Financial Innovation and Borrowers: Evidence from Peer-to-Peer Lending

Transcription

1 Financial Innovation and Borrowers: Evidence from Peer-to-Peer Lending Tetyana Balyuk April 19, 2018 Abstract I study how peer-to-peer lending innovation affects credit provided by banks. I show that banks rely on certification by peer-to-peer lenders when deciding to increase credit to consumers. This credit increase is larger for borrowers who are likely more credit constrained due to shorter credit histories and lower credit scores. Changes in demand or creditworthiness cannot explain this finding. I find no evidence that increased credit access leads to more delinquencies. These results are consistent with information spillovers when multiple lenders make credit decisions sequentially. I conclude that financial technology innovation can increase access to credit even from existing lenders. Keywords: access to credit, credit rationing, financial innovation, financial technology, FinTech, consumer finance, peer-to-peer lending JEL Classification Numbers: G21, G23, D14, D45, D82 I would like to thank Pat Akey, Allen Berger, Claire Celerier, Yongqiang Chu, Peter Cziraki, Michele Dathan, Sergei Davydenko, Olivier Dessaint, Craig Doidge, Alexander Dyck, Rohan Ganduri, Mariassunta Giannetti, Will Gornall, John Hackney, Michael King, Jiro Kondo, Lisa Kramer, Marina Niessner, Aleksandra Rzeznik, Berk Sensoy and seminar participants at the SHoF FinTech Conference, AFA PhD Poster Session 2017, NFA Conference 2016, USC Marshall PhD Conference in Finance, University of Toronto, University of Lausanne, Wilfrid Laurier University, Queen s University, McGill University, Rice University, University of Houston, University of Rochester, University of New South Wales, Indiana University, University of British Columbia, Emory University, University of South Carolina, University of Notre Dame, and INSEAD for their helpful comments and suggestions. I also thank Gabriel Woo and Paul Sy from the Royal Bank of Canada and Mark Engel, Daniel Bonomo, and Dina Duhon from Scotiabank for helpful discussions on consumer credit. All errors are my own. The latest version of the paper and Online Appendix are available from Goizueta Business School, Emory University, tetyana.balyuk@emory.edu, tel. (404)

2 Savers have never had a worse deal but for most borrowers, credit is scarce and costly. That seeming paradox attracts new businesses free of the bad balance sheets, high costs and dreadful reputations which burden most conventional banks. (Economist, 2014) 1 Introduction The consumer credit market is one of the largest and most important credit markets, with outstanding credit of $3.5 trillion in the U.S. in 2015 (FED, 2016). Yet, financing frictions create distortions in this market. Sources of imperfections include information asymmetries and adverse selection (Stiglitz and Weiss, 1981), high costs of debt refinancing (Brito and Hartley, 1995), and imperfect competition (Parlour and Rajan, 2001). These frictions are indicated by certain features of this market such as high and similar rates on credit cards, and credit rationing (e.g., Stango and Zinman, 2009). This paper examines whether financial technology (FinTech) innovation can reduce information-related inefficiencies, focusing on an important but so far neglected innovation: peer-to-peer (P2P) lending. A number of FinTech-driven innovators have recently focused on entering the consumer credit market as they see frictions they can overcome and potential for profitable entry. In the P2P segment alone, FinTech lending has grown rapidly to over $16 billion in loan originations by P2P lending accounts for 6.7% of the credit outstanding in the personal unsecured loan market, and it is expected to grow to $150 billion per year by 2025 (PwC, 2015). Despite the exponential growth of these alternative lenders, little is known about their impact on borrowers. The existing literature on technological innovation and financial markets focuses primarily on information acquisition and the competition effects of improvements in screening technology (e.g., Genrig, 1998; Mishkin and Strahan, 1999; Hauswald and Marquez, 2003; Banerjee, 2005; Philippon, 2016). However, the literature provides little evidence of how FinTech affects borrowers in general and the credit supply in particular. This paper studies how borrowers are affected by FinTech innovation in the form of P2P lending. I investigate this impact by examining the credit provided by traditional financial intermediaries (e.g., banks) after consumers use P2P lending platforms. First, I examine how obtaining a P2P loan affects the consumer s access to bank credit, and I explore the heterogenous effects of this innovation 1

3 on borrowers who have different ex ante probabilities of being credit rationed. I then investigate whether changes in access to credit can be explained by changes in consumer borrowing patterns or by changes in consumer credit scores. Finally, I consider whether increased access to credit results in excess borrowing and delinquencies. The main contribution of this paper is to show that FinTech innovation can reduce financing frictions in the consumer credit market by improving the information environment. I show that P2P lending causes increased access to credit from existing lenders outside the P2P platforms. These results are consistent with banks relying on certification by P2P lenders. I focus on P2P lending because it is by far the most successful FinTech lending model in the consumer credit market. Its main innovation is the direct matching of borrowers and lenders through online lending platforms. Borrowers request a loan online, then multiple investors screen loan applications and decide how much to lend. 1 These FinTech companies have created a public market for consumer debt that is similar to the bond market and did not exist before This innovation has implications for how borrower information is processed. Another innovative feature is the use of fully automated algorithms to price and underwrite loans in order to lower screening costs. Thus, P2P lending is an innovation in lending technology rather than in the product space. When lending decisions are made sequentially and when other lenders can observe loans from FinTech companies, these improvements may lead to information spillovers from FinTech lenders to traditional intermediaries. I use application-level data from Prosper Marketplace (Prosper), one of the largest P2P lenders in the U.S. This data set contains detailed hard information on approved and rejected loan applications. The main sample is from The median applicant has a strong borrower profile (i.e., a high credit score, high income, and a long credit history) but lacks collateral and the capacity to take on more debt. While she is a prime borrower, she appears to be risky and credit constrained. Three features of Prosper s platform facilitate identification of the causal effect of P2P lending. First, Prosper tracks repeat borrowers (i.e., those who submit applications several times), which allows me to construct a panel of borrowers and observe changes in their financials after a P2P loan. 1 P2P loans are unsecured amortizing loans. The median loan size is $12,000, and maturities are 3 and 5 years. 2

4 Second, the loan amount and the interest rate are set before funding. Thus, it is funding by investors that determines whether a borrower obtains a loan. Third, a P2P loan can be extended only if investor commitments surpass 70% of the requested amount. I use this threshold for identification. The methodological approach in this paper utilizes the panel structure of data constructed from application-level observations. I make use of within-borrower variation in credit outcomes, depending on whether borrowers receive P2P loans or are rejected, in OLS regressions with borrower fixed effects. I then use the unique setting of the Prosper platform to deal with endogeneity due to the omitted variable bias. This bias may arise if borrowers who obtain P2P loans are systematically different from those who get rejected. It might be that the unobservable borrower risk is driving the results rather than P2P lending. I identify the causal effects of P2P lending by exploiting the random treatment of borrowers within the vicinity of the funding threshold in a regression discontinuity design (RDD) analysis. I use the fuzzy version of RDD because not all borrowers who cross the threshold eventually receive P2P loans, because some applications are canceled or withdrawn. 2 I first examine whether P2P lending affects access to bank credit. In a world with complete markets and no frictions, financial innovation is irrelevant, since it does not change the fundamentals. In such a setting, P2P lending does not affect the overall credit supply to borrowers, because increased lending from P2P platforms is perfectly offset by a reduction in lending by other financial institutions. This happens because borrowing from a new lender increases indebtedness and default risk, which imposes negative externalities on existing lenders. Predictions change if one allows for imperfections (e.g., Dell Ariccia, Friedman, and Marquez, 1999; Marquez, 2002). One of the most prevalent imperfections in credit markets is that asymmetric information leads to market failures, such as credit rationing when some consumers are denied credit (Stiglitz and Weiss, 1981; Bester, 1985; Arnold and Riley, 2009). If P2P lending does not change the fundamentals, existing lenders should decrease access to credit for P2P borrowers or, at best, P2P lending should not change access to bank credit if the probability of sequential borrowing is incorporated ex ante into credit decisions (as in Bizer and DeMarzo, 1992). In contrast, if P2P lending improves the information environment, 2 I conduct multiple robustness checks to support the assumptions of the RDD and the validity of the analysis. 3

5 banks may rely on a signal from P2P lenders and increase access to credit after observing the granting of a P2P loan as information spills over through multiple lending relationships. The tests below seek to differentiate between these alternative views of P2P lending. I find that not only do P2P borrowers expand their credit through this channel, but banks increase their credit supply to these customers. The key variable of interest is the quantity of credit provided by banks, as proxied by credit limits on revolving accounts (e.g., credit cards, lines of credit). I focus on revolving accounts because they allow for differentiating between the credit supply (as measured by limits) and the credit demand (as measured by balances). 3 Using fixed effects regressions, I show that P2P lending is associated with an increase in credit limits of $1,020, or 2.6% relative to the mean on the first application. The size of this effect is more than half of the effect of home ownership, and it is a nontrivial increase. In similar tests, I observe that the increase in the number of revolving accounts is much smaller (1.3%), suggesting that this increase comes primarily from existing lenders. The RDD results support the findings above. I find that P2P lending leads to a 57.7% increase in credit limits for marginally funded borrowers. A much larger increase for this category of borrowers can be explained by the local average treatment effect (LATE), since I identify this effect out of a small subsample of borrowers who differ from other P2P borrowers. While the results from this analysis are cleaner and provide for identification, the treatment is heterogeneous, and the magnitude of the coefficient should be interpreted with caution. Nonetheless, the RDD shows that the effect on credit limits moves in the same direction as in the OLS regressions. These results are novel and suggest that banks take P2P lending into account when making decisions to change access to credit. But do P2P lenders reduce information asymmetries? Given the P2P process that I describe below, it appears that they do not generate any new soft information that is not available to existing lenders. Rather, if they do improve information, it is likely the result of improved accuracy of screening based on hard data. To rule out an explanation that these results are driven not by improvement in information but by reduced screening costs, I investigate whether the effect is differ- 3 P2P loans are treated as installment debt, so they are not part of revolving accounts. 4

6 ent for borrowers who have different ex ante likelihoods to be credit rationed by banks. I find that the increase in access to bank credit following a P2P loan is larger for borrowers with shorter credit histories and with lower credit quality (5.21% increase), which is in line with the information story. Next, I examine whether these findings can be explained by changes in consumer borrowing patterns and credit scores. Since 70% of P2P loans are taken with the purpose of debt consolidation, it may be the case that banks react to this reduction in revolving credit, which improves borrower creditworthiness. The results of the regressions with borrower fixed effects show that P2P lending is associated with a 7.6% decrease in revolving balances and a 10.4% decrease in revolver utilization. This finding suggests that P2P borrowers mostly refinance their expensive credit card debt using P2P loans, which is consistent with credit repricing. 4 The RDD analysis, however, shows that P2P lending does not lead to lower revolving balances and lower revolver utilization for marginally funded borrowers. It appears that the average borrower consumes the P2P loan. Given that marginal borrowers do not decrease their revolving debt while their access to bank credit still increases, I deduce that higher credit access cannot be explained by changes in borrowing patterns from banks after a P2P loan. I also show that expanded access to credit is not driven by changes in credit scores. To provide additional evidence for the channel behind my findings, I explore whether reduced demand for bank debt differs between borrowers who have different credit scores. If P2P lending uses marginal pricing, the extent of substitution should be related to the pre-existing costs of pooling, and borrowers with the best credit quality should benefit most from the change. This is precisely what I find. The decrease in revolving balances is stronger for borrowers with higher credit quality. Finally, it may be possible that P2P lending does not improve fundamentals but rather offers a new way for financial intermediaries to take advantage of behavioral biases in borrowers decision making (e.g., Laibson, 1997). To examine this possibility, I track the total debt and delinquencies of borrowers on any credit products, including bank debt. I find that the total debt increases by around 4.5% for all borrowers, with no change for marginally funded borrowers. However, I do not find any 4 Although I do not observe wealth or consumption, I find that the total debt of borrowers with P2P loans increases. I consider the reduction in credit balances to be evidence of refinancing, since debt-financed consumption likely increases. 5

7 evidence that increased access to credit leads to higher delinquencies. This result is inconsistent with overborrowing, and it suggests that borrowers who are screened in by P2P lenders are creditworthy. The findings in this paper give rise to two main alternative explanations. The first one is that banks respond to competition from P2P lenders. I provide suggestive evidence that this is unlikely. 5 Another concern is that this paper focuses on repeat borrowers, since outcomes for one-time borrowers are unobservable. This attrition may bias the results if it were borrowers of higher credit quality who return for another loan. I show that the opposite is true. Collectively, the interpretation that best explains the totality of my findings is that FinTech innovation in the form of P2P lending gives rise to certification as information spills over through multiple lending relationships, and this innovation could reduce financing frictions in the consumer credit market. This paper has important policy implications. The recent regulatory debate calls for stricter regulation of FinTech amid concerns of lax screening and rising delinquencies in P2P lending. My results suggest that P2P lending may facilitate access to cheaper credit and generate a feedback effect on the supply of credit by banks without leading to overborrowing. This paper contributes to the growing literature on FinTech and P2P lending (e.g., Butler, Cornaggia, and Gurun, 2016; Duarte, Siegel and Young, 2012; Hertzberg, Liberman, and Paravisini, 2018; Ravina, 2013; Lin, Prabhala, and Viswanathan, 2013). To date, the literature mostly focuses on the determinants of funding on P2P platforms. In contrast, this paper explores the effect of P2P innovation on access to credit from banks. This paper also relates to the literature on financial innovation (e.g., Boot and Thakor, 1997; Hauswald and Marquez, 2003; Keys et al., 2010). Whereas most of the literature focuses on financial innovation in the product space, I examine the technology side of financial innovation. This paper shows that FinTech innovation can impact borrowers by increasing their access to credit and by reducing financing frictions. To the best of my knowledge, this is the first paper showing that FinTech innovation can alleviate personal financing constraints. 5 I address this explanation by examining borrowers who reside in localities that have different levels of bank competition, since banks have stronger incentives to respond to new entrants in less competitive markets. The effect of P2P lending on access to credit does not differ across localities, suggesting that the results are unlikely due to competition. 6

8 2 Peer-to-Peer Lending 2.1 The Innovation of Peer-to-Peer Lending Peer-to-peer (P2P) lending is often described as a financial disruptor in the consumer credit market. 6 While the question of whether P2P lending indeed disrupts traditional financial intermediaries (e.g., banks) remains largely unexplored, 7 there seems to be little disagreement about the fact that P2P lending is one of the most prominent innovations in consumer finance in the past decade, and its impact is comparable to microfinancing or the introduction of the ATM. 8 The two largest P2P lending platforms in the U.S., Prosper and Lending Club, originated $16.1 billion in loans to 1.16 million borrowers from Q to Q These numbers compare to $924.2 billion in revolving credit and $2.57 trillion in non-revolving credit outstanding in the U.S. in Q (FED, 2016). The outstanding amount of personal unsecured loans was approximately $241 billion in 2015, and P2P loans are most comparable to this non-revolving credit category. I find that P2P lending has been growing exponentially since 2013, and I estimate that the market was at least $2.75 billion in Q One of the most conservative growth expectations is that the market will grow to at least $150 billion per year by 2025 (PwC, 2015). The recent rapid growth of P2P lending in the U.S. is even more exciting, because the disintermediation of lending through P2P transactions seems counterintuitive given the perception that financial intermediation itself emerged in response to credit market imperfections (Diamond, 1984; Boyd and Prescott, 1986). To understand the role of FinTech in lending more fully, I describe the innovation of P2P lending, and I discuss its potential to resolve credit market imperfections that banks presumably could not resolve. The innovative approach to credit origination introduced by P2P lending platforms is the direct 6 Jeffery and Arnold (2014) discuss FinTech lending as a disrupting innovation that threatens traditional banking. In contrast, King (2016) considers FinTech to be both a disruptor and an enabler for banks. 7 Bachmann et al. (2011) provide a comprehensive literature review of papers published from the launch of P2P lending in 2005 to Morse (2015) gives a survey of more recent literature on P2P crowdfunding. 8 The New York Times, for example, characterizes P2P lending as a... rare thing, scarcely seen in the financial world since the debut of the A.T.M. or microfinancing: an innovation to help regular people (NYT, 2014). 9 These two platforms accounted for 98% of the P2P loan market in 2014 (Economist, 2014). The cumulative amount demanded on these two platforms over the same period was $93.9 billion in loans requested by 7.1 million consumers. 7

9 matching of borrowers and investors in consumer loans. This model is unlike banks, which pool deposits from investors and then allocate these pooled funds toward loans. By performing this direct matching, P2P platforms created a technology-driven public market for consumer debt similar to the corporate bond market, which did not exist before The main idea behind the P2P loan market was that borrowers can request a loan online and investors can crowdfund the loan by deciding whether and how much to invest. This lending process may have implications both for how information is processed and for the quality of screening. P2P lending platforms could have developed toward increasing screening accuracy either through the wisdom of the crowd, which relies on collective investor screening, 10 or through re-intermediation due to improved screening ability of the platform itself. The evolution of P2P lending platforms seems to be gravitating toward the latter direction. Although P2P lending started with unsophisticated investors as lenders, the market had attracted institutional investors by 2013 (NYT, 2014). These sophisticated investors use their proprietary algorithms to screen borrowers, and they are well positioned to have insights into credit market conditions and characteristics of local markets through leveraging additional local and macroeconomic information that is not available to banks (e.g., see WSJ, 2016). What kind of information could P2P lenders use in their pricing and screening process? Several papers report evidence consistent with additional soft information that was previously extracted by P2P investors and evidence consistent with certification mechanisms that arise endogenously in P2P lending. These papers include Duarte, Siegel, and Young (2012); Iyer et al. (2016); Larrimore, Jiang, Larrimore, Markowitz, and Gorski (2011); Michels (2012); Lin, Prabhala, and Viswanathan (2009); and Hildebrand, Puri, and Rocholl (2017). This additional soft information is no longer available in P2P lending, and investors predominantly rely on hard information in screening loan applications. Therefore, it appears that P2P lending platforms do not generate any new information that is not available to existing lenders. Rather, any improvements in screening likely come from more accurate processing of hard information. New theoretical and empirical evidence (Vallee and Zeng, 10 For example, Allen and Gale (1998) argue that public markets can be superior to financial intermediaries in providing funding, because diversity of opinion is valuable when information is inexpensive. 8

10 2018 and Balyuk and Davydenko, 2018) suggests that the participation of sophisticated investors created incentives for P2P lending platforms to increase the quality of screening over time. These improvements in information processing due to feedback from investor screening cannot be replicated by banks using their credit intermediation model. Another innovative feature of P2P lending is the use of fully automated algorithms to price and underwrite loans to lower screening costs. P2P lending platforms may offer more cost-effective borrower screening by eliminating fixed investments into the branch network and reducing variable costs by replacing loan officers with algorithms and lowering overhead costs. The automation of the entire application, verification, and funding process also lowers screening costs for investors. Anecdotal evidence suggests that the costs of P2P lending platforms are only one third of those of commercial banks (Economist, 2014). Therefore, P2P lenders may be better positioned to screen small loans that banks may not find profitable to screen. Overall, the way P2P lending platforms operate suggests that banks can learn more about borrower credit quality from observing P2P loans, and they can achieve more precise screening by augmenting their credit models with a signal from P2P lenders. 2.2 Prosper s Lending Platform This paper focuses on Prosper s P2P lending platform because its public data are much richer than similar data provided by Lending Club. 11 The platform s design also possesses several institutional features that are attractive from a research standpoint. I describe these features below. The lending process on Prosper s P2P platform, launched in February 2006, starts with a loan application. Only consumers who have a credit score (FICO) of 640 or above are eligible to apply, with some exceptions. A prospective borrower is also required to have nonzero income and a bank account, which means that P2P loans are not provided to the unbanked population. 12 The borrower requests a 11 After completing its registration with the U.S. Securities and Exchange Commission in July 2009, Prosper was obliged to make information on all loan applications public because each application listed on its lending platform is regarded a separate security (i.e., a borrower-dependent note). Prosper provides more detailed information on loan applications and outcomes of originated loans through its website. 12 This means that P2P lending does not directly affect borrowers with subprime credit quality, the poor, or the young. 9

11 specific loan amount and submits information on her income and employment. The lending platform then checks the borrower s credit history and requests her report from a credit bureau. The credit report includes information on the borrower s credit scores, the number of accounts in the borrower s name, utilization ratios, and balances on revolving, installment, and mortgage accounts. 13 Prosper verifies self-reported data for the majority of borrowers. Prosper s business model was initially an auction-type model in which prospective borrowers choose the maximum rate they are willing to pay and investors bid on loan applications listed on the platform. In December 2010, Prosper switched to a model with preset interest rates, in which investors post their commitments to fund the entire loan, or a fraction of the loan, at an interest rate set by the platform. 14 This model is similar to Lending Club s model. Prosper uses a proprietary credit risk algorithm to assign a risk measure (referred to as the estimated loss rate) to the borrower and automatically generate an application-specific interest rate at which the loan is provided, if it is originated. This interest rate appears on the listing, and it is binding for investors who fund the application since they cannot change this rate. The preset interest rate environment is a useful feature of the institutional setting, from a research perspective. This is because the equilibrium in the P2P loan market (in terms of whether a loan is provided) directly depends on funding commitments from P2P investors rather than on the outcome of negotiations between borrowers and lenders. The priced loan application is randomly allocated to one of three funding channels. The Note Channel allows high net worth retail investors to commit to purchasing all or part of the loan with a minimum investment of $25. The listing is crowdfunded for 14 days or until investors commit to funding 100% of the requested amount (there is no oversubscription), unless the listing is withdrawn There is a vast literature, however, that studies the credit behavior and credit availability of high-risk consumers. For example, see Morse (2011) for the effect of payday loans on borrowers, and see Banerjee and Mullainathan (2010) for the effect of temptations on demand for credit by the poor. 13 Consumer credit is divided into two major types: revolving and nonrevolving. Revolving credit allows consumers to borrow up to a prearranged limit and repay the debt in one or more installments. Credit card debt is an example of revolving credit. Nonrevolving credit consists of installment credit and mortgages. Installment credit is closed-end credit extended to consumers that is repaid on a prearranged repayment schedule. Examples of installment credit include motor vehicle loans, education loans, and personal loans. Mortgages are a type of nonrevolving credit secured by real estate. For more details, refer to FED (2016). 14 See Liskovich and Shaton (2017) and Wei and Lin (2016) for evidence on how this model switch affected the borrower pool, interest rates, and funding probability on the platform. 10

12 by the borrower or canceled by the platform. Along with the change in its model, Prosper introduced a partial funding option, in which a loan may originate if investors collectively commit to funding at least 70% of the loan amount. Loan applications subject to partial funding accounted for 93.5% of all Prosper listings from late December 2010 until September In April 2013, Prosper launched the whole loan program, which resulted in the creation of two additional funding channels. The Active Loan Channel allows institutional investors to actively screen loan applications and purchase 100% of the loan. 15 The Passive Loan Channel reserves the loan for sale to institutional investors who precommit to purchasing loans that have specific characteristics based on their risk return requirements. Loan applications in this channel automatically receive full funding. The existence of the funding threshold in the Note Channel is another useful feature of the institutional setting on Prosper s platform, as it allows for identification using regression discontinuity design (RDD). While I use data on loans originated in all channels in baseline regressions, I restrict my analysis to the Note Channel in the identification part of the paper. As mentioned above, not all loans that receive enough funding commitments from investors eventually originate; some are canceled by Prosper as part of the platform s screening efforts or they are withdrawn by borrowers. The lending platform may cancel a loan application and all respective funding commitments if (a) the application contains any personally identifiable or prohibited information, (b) the borrower-reported data cannot be verified, or (c) Prosper deems that the borrower risk is materially greater than the risk reflected in the interest rate. Borrowers may withdraw their application at any time before expiration, and they are required to do so if the information they submitted has changed. During the period from January 2011 to September 2015, 66.2% of applications were converted to loans, 30.0% were canceled by the platform (whether funded or not), 1.1% did not receive enough commitments from investors and expired, and 2.7% were withdrawn by borrowers. If the funded application passes the platform s screening filters, a P2P loan originates. In case of partial funding, the loan originates in the funding amount it received. Prosper reports information 15 An application that does not receive funding within 45 minutes in this channel is reallocated to the Note Channel. 11

13 on loan origination and its subsequent repayment to the credit bureau. There is an initial soft check, in which the loan application is reported to the credit bureau. However, if a P2P loan is provided, this information is submitted to the bureau. Any further loan payments or defaults are also reflected in credit bureau files. Although new loans are reflected in FICO scores with a substantial lag (see Keys, Mukherjee, Seru, and Vig, 2010, for a discussion), P2P loans are reflected separately in credit reports within approximately one month. This reporting makes information about a P2P loan promptly available to banks. Thus, banks may infer changes in the credit quality of their clients by observing a P2P loan, and they can use this information in their decision-making if it provides a valuable signal of borrower creditworthiness. Importantly, banks do not have information on whether a P2P loan was fully or partially funded, and they cannot infer any information based on how much the loan received in funding commitments. Prosper restricts repeat borrowing on its platform. Borrowers are allowed to have a maximum of two Prosper loans outstanding at any time, and the aggregate outstanding principal must not exceed the maximum loan amount allowed on the platform. 16 This likely explains the fact that the median number of applications submitted by repeat borrowers (i.e., those who apply more than once) is two. Since the purpose of the analysis in this paper relies on observing P2P borrowers over time, most of the empirical tests focus on repeat borrowers on Prosper. 3 Data and Sample 3.1 Sample Construction Prosper s data set contains anonymized application-level data from November 2005 (before its public launch in 2006) to September The uniqueness of this data set lies in its scope and richness. It contains observations on both rejected and approved loan applications, so I can contrast outcomes for borrowers who received P2P loans and those who were rejected. Prosper provides a unique identifier 16 Likewise, borrowers may not apply for an additional loan if they no longer satisfy the platform s eligibility requirements or if they fail to pay down the Prosper loan in full. 17 I used non-authenticated queries to Prosper s public application programming interface (API) to access the data. These data are also available to investors for download at 12

14 for each borrower, which allows me to track those consumers who repeatedly apply for Prosper loans, regardless of whether they receive a P2P loan on their first application or they are rejected. I use this structure of the data to construct a panel in which I can observe the cross-section of borrowers and the time series of financials for repeat borrowers. The data set includes a number of loan-specific variables, such as the loan amount and the interest rate, as well as more than 500 borrower characteristics. These borrower characteristics include selfreported information such as income and employment, and credit bureau data that are supplied at the time of the application. The minimum loan amount allowed on Prosper is $2,000, and the maximum amount is $35,000. P2P loans are fixed-interest, fully amortizing loans with maturities of 12, 36, or 60 months. (The 12-month maturity option was discontinued in 2013.) Interest rates range from as low as 5% to as high as 35%. Prosper also charges loan origination fees, and these fees are higher for lower quality borrowers. The main sample in this paper consists of data from December 20, 2010 to September 30, I exclude data from the platform s early years because the funding threshold that I use in my identification strategy was introduced after the shift in Prosper s model. This restriction also ensures that the difference in the platform s funding models is not driving the results. This restriction does not significantly affect the number of observations, given the exponential growth of P2P lending after 2013 and the relatively low loan volume in its early years. However, this sample excludes the crisis period. 18 I also restrict the sample to borrowers who have their first P2P borrowing experience after 2010, except for tests in which this restriction reduces the number of observations significantly and affects the statistical power. I extend the sample back to November 2005 for placebo tests. I do not identify the effect of P2P lending on those borrowers who never applied for a P2P loan. Since the data I use are anonymized, I cannot match them to any other borrower-specific data to explore the effect of P2P lending on these borrowers. 18 Data from the crisis period could help examine the effect of P2P lending on access to bank credit across the business cycle but for the structural break and multiple policy changes on the platform in the early years. 13

15 3.2 Descriptive Statistics Who borrows from P2P lending platforms, and how do these loans compare to traditional consumer credit products? P2P loans are small, unsecured personal loans with a median loan size of $12,000. The vast majority of these loans is taken to refinance existing debt. The proportion of loans taken with the purpose of credit card repayment or other debt consolidation has grown from under 60% in 2007 to more than 80% in 2015, both in terms of the number of originated loans and their dollar amounts. Therefore, P2P loans can be thought of as a substitute for credit card debt, perhaps because these borrowers cannot otherwise refinance their credit card balances having been denied an unsecured personal loan from banks. There is, however, no enforcement of the intended use of proceeds from P2P loans. Lenders on P2P platforms cannot enforce the requirement that borrowers use the loan proceeds for the stated purpose, and there is no other mechanism to ensure accountability. Thus, it is unclear whether P2P borrowers indeed use the proceeds for debt repayment. The median interest rate on P2P loans including fees (APR) is 17.1%. The interest rate is the marginal interest rate, i.e. it is application-specific and is based on the credit quality of a loan applicant at the time of application. Loan APRs range from 6.4% to 36.0% (Figure 1). These rates compare favorably to the credit card rates, which increase even more after a missed payment, according to anecdotal evidence. The pricing of P2P loans is much more granular that the pricing of credit cards, where borrowers of diverse credit quality are generally charged similar rates. A significant portion of P2P borrowers are paying interest rates well above 20%, which are presumably higher than the rates they could receive on personal bank loans if they were not credit rationed. This suggests that while some borrowers may be using P2P lending because it is a less expensive alternative to bank credit, the platform also attracts borrowers who are credit constrained. The funding rate of applications is 96.6%, but only 64% of applicants obtain P2P loans, mostly because they are screened out by the lending platform (Panel A of Table 1). I classify 18.1% of borrowers as repeat borrowers, i.e., they submit more than one loan application to the platform. I observe a median number of two loan applications for repeat borrowers. 14

16 A representative borrower on Prosper has a strong borrower profile (Panel B of Table 1). Based on her credit score (FICO), 19 she is considered a prime borrower. The median FICO across loan applications is 690, which is close to the minimum eligibility requirement on Prosper, although this still places P2P borrowers in the prime category. The median ScoreX, another aggregate measure of borrower creditworthiness, is 713. The FICO distribution on the lending platform is skewed toward riskier types of borrowers compared to the distribution of borrowers in the general population (Figure 2). This suggests an adverse selection of consumers to P2P lending platforms, as the borrower pool is likely dominated by borrowers who were credit rationed by banks, i.e., they were either denied credit or were given lower access to credit than they demanded. The mean annual income of Prosper borrowers is $73,889, which is more than twice the per capita annual income of $30,176 in the U.S. (BLS, 2014). The median annual income of Prosper borrowers is $63,000. A representative P2P borrower has been employed for 6.5 years with her current employer, and she has 17 years of credit history. This evidence suggests that she is a high-income but young individual. If most of her income comes from self-employment, she may lack the documentation necessary to obtain a bank loan. Despite high credit scores and employment, P2P borrowers seem to lack collateral and the capacity to take on more debt. The median borrower on Prosper does not own a home. The mean mortgage balance of P2P borrowers is $103,700, and the median balance is $18,200. This is much lower than the mean and median mortgage/home-equity line of credit (HELOC) balances of $156,400 and $116,000, respectively, in the general population (SCF, 2013). This suggests that P2P borrowers, on average, lack collateral for a HELOC that can be used to refinance existing debt. This may also indicate that an average P2P borrower is not eligible for a mortgage based on criteria other than her FICO score. A representative borrower has a high debt burden. Her median total debt balance is $93,400, which is larger than the median household debt of $60,400 in the general population (SCF, 2013) and thus is much higher than the total debt carried by a median individual in the U.S. Evidence presented in Panel C of Table 1 is also in line with the conjecture that P2P borrowers 19 The FICO score is the most used and most reliable measure of borrower credit quality, although it is a coarse measure. For this reason, I focus on FICO in this paper. However, since FICO is not available in the data throughout the entire sample period, I also report specifications that instead include ScoreX in baseline tests, for comparison. 15

17 may be credit rationed by banks. The debt structure of a representative Prosper borrower contains a large proportion of revolving debt, which is usually unsecured debt with high interest rates, such as credit card debt. For every nine open accounts that P2P borrowers have at the time of application, seven are revolving accounts. A representative borrower utilizes her revolving accounts aggressively. Her median revolver utilization is 48% and her median credit card utilization is 56%. 20 A comparison of the revolving balance to the monthly debt payment suggests that a large portion of revolving debt is rolled over to subsequent periods rather than paid down. The last two columns of Table 1 report mean and median statistics on repeat borrowers. While most differences between the loan applications submitted by one-time and repeat borrowers are economically very small or statistically insignificant, a few notable differences are worth mentioning. Repeat borrowers appear to be riskier from the lending platform s standpoint as the platform assigns a higher estimated loss rate and hence assigns a higher interest rate to first-time applications by repeat borrowers. The platform also seems to screen out more of these applications, presumably because these borrowers appear riskier, as the probability of loan origination to borrowers who subsequently come back is almost twice as low. Investors are also somewhat less likely to fund these applications. Together, this evidence suggests that risker borrowers are more likely to come back for another P2P loan (I discuss this issue, along with a possible selection issue, in Subsection 7.2). 4 Empirical Methodology 4.1 OLS with Fixed Effects I first estimate an OLS regression model with borrower fixed effects to show the impact of obtaining a P2P loan on credit outcomes for the entire sample in the data panel that I construct. This allows me to make use of within-borrower variation in access to bank credit, and it allows me to compare outcomes for borrowers who receive P2P loans to outcomes for borrowers who were denied a loan. I proxy for access to bank credit with limits on revolving accounts. I use revolving accounts since 20 Credit utilization is the ratio of credit balances to credit limits. 16

18 they allow separation between the credit limit set by banks, which proxies for credit supply, and the credit balance, which measures the actual usage of revolving credit by consumers and proxies for credit demand. This separation is not possible for other types of debt, such car loans or mortgages, because the balances I observe are the outcome of the partial equilibrium between the supply and demand in the market for each of these credit products. The main OLS regression with borrower fixed effects measures the marginal impact of obtaining a P2P loan on access to credit from banks after each loan application. The baseline specification is Y ist = βp2p loan is,t 1 + X ist ζ + α i + γ st + u ist, (1) where Y ist represents revolver limits, P2P loan is,t 1 is the lagged indicator for a P2P loan, X ist is a matrix of controls, α i is borrower fixed effects, and γ st is state year fixed effects. The regressor of interest P2P loan is,t 1 is an indicator variable that takes the value of 1 if a P2P loan is originated on the preceding loan application and 0 otherwise. This variable is set to zero for the first application. I control for income, length of employment, home ownership, debt-to-income ratio, and the credit score (FICO or ScoreX) bins. I choose these former controls in addition to the credit score since they represent variables that are typically not updated in the credit bureau files. Therefore, they can be complementary to the credit score in assessing borrower risk, and thus are likely not part of the credit score. I use contemporaneous controls because they better capture the credit quality of borrowers at a particular point in time. I use borrower fixed effects to absorb any time-invariant unobservable borrower characteristics. I use state year fixed effects to remove the effect of local economic conditions or credit cycles in a given year that may be driving the results. The use of states in the definition of fixed effects is warranted because some individuals in the same move from one state to another, and there may be differences in consumer credit markets across states because of local supply and demand factors or state-level regulation. As discussed above, the median repeat borrower on Prosper s platform submits two loan applications, and there is significant attrition in the data panel, with the sample becoming more selected 17

19 after the second application. This is because submitting another application is a choice variable of consumers. 21 In order to mitigate concerns that this sample attrition may be affecting the results, I restrict the analysis to the first and second applications. I relax this constraint in robustness checks. I also check the effect of alternative definitions of the regressor of interest and alternative regression specifications on the results. 4.2 Regression Discontinuity Design (RDD) Endogeneity One concern with the OLS regressions is the potential endogeneity bias that arises due to unobservable borrower risk. This risk may be correlated with consumer credit variables (e.g., credit limits) or with loan origination in P2P lending markets. Banks infer the credit quality of borrowers based on hard and soft information, and they determine the level of credit supply to consumers based on borrower risk. The P2P lending platform also uses borrowers characteristics (including credit variables) when assessing the credit risk of each loan application and when setting the interest rate. In turn, P2P investors make funding decisions based on both the interest rate offered by the platform and their own screening algorithms that incorporate consumer credit variables. The unobservable borrower risk creates a potential omitted variable bias, since it determines whether a borrower receives a P2P loan as well as access to credit from banks. While borrower fixed effects remove time-invariant unobservable differences in borrower risk, any such differences that are time varying cannot be captured by fixed effects. This potential endogeneity may impair the ability of an econometrician to establish causality between P2P lending and credit outcomes for borrowers. I attempt to overcome this issue with regression discontinuity design (RDD) that focuses on the subsample of marginally funded borrowers in which I can identify the causal effect of P2P lending. 21 Selection due to attrition arises because fewer and fewer borrowers return to Prosper because the application number increases and because outcomes for those who do not return are unobservable. 18

20 4.2.2 Local Fuzzy RDD I use the setting in the Note Channel on Prosper in which loan applications are crowdfunded by retail investors and are subject to the partial funding threshold to identify the causal impact of obtaining a P2P loan on access to bank credit. The idea relies on exploiting the way loan applications are funded in this channel: As the application remains outstanding on the platform, each investor can view the application characteristics and contribute to its funding with a minimum contribution of $25. The percentage of the requested loan amount funded by investors at the time of listing expiration (14 days maximum) determines whether the P2P loan can be originated. This is used as a forcing variable in regression discontinuity design (RDD) analysis. I call this variable percent funded, and I use the 70% funding threshold in the Note Channel to examine borrowers who ended just above and just below the threshold at listing expiration. I use the fuzzy version of RDD (e.g., Porter, 2003), since some applications that cross this threshold do not result in loan origination because they are withdrawn or canceled (i.e., in the sample of borrowers above the threshold, some are compliers and some are non-compliers). I instrument P2P loan origination with an indicator variable for the loan application crossing the funding threshold. Focusing on borrowers in the near vicinity of the threshold is close to a randomized experiment if borrowers whose applications just cross the threshold simply get lucky. Consider a P2P investor who observes two similar loan applications that have funding commitments from other investors. Let us also consider that two loan applications are just below the funding threshold when the listing is about to expire, and the P2P investor has the time, or the funds, to invest in only one of these applications, contributing a small fraction of the total. Thus, funding from this pivotal investor brings one application across the threshold but not the other. As a result, only one application is converted to a P2P loan. If there are many small investors who contribute to funding loans in the Note Channel, the RDD will give estimates of the causal effect of receiving the loan. I hand-collect data on the number of investors per loan from Prosper s sales reports to verify my assumption about 19