interest rate. 8 Character C implies current, M implies modified, 3 implies 30 days past due, 6 implies 60 days past due, 9 implies 90 days past due and F implies foreclosed 14

15 C. Descriptive Statistics Issues in my sample consist of 2.2 million loans with a total loan amount of more than 500 billion dollars. This is approximately 18% of the total private-label RMBS securities outstanding in Table I gives the distribution of issues, loans and total loan amounts across the issued years. However, most of the analysis was restricted to only 1.08 million loans since the complete loan histories were not available for the other loans. As can be seen from the figure 1, the pattern in the number of foreclosure initiations in my sample closely follows the pattern in the foreclosure initiations of all the mortgages. Moreover, Agarwal et al. (2011b) show that a year following delinquency, in 2008, about half of the borrowers in foreclosure. In my sample, the foreclosure rate in 2008 is about 55%. Therefore, I believe that the sample in this paper is a reasonable representation of the sub-prime mortgage population. Panel A of the table II reports the descriptive statistics of the loans with complete loan histories. The vintage year is the year in which those loans were securitized. We can see here that the observable quality of the loans securitized have generally declined from 2003 to 2006 while the number of loans securitized increased during the same period. Average FICO score has decreased from 695 to 663 and the percentage of full documentation loans securitized has decreased from 60% to 38%. Panel B of the table II reports the same statistics for the subset of loans that were at least 60 days delinquent at least once during their life cycle. There are 463,359 such loans in the sample and the observable quality of those loans is markedly less than the entire sample. These loans have a significantly low FICO scores and high LTV ratios. However interestingly the average interest rate of those loans is lower compared to the entire sample and this is probably an indication of poor loan pricing by the originators as in Rajan, Seru, and Vig (2008). A more careful analysis is necessary to understand this anomaly in pricing and it is not the focus of this paper. Table III presents the percentage of delinquent loans foreclosed for each of the top 20 servicers for each delinquent year 10. We can observe a large cross-sectional variation of the percentage of delinquent loans foreclosed. In 2007, for example, the minimum foreclosure rate is 18% while the 9 In 2007, the total private-label RMBS outstanding amount was 2,704 billion dollars and this was the highest ever private-label RMBS outstanding amount 10 Delinquent year is the year in which the loan goes in to 60 days delinquent status 15

16 maximum foreclosure rate was 87%. However compared to high average cross-sectional variation (16.42% in ), time series variation is low (8.82% in ). This difference in variation is consistent with the argument that servicer specific factors impact the foreclosure rate of each servicer. IV. Empirical Methodology and Results A. Servicer Impact I begin my analysis by evaluating whether the servicer has an impact on the level of foreclosure initiation rate of the loans they service. However identifying the impact of the servicer on the level of foreclosure is difficult due to unobservable loan characteristics of loans serviced by each servicer. A logistic regression on the observable loan characteristics with servicer fixed effects and evaluating statistical significance of fixed effects rely on the assumption that, conditional on observables, the loan assignment to servicers is random. Generally, this assumption is likely to be violated since some of the servicers specialize in certain types of loans such as Alt-A and some are restricted to a particular set of originators or geographic regions. So the foreclosure rate may be related to unobservable loan characteristics but not the servicer. One can argue that the random assignment assumption is likely to be satisfied for this test if conditioned on a plethora of observable loan characteristics since I am restricting my analysis only to delinquent loans. Which means that the effects of unobservable factors has already incorporated in the default. However, it is possible that this assumption is violated and the estimated coefficients are biased. To alleviate this concern I compare the foreclosure rates before and after servicer changes for two large servicers during the foreclosure crisis. I use the cases of IndyMac Bank and Countrywide Home Loan Servicing where the loans serviced by these institutions were acquired by FDIC and Bank of America respectively. These were two of the largest servicers servicing private label RMBS and in my sample they collectively serviced 26% of loans serviced by the top 20 servicers. 16

17 A.1. Change in the Servicer for Countrywide Mortgages Countrywide Financial, founded in 1968, was by 1992 the largest originator of single-family mortgages in the United States and in 2006 Countrywide financed 20% of all mortgages in the United States. Following large losses suffered by Countrywide, in January 2008, Bank of America announced that it planned to purchase Countrywide Financial for 4.1 billion dollars and completed the acquisition in June With this acquisition, the Countrywide Home Loan Servicing, Countrywide s mortgage banking subsidiary, became the mortgage unit of Bank of America. First, I exploit this servicer change of the mortgages initially serviced by Countrywide in a difference-in-difference setup in the following form to identify the impact of servicer change on the foreclosure level. This analysis focuses on how the foreclosure rate of mortgages serviced by Countrywide changed before versus after the servicer change, relative to mortgages serviced by the other servicers. F oreclosure i = α + β 1 Countrywide + β 2 T ime Jul08Dec08 + β 3 Countrywide Jul08Dec08 (1) I construct the sample of Countrywide loans with the following procedure. I select 6 months periods before and after Bank of America acquired Countrywide s mortgages, from January 2008 to June 2008 and from July 2008 to December For each period, loans that were delinquent at the beginning of each period or loans that went to delinquent state in the first three months of each period were selected 11. There are 4,848 Countrywide loans that match this criteria for the first period from January 2008 to June For the second period, there are 5,615 loans that match the criteria. As the control sample, I construct a matched sample of mortgages serviced by other servicers. The matched sample was generated by matching the FICO, LTV, documentation, MSA, time period, whether the borrower has negative equity and interest rate type of loans using a basic propensity score matching. MSA, time period and negative equity dummy variable were exactly matched. The loans serviced by IndyMac Bank was excluded from the control sample. Countrywide is a dummy variable that takes the value of 1 if the loan was initially serviced by Countrywide. T ime Jul08Dec08 is a dummy variable indicating whether the loan was in delinquent 11 Only the first three months of each period was selected leave enough time for the servicer to make a decision. As shown later in the paper, a servicer takes about 3 months on average during the crisis period to initiate a foreclosure or a modification. All results are robust to changing this criteria 17

18 state when the servicer changes or goes into delinquent state in the first three months after the servicer changes. Mortgages that were delinquent in the first period and re-defaulted in the subsequent period were ignored and, therefore, any given loan belongs only to one of the two periods. Table IV presents univariate evidence. In Panel A, we can see a huge drop in the rate of foreclosure initiations from 49% to 15% after Bank of America acquisition even though the observable loan characteristics are very similar. The statistics for the control sample generated using propensity score matching method are reported in the Panel B. In the matched sample the reduction in the foreclosure rate in June 2008-December 2008 period is significantly lower compared to the reduction of foreclosure rate of Countrywide s loans. Coefficient estimates and the robust standard errors of equations 1-estimated using matched data described above-are reported in tables V. Logistic regression model results are reported in the column (1) and F oreclosure i is a binary variable indicating whether the foreclosure was initiated during each period. For this analysis, I only follow mortgages till the end of each period. F oreclosure i takes the value of 1 if the foreclosure is initiated during the period concerned. There would be right censoring as there is a possibility that the foreclosure was initiated after the end of the period. Columns (2) and (3) report the results of the Cox Proportional Hazard models which account for the bias from right censoring. For these models F oreclosure i is time taken to initiate the foreclosure. Column (3) includes a negative equity dummy variable, change in the unemployment rate in the MSA and interaction of the two as additional control variables. In both the tables, the main interaction variables are significant across all the specifications. This implies a change in the nature of default loan administration activities for these issues when the servicer changes. The negative coefficient for Countrywide T ime Jul08Dec08 implies that the foreclosure rate is significantly lower after Bank of America acquisition. A.2. Change in the Servicer for IndyMac Mortgages IndyMac Bank, which was spun off from Countrywide in 1997, was specializing in originating and distributing Alt-A loans 12. Following a bank run, where more than 1.3 billion dollars was withdrawn by depositors in less than three weeks, FDIC was compelled to take control of IndyMac 12 Alt-A loans are mainly low documentation loans where the credit decision is based mainly on the FICO score. Income and assets are simply stated or not required 18

19 in July After acquisition, FDIC implemented a program to systematically modify the troubled mortgages based on the NPV of estimated future payments and FDIC s objective was to use this as a test case for their push towards greater renegotiation rates. Eventually IndyMac was sold to OneWest Bank in March As in the previous case, the objective here is to evaluate how the foreclosure rate changes when the servicer changes for IndyMac loans relative to loans serviced by other servicers. In IndyMac Bank s case, we have three different time periods. During the period starting from October 2007 and ending in June 2008 IndyMac Bank was servicing the loans. In the second period from July 2008 to February 2009 FDIC was handling the delinquent IndyMac loans while from March 2009 to January 2011 OneWest Bank was acting as the servicer. Same criteria as in the Countrywide case was used here to construct the IndyMac sample and the control sample where loans that were delinquent at the beginning of each period or loans that went to delinquent state in the first three months of each period were selected. There are 1,061 IndyMac loans that match this criteria for the first period from October 2007 to June For the second period, there are 12,226 loans and 20,166 loans satisfied the criteria for the third period. Same propensity score matching method was used to generate the control sample. Univariate evidence presented in table VI show that when FDIC was managing the delinquent loans the foreclosure initiation rate went down by 7% while the foreclosure rate of the similar loans actually increased by approximately the same percentage points in July 2008 to February 2009 time period. When OneWest started servicing IndyMac Loans in March 2009 the foreclosure initiation rate jumps up by 10% while the foreclosure initiation rate of matched loans decreased. The multivariate evidence is presented in table VII. For the IndyMac s case, I have two time dummy variables, one to indicate the period during which FDIC was servicing (T ime Jul08F eb09 ) and the other to indicate the period OneWest Bank was servicing (T ime Mar09Nov09 ). The variables of interest are IndyMac T ime Jul08F eb09 and IndyMac T ime Mar09Nov09. Statistical significant estimates for these coefficients with signs consistent with univariate evidence imply that the servicer has a significant impact on the foreclosure initiation rate. The results in this section are consistent with the findings of Agarwal et al. (2012a) and support 13 See Bovenzi (2015) for a detailed description of the IndyMac Bank s case 14 Number of observations for IndyMac loans is lower for the first period since for most of the IndyMac loans the loan history is only available from April

20 the broad hypothesis that the servicer influences the foreclosure initiation rates of the mortgages they are servicing. B. Information Asymmetry Even though the difference-in-difference regressions outlined in the previous section can shed light on the question of whether servicer has any impact on the foreclosure rate, those do not inform us on the channels through which this potential effect takes place. Testing the information asymmetry hypothesis entails testing of two hypotheses. 1) borrowers are more likely to strategically default when the probability of modification is high and 2) when the information asymmetry is more serious, the servicers increase foreclosure rate to discourage strategic defaulters. B.1. Strategic Default Hypothesis To test the first hypothesis I use the case of IndyMac Bank, where FDIC implemented a systematic modification program which resulted in an exogenous shock to modification rate. Results in section IV.A.2 show that the effect on modification rate is large and statistically significant. The first hypothesis predicts that this would induce more defaults as strategic defaulters try to gain from the expected higher modification rates. For this analysis, I focus on three 8 month time periods, October 2007 to June 2008, July 2008 to February 2009 and March 2009 to November 2009 where loans were serviced by IndyMac, FDIC and OneWest Bank respectively. My main interest here is on how the default rate changes when the probability of modification increases unexpectedly in the second period. I construct the IndyMac sample by selecting the loans serviced by IndyMac that are current at the beginning of each period. As the control sample, I construct a matched sample of mortgages by matching the FICO, LTV, documentation, MSA, time period, whether the borrower has negative equity and interest rate type of loans using propensity score matching method. Table VIII provides a comparison of default percentage, mean LTV, mean FICO, the percentage of full documentation loans and percentage of fixed rate loans for IndyMac loans and loans serviced by other servicers before and after FDIC implemented the modification program. Comparison of Panel A with Panel C reveals that there are notable differences in the full documentation loans 20