Risks of Long Term Auto Loans

Transcription

1 Risks of Long Term Auto Loans Zhengfeng Guo Xinlei Zhao Office of the Comptroller of the Currency First version: May 2016 This version: January 2017 Keywords: Auto loans; loan terms; long-term auto loans; loan pricing; soft information; strategic default. EL classification: G21. Zhengfeng Guo is a Financial Economist in the Credit Risk Analysis Division of the Office of the Comptroller of the Currency. Xinlei Zhao is the Deputy Director of the Credit Risk Analysis Division of the Office of the Comptroller of the Currency. To comment, please contact Xinlei Zhao at Office of the Comptroller of the Currency, th Street SW, Mail Stop 6E-3, Washington, DC 20219, or call (202) ; or Xinlei.Zhao@occ.treas.gov. The views expressed in this paper are those of the authors alone and do not necessarily reflect those of the Office of the Comptroller of the Currency or the U.S. Department of the Treasury. This paper benefits tremendously from the early work by Deming Wu. The authors would like to thank the excellent research support by Andrew Goad, and the valuable comments from Fredrick Andersson, Marcey Hoelting, Min Qi, Lan Shi, Natalie Tiernan, Yan Zhang, Chris Henderson and seminar participants at the Office of the Comptroller of the Currency and the 2016 Interagency Risk Quant Forum. The authors take responsibility for any errors.

2 Risks of Long Term Auto Loans Zhengfeng Guo and Xinlei Zhao Abstract: Among banks and credit unions, long term auto loans have higher default rates than shorter-termed loans during the first four years after loan origination. Further, credit union loans have significantly lower default rates than bank loans. Such differences cannot be explained by observed risk factors, and the differences widen as the loans age. The difference in auto loan default experience from various lenders is likely due to their different use of soft information. Finally, relative pricing does not reflect the relative default risks among a significant proportion of auto loans, especially among long term loans. 1

3 I. Introduction Auto loans have been capturing the media s attention lately because of the recordhigh auto loan balances in the U.S at present, while at the same time, more car buyers are showing signs of struggling to make auto loan payments recently. 1 The striking feature among auto loans nowadays is the increasing loan terms. 2 Figure 1a shows that the proportion of new auto loans with terms beyond five years in the U.S. has been increasing steadily from less than 55% in the first quarter of 2013 to approximately 64% in the first quarter of Such a phenomenon is mainly driven by the more frequent origination of 7-year terms, and the 6-year term is currently the most commonly used term type. Further, our conversation with practitioners suggests that the use of deferral, renewals, or re-writes at the back-end has been increasing, which extends the loan terms even longer. Does the sharp rise in auto loan terms in recent years make auto loans more risky? This question is of valid concern to regulators, policy makers, and anybody interested in the general health of the economy, as some people contend that the major reason behind the recent lengthening in auto loan terms is to lower monthly payments, thus making auto loans more affordable. 3 There is also some argument in the media that the driving force 1 See for example, reports like and See, for example, Introducing the 97-month car loan, Wall Street Journal, April 8, 2013, by Mike Ramsey, and the Wall Street Journal report at 3 See for example the articles Long-term auto loans: Do you know how long is too long? at and at How long should my car loan be? 2

4 behind auto sales is not the economy, but record auto lending, 4 which to some extent mirrors the factors driving the booming housing market leading up to the 2007 crash. On the other hand, there is little insight from the existing academic literature, as none of the small literature on auto loans specifically addresses the risks of long term auto loans. 5 Further, since only a small proportion of auto loans are securitized, 6 any risk unaccounted for might have a direct impact on lenders books in the years to come. Against this backdrop, this paper aims to enhance our understanding of the default risk in long term auto loans. We focus on default probabilities in this paper, as we do not have data on loss recovery. Further, given that the origination loan-to-value ratios are typically well above 100%, the loss recoveries are likely to be low, especially if the usedcar prices retreat from its recent highs. Default probabilities are therefore likely the key driver of credit losses from auto loans. The contribution of this paper is twofold. First, we examine whether long term auto loans are more risky than shorter-termed ones and whether the risks among long term auto loans are priced properly. Second, we assess the value of soft information in credit risk management. There is some debate among researchers on the usefulness of soft information. Some authors argue that soft information can be critical to managing credit losses (for example, Agarwal et al (2011) and Rajan, Seru, and Vig (2015)), while other researchers contend that the development of technology and labor productivity in banks has made soft information less important (for instance, Petersen and Rajan (2002)). 4 See for example, the Wall Street Journal article Subprime flashback: Early defaults are a warning sign for auto sales at 5 For example, Heitfield and Sabarwal (2004), Agarwal, Ambrose, and Chomsisengphet (2008), Yeh and Lee (2013), and Wu and Zhao (2016). 6 Most of the securitized auto loans are subprime auto loans from finance companies. 3

5 We add to this literature by studying whether lenders with different access to soft information might experience different default rates for loans with similar easy-to-collect information on loan and borrower characteristics. Our study is based on a sample of auto loans from a credit bureau that spans over 11 years, from We find that borrowers of short term auto loans (i.e., one to two years) have the lowest credit scores. Although auto loans with terms beyond five years seem to have better credit quality (such as credit bureau scores) than those with terms below two years, there is obvious deterioration in borrower credit quality when loan terms increase beyond five years. Further, bank and credit union loans appear to be quite comparable in terms of the commonly observed borrower characteristics, while borrowers from finance companies seem to be more risky in such dimensions. After accounting for the risk factors commonly used in auto loan underwriting, we find that, among banks and credit unions, long term auto loans have higher default rates than shorter-termed loans during the first four years after loan origination. In addition, bank loans are riskier than credit union loans throughout the loans life time, especially among seasoned auto loans, after accounting for the commonly used loan and borrower characteristics. The stark difference between bank and credit union loans suggest that better access to soft information relative to banks might have helped credit unions in their credit risk management. Finally, although we find bank loans tend to be riskier than credit union loans, controlling for the same risk characteristics, banks actually charge lower annual percentage rates (APRs). Further, we find that APRs vary significantly with credit scores, but there is little differentiation along the term dimension, with APRs slightly 4

6 lower among long term loans originated by banks and credit unions. Therefore, among a significant fraction of the auto lending industry, relative pricing fails to reflect the relative default risk across various types of auto loans. The findings in this paper thus raise serious concerns on the lending practices in the auto lending industry. Our results suggest that a significant number of auto loan lenders do not seem to understand the higher risk among the long term auto loans. Even though credit unions appear to be more careful in underwriting, they still do not account for the higher risk among long term auto loans in their pricing scheme. The rest of the paper proceeds as follows: Section II discusses the data and how various types of lenders have different access to soft information, and Section III discusses the nature of long term auto loans. We investigate the risks among long term auto loan in Section IV and examine loan pricing in Section V. We finally draw conclusions in Section VI. II. Data description, lender types, and summary statistics II.A. Data construction Our data are from a major credit bureau in the U.S. and are constructed as follows. The sampling unit is the unique consumer identifier record (CID). In June of 2005, 0.7% of the 2005 CID universe from the credit bureau archive was randomly sampled, and that constitutes our 2005 data. Once selected, a CID is always reported. Each year thereafter, a new CID entrants list is populated and randomly sampled from to counter attrition and maintain the 0.7% of the CID universe target. For example, to create the 2006 data, the credit bureau first defined the 2006 New Entrant population this population consists of credit records (CIDs) that entered the credit 5

7 bureau universe during the 2006 archive year. Then the credit bureau sampled CIDs from the 2006 New Entrant population using the same 0.7% sampling rate as was used for The 2006 New Entrant sample CIDs were added to the list of CIDs sampled from 2005, and credit records were returned for each CID. This is the basis of the 2006 data. The algorithm is the same for later years. Each year a New Entrant population is defined and sampled. Then credit records are returned for the newly sampled New Entrant CIDs and all CIDs sampled in previous years. As a result, the 2008 data report credit records for all CIDs sampled in 2005, as well as all sampled New Entrant CIDs. Our credit bureau data consist of both attribute and tradeline data. The attribute data file provides annual snapshots of individual characteristics and account level credit file as of June 30 th of the file year from 2005 through The attributes include annual information on the geographic location of an individual including city, state, and zip code, consumer credit risk score (i.e., bureau credit score), and 10 summary credit attributes using tradeline data for each individual. The tradeline data contain annual snapshots of detailed account level characteristics on all credit accounts, inquiries, public records, and collections for all scorable individuals present in the attribute file and some unscorable individuals. 7 We also have information on the account delinquency status over the past 48 months. Therefore, for some loans still outstanding in 2005, we could observe their delinquency status back to The credit bureau data do not have information on the loan-to-value ratio (LTV) of the auto loan, neither at origination nor updated. Additionally, there is no information on borrower income, job, and education, on whether 7 We divide all balances by 2 if the account is a joint account, and the credit score is of the primary account holder. 6

8 the collateral is a new or used car, or whether the loan is directly financed or indirectly financed through a dealership. Our study excludes auto leases. Our base sample consists of auto loans observable in our sample during the period from June 2005 to June 2015, and a loan drops out of our sample once it hits 90+ DPD, or is paid off, or is censored at the end of the sample period. This sample consists of 695,971 individuals with auto loans outstanding, 1,416,600 loans and 3,497,513 loan-year observations. II.B. Lender types and use of soft information We categorize different lenders in the credit bureau data into four types. 8 Group 1 consists of banks and they constitute roughly 22% of the total sample. Group 2 consists of lenders in the financial category in the credit bureau data (which we call the finance companies in the rest of the paper), making up approximately 44% of the sample. 9 Group 3 comprises the credit unions, about 25% of the sample. The credit bureau data have various category names for the remaining lenders, and we group all of them into Group 4, which we term other lenders. This group makes up only around 9% of our sample. We exclude Group 4 lenders in our analysis because of their small market share; loans from these lenders are the most risky. We break our analysis into different types of lenders. First of all, credit union loans may behave differently because of their unique organizational structures and operations. There is no restriction as to which customers banks or finance companies can provide potential services, and customers are usually not owners of these lenders. Banks 8 In the data, we can observe the type of lenders but not the name of the lenders. Therefore, we cannot conduct sub-group analysis to identify any heterogeneity in the age survival function among different types of banks, finance companies, or credit unions. 9 This group includes captives and other auto financing companies and credit bureau does not provide enough information for us to separate captives from other auto financing companies. 7

9 or finance companies are of the mission to make profits, and the net income is available to shareholders but not account holders of the lenders. By contrast, credit unions are member-owned, not-for-profit financial cooperatives that provide savings, credit and other financial services to their members, with the mission to be community-oriented and to serve people instead of making profits. Credit union membership is typically based on a common bond, and the members belong to a specific community, organization, religion, or place of employment. Credit unions pool their members savings to finance their own loan portfolios, rather than rely on outside capital. Furthermore, credit unions are governed in a democratic way, with each member able to run for the volunteer board of directors and cast a vote in elections, regardless of the account size in the credit union. Because all account-holders are members and owners, net income is applied to lowering interest rates on loans, raising interest rates on savings or to new product and service development. Therefore, all borrowers from credit unions share the benefits as well as the costs of the operation, because they are owners at the same time. The underwriting process for a prime auto loan is automated, typically taking less than ten minutes. Within such a short window, it is very difficult, if not impossible, to process and verify important borrower information, such as income, and the underwriting process of prime auto loans is heavily dependent on information readily available from credit bureaus, especially credit bureau scores. As a result, income is rarely among the risk drivers of auto loan underwriting models and it should be deemed soft information in 8

10 auto lending. 10 Additional soft information, such as job security and willingness to pay, is even more difficult to collect. However, only members or owners of the credit union can borrow from the credit union, and a credit union should have obtained some income- or job-related information when a person applies to become a member of the credit union. So the interaction between credit unions and their borrowers is not restricted to just the a few minutes during the underwriting process, and credit unions should be better positioned to collect and verify some soft information via potentially multiple interactions with their borrowers/members. In addition, also because of credit unions membership structure, the interactions between credit unions and their borrowers should continue even after loan origination. Further, the common bond connecting the members or borrowers of credit unions, such as, neighborhood, employment or religion, may help uphold the moral standard of a community, and a borrower may be quite reluctant to default if the default can be observed by his or her peers. The common bond can also make the collection process less costly, which can further hinder defaults not driven by credit reasons. In a word, the common bond is likely able to boost a credit union s ability to monitor on an on-going basis their borrower s willingness to pay, and such soft information is beyond the hard and soft information related to borrowers ability to pay at time of loan origination. Second, while banks and credit unions mainly target at prime borrowers, many finance companies specialize in extending auto loans to lower credit quality borrowers. 10 Banks typically manage different retail products separately instead of linking different accounts of the same borrower. As a result, we rarely see banks using deposit account information in their loan underwriting models. Further, the overwhelming majority of bank auto loans are indirect loans initiated by auto dealers, and such loans definitely do not link with the other debts of the borrower in the same bank. 9

11 Underwriting a subprime loan usually involves intervention from a loan officer, 11 requesting additional information, such as verifiable household income, job tenure, and expenditures, and such underwriting can take hours or days. So the lengthy underwriting process for subprime auto loan borrowers collects much soft information beyond what is available from credit bureaus. Since a significant proportion of finance company auto loans are for subprime borrowers (which we will show later in Section II.C), finance companies as a whole should have heavier use of soft information than either banks or credit unions. However, finance companies do not have the close customer relationship as credit unions do, so they should not be able to monitor their borrowers willingness to pay post loan origination as effectively as credit unions can. In summary, loans from various types of lenders will most likely behave differently, conditioning on the commonly used information available from credit bureaus. A comparison of the behavior of auto loans among different lenders can thus help us better understand how soft information can contribute to risk management in the auto lending segment. II.C. Summary Statistics Table 1 presents the summary statistics of the variables used in this paper. We define 90+ days past due (DPD) as default. The numerator in PTI is the monthly payment of the auto loan in question, while the numerator in DTI includes the total outstanding balance on all auto loans and mortgages of the borrower. 12,13 The denominator in both 11 See Agarwal et al (2011) for more descriptions on loan officer intervention. 12 We have tried adding other consumer credits, such as credit cards, home equity lines, and student loans into the DTI calculation. We find the coefficient estimate of such alternative definitions of DTI to be negative, which says higher consumer leverage is negatively related to the probability to default on auto loans, a finding that is counter-intuitive. The coefficient estimate of DTI is positive if only mortgages and auto loans are used to define DTI. Such results might be driven by the pecking order of defaults in consumer debts. The industry experience has been that consumers tend to default on other non-mortgage 10

12 PTI and DTI is the annual average personal income at the county level, as the credit bureau data do not have income information at the individual level. The data do not report APR and we calculate the APR based on the origination amount, monthly payment and the loan term. We define subprime when the credit bureau score is below Unemployment is at the county level. The used car prices are from Manheim, Inc, which is the largest automobile auction company in the world by volume of trade, and the Manheim used vehicle price index is the most widely used index in the industry. Table 1 shows that the average origination amount of auto loans in our sample is $20K, with monthly payment averaging $397 and DTI averaging The majority of the sample consists of borrowers with outstanding mortgages and roughly one-third of the borrowers are subprime. The sample has a good coverage for a wide range of credit bureau scores, ranging from below 500 to 850. The credit card utilization rate also shows a wide range, with its mean at 33%. Note that we do not have a complete list of loan and borrower characteristics in Table 1. Our conversation with bank executives suggests that banks underwriting model are primarily based on credit bureau scores. Among the variable listed in the table, DTI are hardly ever considered, and PTI and loan-to-value ratios normally enter the underwriting criteria in terms of high limits. Additional important risk drivers, such as income information, are rarely considered during the underwriting process among prime consumer debt before they default on auto loans. As a result, the amount of other types of non-mortgage consumer debt may not be particularly relevant in a consumer s decision to default on auto loans. The pecking order of default in consumer debt is beyond the scope of this paper, and we only keep mortgages in this study to keep our side story simple. Note that the survival functions we uncover in this study holds regardless of how we define DTI. 13 For joint accounts, we follow the convention in the literature and take only half of the outstanding loan amount in the calculation. 14 Throughout the paper, we use the up-to-date credit scores unless the credit scores are specifically noted as loan origination credit scores. 11

13 loans as such information is costly to collect and verify. Therefore, the variables in Table 1 should have captured the variables used in the typical prime auto loan underwriting processes among banks rather comprehensively. We the call the risk factors in Table 1 observed risk factors, and additional important and harder-to-collect information soft information for the rest of the paper. Obtaining information beyond that available in the credit bureau data is practically impossible for academic research. However, this challenge does not compromise the purpose of this paper. Our goal is not to investigate a comprehensive list of determinants of auto loan defaults, and lack of additional information, such as borrower income, job, education, or willingness to pay, does not hinder us from studying whether any knowledge about such soft information adds value to credit risk management. This is because the residual risk from our model after controlling for the commonly used risk factors will reflect the impact from soft information, and we make inferences on the regression residuals. Figure 1b depicts the distribution of terms of newly originated auto loans over time in our sample. It is clear from this figure that terms longer than five year are not new innovations from recent years, and they have been in existence as long as ten years ago. The proportion of newly originated loans with 6-year and longer terms hit a peak in 2007 in our sample, and then quickly declined during the latest financial crisis. 15 However, starting from 2009, the fraction of newly originated long term loans quickly rose again; the proportion of auto loans with terms beyond five years increased monotonically from a low of 29% in 2009 to nearly 50% by June 2015 in our sample. Over the 12 years starting in 2004, the 5-year term lost its popularity, and by 2015, the 6-15 Throughout the paper, we categorize the loans with 7 year and above terms as 7-year term. 12

14 year term is undoubtedly the most commonly used auto loan term by a wide margin. Therefore, our sample clearly demonstrates a pattern of rising terms of auto loans over time, consistent with the pattern in Figure 1a, which is based on the entire portfolio of auto loans in the U.S. Note that the proportion of loans with terms above five years, and in particular, the fraction of loans with terms beyond six years, is lower in Figure 1b than in Figure 1a. Based on the way our sample is constructed, we can conclude that the long term loans are more concentrated among the new entrants post 2005 (most likely the younger generation), rather than the credit bureau population in As younger consumers are more financially constrained, the difference in the two graphs in Figure 1 is the first piece of evidence implying that the long term auto loans are extended to more risky borrowers. III. Who borrows long term auto loans? Given the pattern shown in Figures 1a and 1b, it is natural to ask what types of loans have longer terms. We present these statistics in Table 2. We categorize the sample by length of loan terms, less or equal to 2 years, 3, 4, 5, 6, and 7 years. Panel A reports the statistics over the entire sample. A pattern that clearly stands out is that the origination credit bureau scores monotonically increase for terms up to five years. The credit bureau scores are noticeably lower for 6-year term auto loans; there is clearly an increase of bureau scores from 6-year term loans to 7-year term loans, but the average credit bureau score of 7-year auto loans is still lower than those of 3- to 5-year auto loans. The deterioration of the credit quality as the auto loan term rises from five to six years and beyond can also be seen from the credit card utilization rates at origination in the last column of Panel A. At the same time, both the origination loan amount and DTI increase 13

15 monotonically with terms, and the auto loan monthly payments are also the highest among the longest-termed loans. APR has a general downward trend as the term lengthens, and this pattern could be driven by the declining interest rate and the increasing popularity of the long term loans in recent years. We will investigate loan pricing in Section V. Overall, Panel A suggests that longer terms, and in particular, 6- year terms, are used to buy more expensive cars by consumers who are, to some extent, quite financially constrained to keep their monthly payments at a relatively manageable level. Such findings are consistent with the media comments discussed in the introduction. We break the statistics down by different lender types in Panel B of Table 2. It is clear from this panel that borrowers from finance companies are the most risky. Both banks and credit unions hold higher lending standards and this panel suggests that bank borrowers have comparable credit quality as credit union borrowers in terms of observed risk characteristics. 16 The lower credit quality of long term auto loans can further be seen in Panel B of Table 2 among all three types of lenders, and the decline in credit quality after the 5-year term is the most dramatic among finance company loans. We have also investigated the credit bureau scores of different terms by each year and the finding of lower bureau scores among 6- or 7-year auto loans is robust over time. IV. Empirical results on the risk of long term auto loans IV.A. Loan risk after controlling for observed risk factors 16 These findings are actually confirmed in kernel density plots (not reported to save space but are available upon request). 14

16 We use the discrete hazard models. A loan is deemed default if it hit 90 days past due (i.e., DPD) or more. We use loan-year panel data in the regression analysis, with a loan-year dropping out of our sample if the loan is paid off or if it hits 90+ DPD or defaults. 17 Letting P it, + 1 be the probability of default for individual i at time t, the hazard model can be written as: P = a + b + θ + g it, + 1 log Loan level variables it, loan age it, loan termi macroeconomic variables t 1 Pit, l origination year dummies (1) The dependent variable in the regression is equal to 1 if an auto loan defaults (i.e., becomes 90+ DPD) during the period from July 1, year t to June 30, year t + 1, and 0 for all other non-default loan-years. All the explanatory variables are based on values from either year t (for time-varying variables) or from origination (for non-time varying variables). The coefficient estimates of the interactive terms, θ, are the residuals after controlling for the other risk factors for that particular term-age group. We use the interactive terms to account for the possibility that the age survival functions vary for different origination terms. Further, because various lenders may behave differently, we run equation (1) separately for the three types of lenders. We report regression results in Table 3. All three columns show that mortgage holders and persons with higher credit scores are less likely to default. 18 On top of credit scores, the subprime dummy, which is equal to 1 when credits score is less than 660, 17 Note that we only model defaults in equation (1) without simultaneously accounting for prepayments. In robustness tests, we use a competing risk framework by modeling default and prepayments simultaneously, and results from such an analysis do not change qualitatively from those reported here. Therefore, our findings are not affected by prepayment. 18 We have also tried controlling for the FICO impact via step functions, where each step is separated by 10 FICO points. Results from such an alternative specification are similar to those presented here. 15

17 carries a significantly positive coefficient. Not surprisingly, a higher APR is connected with a higher probability of default in auto loans, and borrowers with higher credit card utilization ratios are more likely to default on auto loans. The coefficient estimates are significantly positive for the county-level unemployment rate and significantly negative for the Manheim used vehicle price index, both results being intuitive. The stand-alone DTI term is positively associated with the probability of default. Since we have an interactive term between DTI and mortgage holders, the coefficient of the stand-alone DTI term suggests that, for those without mortgages, the higher DTI is related to higher probability of default on auto loans. The relation between DTI and probability of auto loan defaults among mortgage holders is the combination of the coefficients of the stand-alone DTI term and the interactive term, and the derived coefficients are still statistically significant in all three columns, suggesting that DTI is also significantly related to the probability of auto loan defaults among mortgage holders. The sign for PTI is significantly positive only in the first column among banks. We find that this result is due to the inclusion of interactive terms with loan terms in the model specification. If we only include loan age dummies but no interaction with loan terms, the coefficients of PTI are statistically positive in all three columns. This is because payment increases with loan terms (as shown in Panel A of Table 2), which creates high correlation between PTI and loan terms. 19 Because of space limitations, we do not report the coefficients of the term-age interactive terms in Table 3. Instead we plot the coefficient estimates in Figure 2, for loans from banks, finance companies, and credit unions, separately. To save space, we 19 We are aware of the noise in PTI and DTI calculation. We have explored a model excluding these two variables and the results on the remaining variables are similar to those reported in Table 3. 16

18 discuss the test statistics here and these test statistics are available upon request. We find from our data that many loans still exist after the supposedly original maturity date. For example, some five year loans still show up in the data after 60 months since origination. One reason for this phenomenon is the no payment option: if the contract allows no payment for the first six months, the maturity date will be pushed back by six months. Another reason is due to loan extensions. We add one more year beyond the maturity date in Figure 2, and there are few observations after one year post the original maturity date. The omitted group in the regression in Table 3 is loans with one-year term. Panel A of Figure 2 shows that the lines of 2-year loans and 3-year loans are quite close to each other, suggesting that, even though the borrower credit quality may be lower for 2-year loans, after controlling for the observed risk factors, 2-year loans are not more risky than 3-year loans. The lines for 4- and 5-year loans lie above the lines for the 2- and 3-year loans during the first three years after origination. However, the coefficient estimates on the 2- to 5-year loan lines in Panel A of Figure 2 are largely not significantly different among themselves. This latter result suggests that during the first four years of their lives, bank loans with maturities below 5 years show similar levels of default risk, after controlling for the observed risk factors. The lines for the 6- and 7-year loans are clearly the highest in Panel A of Figure 2 during the first four years since loan origination, and test statistics show that the coefficient estimates of the 6- and 7-year loans are significantly (at the 1% significance level) higher than those of shorter-terms loans over each of the first four years. 17

19 Therefore, after controlling for the observed risk factors, the long term bank auto loans are significantly more risky than the shorter-termed loans. Most of the coefficient estimates for the interactive terms in Panel B of Figure 2 are not significantly different from zero during the first five years, and the lines for the short- and longer-termed loans cluster together during the first five years, except for the third year of 7-year maturity loans. Therefore, although auto loans from finance companies have the riskiest borrower characteristics (as shown in Table 2), Panel B of Figure 2 says that the observed risk factors can largely explain the default risk among finance company auto loans. After controlling for the observed risk factors, there is little difference in default risk among finance company auto loans with different terms. In Panel C among credit union loans, the lines for the 6- and 7-year loans always lie above the other lines, and the coefficient estimates behind these two lines are statistically significant in the first five years. This result indicates that among credit union auto loans, 6- and 7-year term loans are significantly more risky than shorter-termed loans, after controlling for the observed risk factors. To investigate whether the results in Figure 2 are robust, we run equation (1) for the sub-periods before 2009 and after We find that the finding that long term loans are more risky after controlling for the observed risk factors among banks and credit union loans are robust in both sub-periods. Therefore, results in Figure 2 suggest that long term auto loans are more risky than loans with shorter maturities among bank and credit union loans, and this phenomenon is robust both before and after the latest financial crisis. By contrast, long 20 We do not report these results in the paper because of space limitations and these results are available upon request. 18

20 term auto loans are not more risky than shorter-termed loans among finance company loans. The sharp difference between Panel B and Panels A and C suggest that the lengthy underwriting process for subprime loans (more concentrated among finance company loans) seems to help to reduce the higher default risks of the long term auto loans. This is the first piece of evidence that soft information may help to assess loan default risk. Second, the lines are clearly upward trending in Panels A and B for auto loans with terms higher than four years, and the upward pattern is the strongest among bank loans. By contrast, all lines show strong declining patterns in Panel C among credit union loans. The pattern in Panel C is intuitive and expected. Given the auto depreciation pattern and the loan amortization schedule, the loan-to-value ratios should be falling as the loans age. So the age survival function should be decreasing, and that is the main reason why lenders typically are not very concerned about auto defaults two years post loan origination. Since two-thirds of the auto loans are originated from banks or finance companies, the age survival functions in Panels A and B of Figure 2 will dominate in the overall sample, and we will observe rising default rates among more seasoned auto loans, a finding documented in Wu and Zhao (2015). 21,22 We have also conducted the same analysis during the pre2009 and post 2009 subperiods. The finding of rising default rates among older auto loans is robust among banks and finance companies, while credit union loans always show a declining age survival 21 Wu and Zhao (2015) do not investigate the reasons behind the rising default rates as auto loans age, which we do in Section IV.C. 22 Note that we do observe loan extensions in the data. To examine whether the results here are driven by these loans with extensions, we conduct a robustness test, excluding these extension loans from the sample, and find that the results do not change qualitatively. Therefore, the rising age survival function is not entirely driven by loan extensions. 19

21 function. We do not report these results in the paper because of space limitations and these results are available upon request. The divergence in age survival function among different lenders is interesting, and we will further explore the reasons behind the phenomenon in Section IV.C. Third, the coefficients of the interactive terms for the 4+ year loans are all significantly positive in Panel A. This result suggests that, after controlling for the observed risk factors, there is still a significant amount of default risk among bank auto loans that cannot be accounted for. Only a few of the coefficient estimates of the 6- and 7- year terms are significantly positive in Panel C, suggesting that the proportion of unexplained default risk may be smaller in credit union loans. Figure 2 therefore seems to suggest that bank loans are more risky than credit union loans. To formally compare these two sets of portfolios, we resort to rigorous analysis in the next subsection. We only focus on the comparison of bank and credit union loans in the next section because these two portfolios are similar in terms of observed risk characteristics. IV.B. Comparison of bank and credit union loans 1. Regression results We first run equation (1) by each loan age among the pool of bank and credit union loans with a bank dummy added and report the results in Panel A of Table 4. We run equation (1) by each loan age instead of a pooled regression including all loan ages, because the diverging age survival functions between bank and credit union loans will always drive the coefficient of the bank dummy to be significantly positive in a pooled regression. 20

22 The first column of the panel shows that the coefficient estimate turns to be significantly positive starting from the second year since loan origination. Such a result suggests that, bank loans are significantly more risky than credit union loans, controlling for the observed risk factors. The remaining two columns of the panel report results among the prime and subprime loans separately. The second columns shows that the difference in default rates between the two prime portfolios is minimal during the first three years after loan origination and the performance divergence shows up starting in the fourth year. The last column says that the difference in loan performance between the two subprime portfolios starts in the second year and becomes stronger as the loans age. We do not have space to report results on the observed loan and borrower characteristics and macro factors in Panel A of Table 4, but these results are available upon request. The most notable finding from those results is that the observed risk factors are overwhelming statistically significant in the regressions during the first four years after loan origination. However, starting from the fifth year, many coefficients of the observed risk factors (including the macro factors) become statistically non-significant, while the coefficient of the bank dummy increases in magnitude. Those results suggest that defaults among seasoned auto loans are typically not caused by the observed risk factors. We will investigate the diverging age survival function between banks and credit unions in Section IV.C. 2. Matching results We further examine the performance difference between banks and credit unions using a non-parametric matching method in Panel B, because the linear assumption in the regression framework might be too stringent. We first elaborate on our matching 21

23 mechanism. For each year in our sample, we find similar credit union loans bank loans and investigate their performance in the next 12 months. We adopt exact matches in terms of calendar year, loan age, and whether the borrower holds a mortgage. We restrict matching to be within 10 point difference credit bureau score, 5 percentage point difference in terms of credit card utilization, 1 percentage point difference in county-level unemployment rate, a difference of 0.2 in DTI, and 0.05 in PTI. Such a matching criterion yield 347,397 loans, out of a total of 452,496 bank loans, matched with credit union loans. We have also tried different cutoffs in the matching criteria. Loosening the matching criteria leads to larger matching samples with lower matching quality, while tightening the matching criteria results in smaller matching samples with better matching quality. Results from alternative matching criteria are quantitatively similar to those reported in Panel B of Table 4 and are not reported to save space but are available upon request. The unmatched bank loans are of lower credit quality with lower credit bureau scores and higher credit card utilization at origination. It is clear from Panel B of Table 4 that banks loans always have significantly higher default rates than credit union loans, and the difference in default rates between the two portfolios stays around 0.10 percentage points throughout the first five years after the loan is originated. The difference in default rates between the two portfolios jumps to around 0.83 percentage points in the sixth year and 2.90 percentage points in the seventh year. By contrast, the default rates are 0.70 and 1.53 percentage points in these two years among the matching credit union loans. These results are largely in line with those in Panel A from the regression method. 22

24 The last six columns of this table show that, based on this method, the default rate gap between the bank and matching credit union portfolios is minimal among the prime borrowers and quite large among the subprime borrowers. Further investigation (not reported to save space) reveals that the difference in default rates between banks and credit unions shrink even further among borrowers with credit scores above 700. Therefore, the results in this subsection suggests that, conditional on the same risk factors readily available from credit bureau data, bank loans are significantly more risky than credit union loans among subprime borrowers, and the inference differs on the prime auto loans, depending on the methodology used. As we discussed earlier that credit unions are likely to have more soft information via its multiple interaction with their members/borrowers, and thus the performance difference between banks and credit unions is most likely due to the additional soft information taken into account in the underwriting process of credit unions. So the evidence in Table 4 provide more evidence suggesting that soft information can help to manage default risks ex ante.. IV.C. Age survival function This section investigates further the difference in age survival function among credit unions and non-credit unions in Figure 2. The finding from section IV.B.1 that defaults among seasoned auto loans are typically not driven by the observed risk factors is also intriguing. Are these two findings related? Can both phenomena be driven by the soft information that is more difficult to capture, such as income instability or borrower s unwillingness to pay post loan origination? We do not have information to examine if bank borrowers tend to have lower income or more likely to suffer from income 23

25 instability than credit union borrowers. We examine here whether borrower unwillingness to pay might play a role in the rising age survival functions among bank and finance company loans. Note that borrower unwillingness to pay is a type of soft information that is very difficult to observe and collect. One type of unwillingness to pay is a borrower s choice to default on certain debt but keeps on making payments to her other debts. 1. Default choice between auto loans and credit cards We first explore whether a consumer defaults first on auto loans or credit cards when feeling financially constrained and has to miss payments on some of her debt obligations. Because we investigate consumer choices when financially constrained, we include in this analysis only consumers who have defaulted on either their auto loans or their credit cards, but not both or neither. The regression specification is as follows: Pit, + 1 = exp( a b Loan level variables it, l origination year dummies) (2) Note that this choice is made by the same individual, so the decision is not driven by the consumers credit quality. Therefore, we do not include individual level characteristics, such as credit bureau scores, or unemployment rate in this analysis. The left-hand-side variable is equal to 1 if the consumer defaults on credit card and 0 if the consumer defaults on auto loan. The regression results are reported in Table 5. The constant term is large and positive in all three columns, while the magnitude of the negative coefficients of the age dummies is much smaller than the constant term. Such results suggest that consumers are always more likely to default on credit cards than on auto loans, regardless of the age of 24

26 auto loans. This finding is consistent with the general understanding in the financial industry. The coefficients of the sixth and seventh year dummies in the first two columns are significantly lower than those of the other age dummies, and this result points towards an increase in the probability of defaulting on auto loans before defaulting on credit cards when the auto loan is more than five years old. Such a finding thus suggests that the default pecking order weakens as the auto loans age. By contrast, the coefficient estimates for the sixth and seventh year dummies are not significant in the third column. Therefore, when auto loans are issued by banks or finance companies, the age of the auto loan affects a consumer s decision on whether to default on the auto loan or the credit card debt. As the auto loans age, the consumer becomes increasingly less likely to default on credit cards and more likely to default on auto loans, if the auto loans are from a bank or a finance company. However, when the auto loan is from a credit union, the seasoned auto loan does not make the consumer more (less) likely to default on auto (credit card) debt. 2. Choice of which auto loan to default by the same borrower We next investigate in Table 6 borrowers choice on which auto loan to default if they can only afford to keep making payments on one or some of their auto loans but not all of them. Again, as this choice is made by the same individual, the decision is not driven by the individual s credit quality. We include in this analysis only borrowers who defaulted on one or more of their auto loans but kept their other auto loan(s) current at the same time. We calculate the differences in age, payment amount, and the outstanding balances between the loan and 25

27 the remaining auto loans of the same individual. 23 The dependent variable is equal to 1 if the consumer defaults on the auto loan and 0 otherwise. In total, we have 24,308 consumers and 55,930 auto loans in this analysis. After controlling for the difference in APR, payment amount, outstanding balance and loan age, the coefficient of the difference in the age between the loan in question and the mean age of the remaining auto loans of the same individual is significantly positive in the first two columns among bank and finance company loans, but significantly negative among credit unions loans. Note that this relation is not contaminated by nonlinearity. We have used step functions to indicate age differences by year from -3 years to 3 years, and the coefficients of the step functions increase (decrease) monotonically as the loan age difference rises among the bank and finance company (credit union) loans. Therefore, when the older loan is issued by a bank or a finance company, the borrower is more likely to default on the older auto loan if she cannot make payments on all her auto loans. However, when the older loan is issued by a credit union, the consumer is more likely to default on the newer auto loan. 3. Discussion Tables 5 and 6 both show sharp difference in consumer choice between credit unions and non-credit unions. As we have discussed earlier, credit unions may enjoy an informational advantage in on-going monitoring because their borrowers are required to have other banking relationships with them. So credit unions should be able to continue to collect valuable information on borrowers ability to pay, and thus, default due to 23 We do not include variables, such as credit scores or the homeowner dummy, in the results reported here because these variables do not show statistical significance. Such results are not surprising because all individuals included in this analysis are those that have defaulted in their auto loans, and these variables may help to predict which individuals may default but not which loan to default. 26