Judging Borrowers by the Company They Keep: Social Networks and Adverse Selection in Online Peer-to-Peer Lending

Transcription

1 Judging Borrowers by the Company They Keep: Social Networks and Adverse Selection in Online Peer-to-Peer Lending Mingfeng Lin Siva Viswanathan Decision, Operations, and Information Technologies Department Robert H Smith School of Business University of Maryland, College Park, MD N. R. Prabhala * Finance Department Robert H Smith School of Business University of Maryland, College Park, MD August 2009 ABSTRACT We study the online market for peer to peer (P2P) lending in which individuals bid on unsecured microloans sought by other individual borrowers. Using a large sample of consummated and failed listings from the largest online P2P lending marketplace - Prosper.com, we test whether social networks lead to better lending outcomes, focusing on the distinction between the structural and relational aspects of networks. While the structural aspects have limited to no significance, the relational aspects are consistently significant predictors of lending outcomes, with a striking gradation based on the verifiability and visibility of a borrower's social capital. Stronger and more verifiable relational network measures are associated with a higher likelihood of a loan being funded, a lower risk of default, and lower interest rates. We discuss the implications of our findings for financial disintermediation and the design of decentralized electronic lending markets. Corresponding Author: Siva Viswanathan can be reached at sviswana@rhsmith.umd.edu. The authors thank Ethan Cohen-Cole, Jerry Hoberg, Dalida Kadyrzhanova, De Liu, Vojislav Maksimovic, Gordon Phillips, Kislaya Prasad, Galit Shmueli, seminar participants at the 2008 Summer Doctoral Program of the Oxford Internet Institute, Oxford University, participants at the 2008 INFORMS Annual Conference, as well as participants at the 2008 Workshop on Information Systems and Economics in Paris, France, for their valuable comments and suggestions. Mingfeng Lin is also thankful to the Ewing Marion Kauffman Foundation for the 2009 Dissertation Fellowship Award, and to the Economic Club of Washington D.C. (2008) for their generous financial support. We also thank Prosper.com for making the data for the study available. The contents of this publication are solely the responsibility of the authors.

2 1. Introduction The ability of online markets to efficiently bring together buyers and sellers has transformed businesses, spawned several success stories, and redefined the roles of traditional intermediaries. In this paper, we study the online market for peer to peer (P2P) lending, where individuals make unsecured microloans to other individual borrowers. This market was virtually nonexistent in 2005 but has experienced significant growth since then with the biggest of them, Prosper.com, logging over 200,000 listings seeking $1 billion in funding since its inception. We study a large sample of consummated and failed listings from Prosper.com. We examine the relation between lending outcomes and online social networks, focusing on the potentially different effects of structural and relational aspects of social networks. We show that the social capital generated by relational aspects of networks matters. It mitigates adverse selection in the P2P lending and generates better lending outcomes, provided it is visible to and verifiable by potential lenders in the marketplace. We analyze lending decisions made by individuals in the largest online P2P network, Prosper.com. This is a highly decentralized marketplace in which there is little opportunity for face-to-face contact between potential borrowers and lenders, so the asymmetric information problems in conventional lending are amplified in this setting. Our study examines whether social networks alleviate information asymmetry, and if so, what aspects of these networks help. We observe two types of networks in the prosper setting. Borrowers can become members of a group, an affiliation that cannot be undone before a loan has been repaid in full. Additionally, a unique feature of the online P2P lending marketplace is that participants can create and maintain social networks of friends. Prosper.com members can be friends with other members. Alternatively, a member can invite non-members to join the online network and become a friend. The friends generated by a borrower can endorse a borrowing request and the endorsement is visible to other bidders for a loan. A friend may also choose to not endorse a borrower and the number of friends who do not endorse a borrower is visible to potential bidders. Additionally, we can infer the roles played by the friends of a particular borrower. We can observe, for instance, whether a friend has elected to undergo verification as a lender on prosper.com. If a friend is registered as a lender, we can also observe whether the friend has bid on the borrower's listing or not. We also extract historical data on whether the friend has bid on other listings in the past or not and whether these bids were successful. The nature of a borrowers social network, graded by the nature of the roles on Prosper.com and on the particular borrower's listing, constitutes the focus of our study. We examine whether the social network variables are related to the probability of a listing becoming successfully funded, the interest rate on the listing, and the ex-post loan default hazard rates. The theoretical motivation for our tests comes from research on social capital, social networks and economic outcomes. The broad topic of social capital has foundations in the field of sociology but has been of considerable interest outside, especially in economics, finance, and management. Social networks have been of particular interest in this context. As Durlauf and Fafchamps (2004) suggest, networks are most promising avenue for disentangling the effects of social capital on economic exchange. Research on the economic impact of social networks emphasizes the distinction between their structural nature and their relational aspects, which can be further classified as either arm's-length or strongly embedded depending on the extent to which economic relationships are intertwined into the social network. Granovetter (1972, 2005) points out that both forms of social relationships have the potential to create value. Weaker arm's length relationships create value because they provide the entities with access to diverse and heterogeneous sources of information and resources. On the other hand, stronger embedded relationships also create value, by facilitating the sharing and transfer of private resources and promote self-enforcing governance. The relative importance of structural and relational aspects of social networks that are formed online in P2P lending markets, and the importance of embedded relations in influencing positive lending outcomes are the empirical questions examined here. Our sample covers 205,132 listings on Prosper.com from January 2007 to May 2008 seeking an aggregate amount of $1.7 billion. Of these, 16,500 listings for $114 million are successfully funded. We first analyze the probability that a listing seeking a loan is funded. In the probit model, the hard credit variables have the expected signs. Borrowers with better credit grades are more likely to be funded, as are less stretched borrowers who have lower

3 debt to income ratios and borrowers with fewer credit inquiries in the six months prior to a listing. Several other control variables also yield results of interest in their own right. Borrowers under the jurisdiction of usury laws limiting maximum interest rates are less likely to attract funding. The maximum rate a borrower is willing to pay has an interesting non-linear effect. Both low and high maximum rates are less likely to be funded. Low rates indicate low profitability for potential lenders, while the willingness to pay high rates is interpreted as a sign of financial strain, leading to greater likelihood of unfunded listings. The quadratic shape is remarkably consistent with the theoretical model of credit rationing by Stiglitz and Weiss (1981). Likewise, we find that bankcard utilization also displays quadratic curvature. Our findings line up well with received economic theory, even ones such as credit rationing that involve some degree of extrapolation and sophistication. It is interesting that economic theory works reasonably well even in a marketplace where lenders making loans are not sophisticated institutions but atomistic investors putting small sums of money to work. The main question pursued here is the relation between social network variables and lending outcomes. We find little evidence that the structural aspects of a borrower's network, measured as the centrality, betweenness, and closeness, matter for lending outcomes. On the other hand, relational aspects matter and are consistently statistically significant. In particular, there is a striking gradation effect of verifiability and visibility to potential lenders. Figure 1, which maps the hierarchy of a borrower's social network, illustrates our empirical strategy. At the top level is the total number of friends of a borrower regardless of who these friends are. The next level distinguishes between friends with or without a role. Those with a role actually signed up as a lender or borrower on Prosper.com, a step that requires individual screening including verification of personal details such as the social security number, the address and the individual's bank accounts. Further down the hierarchy, we can distinguish between persons who actually participate in Prosper.com as lenders or borrowers. At a finer level, we can observe whether a friend participates in bidding for the specific loan on hand, and at a still finer level, we can observe whether the friend wins in the listing or not. The lower the level in the hierarchy in Figure 1, the greater the relation specific social capital verifiable and visible to lenders on the Prosper network. When we distinguish a borrower's friends on the basis of the roles, identities, and actions, the stronger a tie and the more verifiable and visible it is to lenders, the greater the probability of attracting funding. As interestingly, nonactions by particularly verifiable relationships lead to less favorable outcomes. For instance, consider a borrower whose social network contains several friends but several choose not to bid on a particular loan. Such a listing is significantly less likely to get funded, consistent with the view that negative social capital (Portes 1998) is generated by verifiable bidders who do not bid. Conversely, consider the less verifiable information conveyed by endorsements received by borrowers. This variable turns out to be not statistically significant. We also construct social capital measures using group affiliations rather than friendship networks. Once again, verifiable antecedents matter. We find that listings by borrowers who are linked to verifiable groups such as affiliation with alumni groups of universities or common employers are more likely to be funded. In sum, the results suggest that it is the relational aspects of social networks that matter provided they are verifiable and in the information set available to potential lenders. The results on the funding probability indicate that lenders use social capital in determining whether to fund loans. We next test whether it has information on default probabilities beyond that contained in traditional credit variables. We estimate survival models in which social capital variables are incorporated as covariates to predict default. We find that borrowers who have strong and verifiable relationships are less likely to default. As we discuss later, the existence of these social network peer effects is consistent with work in economics, sociology, and theories of normative and informational influence in social psychology. We also conduct other analyses and these reveal additional insights. Petersen and Rajan (1994) emphasize that there can be distinct quantity (or supply) and price effects in credit markets. In the P2P lending context, our funding probability results represent evidence of quantity effects. To assess price effects of social capital, we estimate a Cragg (1971) censored regression model in which social capital potentially has separate price and quantity effects through different coefficients on the probability of funding and on the loan rate conditional on the loan attracting funding. Once again, social network variables have beneficial effects: they reduce the interest rates on funded loans. We also consider the effect of unobservable factors using two-stage estimates with instruments.

4 We find that a greater probability of funding lowers both interest rates and ex-post default rates, consistent with the hypothesis that borrowers repay trust reposed in them by lenders. We also examine the role of quasi intermediaries who emerge in this network in the form of members who borrow as well as lend and find that they outperform pure borrowers. We also find some evidence of contagion, in the sense that defaults are less likely for neighbors of borrowers whose friends are less likely to default. We conduct other robustness tests that are more fully described in the paper. Our findings have implications for research and practice. One view of our study is that it represents data from a natural experiment with a particularly severe Akerlof (1970) style lemons problem. Our findings can then be interpreted as evidence that even atomistic individuals with small sums of money at stake in individual transactions act in ways to mitigate adverse selection in ways remarkably consistent with predictions of economic theories. An equally valid perspective of our study is that it is an analysis of the economic value of social networks. While it is widely accepted in economics and sociology that networks matter, especially to the sets of individuals forming the networks and the organizations that employ them, our study provides quantification of its value and puts additional boundary conditions to the claim. We find that networks are valuable not merely to individuals or organizations forming or containing them, but also to third-party outsiders, by helping mitigate informational asymmetry and adverse selection problems between the individuals in the network and outsiders. While we empirically support the relevance of the relational embeddedness dimension of social capital (Granovetter 1972), it is also important that such relational embeddedness be visible and verifiable to outsiders such as lenders, in order to unleash its value. Our results suggest a novel role for social capital in economic transactions. Social capital is conventionally conceptualized as a group attribute of a collection of individuals that enhances the efficacy of transactions between the individuals for economic gain. Examples from the sociology literature include Coleman (1988), Mizruchi (1992), or Putnam (1993) while recent applications in finance include Hochberg, Ljungqvist, and Yu (2007) and Cohen, Frazzini, and Malloy (2008a, 2008b). Alternatively, social capital can be viewed as an individual attribute that adds to a person's physical and cognitive ability and generates an economic benefit (Granovetter 1972). Under either view, social capital is beneficial in facilitating transactions between individuals or within or between organizations that employ them. Our evidence suggests yet another role for social capital, viz., its role in facilitating transactions with third parties through an informational externality. Social capital between individuals can be harvested to facilitate transactions with outsiders, such as lender in financial markets. Our study is also of separate interest because of its focus on the economic value of online social networks, an area on which there is little prior work. While online networks may be as valuable as their offline counterparts, this is not obvious given the relative ease of creating and building them. The P2P lending marketplace is an especially interesting context to study online networks because these types of networks are integral to the marketplace. Furthermore, one limitation of the received work on online networks is the difficulty of quantifying economic outcomes or the strength of ties, necessitating costly methods such as surveys or interviews (e.g. Karlan 2007; Moran 2005; Uzzi 1999) or subjective measures of outcomes (e.g., Bagozzi and Dholakia 2006; Uzzi and Lancaster 2003). In our study, both the network itself and the associated economic outcomes are quantifiable using relatively objective measures such as funding probability or interest rates. Our findings also contribute to the current debate over financial market reform and market design. A wellestablished literature dating back to at least Fama (1985) argues that financial intermediaries such as banks are special because they produce soft information. This is fuzzy, hard-to-quantify information about borrowers that is often critical to successful lending outcomes (Petersen and Rajan, 2002; Rajan, 2002). A natural concern with electronic markets supplanting intermediaries is that such soft information will no longer be produced, perhaps adversely affecting market quality. Our study suggests that these concerns may be at least partially mitigated. While evolution in information technology certainly poses challenges to the conventional information gathering and processing activities of conventional financial intermediaries, there is a mitigating effect. Digitization also makes it feasible to extract, codify, and convey softer information such as a borrower's social capital, enabling its use in conditioning lending decisions, provided it is credibly verifiable by outside bidders, as emphasized in our study. We also uncover natural variation in the data that has an interesting parallel to securitization in financial markets. When group leaders have incentives akin to those of originators in securitized financial markets,

5 originations tend to be of lower quality, supporting the need to re-evaluate the originate-sell model of securitization in financial markets. The rest of the paper is organized as follows. Section 2 provides the theoretical background and reviews the related literature on social networks and microfinance. Section 3 provides an overview of the research context followed by a description of the data used in the study. Section 4 describes the empirical methodology, and Section 5 contains the results of the study. Section 6 discusses the implications of our results and concludes. 2. Theoretical Motivation and Literature Review We begin by discussing research that provides a theoretical underpinning for our study. We draw on and contribute to multiple streams of research including work in economics, finance, management, and sociology. To place our findings in perspective, we review the related literature. Theoretically, the relation between social capital and economic outcomes has been the subject of much attention in economics. Granovetter (2005) overviews applications in such diverse areas as labor economics, price setting, production, financial innovation, and entrepreneurship. An individual's social capital is typically identified using her social networks. Burt (1992, page 9) describes an individual's social capital as friends, colleagues, and more general contacts through whom you receive opportunities to use your financial and human capital. Portes (1998, page 6) adds that there is growing consensus on social capital being the ability of actors to secure benefits by virtue of membership in social networks or other social structures. In a careful methodological survey of social capital research, Durlauf and Fafchamps (2004) argue that the most successful theoretical work and the most compelling empirical work is the role of networks in facilitating economic exchange. Our work, which studies how social capital is leveraged through networks, has this focus. The literature also offers a useful differentiation between the different dimensions of social capital (e.g. Granovetter 1992; Moran 2005). Structural embeddedness refers to the position of an actor in the network while relational embeddedness refers to the quality of the relationship among actors in the network. Empirical evidence on structural aspects of networks includes work on venture capital by Shane and Cable (2002) or Hochberg, Ljungqvist and Yu (2007) and that of Cohen, Frazzini, and Malloy (2008a, 2008b), who find that educational network ties influence fund manager investments and the quality of analyst recommendations. Studies in sociology and management for instance, have shown that on a given network, the position of an agent can often reflect his or her resources. Some positions in a network endow control over the flow of information and other resources through the network, e.g., individuals who are in hubs, those with weak ties (Granovetter, 1972) or those occupying structural holes (Burt 1992). Bampo, Ewing, Mather, Stewart, and Wallace (2003) investigate the structure of digital networks on the performance of viral marketing campaigns. Relational aspects of networks can also influence economic outcomes. Examples of studies of relational embeddedness include Grewal, Lilien, and Mallapragada (2006), who study open software projects and Cowan, Jonard, and Zimmerman (2007), who examine embeddedness in a network of collaborators on innovation. Whether structural or relational aspects of networks matter in lending and the P2P lending market remains an open question. The question is of special interest given the decentralized nature of the network with independent lending decisions made by small investors on an arm'slength basis to borrowers. Theory suggests avenues by which peers can influence the perceptions and behavior of individuals linked through social networks. Social psychologists identify two psychological channels, informational influence and normative influence (e.g., Campbell and Fairey 1989; Cialdini and Goldstein 2004). The idea of informational influence is comparable to the theory of informational cascading in economics (Bikhchandani, Hirshleifer, and Welch, 1992). Normative influence, on the other hand, refers to conformity to the positive expectations of others, motivated by the desire for approval and to avoid rejection (Manstead and Hewstone 1995). It is often associated with sanction mechanisms such as shame, embarrassment and guilt (Sabini and Silver 1978; Campbell and Fairey 1989; Karlan 2007), or stigma due to blemished character (Goffman 1963). Avoidance of these consequences is sometimes exploited in microfinance as a substitute for physical collateral (Ledgerwood, 1999). It is also found to influence individual's bankruptcy in developed countries (Cohen-Cole and Duygan-Bump 2008).

6 Explicit economic incentives represent another channel for peer influence. Reward schemes conditioned on group performance exist, for example, in joint-liability group lending programs in microfinance (Morduch 1999). Alternatively, peer influence can arise out of the indirect economic effects. Karlan (2007) argues that the pressure to repay microloans can come from a desire to protect their social connections and avoid any repercussions such as reduced trading partners for one's business (Karlan 2007, page F58). This is echoed by Granovetter who proposes that individuals with whom one has a continuing relation have an economic incentive to be trustworthy, so as not to discourage future transactions (Granovetter 1985). Such economic incentives can further be intertwined with social norms, which carries strong expectations of trust and abstention from opportunism (Granovetter 1985, page 490). Hence, economic and social motivations, both related to the presence and action of others, combine to discourage what Granovetter calls malfeasance (1985) or opportunistic behaviors. The positive effects of such social relationships represent a second notion of social capital, one that applies to entire groups rather than individuals (e.g. Coleman 1988). Latane (1981) argues that stronger degrees of mutuality and interdependence lead to greater levels of normative influence and economic incentives the stronger the tie, the greater the strength of the peer influence. The mere existence of social capital however, is not sufficient to ensure its relevance to lending outcomes. Our study emphasizes that is necessary for the capital to be visible to lenders so it forms part of their information set, but more critically, it should be credibly verifiable by potential lenders to have proper effect. A different issue arises because of the relative ease of forming online networks, which may reduce their credibility as means of reducing opportunistic behavior by borrowers. Verifiability gains increasing importance in this context. The verifiability issue is also important from a behavioral and marketing perspective, rather than a purely economic one. As noted by Rosenthal (1971), a message must be testable by means independent of its source and available to its receiver to be verifiable. Such requirement of verifiability applies to online social networks as well. An alternative lens on the issue of verifiability can be found in the advertising literature. The degree of pre-purchase verifiability is the basis for researchers to categorize advertising claims as search, experience, or credence (Nelson 1970; Darby and Karni 1973). Only search attribute claims, or claims that can be verified prior to purchase, are credible to consumers (Nelson, 1970; Calfee and Ford, 1988). As in advertising, a borrower's social network is also subject to skepticism from lenders; it is credible only to the extent it is verifiable. Our study also adds to an extensive literature in finance and economics that studies credit markets. A key theme in this literature is information asymmetry, which presents itself through ex ante adverse selection and ex post moral hazard. The ex-ante information asymmetry essentially considers the Akerlof (1970) style adverse selection problems in lending. As pointed out in a review by Gorton and Winton (2003), the finance literature models intermediaries as resolving borrower-lender informational asymmetry. Gorton and Winton argue that lenders at the end of the chain face two key issues. One is reliability. While a small number of agents can produce information about the borrowers and sell it to other lenders (thereby economizing on the cost of information production), these producers cannot credibly ensure that the information produced is valuable (Gorton and Winton 2003, page 444). In addition, even if the information producer can assure other agents that the information is valuable, purchasers of the information can always easily resell such information or it could be reflected in prices so it becomes uneconomical to produce information in the first place (Grossman and Stiglitz 1980). Social networks can provide information relevant to lending outcomes. If someone who knows the borrower personally can attest to his or her credit-worthiness, or even better, participate in lending to the borrower, the loan should be relatively less risky. Obtaining and transforming such information into a usable format was difficult in traditional lending, but digitization and information technologies have helped overcome this constraint. The key issue with the use of such information is its reliability, which can be mitigated if the marketplace has credible verifiability standards. An alternative view of the use of social networks is based on search costs. In general, it is costly to set up a market where multiple borrowers and lenders are matched in a decentralized fashion. Intermediaries such as banks could be viewed as adding value by economizing on the matching costs by centralizing these efforts. However, the Internet significantly reduces transaction costs (for instance, see Malone, Yates and Benjamin 1987). Thus, technology shifts the tradeoff between search and intermediation at the margin in the direction of making matching a more viable option. An additional effect of digitization and Web 2.0 technologies is that they alter the very way in which users interact and connect with each other, resulting in

7 generation of alternative sources of information and social capital, and new means of transmitting the information efficiently. The net effect of these forces is to facilitate the growth of lending networks that are decentralized. Few empirical studies explicitly deal with social networks in financial lending. Much of the received work focuses on the dyadic relationship between banks and the borrowing firms. Using survey data on small business loans, Petersen and Rajan (1994) find that bank or supplier relationships increase the supply of credit to small firms. In related work, Agarwal and Hauswald (2007) argue, private information drives relationship-debt transactions, whereas public information facilitates arm's-length lending. Our study contributes to this body of work by teasing out the impact of the relational and structural aspects of social networks in lending. Drucker and Puri (2007) provide evidence that sales of loans are associated with durable relationships. Organizational researchers apply social network theories to the banking sector. Using a social embeddedness approach, Granovetter (1985) and Uzzi (1999) study how bank-borrower relationships affect a firm's acquisition and cost of capital and introduce the idea of networks in these papers. Our study focuses explicitly on social networks, illustrates how technology hardens it into usable form for lenders, and shows that social networks can act as new sources of soft information about borrowers. Besides ex-ante information production to alleviate information asymmetry, the finance literature also focuses on ex-post monitoring after a loan is consummated. Diamond (1984) argues that banks are established to economize on the costs of ex-post monitoring activities. But this leads to the problem of monitoring of the monitors, which may be resolved by diversification (Gorton and Winton 2003, page 443; Williamson 1987). Ex-post monitoring is also an important issue in online peer-to-peer lending. Peer-to-peer lending websites provide some of the administrative aspects of monitoring such as obtaining and disbursing monthly repayments and recording defaults with credit bureaus and collection agencies. However, the advancement in information technology does not solve the fundamental issue of monitoring costs. It is inefficient for each investor to monitor borrowers and its delegation still leaves a residual moral hazard problem. Embedding social networks can mitigate some portion of this problem, especially when friends participate in a borrower's loan. The social stigma or reputational costs of default now matter more, because the news of default is transmitted to a borrower's social network, especially when the members of a social network are also lenders. The resulting peer pressure may mitigate the risk of default due to lost social capital, stigma, or reputation in the spirit of Diamond (1991). There is a small but growing body of research on peer-to-peer lending that focuses on the personal characteristics of borrowers to test theories of taste based discrimination (Pope and Sydnor 2008; Ravina 2008). Pope and Sydnor examine loan listings between June 2006 and May 2007 and find evidence of partial taste-based discrimination of Becker (1971), with directionality against blacks, the overweight, and the elderly, and in favor of women and military affiliations. Ravina examines listings for a one-month period between March 12, 2007 and April 16, Consistent with taste based discrimination based on beauty, borrowers rated as being more beautiful are more likely to get funded and pay less but these borrowers are not less likely to become delinquent. While both studies focus on personal characteristics, our work addresses social capital, and more specifically the relational aspects of social networks. Our study supports the complementary view that social capital could be a positive tool for statistical discrimination in the style of Arrow (1972) or Phelps (1972), and thus helps mitigate adverse selection problems in the P2P market. Our findings that lenders extract and use lending outcome-relevant information is similar in spirit to Iyer, Khwaja, Luttmer, and Shue (2009), who find that subjective information extracted by lenders can account for about one third of the statistical content of the true credit score of borrowers. Online P2P lending can be viewed as a digitized and somewhat modified version of traditional microfinance programs (see, e.g., Morduch 1999 for a review of this work). Like peer to peer lending, microfinance is typically collateral-free because the targeted borrowers are typically poor or ultrapoor. In both settings there are similar information asymmetry problems. Our findings suggest a new parallel between P2P lending and microfinance, in that both rely on soft collateral or group attributes for repayment (Ledgerwood 1999). On the other hand, traditional microfinance programs face a tradeoff between financial sustainability and scale. While many microfinance programs work well in smaller communities or regions, informational issues get magnified when applied to larger areas making these programs financially unsustainable. The online P2P lending programs in our sample display scalability, involving larger networks of geographically disperse individuals rather than say, village

8 collectives or cooperatives. Our findings suggest that digitization and technology can mitigate the problems that limit the scaling of traditional microfinance programs. The term peer-to-peer is used to contrast client-server configurations in computer networks. Whereas in clientserver architecture, one server occupies a central position in the network, in a peer-to-peer architecture all computers are on equal footing there are no hierarchies and all computers are peer nodes. The decentralized, distributed nature of P2P networks is usually considered an advantage in computer networking. Over time, the idea of decentralized peer-to-peer networks has been adapted to wider contexts. A well-known example is peer-topeer music sharing (e.g., see Leibowitz (2006) or Bhattacharjee, Gopal, Lertwachara, Marsden, and Telang (2007) for recent work). These studies typically investigate the impact of the growth of P2P music sharing networks on sales and consumer welfare. The structural or the relational aspects of these networks or their impacts on outcomes typically do not receive much attention in these studies. What makes the study of peer-to-peer lending more interesting relative to this work is the availability of objective transactional outcomes that are associated with both the structural as well as relational aspects of the network. 3. Institutional Background On P2P Lending Our data come from a leading online peer-to-peer lending website, Prosper.com. Prosper.com opened to the public on February 5th, At the end of 2008, it had 830,000 members and more than $178 million in funded loans. We describe institutional information about the lending process on Prosper.com and the information provided by participants on the network. 3.1 Lender and Borrower Verification Potential users initially join Prosper.com by providing an address, which is verified by the website before membership is initiated. However, to actually engage in a transaction, users must go through a more formal verification process that varies based on whether the participant wishes to borrow or lend. As we describe below, the verification process is more extensive for borrowers. Any person residing in the US with a valid social security number and a minimum FICO (Fair Isaac Credit Organization) credit score of 520 can register as a borrower on Prosper. Registered borrowers must provide a social security number, a driver's license number, an address and other information. The information is verified, and data are pulled from the borrower's credit report, which is provided by Experian, one of the major credit reporting agencies in the US. A borrower must also provide a valid bank account number. Loan proceeds are credited to this account and funds withdrawn automatically for monthly loan repayments. At the time of the study, borrowers were allowed to borrow a maximum of $25,000, and a maximum of 2 concurrent loans. All loans that originate on Prosper are 36-month loans. A lender who joins Prosper is also subject to a verification process. A social security number, driver's license number, and bank account information are necessary before a lending transaction. A lenders' credit information is not revealed to a borrower. However, before a lender can lend, she needs to first transfer the funds from her bank account to her Prosper (non-interest bearing) account. To protect the privacy of borrowers and lenders, their true identity is never revealed. Each member has a username that is chosen when signing up. 3.2 Listing After being formally allowed to create a listing, a borrower must first create a profile that contains a personal description. When the borrower requests a loan, he creates a listing, which is a request for a loan. Figure 2 provides a screenshot of a sample listing. In the listing, the borrower specifies the amount that he would like to borrow and the maximum interest rate that he is willing to pay. We control for this variable in our models. The borrower also chooses one of two auction formats: a closed auction, also called immediate funding, or an open format. In the immediate funding alternative, a listing closes as soon as the bid cover ratio (the ratio of the total

9 amount bid to the total amount sought) reaches 1.0. The interest rate of the loan is the borrower's asking rate. In the open auction format, the auction remains open even after the entire amount requested is funded. Lenders can continue to bid down the interest rate of the loan. The duration of the auction is now standardized to 7 days. We include controls for the auction format in our specifications. As Figure 2 shows, a listing displays information from the borrower's credit report, including the number of credit inquiries, his debt-to-income ratio, and a letter credit grade, which is a coarse version of the borrower's FICO score. While lenders can see a borrower's credit grade, they cannot observe the actual FICO score. Letter grades range from AA (high quality) to HR (low quality) high-risk borrowers. The correspondence between letter credit grades and the actual FICO score is shown in Table 1. The purpose of a loan is also specified in listings and could be a business loan, an auto loan, mortgage, or a student loan. We control for the loan purpose in our models. If a potential lender has a question about the listing, she can send the borrower an electronic message by clicking a hyper-link on the listing. The borrower can choose to not respond, respond privately, or make answers available to other potential lenders. We include the total amount of text included in a listing in our analysis. The listing also contains information about the borrower's friends and groups to which he belongs. These variables are an important part of our analysis and are discussed separately in the next section. With all the above information in place, the listing can go live to solicit bids from lenders. 3.3 Bidding When a lender sees a listing, she can decide whether or not to lend to the borrower. An important feature of online peer-to-peer lending is that an individual lender does not have to finance the entire loan request. The minimum amount she can lend is $50. A lender interested in a borrower's listing can bid an amount that she's willing to lend and specify the minimum interest rate. The actual bidding process uses a proxy bidding mechanism. More specifically, if the loan has not yet been funded 100%, the ongoing interest rate will be the borrower's asking rate, even if the lenders' minimum rate is lower. Once 100% of the requested funding has been reached and the format of the auction is open, the ongoing rate decreases as the lender with the highest rate-bid is competed out. In a sense the auction is similar to a second-price auction. All bids are firm commitment bids and no withdrawals are allowed. From a lender's viewpoint, a bid could win or be outbid, in which case the lender can place a second bid to rejoin the auction. From a borrower's perspective, a loan is either fully funded or not, in which case the auction is deemed to have failed and no funds are transferred. 3.4 Post-Bidding: Funding and Repayment Once the bidding process ends, the listing is closed and submitted to Prosper staff for further review, and sometimes additional documentation is required of borrowers. Once the review process is completed, the loan is originated, and the funds are collected from the winning bidders' accounts and transferred to the borrower's account, after deducting fees. These fees include borrower closing costs and lender service fees. The actual fee depends on the credit grade of the borrower, and ranges up to 2% of the loan amount. Loans on Prosper.com have a fixed maturity of 36 months with repayments in equated monthly installments. The monthly repayment is automatically deducted from a borrower's bank account and distributed to lenders' Prosper accounts. If the borrower pays his monthly amount due in time, the loan status for that month (or payment cycle) is considered current. If a monthly bill is not paid, the loan status will be changed to late, 1 month late, 2 months late, etc. If a loan is late for 2 months or more, it is sent to a collection agency. The delinquency is reported to the credit report agencies and can affect borrowers' credit scores. 3.5 Social Networks Any members who signs up with Prosper.com and has a verified account can create or join a social network. We consider two types of social networks, a friendship network and groups. In a friendship network, a member can be a friend with other members who already have a valid user ID on Prosper.com. Alternatively, the member can ask offline friends to join Prosper.com and become an online friend on the network.

10 Friendships are typically created through an automated process using a dedicated section of Prosper.com. After the inviting member fills out the friend's address and a short message, Prosper.com generates an message on behalf of the member, which contains a link to join Prosper. The recipient can click on the link contained in the to sign up, or use the link to establish friendship with an already-registered account on Prosper. As of the time of writing, invitations can only be sent through Prosper.com. While both borrowers and lenders can have friends, members can also elect to not go through the more extensive process of verification to become borrowers or lenders. These individuals can still be friends with others. This fact introduces interesting variation in the quality of information about a friend, which we exploit in our empirical tests. In addition, friendship ties are bi-directional. From an empirical viewpoint, a member's friendship network is visible on the profile page or a listing page. Other members can click through the link to see the profile information of those friends. In our sample, 56,584 listings report friends. A second type of social network on Prosper.com is a group. There are 4,139 groups as of the writing of this paper and 41% of listings in our sample are associated with a group. Any member can create a group and a member can typically join any group whose membership criteria are met. However, as of the time of writing this draft, each individual can be a member of only one group at a time. Entry or exit into a group is free but this bar is raised for borrowers. If a borrower is a member of a group when requesting a loan, the borrower cannot leave the group or join any other group until the outstanding loan is repaid in full. The leader of each group can determine the rules regarding who can become group members and how others may join. Some groups, such as alumni groups, typically require a high degree of verification. Applicants must prove that they are indeed affiliated with the institution. Other groups require very little verification. Before October 2007, if members of a group were able to obtain a loan from Prosper, the group leader could receive a commission (up to 2%), known as the Group Leader Rewards. These group leader rewards were terminated after October Data Our sample comprises all listings that seek funding on Prosper.com between January 2007 and May We obtained information regarding borrower's credit history, their unique identifiers (not their social security numbers), the social network variables, features of their auctions, and outcome of their loan listings using an API provided by Prosper.com website. The information on the website is updated in real time, so caution is needed to ensure that the downloaded data is measurable as of the listing date and in the information set of potential lenders. The following paragraphs describe the variables used in our analysis and discuss some descriptive statistics. 4.1 Hard Credit Information We define hard credit information using the standard indicators adopted by banks in their loan decisions. As mentioned earlier, Prosper.com provides a letter grade for each borrower ranging from AA to HR. The AA grade is awarded to borrowers with credit scores exceeding 760, grade A is granted for scores between 720 and 759, and so on. The lowest grade, HR (high risk), corresponds to a credit score between 520 and 559. The underlying numerical scores roughly correspond to the familiar FICO score, modified by Prosper.com using a proprietary model called Experian Scorex Plus. In addition, the website gives other hard credit information, including a borrowers' debt-to-income ratio and the number of credit inquires in the previous six months prior to the listing. We include these variables as additional credit indicators to allow for the possibility that the letter grade itself is not a sufficient statistic for credit risk. One empirical question is whether we should include numerical transformations of the letter grade [such as AA=1, A=2 (or 4), B = 3 (or 9), etc.] or dummy variables for the individual letter grades. The former has the advantage of parsimony, as it requires one coefficient. Including a full set of dummy variables is more complete but requires estimation of six coefficients. Because of a large sample size (205,132 listings), we elect to include the full set of dummy variables in our models.

11 4.2 Social Networks We use both the friendship network and group-based affiliations to construct measures of the social network information. These variables capture soft information: the type of information not used in a bank's traditional quantitative metrics but could be relevant to lending outcomes. As Petersen (2002) discusses, technology could potentially find ways to harden soft information, and our study provides one example where it does so. Our study distinguishes between structural and relational aspects of networks. The commonly used metrics of a network structure include degree centrality, betweenness centrality, coreness, effective size, and efficiency (Hanneman and Riddle 2005). Degree centrality is simply the number of members, or friends, that a borrower is connected to. Betweenness centrality measures the frequency that a member is on the shortest paths (geodesics) between any two other members on the friendship network. Coreness measures how close a member is to the core of a component, instead of being a peripheral member of the network. The effective size of a member's ego network is the number of alters (friends) of a member minus the average number of ties that occur among these friends, which is a measure of the structural hole (Burt 1992). Efficiency is the effective size normalized by the actual size of the network. While degree centrality can be easily computed, other metrics require specialized social network analysis software. To calculate these metrics, we extract all components in the friendship network of size 3 and up using the software package Pajek. We import these components into UCINET, which provides these metrics but does not handle large networks. While we do report some results based on some of these metrics, they are largely insignificant. Structural aspects of networks matter less; rather, relational aspects are significant and we turn to these variables next. The relational social network measures concern the roles and identities of the members in the network. Figure 1 discusses the hierarchy underlying our empirical strategy. In Level 1, we distinguish friends according to whether their identities are verified on Prosper, i.e., individuals who have elected to undergo verification of identities and bank accounts versus individuals who have merely registered and are thus little more than persons with a verifiable address. At level 2, we categorize the verified friends of borrowers based on their specific roles whether these friends are borrowers or lenders. Lenders are individuals with extra financial capital while borrowers are likely to be facing financial constraints. On the other hand, borrowers are subject to greater scrutiny as they have assigned credit grades that form a backbone of their listings. Level 3 further differentiate between real lender friends those who have lent prior to the current listing; and potential lender friends those who have provided enough information to Prosper to be listed as lenders but have yet to participate in a loan. Level 4 differentiates real lender friends according to whether they bid on the specific borrower's listing or not. Here, we can have two effects. Having friends who participate can enhance lending outcomes. However, having several friends who do not participate in a borrower's auction can send negative signals about the borrower, which can potentially lead to negative outcomes if inaction is interpreted by other lenders as a negative signal. Level 5 is the finest classification. Here, we distinguish further between lender friends who bid on the borrower's listing and won (i.e. actually lent money to the borrower), and those who bid but did not win. Overall, as we progress from higher to lower levels of this friendship hierarchy, the relationship between the borrower and lender becomes more actionable, verifiable and more strongly embedded. In addition to friendship networks, we also consider membership of groups. There are many different types of groups on Prosper.com. For our analysis, we manually coded all groups that have at least 3 members and are active in the generation of loans. We categorize all groups into one or more of several categories. One set of groups, for instance, are based on religion. In this group, members can identify themselves as belonging to a particular religious or church denomination. A second category of groups uses alumni affiliation, based on whether an applicant graduated from a particular university or whether an applicant has been a past employee at a company. These antecedents tend to be verified by group leaders. Other categories include geography-based groups, in which members live in a particular region (e.g., greater Washington DC area), which is verified based on applicant addresses. We also found group categories based on military membership, groups with medical needs,

12 demographics (such as being of Hispanic or Vietnamese origin), groups based on hobbies, and groups affiliated with business development or those with a generalized purpose of helping members financially. Each group has admission criteria of varying stringency and verification required for membership. We include dummy variables for the types of groups to which applicants belong, testing whether group affiliations matter for lending outcomes, particularly those involving verifiable antecedents. The correlation among groups is fairly low although a few groups can span multiple categories. In addition to group dummies, we also include controls for group size. A priori, the sign of group size for lending outcomes is not clear. Larger groups may involve more peer pressure to repay and are perhaps associated with better loans. On the other hand, larger groups may be subject to less overall oversight. The actual sign is thus an empirical issue. Finally, we include a dummy variable for whether group leaders have financial incentives for funded loans, a practice discontinued by Prosper.com since October Other Explanatory Variables In addition to the soft and hard credit information, we also gather information on the auction characteristics. Prosper listings vary according to the type of auction conducted for the listing. One characteristic is whether the auction is conducted via the open or closed format. The latter closed as soon as it is funded 100% and perhaps indicates borrowers with more urgent financial need. We include a dummy for the auction format in all our specifications. We also considered maximum auction duration. This can range from 3 to 10 days but more recently it has been standardized to a 7-day duration format. This variable showed little significance in any of our models and we omit it. Some states in the US have usury laws that enforce a cap on the allowable interest rate on consumer loans. In principle, usury laws are intended to protect customers. On the other hand, these laws could negatively affect the likelihood of loans being funded, if the supply curve for funds intersects the demand curve at a rate above a state's usury limit. Whether the laws have this bite or not is an empirical issue that we examine. After April 15th 2008, Prosper started collaboration with a bank in an effort to circumvent that limit. Our sample spans both periods, so we include a control for usury laws. Each borrower must also indicate a maximum borrowing rate that she is willing to pay. We include a control for this variable, in part to test the credit rationing model of Stiglitz and Weiss (1981), whose insight is that there could be rationing at very high rates, when the greater profitability from higher rates is swamped by the greater likelihood of default for borrowers willing to pay high rates. The setting and the type of reasoning by players in Stiglitz and Weiss is probably far from our setting of atomistic lenders bidding for a small piece of a listing. It is thus interesting to test whether the data supports the signaling equilibrium from relatively elaborate economic models. Thus, we include both the maximum rate and a squared term to test for the non-linear effect. To control for broad lending rates, we could use a proxy such as the LIBOR, the Fed Funds rate, or prime rates, many of which serve as floating rate indexes used by banks to set interest rates. But these rates are noisy, since they do not incorporate credit spreads for different quality borrowers and any regional variation in the spread. Moreover, they do not incorporate the shape of the term structure of interest rates since the rates are short-term while Prosper.com loans are of 36-month duration. We purchased a proprietary dataset from a professional company that collects information on interest rates in different US markets. This dataset gives us the average interest rate for borrowers in each credit grade, in each regional market, for a given month. The term of loans is 36 months, consistent with Prosper loans. In our analysis, this variable serves as a proxy for the outside option of borrowers. In terms of other controls, we follow the finance literature and control for loan purpose. We read listings to assess the purposes for which Prosper.com loans are sought. Borrowers indicate several types of needs, including debt consolidation, home improvement, business loans, personal loans for a variety of purposes (including vacations), or student or auto loans. The loan purpose is self-indicated by borrowers and can thus be thought of as cheap talk especially if the listing does not have adequate detail. However, it is often the case that potential lenders

13 communicate with borrowers during the auction process and seek more tangible details during this back and forth. In balance, we thought that it is likely that there is some information in the loan purpose, so we categorize and include this in the regression. As a new business model, Prosper.com has received significant media exposure since its inception. Articles in the media make it more likely to attract new borrowers and lenders to the website after their publication. To control for the potential influence of such news, we include an additional variable to absorb these exogenous shocks to this marketplace. We download the search volume on Google for Prosper.com and construct a dummy variable based on whether there is a significant change in search volume, which we call spikedays. This variable also serves as an interesting instrument to identify the probability that a listing will be funded in two-equation systems. Spikes in search volumes could affect probabilities of being funded but should not affect the probability of a loan default, which is a function of borrower's credit circumstances. Finally, we include quarterly dummy variables to control for unobserved time effects in the marketplace. 5. Results and Discussion Our sample has a total of 205,132 listings with an average funded loan amount of $6,973. Of the sample, 56,584 or 27.58% of the listings report friends while 148,548 (71.42%) of the listings report no friends. The percentage of listings in which borrowers have friends is spread across Prosper.com credit grade spectrum. For instance, of the 6,523 AA listings, 1,881 or 28.84% have friends, while of the 33,068 D grade listings, or 28.61% have friends. In the high risk, or HR category, 22,556 out of 62,904 listings, or 26.39%, are associated with a friend. Listings in which borrowers have no friends have mean debt-to-income ratios of 58% while listings of borrowers with friends have debt to income ratios of 57%. Borrowers with no friends have about 4.17 credit inquiries in the six-month period prior to the listing date, while borrowers with friends have about 4.22 inquiries in the previous six-month period. In the overall sample, 16,500 listings successfully attract funding so the unconditional probability of a successful listing is 8.04%. Our other social network variable is group membership % of all listings, or 59,978 listings indicate a group affiliation. Of this sample, 28,006 listings, or 46.63%, are associated with an incentive structure in which group leaders are rewarded with a fee for successful listing. In terms of univariate statistics, in the sample of listings where borrowers have no friends, 10,410 out of 148,548 listings, or 7.04% are successfully funded, while 6,090 or 10.76% of listings where borrowers have friends are successfully funded. Likewise, 7.09% of listings not affiliated with a group are successfully funded, while 10.36% of listings with a group affiliation are funded. These univariate differences in funding probability could, of course, reflect differences in the funding probabilities across the two groups or reflect differences in other characteristics of the borrowers described above. To assess this issue, we estimate econometric specifications for the probability that a listing will be funded. Section 5.1 reports the results. We turn to models of the default probability and the interest rate for funded loans in Sections 5.2 and 5.3. Table 2 describes the list of the explanatory variables. Table 3 describes the variable sets used in the specification while Table 4 describes the corresponding mnemonics used to identify the estimated specifications in the econometric models Funding Probability This section examines the probability that a listing on Prosper.com is successfully funded using a probit model. α α α ε Probability (funding) = 1 HardInfo + 2 SocialNetwork + 3 Other Variables+ i Table 5 displays the probit results. To control for unobserved changes in Prosper.com policies, we include quarterly fixed effects. We report six sets of results that vary based on how they specify the social network 1 While the main results are reported in Tables 5-7, as discussed in the next paragraphs, we also emphasize that we conduct several robustness tests. For instance, we control for multiple listings, include a continuous measure of funding as a percentage of the amount sought (rather than a 0/1 variable denoting funding status), estimate models in which the mechanical effect of friends' bidding is eliminated, and specify a model for repayment of trust. For brevity and because these results do not alter our main conclusions, we do not include the tables containing them. These are (readily) available upon request from the authors.

14 variables. All six specifications include a common set of variables and controls. We discuss coefficient estimates for these controls first, after which we turn to the results related to the social network variables Hard Credit Variables The first variables in our model represent hard credit information. We include several dummy variables, one for each rating category. The baseline dummy variable is the top credit grade AA, so that the estimated coefficients give the incremental probability of a listing being funded relative to an AA listing. All credit grade variables are significant in Table 5 and have the expected negative sign; worse credit rated borrowers are less likely to be funded. As the borrower's credit grade decreases from A to HR, we observe that the coefficients on these dummies also decrease and the differences between the coefficients are statistically significant. Thus, the lower the credit grade, the lower the probability of the loan being funded. In addition to the credit grades, we include data on bankcard utilization and its square. In Table 5, both are significant with opposite signs. High credit utilization is viewed as being undesirable by lenders, consistent with borrowers being financially strained. These borrowers are likely to have limited access to new credit and are thus vulnerable to any negative shocks to income. On the other hand, a very low level of utilization is viewed by lenders are a negative signal too, perhaps because it indicates little credit history of substance. Neither extreme appears to be desirable from the lenders' perspective. We also include the number of credit inquiries in the last six months. A large number of credit inquiries suggest that the borrower might be in urgent need of funding. More importantly, it might indicate that the borrower has been rejected before, which casts doubt on creditworthiness. In Table 5, we find that more inquiries decrease the probability of the loan being funded. The debt to income ratio is the ratio of the borrower's total debt to his income. A higher debt-to-income ratio suggests that a borrower is less able to take on additional debt, and is therefore more likely to default in case of shocks such as unemployment. These borrowers are less likely to be funded Auction Characteristics As discussed earlier, borrowers can choose whether to use an open format or a closed format in which the auction closes as soon as the loan is fully funded. Choosing a closed auction format can lessen bidder competition and encourages aggressive early bidding, which should increase the probability that a loan is funded. The results in Table 5 are consistent with this view. We note that this result is obtained even after controlling for the amount sought in the listing. Greater amounts sought are less likely to be successfully funded, but even conditional on amount, closed auctions are more likely to result in a successful listing. On the other hand, a closed auction could also be interpreted as a signal of the willingness of the borrower to forgo competition, which may reflect a borrower who is more financially strained. This could result in greater default rates and higher interest rates if market participants anticipate this outcome. We test these propositions in Sections 5.2 and 5.3. Our models also control for auction categories, or funding purposes. The results in Table 5 suggest that three categories of funding variables stand out. Business loans (listingcatg4) appear to be perceived as being riskier by lenders, and therefore have lower probabilities of funding. Debt consolidation loans (listingcatg2) are more likely to be funded than other loans. Evidently, lenders appear to value the fact that borrowers are using Prosper.com to shop interest rates, thus limiting their total indebtedness. We do note, however, that there may be a time consistency issue, because borrowing from Prosper.com does not obligate borrowers to necessarily maintain lower borrowing levels for the term of the loan, even if this is a stated purpose at the time of a loan. Perhaps the most interesting finding is that specifying some purpose, category (listingcatg1), increases the overall probability of funding. Lenders in the P2P marketplace appear to value getting some information over getting none about potential uses of proceeds.

15 5.1.3 Social Networks: Friendship Network We start by considering structural measures of networks. The degree centrality measures a borrower's position in the friendship network. Specification P1 in Table 5 shows that degree centrality is positively related to the probability of being funded. As discussed below, this relation masks a more extensive relation caused by the roles and identities of the members of the friendship network. We included other structural network measures such as coreness, effective size of network, and efficiency but none of these alternative metrics have any significant effects on the probability of funding. Specification P2 in Table 5 distinguishes friends according to whether their identities are verified on Prosper or not. This process effectively decomposes degree centrality into two orthogonal pieces, friends who are verified and those who are not, with both summing up to the total number of friends. We find an interesting pattern. If a borrower is connected to friends who are not verified, the loan is less likely to be funded. Lenders apparently view the presence of friends who do not choose to initiate verifiable roles in Prosper.com as a negative signal, so connections that merely verify addresses but do no more have no economic value. To the extent they are connections who choose to have no roles on Prosper, they even harbor negative information about the quality of the social capital of a borrower. In contrast, specification P2 also shows that having friends with roles on Prosper (the coefficient for TTLROLE) is positive and significant at 1%. These results constitute the first evidence that it is not numbers alone, but rather that roles and identities matter. To delve deeper into the role and identity issues, we categorize the friends of borrowers based on their specific roles. Once again, we decompose the total number of verified friends into two orthogonal and additive pieces, the friends with roles as borrowers and roles as lenders, so that the total adds up to the total number of friends with roles. In addition to these two components of friendship, we also include the total number of friends with no roles. The results are displayed in specification P3 of Table 5. The number of friends with no roles continues to have a negative coefficient, as before. We find that being connected to borrowers has no impact. However, having additional lender friends increases the probability of the loan being funded. To probe the number of lender friends in greater detail, we further differentiate between real lender friends those who have lent prior to the current listing; and potential lender friends those who are yet to lend prior to the start of the current listing, and the real and potential lender friends add up to the total number of lender friends. We continue to include the excluded variables, all friends with no roles, and friends who are borrowers but not lenders as control variables. Specification P4 reports the results. There is a continued gradation of the friendship effects based on the nature of a friend's role and its verifiability and visibility to outside lenders. Having just potential lender friends has little effect. Our results show that having potential lender friends does not have a significant impact on any of the transactional outcomes. In contrast, having real lender friends increases the probability of the loan being funded and the coefficient almost doubles relative to that for the total number of lender friends. At the next level, corresponding to Level 4 in Figure 1, we now differentiate real lender friends further according to whether they bid on the specific borrower's listing or not. At this level, it is also possible that a potential lender who has not lent in the past now chooses to initiate bidding with the current loan. Thus, we decompose the number of real (past) lenders who bid on the current listing and those who do not, as well as the number of potential lenders who bid or do not bid on a borrower's current listing. Specification P5 in Table 5 reports the results. We find effects for both the potential lenders and the real lenders. While the total coefficient for potential lenders is insignificant (as in specification P4), this appears to merely disguise variation based on whether potential lenders bid or not. If a borrower has more potential lender friends who do not bid, the borrower is less likely to generate funding for a listing. On the other hand, the greater the potential lender friends who bid, the more likely is the listing to get funded. We see similar and even stronger effects for friends who are real (past) lenders. A greater number of real lender friends who bid on a listing make the listing more likely to be a success. More strikingly, if the borrower has a real lender friend but this friend chooses not to lend to this particular borrower, this inaction is interpreted by other lenders as a negative signal, thereby decreasing the probability of the loan being funded.

16 At the finest level of role and identity is whether a real lender friend whose bid actually won or lost in the listing. Real lender friends who win a bid not only signal their willingness to lend to the friend but also signal their willingness to compete and win in the auction. This type of behavior may serve as a more positive signal to outside lenders to participate in the loan listing aggressively. Specification P6 in Table 5 examines this final decomposition. We find that the number of real lenders bidding on a friend's loan listing has beneficial effects on the funding probability both when the friends win and do not win in the auction, although the coefficient for winners is about twice that of the coefficient for lenders who do not win. The other coefficients in the specification are not affected. Specifically, a greater number of potential lender friends who participate in bidding continue to increase the odds of a successful funding. Likewise, the number of potential and real lenders who do not bid continues to decrease the chances of listing success, consistent with non-participation by friends able to bid serving as a negative signal to outside lenders. In sum, our results establish that social capital matters in attracting outside financial capital. Furthermore, the structural aspects of the social network are not necessarily critical. Rather, the role and identity of the members of a social network matter. In this context, verifiability is critical. Social capital that is verifiable and visible to outside lenders has the capacity to influence outside lender behavior. It is interesting that the results are obtained even in a setting in which these outside lenders are atomistic individuals participating in arm's-length transactions with the individuals possessing the social capital Social Network: Groups A second aspect of a social network that we are interested in is the group that an individual borrower is affiliated with. As a first control, we include the natural logarithm of the group size. The results for group size are mixed. A large group size is less likely to result in a successful listing in specifications P1 through P4; the larger the group, the less valuable the value of the affiliation perceived by lenders, suggesting that membership of a group loses its informational relevance when it is large. Perhaps the effect of default of a group member on the overall group credit quality declines when the membership is very high. Additionally, members who choose to belong to a larger group are voluntarily foregoing membership of a smaller group, recognizing which potential lenders may become less willing to fund a listing. The group size variable loses significance in specifications P5 and P6. We also test for non-linearities in group sizes, and find no significant effects. At one point of time, Prosper.com allowed groups to be set up with the provision that the group leaders get rewards when a member of their group obtained a loan. This incentive could encourage members to form groups, become leaders, and perhaps increase the marketing of a loan and thus the probability of a successful listing. We find evidence consistent with such a hypothesis. When a group leader is rewarded, loan listings by the members of the group are more likely to get funded in all specifications. A related question is whether the group leader is more likely to attract funding. Being a group leader may convey some positive information about a member's social and/or financial status. We find some evidence for this view. Being a leader is more likely to result in a listing being funded at 10% significance in three specifications and 1% in three others. We also consider the group type. Here, we again draw a distinction between group memberships that are less or more verifiable based on the criteria imposed for joining a group. We find two categories of groups where there is a relatively high bar on verifiability, alumni memberships based on university or former or current employers, and geography based groups. For both variables, group membership results in a greater chance of listings being funded in all six specifications. Interestingly, being affiliated with religious groups also matters. Other things being equal, borrowers indicating a religious group affiliation are more likely to draw funding. It is not readily clear why this result should be obtained. One possibility is that individuals with religious affiliations may default less; alternatively, some lenders may have a taste for lending to people with religious affiliations, in the spirit of the taste based discrimination hypothesis (Becker 1971). In a later section, we test these (and other) hypothesis by examining ex-post default rates on loans.

17 5.1.5 Other Variables Credit rationing arises when a borrower's willingness to pay high interest rates is considered a negative signal of his ability to repay. A well-known article by Stiglitz and Weiss (1981) formalizes this intuition. We can test this hypothesis in the P2P lending networks in a setting close to that modeled by Stiglitz and Weiss. Borrowers post an offer rate and lenders decide whether to lend or not. Our results show that while the linear term in the borrower's maximum interest rates has a positive coefficient, the quadratic term has a significant negative coefficient on funding probability. Thus, individual lenders behave in a manner consistent with the theoretical arguments for credit rationing. This is an interesting finding as it demonstrates that credit rationing is more prevalent than previously thought. While prior work emphasizes rationing in the context of an intermediated finance system, it also emerges organically even when lending decisions are made in a decentralized market by small individuals. When a state enforces a usury law, the interest rate cap prevents riskier borrowers from obtaining loans in their home state. We examine the impact of the usury laws on the probability of being funded. We find that for borrowers whose equilibrium interest rate is below this ceiling are less likely to get funded, perhaps because lenders perceive these borrowers as being pooled with a set of riskier borrowers who would otherwise not be funded or be funded at higher rates. Consistent with the supply side arguments against usury rate caps, usury laws have a negative effect on attracting lending capital from outside investors. Finally, we consider the length of the descriptions in the profile, the text in the listing as well as the text in the description of the groups to which a member belongs. A priori, it seems likely that longer text descriptions increase the probability of funding as they increase the amount of information available to potential lenders. However, this is by no means clear, borrowers could use text to disguise information so more could actually be less. Empirically, we find evidence for the former effect. The natural log of the total length of the text in the loan description increases the probability that a loan will be funded. 5.2 Loan Defaults Prosper.com records the status of loans in each month, or payment cycle. If the borrower pays off the monthly amount due, the loan status is listed as current for that month. Otherwise, the loans can be late, 1 month late, 2 months late, etc. Consistent with the practice of prosper.com, we create a dummy variable DEFAULTED if a loan is late for 2 months or more. In this section, we model the default hazard as a function of hard credit variables, interest rates, and soft information, an approach taken, for instance, by Gross and Souleles (2002) in modeling consumer loan defaults. As the duration literature points out, survival models are appropriate statistical techniques given the right censoring in the data. Specifically, if all loans in our completed their lifecycles, ending either in full repayments or a defaults, we could use a simple probit or logit model for loan performance. However this is not the case: the performance data is still highly censored because our observation of defaults ends in October Thus, even if we observe that a loan is current as of October 2008, it does not mean that loan will not default later. In fact, this data structure determines that the appropriate empirical strategy is to use a survival model Cox model of loan performance The key dependent variable in the survival model is the time-to-default, or the number of payment cycles after which a loan defaults. Survival models estimate the hazard function, or the probability of surviving for the next instant of time given that a subject has survived until time T. The hazard function h(t) is defined as h(t)=pr(t T t+ t t T) Different survival models vary based on how they specify the survival function. Preliminary diagnostics indicate that the baseline hazard increases and then decreases at a slower rate over time, suggesting that either a parametric log logistic model or a Cox model is appropriate. We employ the Cox model (see, e.g., Cleves, Gould, Gutierrez and Marchenko 2008), which specifies the hazard as 0 h( t x ) = h ( t)exp( x β ) j j x

18 where h ( ) 0 t is a baseline hazard rate, and j x represents the explanatory variables. For easier interpretation, the results we report for each covariate x j in the Cox model is not the coefficient β j, but rather its exponential form, which is called the hazards ratio. The significance of this hazards ratio must be gauged by comparing it to 1.0 rather than zero. The standard errors of the exponentiated coefficients are obtained by applying the Delta method to standard errors of the coefficients (Cleves et al 2008, page 133). This allows direct and intuitive interpretation of the results. For instance, if we have a dummy variable of whether the borrower is from a usury state, and the hazards ratio is 1.2, it means that people from the usury states are 20% more likely to default than those from a state without the usury law. To facilitate interpretation, we scale the ratios in our Cox model (including bank rate, interest rate on the borrowing, and debt to income ratios) so the hazards ratios in the Cox model are interpreted as the increase in risks with a percentage increase of the covariate. This transformation does not alter significance levels but helps interpret the results of the Cox model. The statistical results are in Table 6. More credit inquiries and higher debt to income ratios have hazard ratios exceeding 1, so they increase the probability of default. For instance, the hazards ratio on credit inquiries is 1.037, indicating that one additional credit inquiry is associated with an average of 3.7% increase in the loan hazard rate. Bank credit utilization displays quadratic curvature. The hazard ratio for bankcard utilization is 0.996, and is significantly less than 1, so some bankcard utilization decreases the risk of default. However, the squared term has a hazards ratio significantly greater than 1, suggesting that excessive card utilization increases the probability of default. Hard credit variables that are associated with higher funding probabilities are also the covariates likely to result in lower default rates. Auction characteristics yield results of mixed statistical significance. Closed auctions, which are more likely to result in funding, might signal more financially constrained borrowers who are more vulnerable to risk, in which case they should result in greater default. However, the hazard ratio for auction format is not significantly greater than 1. The auction category results are not significant. Borrowers with business loans are about 24% more likely to default, but the significance level is only marginal. Debt consolidation loans and loans indicating a purpose are about as likely to default as other loans, with hazard ratios of 0.99 and 0.98, indistinguishable from 1. However, larger loan amounts requested are more likely to result in default, as indicated by a significant hazards ratio of Turning to the social network variables, we find that the total number of friends is insignificant as a predictor of default in Table 6 (Specification C1). In specification C2, we decompose the friends into those with verified identities on Prosper, i.e., those who classify themselves as lenders or borrowers, and friends with no verification. As in Table 5, we find a consistent pattern, other things equal, having more friends without verified identities actually increases the odds of default with a hazards ratio of 1.05 while friends with verified identity decrease the odds of default. Neither variable is significant. In specification C3, we find sharper effects for the number of Prosper verified friends who are potential lenders. Having lender friends decreases default risk, with a hazards ratio of 0.91 significant at 1%, while having borrower friends (with similar financial needs) is insignificant. Specification C4 includes the number of lender friends but this time controlling for whether they actually participate in lending. Having real lender friends further decreases the hazards ratio to 0.88, indicating even lower probability of default, and the hazards ratio decreases to 0.86 when we consider friends who bid on a listing. The coefficients are significant at 1%. Similarly, friends who bid on and win a listing, these types of listings have still lower hazard rates of 0.79, significant at 1%. The odds of default are significantly reduced when friends have a personal stake. Financial stakes taken by friends appear to be the strongest information signal for outside lenders that a borrower is credit worthy. Alternatively or additionally, peer pressure is generated when friends take stakes in a borrower's listing, generating a positive externality to not default on loans. However, the data suggest that this is not a first order force because the median contribution of friends is less than 1%. Thus, the positive social capital communicated by the existence of verifiable friends appears to be the major reason why social networks matter in the Prosper.com setting. In terms of group characteristics, Table 6 shows that only two matter for loan performance, alumni groups and geography-based groups. Interestingly, of the various groups considered in our study, only these two groups

19 contain verifiable information about members, borrowers need to prove that they were actually part of the relevant organization before they can join alumni groups (such as universities or companies), and geography information is verified during the registration process. Being members of these two groups increases the probability of the loan being funded and decreases the risk of default. None of the other groups impacts the risk of default. Controlling for the type of the group, we find that group size does not affect the loan performance or the interest rate of funded loans. We also test for non-linearities in group sizes, and find no significant effect on the interest rate or risk of default. Group leader rewards might increase the marketing for a listing, so listings with these reward features are more likely to be funded. However, it is not clear whether these efforts result in better or worse quality listings because the rewards are contingent only upon origination and not ex-post performance. Similar arguments have been made for the origination-sale models of loan securitization. If banks originate loans that are sold off after origination, the result is a lower quality pool of loans originated by the bank. Such incentives and transactions have been argued to be at the heart of the 2008 financial crisis, because they encourage loan generation without regard to repayment. These distortions are evident in our results as well: we find that loans with group leader rewards are about 15% riskier than others, even though the loans were more likely to be funded and the interest rates were lower. Not surprisingly, such group leader rewards were discontinued by Prosper. Of the remaining variables, interest rates have a positive effect on default in Table 6. In states with usury law limits on interest rates, riskier borrowers may elect to come to Prosper.com because they cannot obtain loans from their home states, in which case their loans could be riskier. While the point estimate for the usury law state coefficient exceeds 1, the difference is not significant. The length of the text description increases the probability of funding but in Table 6, it has no appreciable effect on default. The Cox hazards model estimates can be used to recover estimates of the baseline hazard function (Kalbfleisch and Prentice 2002; Cleves et al 2008). As shown in Figure 3, the baseline hazard of default increases sharply at the beginning and the slowly, reaching a peak at about 10 months, and then slowly wears off. This pattern is remarkably consistent with the analysis of consumer lending delinquency patterns in Gross and Souleles (2002, page 327). 5.3 Interest Rate (Price) Effects Our third specification examines the interest rate at which a loan is funded. The motivation for this model is straightforward. Sections 5.1 and 5.2 show that social network variables have beneficial quantity effects, they increase the likelihood of loans being funded, and are associated with lower default rates. The question is whether social network variables have complementary price effects for borrowers. We examine this issue by regressing interest rate spreads on loans on social network variables plus controls. The observed interest rates on loans come with a selection bias because we observe interest rates only when listings are successfully funded. Although Prosper.com provides an ongoing interest rate for listings that are not fully funded, it is usually the asking interest rate posted by the borrower, and does not reflect the reaction of the lenders to such requests. The actual interest rate for listings that do not get funded remains unobservable. We account for selectivity by using the Cragg (1971) model that specifies an equation for the probability of being funded as a function of observables and an unobserved error, and a second equation for the interest rate, which is observed only if a listing is successful. We use the two-step method of Heckman (1979) to estimate the coefficients (see Cameron and Trivedi 2005). The model can be identified through exclusion restrictions or non-linearity intrinsic to selection models, the latter effectively identifying the model through functional form. We obtain substantively similar results through both methods. We implement the exclusion restriction by using SPIKEDAYS in the probit model as an instrument that affects probability of funding but not the interest rate. SPIKEDAYS passes the strong instrument test of Staiger and Stock (1997). Its F-statistic exceeds 50, well above the strong instrument cutoff of 10 suggested by Stock and Staiger. The results with the instrument included are reported in Table 7.

20 In Table 7, all credit grades have positive and significant coefficients, and the coefficients increase as the credit quality worsens. The point estimates suggest that there is a premium of 4.9% for worse credit grades. A large number of credit inquiries increase the interest rate of funding. Bankcard utilization shows a curvature similar to the other models estimated before. The level has a positive coefficient and the quadratic term a negative coefficient and both are significant. Thus, greater credit card utilization first decreases the interest rate and then increases it, consistent with lenders' valuing some but not excessive utilization of credit card limits. As before, higher debt-toincome ratios lead to worse outcomes, i.e., higher interest rates on funded loans. The results for the auction type reveal a tradeoff in choosing an auction type. While Table 5 shows that a closed auction format increases the probability of a loan being funded, Table 7 reveals that it comes at a cost, closed auctions are likely to result in higher interest rates. With regard to loan purpose, the results for interest rates are consistent with those for the funding probability and default rates. Business loans are perceived as being riskier and command higher interest rates, while debt consolidation and specifying some listing category versus specifying none result in lower interest rates. Other things equal, greater loan sizes require higher interest rates. This is not a pure supply side effect since individual loans are small parts of the overall supply of available funds, and because bigger sized loans are more likely to default, as seen in Table 6. We consider the remaining control variables. We include both the linear and the squared interest rate terms in the specification for empirical consistency with the other models, but the theoretical motivation for the non-linearity in interest rates is less obvious. Credit rationing theories argue that borrowers who are willing to pay high rates are more likely to be rationed but it is not obvious that there is a similar implication for the loans that are actually funded. The linear interest rate term is negative in four specifications and positive in two others while the squared term is consistently positive in all models. In unreported results, we examine models with the linear term alone and report a positive and significant coefficient. Usury law state borrowers have a positive coefficient, indicating that they face higher interest rates. The results relating the social network variables to interest rates are remarkably consistent with those for funding probability and default rates. None of the structural variables including centrality, coreness, effective size, and efficiency affect loan interest rates. On the other hand, the variables reflecting role and identity matter, with direction and gradation consistent with the results for funding probability and ex-post default. Connections to friends not verified by Prosper.com increase interest rates, while more connections to verified friends has the opposite effect of decreases interest rates. Connections to friends who can lend decreases interest rates by 60 basis points, while connections to borrower-friends have no effect. Both connections to real lenders and those to potential lenders matter, and the real lender coefficient results in a greater decrease in interest rates to a 70 basis point effect. These effects and their significance remain quantitatively similar when we consider real lenders who win or lose on the specific borrower auction. Interestingly, having potential lender-friends who do not bid on a borrower's listing hurts, increasing loan spreads by about 20 basis points. The group variables also explain interest rates in a fashion largely consistent with the funding probability model. Belonging to a group that has a religious motif lowers the interest rate on loans significantly, although it has no effect on default rates. Business or university alumni affiliation groups show an even stronger effect, lowering interest rates by close to 120 basis points, consistent with the lower ex-post default rates for such loans. The rewards for group members show an interesting effect, individual lenders are more likely to fund loans in which group leaders are rewarded and the interest rates are also lower by about 50 basis points for such listings. The market does not appear to internalize the potential distortion in group-leader incentives. Geography based groups have insignificant interest rate effects and about a 10% significance in ex-post loan defaults. The group size itself has marginal statistical significance and little economic significance in explaining interest rates; similar in spirit to the negligible effects on ex-post default rates for this variable in Table 6.

21 5.4 Additional Robustness Tests To examine the robustness of our findings, we conducted a number of tests in addition to the main models presented and discussed above. A full set of results from these robustness checks is available from the authors upon request One Borrower, Multiple Listings We examine robustness to alternative econometric specification. We consider potential panel data effects in the probability of funding equation. Borrowers whose listings have failed can relist on Prosper.com with a fresh request. Thus, there can be correlations among listings created by the same borrowers. These can be properly accounted for using panel data methods. We note that ours is a highly unbalanced panel dataset, with over 53% of members posting only one listing, rendering fixed effect models unviable. We use a random effects probit model to estimate the probability of funding using a panel data setup, where each borrower is a unit. The results from this panel data model are similar to the previous results for the funding probability. On the other hand, the standard panel data setup may not be appropriate because of Prosper.com rules, under which a borrower can have only one listing at a time and cannot have more than two loans concurrently outstanding. Thus, it may be likely that borrowers have multiple listings because previous listings were not successfully funded. As an alternative approach, we model the number of listings that a borrower needs to post before getting funded. Reversing the terminology of survival analysis, a failure is considered to occur when the borrower stops requesting loans, and the data is considered censored if he has not yet successfully funded his loan. Our primary dependent variable here is the time to failure, or the number of listings that a borrower creates before he obtains funding. Again, we estimate it using a Cox proportional hazards model. Our main results do not change. For instance, having one additional unverified friend decreases the probability of getting funded by 5.6%, while having a real lender friend who bids and wins increases the probability of getting funded by 33% Continuous Measure of Funding We further test other specifications for the funding probability model. One possible model is to use a continuous measure of funding as the dependent variable, the percentage of funding obtained. This is the ratio of committed bids from lenders to the amount requested by borrower. We downloaded the data from Prosper.com and used it in the econometric specification. The data show that this variable has significant spikes at 0 and at 1. We fit a double censoring Tobit model to estimate it. These results are similar to those in Table 5. The specification also provides an additional avenue to rule out the proposition that the relation between friends' bidding and funding probability could be positive merely mechanically. We eliminate all the bids made by a borrower's friends and reestimate the Tobit model. The results remain unchanged. We also reestimated the econometric model for the probability of funding (rather than the amount). Once again, the results are similar, which is unsuprising because friends account for a relatively small portion -- 1% -- of the overall bid. It appears that friends matter not for the physical capital they bring to a listing but for the positive signals that having verifiable friends sends about the social capital of a borrower Auction format Another potential confounding factor is the choice of auction formats. Unobservable borrower characteristics could both influence their choice of the auction format, as well as affect the outcome of the listing. To account for this possibility, we use an instrumental variable (IV) method. The instrument should capture the urgency of the borrower to obtain loans, urgent borrowers are more likely to use a close type auction, but this will not affect the outcome directly. One such candidate is the number of days between registration date and the first listing date. An additional complication, however, is that both the potentially endogenous variable (auction format) and the outcome variable (whether the listing is funded) are both binary variables. As Wooldridge (2002) points out, using an IV probit method is inappropriate in this situation. An alternative, according to Maddala (1983), is to use a multivariate probit model. We use a multivariate probit specification (Cappellari and Jenkins 2003) to estimate our

22 model. The estimation produces similar results to the ones reported here. We also test for potential effects of some other variables not included in our primary models. These variables yield results that have marginal significance at best. An interesting variable is the number of years since a borrower's first credit line, a proxy for the borrower's age and credit experience. It has small effects on the funding probability and interest rate and no effect on default Repayment of trust The subject of trust has attracted a lot of attention in the recent economics and finance literature, which examines the relation between this dimension of social capital in facilitating economic and financial transactions. The P2P market provides an interesting avenue for examining trust. Given the arm's-length relation between the borrowers and the many small lenders, funding a loan can be viewed as an act of trust by a lender in a borrower. We test whether the trust is repaid by analyzing the relation between long-term default performance and the probability of getting funded. This is equivalent to estimating the system of equations π =Probability (funding) = α 1 Hard Info+α 2 Social Network + α 3 Other Variables+ε i h(t) = γ 1 Hard Info+γ 2 Social Network + γ 3 Other Variables+γ 4 π +η i We use an approach similar to instrumental variable methods. More specifically, we use the competition among borrowers (calculated as the number of auctions that begin in the same week as any given listing) as an instrument variable, because theoretically such competition should only have an effect on whether the borrower can obtain loans, but not on how the borrower repays. Using the logic of two-stage least squares, we include this instrument to predict the probability that a listing can be funded. The instrument satisfies the Staiger and Stock (1997) strong instrument condition. We include the predicted probability as an additional covariate in the Cox proportional hazards model. Our results show that trust is indeed repaid, listings that are more likely to be funded are also more likely to be repaid. In fact, a 1% increase in the funding probability translates into 1% lower default probability Exploratory network analysis and other network metrics We also conducted an exploratory social network analysis of the online network of friends. This yields some interesting results not visible otherwise. As an overview, the friendship network consists of many disjoint segments, or components. Among them, the largest component has 403 members. There is no significant overlap between the friendship network and the group network. Using the largest component as an example, very few friends in the same component belong to the same groups. This suggests that we need not worry about high correlation between friendship metrics and group metrics, therefore it is appropriate to include both group and friendship measures simultaneously. Despite their prominent positions on the friendship network, some members actually have no roles. Some members who have no roles are actually cutpoints in the friendship network; in other words, removing them will increase the number of components. This further strengthens our argument that it is important to look at the roles and nature of friends before we can argue for the effect of friendships. Finally, Coleman (1988) argues that the enforcement of norms depends on network closure; therefore closures can help reduce unethical and opportunistic behaviors (Burt, 2005). A closure exists when two friends of an ego are connected to each other, thereby reducing the power of the ego. In our exploratory analysis however, we found that many components of the friendship network are actually star-shaped, with very little closure. In terms of using other social network variables, one variable that does matter is the number of friends' defaults in a borrower's neighborhood (ego network). The results indicate that a higher number of defaults in a neighborhood of a borrower is associated with higher risk of the ego's loan. We also include the number of endorsements. As noted earlier, friends can also endorse or recommend borrowers but this is relatively cheap talk and should add little to the results based on friendship network or actual bidding by friends. Our results show that having endorsements has no impact on loan performance. This result holds for both having/not-having endorsements as well as for the number of endorsements received. Our previous work examines bids from friends. It might be possible that bids from other members in the same group as the borrower (peer group members) can help reduce the risk of loans. Our results show that this is not true, once we control for other variables that we used (hard credit information, other social network information and so on), bids from peer group members have no effect on any of the transactional outcomes.

23 5.4.6 Other variables In an earlier regime, Prosper.com allowed users to choose the duration for their auctions, but has lately standardized the durations at 7 days. These types of variables as well as other unobserved changes in the website should be subsumed by the time period quarterly effects. As anticipated, once we include quarterly fixed effects, we no longer observe any significant effects of this variable. Prosper.com provides to lenders other additional hard credit information available from the borrower's credit report, including number of public records, amount delinquent, and the number of revolving credit lines. These variables are highly correlated with the hard credit variables used in our analysis, and adding them does not change our findings either. 6. Conclusions and Implications Recent developments in digitization and Web 2.0 technologies have dramatically altered the way individuals can interact and connect with each other. These changes have spawned significant growth in online social networks such as Facebook, LinkedIn, and MySpace. These online social networks serve as the cornerstone for new business models that promise to transform traditional commerce. Our study looks at one of the first attempts to build such businesses, online peer to peer lending without financial intermediaries. One view of our study is that it is an analysis of the economic value of social networks and more generally, social capital. Our results support the view that social capital matters in influencing economic outcomes, consistent with the view of prior work in economics and sociology. While the literature in these areas emphasizes the relevance of social capital to individuals in social networks or the organizations that employ them, our work suggests that it can also be harvested by individuals in transactions with outside third parties such as the anonymous lenders that populate the P2P marketplace. The realization of these benefits depends on the extent to which the networks exert social influence, and equally importantly, the extent to which they are visible to and verifiable by potential lenders. A second perspective of our study is that it represents data from a natural experiment in which there is a particularly severe problem of adverse selection. Peer-to-peer lenders lack the sophisticated risk assessment methodologies, scale economies, or soft information in lending from a broader vector of banking relationships available to banks and centralized lending institutions. Social networks appear to help mitigate the adverse selection problem, acting as a new source of soft information that is useful in determining lending outcomes. Alternatively, or additionally, our study can be viewed as characterizing trust. As pointed out by Sapienza, Toldra and Zingales (2007), or Carlin, Dorobantu and Viswanathan (2007), trust is a critical driver of economic transactions when the environment is unstructured. In fact, Guiso, Sapienza and Zingales (2004) remark that financial contracts are the ultimate trust-intensive contracts. From this viewpoint, our results suggest that social networks facilitate economic transactions by building trust that is repaid by borrowers. Our results have other implications. With regard to the design of businesses based on online networks, our results suggest these markets should incorporate and facilitate ease of using multiple (structural and relational) social network metrics for end-users. In fact, one could make a reasonable case that P2P markets should incorporate functionalities that not only promote interactions among members, but also enable borrowers to credibly signal their embeddedness in their social networks to lenders. Indeed, Prosper.com has taken steps in this direction in more recent listings where social network information is given greater prominence in listings seeking funding. In a similar vein, our results show that in addition to friendship networks, groups can also play a valuable role in reducing information asymmetries. Thus, increasing the interdependence among group members and making these ties verifiable can facilitate capital flows, thus enlarging the scope and applicability of microfinance-style mechanisms. It is also interesting that while technology may be viewed as disintermediating financial markets by breaking relationships between lenders and borrowers, there is also a mitigating effect. Technology also facilitates the advantages of relationships by enabling generation, transmission, and hardening of soft information that is conventionally inaccessible or unavailable. From the viewpoint of evaluating the market value of social network based business models, our evidence suggests an implication that is perhaps not obvious at first sight: merely having several friends does not assure economically useful social capital. It is not the size of a borrower's network or the number of social ties, but rather its quality, as

24 judged by interactions and embeddedness that determines positive social capital. Consequently, the nature of the social ties and the relational aspects of the network need to be carefully considered in evaluating their effectiveness. While our work focuses on borrowers in online markets, an equally important and intriguing question is the role of social networks to lenders. It is interesting to understand how a lender's social network affects her investment strategies and their performance. We leave that for future research. Some of our results have interesting parallels in the 2008 financial crisis. We uncover an interesting analogy to the originate-sell model used by banks to originate and finance mortgage backed securities. Banks that originate, bundle, and sell of mortgages through securitization have little incentive to pay heed to loan repayment, which could result in moral hazard problems. These moral hazard problems have been held culpable for the 2008 financial crisis. There is a similar natural experiment in our data when group leaders were paid fees for originating loans. We find evidence that there is more aggressive marketing, which results in loans being more likely to be funded. However, these loans turn out to be ex-post riskier because they are more likely to default. The experiment offers yet another piece of evidence for the moral hazard effects of the originate-sell model underlying mortgage securitization. From a normative viewpoint, our findings suggest that social networks could be harnessed as valuable tools in dealing with mortgage foreclosures. Targeting incentives and programs at borrowers who can demonstrate verifiable support from individuals in their social networks could result in better-targeted mortgage renegotiation programs by mitigating the adverse selection costs of an across-the-board program. The alternative approach may even be cost effective by using electronic methods similar in spirit to online P2P lending. REFERENCES Agarwal, S., & Hauswald, R. B The choice between arm's-length and relationship debt: evidence from eloans, Akerlof, G. A The market for "lemons": quality uncertainty and the market mechanism. The Quarterly Journal of Economics 84(3): Arrow, Kenneth J Models of job discrimination. In A.H. Pascal, (eds) Racial discrimination in economic life. Lexington, MA: D.C. Heath,: Bagozzi, R. P., & Dholakia, U. M Open source software user communities: a study of participation in linux user groups. Management Science, 52: Bampo, M., Ewing, M. T., Mather, D. R., Stewart, D., & Wallace, M The effects of the social structure of digital networks on viral marketing performance. Information Systems Research, Vol. 19: Benmelech, E., & Moskowitz, T. J The political economy of financial regulation: evidence from U.S. state usury laws in the 18th and 19th century: SSRN. Becker, Gary S., 1971, The economics of discrimination. University of Chicago Press, Chicago. Bhattacharjee, S., Gopal, R. D., Lertwachara, K., Marsden, J. R., & Telang, R The effect of digital sharing technologies on music markets: a survival analysis of albums on ranking charts. Management Science, 53: Bikhchandani, S., Hirshleifer, D., & Welch, I A theory of fads, fashion, custom, and cultural change as informational cascades. Journal of Political Economy 100(5) Burt, R Structural holes: the social structure of competition. Harvard University Press, Cambridge, MA. Burt, R Brokerage and closure: an introduction to social capital. Oxford University Press. Calfee, J. E., & Ford, G. T Economics, information and consumer behavior. Advances in Consumer Research, 15, Campbell, J., & Fairey, P Informational and normative routes to conformity: the effect of faction size as a function of norm extremity and attention to the stimulus. Journal of Personality and Social Psychology, 57(3): Cappellari, L., & Jenkins, S Multivariate probit regression using simulated maximum likelihood. The Stata Journal, 3(3):

25 Carlin, Bruce, Florin Dorobantu, and S. Viswanathan, 2007, Public trust, the law, and financial investment, Journal of Financial Economics (forthcoming). Cialdini, R. B., & Goldstein, N. J Social influence: compliance and conformity. Annual Review of Psychology, 55(1): Cleves, M., Gould, W., Gutierrez, R., & Marchenko, Y An introduction to survival analysis using Stata. Stata Press. Cohen-Cole, E., & Duygan-Bump, B Household bankruptcy decision: the role of social stigma vs. information sharing, working paper: Federal Reserve Bank of Boston. Cohen, Lauren, Andrea Frazzini, and Christopher Malloy, 2008a, The small world of investing: board connections and mutual fund returns, Journal of Political Economy 116 (5), Cohen, Lauren, Andrea Frazzini, and Christopher Malloy, 2008b, Sell-side school ties. Unpublished working paper, Harvard University. Coleman, J Social capital in the creation of human capital. American Journal of Sociology, 94(S1): 95. Cowan, R., Jonard, N., & Zimmermann, J.-B Bilateral collaboration and the emergence of innovation networks, Management Science, 53: Cragg, J Some statistical models for limited dependent variables with application to the demand for durable goods. Econometrica, 39(5): Darby, M. R., & Karni, E Free competition and the optimal amount of fraud. Journal of Law and Economics, 16(1): 67. Diamond, D Financial intermediation and delegated monitoring. Review of Economic Studies, 51(3): Diamond, D Monitoring and reputation: the choice between bank loans and directly placed debt. Journal of Political Economy, 99(4): 689. Drucker, S., & Puri, M On loan sales, loan contracting, and lending relationships. Review of Financial Studies, forthcoming. Durlauf, S. N., & Fafchamps, M Social capital, forthcoming in Handbook of Economic Growth, Steven Durlauf and Philippe Aghion, Editors. Fama, E What s different about banks. Journal of Monetary Economics, 15: Goffman, E Stigma: Notes on the management of spoiled identity. Englewood Cliffs, NJ. Gorton, G., Winton, A., G.M. Constantinides, M. H., & Stulz, R. M Chapter 8 Financial intermediation, Handbook of the Economics of Finance, Volume 1, Part 1: : Elsevier. Granovetter, M The strength of weak ties. American Journal of Sociology, 78(6): Granovetter, M Economic action and social structure: the problem of embeddedness. American Journal of Sociology, 91(3): 481. Granovetter, M Problems of explanation in economic sociology. Networks and Organizations: Granovetter, M The impact of social structure on economic outcomes. Journal of Economic Perspectives, 19(1): Grewal, R., Lilien, G. L., & Mallapragada, G Location, location, location: how network embeddedness affects project success in open source systems, Management Science. 52: Gross, D. B. and N. S. Souleles An empirical analysis of personal bankruptcy and delinquency. Review of Financial Studies 15 (1), Grossman, S. J., & Stiglitz, J. E On the impossibility of informationally efficient markets. American Economic Review, 70(3): 393. Guiso, G., Sapienza, P., & Zingales, L Role of social capital in financial development. American Economic Review, 94(3): 31. Hanneman, Robert A. and Mark Riddle Introduction to social network methods. Riverside, CA: University of California, Riverside Hochberg, Y. V., Ljungqvist, A., & Lu, Y Whom you know matters: venture capital networks and investment performance. Journal of Finance 62 (1): James, Chris, Some evidence on the uniqueness of bank loans, Journal of Financial Economics 19, Kalbfleisch, J. D., and R. L. Prentice The statistical analysis of failure time data. 2nd ed. Wiley, New York Karlan, D Social connections and group banking. The Economic Journal, 117(517):

26 Latane, B The psychology of social impact. American Psychologist, 36(4): Ledgerwood, J Microfinance handbook: The World Bank. Liebowitz, S File sharing: creative destruction or just plain destruction? The Journal of Law and Economics, 49(1): Malone, T. W., Yates, J., & Benjamin, R. I Electronic markets and electronic hierarchies. Communications of the ACM, 30(6): Manstead, A., & Hewstone, M The Blackwell encyclopedia of social psychology: Blackwell. Mizruchi, M. S The structure of corporate political action: inter-firm relations and their consequences: Harvard University Press. Moran, P Structural vs. relational embeddedness: social capital and managerial performance. Strategic Management Journal, 26(12): Morduch, J The microfinance promise. Journal of Economic Literature 37 (4): Nelson, P Information and consumer behavior. Journal of Political Economy 78 (2) Phelps, Edmund A., The statistical theory of racism and sexism. American Economic Review 62, Petersen, M Information: hard and soft. Unpublished working paper, Northwestern University Petersen, M., & Rajan, R The benefits of lending relationships: evidence from small business data. Journal of Finance, 49: 3-3. Petersen, M. A., & Rajan, R. G Does distance still matter? the information revolution in small business lending. Journal of Finance, 57(6): Pope, Devin and Justin Sydnor, 2008, What s in a picture? Evidence of discrimination from prosper.com, Unpublished Working Paper, University of Pennsylvania. Portes, A Social capital: its origins and applications in modern sociology. Annual Review of Sociology, 24: Putnam, R The prosperous community: social capital and economic growth. American Prospect, 4(13): Ravina, Enrichetta, 2008, The effect of beauty and personal characteristics in credit markets. Unpublished working paper, Columbia University. Rosenthal, P. I Specificity, verifiability, and message credibility. Quarterly Journal of Speech, 57: Sabini, J., & Silver, M Moral reproach and moral action. Journal for the Theory of Social Behaviour, 8(1): Sapienza, P., Toldra, A., & Zingales, L Understanding trust. NBER Working Paper. Shane, S., & Cable, D Network ties, reputation, and the financing of new ventures, Management Science, 48: Stiglitz, J. E., & Weiss, A Credit rationing in markets with imperfect information. American Economic Review, 71(3): Uzzi, B Embeddedness in the making of financial capital: how social relations and networks benefit firms seeking financing. American Sociological Review, 64: Uzzi, B., & Lancaster, R Relational embeddedness and learning: the case of bank loan managers and their clients, Management Science, 49: Williamson, S. D Costly monitoring, loan contracts, and equilibrium credit rationing. Quarterly Journal of Economics, 102(1):

27 Figure 1 Hierarchy of Friends # Friends Level 1 # Friends with no roles # Friends with roles Level 2 # Pure borrower friends # Lender friends Level 3 # Potential lender friends # Real lender friends Level 4 # Potential lenders friends no bid # Potential lenders friends who bid # Real lender friends who bid # Real lender friends who do not bid Level 5 # Real lender friends who bid & win # Real lender friends who bid but lose Figure 2 Screenshot of a Prosper listing