The impact of credit scoring on consumer lending

RAND Journal of Economics Vol. 44, No. 2, Summer 2013 pp. 249?274

The impact of credit scoring on consumer lending

Liran Einav Mark Jenkins and Jonathan Levin

We study the adoption of automated credit scoring at a large auto finance company and the changes it enabled in lending practices. Credit scoring appears to have increased profits by roughly a thousand dollars per loan. We identify two distinct benefits of risk classification: the ability to screen high-risk borrowers and the ability to target more generous loans to lower-risk borrowers. We show that these had effects of similar magnitude. We also document that credit scoring compressed profitability across dealerships, and provide evidence consistent with the view that credit scoring may have substituted for varying qualities of local information.

1. Introduction

Over the last two decades, consumer lending has become increasingly sophisticated as lenders have moved from traditional interview-based underwriting to a reliance on data-driven models to assess and price credit risk. This article presents a snapshot of this transition. We describe the magnitude and channels by which the adoption of credit scoring affected loan originations, repayment and defaults, and profitability at a large auto finance company. Although the study, by design, is focused on a single company, and its experience surely has idiosyncrasies, we suspect that many of our findings may be illustrative of similar transitions at other companies, which taken together have revolutionized markets for consumer credit.

As late as the early 1990s, most lenders were still using a single "house rate" and relied on interview procedures to screen borrowers (Johnson, 1992). As data storage and computing costs

Stanford University and NBER; leinav@stanford.edu, jdlevin@stanford.edu. University of Pennsylvania; mjenk@wharton.upenn.edu. We thank Luke Stein for excellent research assistance, Will Adams for his contributions to this project, and two anonymous referees, the editor, Chris Knittel, Ulrike Malmendier, Vikrant Vig, and seminar participants at IO Fest 2008 at Stanford, the 2009 AEA annual meeting in San Francisco, the 2011 AEA annual meeting in Denver, and the 2009 NBER IO summer meeting in Cambridge, Massachusetts, for helpful comments. We acknowledge support from the Stanford Institute for Economic Policy Research, the National Science Foundation (Einav and Levin), and the Center for Advanced Study in the Behavioral Sciences (Levin). Earlier drafts of this article were circulated with the title "The Impact of Information Technology on Consumer Lending."

Copyright C 2013, RAND.

249

250 / THE RAND JOURNAL OF ECONOMICS

fell, and underwriting technology improved, lenders increasingly began to use estimates of default risk to price individual loans. Today, automated credit scoring has become a standard input into the pricing of mortgages, auto loans, and unsecured credit. Using data from the Survey of Consumer Finances, Edelberg (2006) documents the extent of this transformation. She finds that as a result the correlation between loan pricing and estimated and realized default risk has sharply increased. Grodzicki (2012) documents a similar pattern in the credit card industry and ties it specifically to lenders' investments in information technology. Other articles provide related although more indirect evidence of these effects in the context of small business lending by banks (Frame, Srinivasan, and Woosley, 2001; Petersen and Rajan, 2002; Akhavein, Frame, and White, 2005).

These studies rely either on aggregated data or survey measures of realized loans that allow us to see how the correlation of interest rates and default risk has increased over time. However, whereas the near-universal adoption of credit scoring techniques indicates their value to lenders, there is relatively little specific evidence on exactly how the benefits are realized, the size of the effects, and their organizational impacts. By focusing more narrowly, we are able to complement existing studies by using detailed applicant- and loan-level data to identify the specific channels by which credit scoring impacts loan originations and outcomes, as well as the magnitude of these effects.

We begin in Section 2 by describing the setting of our case study. The data come from an auto finance company that specializes in the low-income, high-risk consumer market. The market is particularly well suited for studying informational problems facing lenders. Default risk is high and recovery values are low, so profitability hinges on identifying better risks in the applicant pool (Adams, Einav, and Levin, 2009; Einav, Jenkins, and Levin, 2012). Loan applicants also vary substantially in their risk of default, and their characteristics and credit histories provide prospective information about this risk. The potential value from stratifying borrowers can be seen in the fact that the top third of borrowers in terms of predicted risk are about 20 percentage points more likely to default than the bottom third.

Until 2001, the company relied on uniform loan pricing and traditional interviews to screen borrowers. The company then contracted with an external credit scoring company that used credit bureau reports and historical data from the company to provide estimates of default risk that could be used to price loans. Starting in June 2001, the company shifted to a centralized risk-based pricing regime, in which new loan applicants were assigned a credit score, and the score determined the minimum down payment required for purchase and the set of cars for which financing would be available. Our empirical analysis in this article focuses on describing the short-run effects of this change, using applicant-level and loan-level data about loans originated a year before and a year after the date when credit scoring was implemented.

In Section 3, we present and calibrate a stylized two-period model, which helps guide our subsequent empirical approach. The model illustrates two distinct responses that result from being able to classify applicants as higher or lower risk. When faced with a high-risk applicant, the lender optimally increases the down payment and reduces the quality of the car, and thus the loan amount. Both effects lead to a fall in the probability of sale and a rise in the repayment rate. When faced with a lower-risk applicant, the lender optimally lowers the down payment and raises car quality, increasing the probability of sale and the amount of credit extended. In each case, the profit per loan and overall expected profit increase. These results motivate us to focus on the heterogeneous effect of credit scoring across applicant pools of different risks.

In Section 4, we present the empirical analysis. The availability of detailed transaction-level data from before and after the adoption of scoring allows for a straightforward empirical approach. We first classify potential borrowers by assigning each loan applicant to a credit category using a rule that mirrors the lender's assignment following adoption. We then construct measures of profitability and related performance metrics--"close rates" on auto purchases, car choices, financing decisions, repayment behavior and recoveries--and compare these metrics, both on aggregate and for the stratified groups, before and after the advent of credit scoring. Finally, we translate the changes into dollar terms by decomposing profits into separate components: the

C RAND 2013.

EINAV, JENKINS, AND LEVIN / 251

probability the applicant becomes a borrower, the size of the investment in each borrower, and the return in terms of loan payments actually made.

We find that the adoption of credit scoring, and the changes it enabled in lending increased profits by roughly 1,000 dollars per loan. The effect is substantial: at the time, the average loan principal was around 9,000 dollars. We also analyze an alternative measure of profitability, the profit (or "net revenue") per loan applicant. After the adoption of credit scoring, loan originations fell, but the profit per applicant still increased, from $751 to $1,070, or by roughly 42%.

Consistent with the theoretical model, we identify two distinct channels through which better information improved loan profitability. First, credit scoring allowed the lender to set different down payment requirements for different applicants. High-risk applicants saw their required down payment increase by more than 25%, creating a higher hurdle to obtain financing. Close rates for this group fell notably, and also default rates, consistent with the idea that higher-risk borrowers were screened out by the higher down payment requirement. Translating this into dollar terms, we find that improved loan repayment was largely responsible for what we measure to be about a 1,200 dollar increase in profit per high-risk loan.

We estimate a similar increase in profitability for lower-risk loans, but the mechanism is different. Required down payments and close rates changed little for lower-risk applicants. Instead, consistent with the model, we observe that car quality and average loan sizes increased substantially. Default rates did not change much, and hence the larger loans had a substantial profit impact due to the high interest rates charged in this setting. For lower-risk loans, the increased "size" of each investment is largely responsible for the dollar increase in profit. Hence, the two channels through which credit scoring theoretically increases profitability in the model both appear to be operative and substantial in the data.

A useful feature of the episode we study is that most salient features of the lending environment, such as advertising, car pricing, sales force incentives, and the composition of the applicant pool, remained stable during the periods before and after credit scoring was adopted. This makes for a relatively clean observational setting. At the same time, concerns about identification can be raised for any before-and-after study, and given that we compare outcomes before and after a single change in company policy, we cannot rule out definitively that there was some underlying confounding change in the environment. A variety of robustness checks, however, support the interpretation we have outlined. In particular, we show that the inclusion of controls for applicant quality and local economic conditions has little effect on any qualitative conclusions one might draw. Our conclusions about the effects of down payment requirements and loan sizes are also consistent with results in Adams, Einav, and Levin (2009) and Einav, Jenkins, and Levin (2012), which use data from the same lender but rely on more recent data and a different identification strategy that relies on sharp changes in pricing schedules for different groups of loan applicants.

The last section examines the differential impact of credit scoring across dealerships in order to gauge its organizational implications. Research by Stein (2002) and others suggests that automated loan underwriting might involve a trade-off, with the increased use of "hard" information crowding out the production and use of "soft" information (see also Berger et al., 2005). This line of thinking indicates that credit scoring might reduce profitability differences across dealerships, particularly if, in the absence of scoring, dealers differ in their ability to tailor loan terms to buyers.1 We show that prior to credit scoring, there was in fact dramatic variation across dealerships in profitability, related primarily to differences in default rates and the matching of cars to borrowers. The advent of credit scoring compressed this variation, as one might expect from the increased reliance on companywide guidelines. Although almost all dealerships became more profitable, the relative improvement was greater for dealerships that had higher default rates and less pronounced matching of cars to borrowers of different risks, the two dimensions that credit scoring tried to address.

1 Bloom et al. (2011) provide an interesting analysis of the multiple possible effects of information technology adoption on organizations.

C RAND 2013.

252 / THE RAND JOURNAL OF ECONOMICS

2. Data and environment

The lending environment. The company we study specializes in making auto loans to consumers with low incomes or poor credit records. During the period we study, the company's average loan applicant had an annual household income of around 28,000 dollars, which would put him at around the 33rd percentile in the United States (Current Population Survey, 2001). Almost a third of the applicants had no bank account, and only 14% owned their own home. A large majority of loan applicants had a FICO score below 600, which is the 35th percentile in the U.S. population and would not qualify for a prime mortgage. Low FICO scores frequently reflect a history of loan delinquencies or defaults, which is consistent with the credit histories of the loan applicants in our data. Over the six months prior to their loan application, more than half of the company's applicants were delinquent on at least 25% of their debt. This type of credit history makes it highly unlikely the applicants in our data could obtain a standard "prime" auto loan.

The lending process in the market operates as follows. Consumers fill out an application when they arrive at a dealership. They work with a sales representative and the dealership manager to select a vehicle and discuss financing terms. About 40% of the loan applicants we observe purchase a car. The purchased cars typically are five to seven years old, with odometer readings in the 65,000 to 100,000 mile range. The average sale price is 8,000 or 9,000 dollars, which represents a notable markup over the dealer cost (see Table 1). Buyers are required to make a down payment but usually finance about 90% of the purchase price. The financing terms are relatively standard across our sample. Buyers are expected to make regular payments at the dealership for a fixed term, typically around three years, and interest rates are high, reflecting the risk of the borrower pool. Annual interest rates average close to 30% in our sample.

A central feature of the market is that consumers tend to be tightly cash constrained. In earlier work, we use abrupt changes in the pricing schedule to estimate demand elasticities (Adams, Einav, and Levin, 2009). A striking finding was that every hundred dollar increase in the minimum down payment reduces the purchase probability of an applicant by two to three percentage points. Moreover, more than 40% of buyers pay exactly the minimum amount down, and these "marginal" purchasers represent substantially worse default risks than buyers who pay more than the minimum down (Einav, Jenkins, and Levin, 2012).

The role of the down payment in screening out marginal buyers is important for understanding how risk-based pricing affects loan originations. In the period prior to the adoption of credit scoring, all buyers were required to make a down payment of at least 600 dollars. After credit scoring was put in place, minimum down payments were held constant or even modestly decreased for lower-risk borrowers but increased to as much as 1,500 dollars for high risks. As we will see, this increase helps explain why the fraction of applicants purchasing a car, and the subsequent default rate, fell in the period after credit scoring was adopted.

As can be seen in Table 1, defaults during the repayment period are common and tend to occur relatively early in the repayment period. About 35% of loans default during the first year of repayment. Less than 40% are repaid in full.2 Following a default, the lender attempts to recover the car, and generally succeeds, but frictions in the recovery process result in a relatively low dollar value of recoveries after expenses are netted out (Jenkins, 2010). The average recovery in our sample was around 1,200 dollars, or around 25% of the original dealer cost of the car prior to the transaction.3

The combination of early defaults and low recoveries means that transaction outcomes have a bimodal pattern. Early defaults tend to result in losses, whereas fully paid loans can be quite

2 These are significantly higher default rates than those reported by Heitfield and Sabarwal (2004) in their study of securitized subprime auto loans, reflecting the relatively poor credit quality of the borrowers in our sample even compared to other subprime populations.

3 This is for several reasons. In more than a quarter of defaults, for instance, it is hard to find the borrower, leading to a lengthy and costly recovery process. About a third of defaults are directly associated with a decrease in car value, such as mechanical breakdowns, car theft, and accidents (without maintaining appropriate insurance). See Jenkins (2010) for more details.

C RAND 2013.

EINAV, JENKINS, AND LEVIN / 253

TABLE 1 Summary Statistics

January?December 2000

July 2001?June 2002

Standard

Standard

Mean Deviation 5%

95% Mean Deviation 5%

95%

Applicant characteristics Applicant demographics

Monthly income Residual monthly income Debt-to-income ratio Car purchased Transaction characteristics Buyer characteristics Monthly income Residual monthly income Debt-to-income ratio Car characteristics Car cost Car age (years) Odometer (miles) Inventory age (days) Lot age (days) Purchase characteristics Sale price Down payment Loan term (months) APR Monthly payment Loan performance Outcomes Default Fraction of payments made Loan payments excluding

down payment Recovery (all sales) Recovery (all defaults) Components of profits Gross operating revenue Total cost Net operating revenue

2,214 1,715 0.26 0.43

2,319 1,723 0.32

4,954 6.4 88,668 68 40

8,370 740 34.1 0.288 362

0.67 0.57 6,113

691 1,032

7,557 5,810 1,746

N = 1.00

973

1,204

985

748

0.16

0.03

N = 0.43

4,000 3,525 0.48

2,256 1,843 0.25 0.37

973 1,079 0.13

1,300 753 0.15

4,088 3,800 0.49

2,410 1,859 0.32

863 1.8 17,822 62 57

3,571 4

57,746 13 1

6,346 9

113,856 178 145

5,273 5.5 81,810 72 43

930 451 3.0 0.019 65

6,907 200 30.0 0.259 298

9,795 1,500 37.0 0.299 421

9,368 1,003 36.6 0.284 374

0.37 3,916

951 999

3,530 965 3,401

0.62

0.05

1.00

0.59

653 11,837 7,146

0

2,530 923

1

2,848 1,483

2,284 4,301 -3,434

12,706 7,378 6,144

9,084 6,193 2,891

N = 0.88

975 1,024 0.12

1,238 824 0.10

4,000 3,750 0.45

N = 0.32

984 1,122 0.10

1,360 790 0.16

4,286 4,018 0.47

1,015 1.7 18,048 63 58

3,717 3

50,242 13 1

6,944 9

108,381 184 152

1,297 502 3.9 0.026 42

7,307 600 32.0 0.219 306

11,495 1,900 42.0 0.299 442

0.37 4,441

1,216 1,243

3,901 1,099 3,727

0.06

1.00

766 13,636

0

3,224

73

3,665

3,013 4,518 -3,005

14,744 8,012 7,620

Note: Residual monthly income = Residual monthly income after debt payments. To preserve confidentiality of the company that provided the data, the number of observations is normalized by the number of applicant in year 2000, N (N >> 10,000). Loan payments, recovery amount, gross operating revenue are in present value (PV). Total cost includes car cost, taxes and fees, and shortfalls when value of trade-in does not cover down payment. Net operating revenue equals gross operating revenue minus total cost.

profitable. Figure 1 documents this pattern by showing the distribution of transaction-level returns. For each sale, we computed the present value of borrower payments--the down payment, loan payments, and recovery in the event of default--discounted back to the date of sale. We use a 10% discount rate, which seems to be in line with industry standards. Neither the calculation here nor similar calculations later in the article are very sensitive to using a somewhat higher or lower number.4 We then divided the present value of borrower payments by the dealer cost of the car, providing an overall rate of return on each transaction. The striking bimodal distribution of returns presented in Figure 1 illustrates the benefits of being able to identify the more creditworthy applicants from those who are relatively more likely to default.

4 Specifically, we ran all the analyses using discount rates of 5% and 15%, and the results hardly change.

C RAND 2013.

 ................
................

In order to avoid copyright disputes, this page is only a partial summary.

Google Online Preview   Download