The Geography of Lender Behavior inPeer-to-Peer Loan Markets∗

Garrett T. Senney

Office of the Comptroller of the Currency†

Original Draft: November 10, 2014Current Draft: February 20, 2017

Abstract. Online platforms promise to overcome the traditional geographic constraint oninvesting due to the cost reduction of online search. Using transaction data from a peer-to-peerlending site, I find strong evidence that lenders are better informed about local projects andthe effect is not driven by "friends and family". My results are consistent with a model ofsocial learning with heterogeneous informed agents. Listings with more early local bidding willattract more lenders, leading to a higher probability of funding and a lower final interest rate, iffunded. This work supports the idea that the behavioral difference between lenders is driven mostlyby informational frictions. Localness embodies some incidentally obtained initial informationaladvantage which rationally causes a disparity in information due to the presence of nonzero searchcost. This asymmetry contributes to explaining why geographic-based frictions are still presentand relevant in online loan markets.

Keywords: Geography, Peer-to-Peer Lending, Informational Frictions.

JEL Classification: D82, D83, L10, L86.

1. Introduction

It is well established that most of the inefficiencies in the credit market are due to theasymmetric information between lenders and borrowers (Stiglitz and Weiss, 1981). Ho-wever, innovations and technological advancements are radically altering the delivery offinancial services. The Internet has enabled people to share more information and developlarger networks. Empirical work has consistently found that the Internet can overcome ge-ographic isolation1 and that search costs are lower online than offline.2 Utilizing these newand cheaper online connections in conjunction with emerging crowdfunding platforms,

∗Special thanks to Jason Blevins for his guidance and insight, and Efraim Berkovich for providing the data.†The views in this paper are those of the author and do not reflect those of the Office of the Comptroller of

the Currency or the Department of the Treasury.1 Forman, Ghose, and Goldfarb (2009); Choi and Bell (2011).2Bakos (1997); Baye, Gatii, Kattuman, and John (2009).

1

individuals are starting to find novel ways to address and overcome the asymmetry in thecredit market. In this paper, I explore the effectiveness of online peer-to-peer lending toaddress the informational frictions that traditional plague the credit market.

Peer-to-peer (henceforth P2P) lending is a mechanism for investors to lend moneydirectly to individual borrowers. This arrangement creates the possibility that borrowerscan obtain loans at lower interest rates than they would get on a credit card or a normalloan without collateral. Individual lenders get the opportunity to invest in short-durationassets with higher rates of return than would be available on bonds or money marketaccounts, all due to the cost savings arising from removing the intermediary.

Additionally, P2P lending has been heralded as a way of leveling the playing fieldby providing financial access to more people. A potential blend of altruism and theallure of higher yields has attracted investors to P2P lending, especially in the aftermathof the 2008 financial crisis, when interest rates and banks’ willingness to lend have hithistoric lows. Compared to traditional banks, online lenders enjoy the advantage thatonline marketplaces currently face fewer regulatory constraints. Moreover, P2P sites arestructured to facilitate faster transfers of capital to the borrowers, and they have loweroverhead costs than companies with physical locations. Therefore, P2P firms operatemore profitably in a market traditionally viewed as fairly risky with low margins (Cowley,2015). P2P lending has established itself as an emerging alternative source of financing forstartups and people with particularly limited access to traditional means of financing.3

Using data from a leading American P2P lending site, I examine whether the Internethas loosened the geographic constraints on investing. My analysis shows that lendersbehave differently based with local lenders being better informed. Furthermore, I can ruleout the "friends and family" and preference explanations to conclude that informationalasymmetry arises primarily due to nonzero search cost. This work expands the rationalherding finding of Zhang and Lui (2012) by providing a theoretically founded mechanismto explain the persistence of information asymmetries across lenders over time.

Early work on P2P lending has focused on the determinants of a listing being fundedand the final interest rates. The consensus is that soft factors such as demographic andnetwork effects are second order in importance,4 after hard factors, such as the verifiedfinancial information.5 Because P2P lending is still in its infancy, its full potential as analternative or supplement to the traditional banking industry is still an open empiricalquestion.

A long literature documents the importance of distance in social and economic behavior.

3Belleflamme, Lambert, and Schwienbacher (2010); Mollick (2014).4Herzenstein, Andrews, Dholakia, and Lyandres (2008); Pope and Sydnor (2011); Ravina (2012); Duarte,

Siegel, and Young (2012).5Klafft (2008b,a); Iyer, Khwaja, Luttmer, and Shue (2009).

2

Theoretical work on early investing and capital ventures predicts that investors and lendersare sensitive to their distance from the borrower, which is especially true in the early stagesof the venture when there is little to no observable history.6 The usual explanation is thatthe cost of gathering information, as well as monitoring, generally increases as the distancebetween lender and borrower grows. The predicted informational asymmetry betweenlocal and nonlocal investors may derive from the informational opacity of startups andsmall firms. Potential investors collect a sizeable amount of information before deciding toinvest. The geographic proximity of local lenders facilitates cheaper and easier access tothis information. Empirical evidence supports these findings.7

The informal capital markets, alternative sources of financing outside of the traditionalmainstream credit market, have not to date been the focus of much scholarly analysis.However, it is generally accepted, as noted by Harrison, Mason, and Robson (2010), thatinformal capital markets comprise a series of potentially overlapping local markets ratherthan one fully integrated national market. Recent empirical work has shown that theInternet has the potential to allow individuals and firms to overcome many traditionalbarriers that have fettered offline markets by mitigating some geographic frictions.8 Otherresearch has found that online platforms might reduce some, but not all, of the distance-related frictions.9 With e-commerce being omnipresent nowadays, it is important toexamine whether this shift in the commercial paradigm coupled with new online tools hasbeen strong enough to alter or even invalidate the theoretical predictions of the geographiceffects on investing.

Although technology makes it easier and cheaper for individuals to gather and processinformation about market conditions, it is unclear if informational frictions such as searchcosts persist in the online marketplace. I examine this ongoing empirical question ofwhether geographic frictions exist on the online credit market, and if so what factorscontribute to their existence. I focus my analysis on the following questions: (1) do locallenders bid different amounts, (2) do local lenders evaluate and price the risk of the listingsdifferently, and (3) do local lenders tend to bid at different times during the auction.Additionally, if distance-related frictions do exist, (4) does the difference in behaviorarise from informational asymmetry or simply a preference on the part of local lenders,and lastly, (5) how does the presence of local lenders in the market affect other lenders’

6Arthur (1988, 1990); David and Rosenbloom (1990)7Florida and Smith (1988); Sohl (1999); Sorenson and Stuart (2001); Powell, Koput, Bowie, and Smith-Doerr

(2002); Mason (2007); Wong, Bhatia, and Freeman (2009).8Ratchford, Pan, and Shankar (2003); Brynjolfsson, Hu, and Rahman (2009); Goldfarb and Tucker (2011);

Lendle, Olarreaga, Schropp, and Vézina (2013).9Blum and Goldfarb (2006); Hortaçsu, Martínez-Jerez, and Douglas (2009); Agrawal, Catalini, and Goldfarb

(2011); Lin and Viswanathan (2014).

3

decisions to enter and their behavior after entering.Using bid- and listing-level transaction data from lending auctions on Prosper.com, I

estimate that local lenders tend to bid earlier and larger amounts than nonlocal lenders.Furthermore, local lenders also seem better able to evaluate the riskiness of loan requests.They tend to ex-ante bid larger interest rates when the loan ex-post defaults and lesswhen the loan ex-post pays back in full. This is true even when controlling for thepotential peer effects. Thus reconciling theory and previous empirical work, my resultsdemonstrate strong evidence for the informational-based nonzero search cost explanationof the observed behavioral differences between lenders.

2. Overview of Peer-to-Peer Lending

P2P lending is a process through which an individual or firm attempts to obtain financingby soliciting for usually small contributions from a large number of online investors. Thesenew financing platforms are the result of a social movement that arose in reaction tothe emergence of new technologies that enable new and cheaper ways of forming socialnetworks (Adams and Ramos, 2010). The financial technology sector has already growthwhich large; crowdfunding, which P2P lending is a part of, accounted for more fundingthan venture capital in 2016 (Massolution.com, 2015). The two current largest domesticplayers combined issued more tha $9 billion in 2015 (NSR Invest, 2016). In this paper, Ifocus on the P2P lending site Prosper.com, which has been called the "eBay of loans".10

Prosper provides unsecured, 36-month, fixed-rate personal loans ranging from $1,000to $25,000 to legal U.S. residents. The members’ true identities, addresses, and othercontact information are never publicly disclosed by the site. For privacy, the borrower isprohibited from releasing that information to lenders. However, the borrower’s state ofresidence is publicly displayed on the listing.

A borrower creates a listing requesting a specified amount of money and the maximuminterest rate that the borrower is willing to accept.11 Borrowers specify a purpose of theloan (Debt Consolidation, Business, etc.) and write a brief description supporting theirrequest. Each listing displays a credit profile about the borrower including debt-to-incomeratio, credit grade (40-point bands of the borrower’s Experian credit score), and homeownership status.

10The market description in this section refers to market conditions and mechanisms that were in placewhen the data was collected. Some policy and regulation changes have since occurred, making the currentmarket slightly different.

11After some institutional and legal restructuring in 2008, Prosper started to collaborate with a Utah-chartered industrial bank, to utilize a 1978 Supreme Court decision allowing all Prosper borrowers to avoidtheir individual state’s usury law and face a uniform legal maximum borrower rate of 36%.

4

The funding mechanism is a descending-price uniform-share auction. The lenders’bids have two components: quantity and price. The bid amount ($50 minimum) is thequantity demanded and bid interest rate is the smallest interest rate that the lender iswilling to accept for a particular loan. The bidding process is proxy bidding. Each bidis considered independent, so a lender may bid multiple times with potentially differentinterest rates. These price-quantity pairs form the supply curve of available funds for theloan requests. The auction is partially open; lenders always see the number of bids andthe quantity of money pledged of each bid. The interest rate submitted by the lenders isonly shown for losing bids; accordingly, before the loan is fully funded, lenders only seethe borrower’s maximum rate.

Successfully submitted bids cannot be rescinded. Given that this platform is a collectionof individualized markets for each individual loan, each market clears when the amountof pledged funds is at least as large as the requested amount, with the interest ratedetermined by the auction. In the case of ties, bids placed earlier take precedence overlater bids. When the auction closes, the bids are sorted by bid interest rate. The bids withthe lowest rates are bundled until the total loan amount has been reached and are thencombined into a single loan. Each winning lender receives the same interest rate, which isdetermined by the marginal losing or last winning bid, depending on the auction. Winningbids are either fully or partially participating in the loan; a partially participating bidmeans that the lender is allocated a smaller share of loan than his or her stated demand.If the loan request is not fully funded by the listing’s close date, the listing is closed anddismissed. Successful loan requests get further review by the site to initiate the neededlegal documentation for the loan to be originated. Borrowers who default on their Prosperloans are barred from using the site again.

3. Data

In this work, I use Prosper data containing all loan requests using the open auction formatposted by borrowers with FICO credit scores greater than or equal to 560 that were activefrom January to October 2008. During the period in my sample, 42,657 loan requestsand 2,022,910 bids were posted. A total of 9,624 listings were fully funded. The publiclyavailable characteristics of the borrowers are: debt-to-income ratio (henceforth DTI, thevariable is top coded at 10.1), credit grade, homeowner status, whether the borrower is ina Prosper group, and state of residence.

Klafft (2008b) confirms that credit grade and DTI are the two most important variablesin determining the financial outcomes. The site provides members the ability to join groupsthat are designed to develop and foster a community of lenders and borrowers, akin to

5

what occurs in relationship lending, contract enforcement via reputation, and peer effects.Prosper groups were mostly centered on shared interests (i.e., college alumni, sports fans,and military members). These groups never gained wide traction. Group membershipwas not generally found to decrease default rates, as anticipated. The ineffectivenessof group networks accounted for many users’ complaints and lack of the expansion ofgroup membership, (Jaworski, 2007). The site now de-emphasizes the role of groupsaltogether. However, since peer effects are known to matter, I collect information on thissocial connection. In the loan-level analysis, the In Group variable indicates whether aborrower is in a Prosper group; and in the bid-level analysis, the In Group variable indicateswhether that particular lender is in the borrower’s specific Prosper group.

In addition to the within-auction variables directly from the data, I construct twovariables –Total Competition and Credit Grade Competition –to measure the competition thateach listing faces. The competition measures are the number of current listings that areactive for at least one full day at the same time that the particular listing is also active. TotalCompetition is the number of total listings, regardless of credit grade, while Credit GradeCompetition is the number of listings from the same credit grade. At the bid level, I recreatethe auction to determine the current standing interest rate, money pledged, and bid countthat existed in the auction at the exact moment that each lender bid on a particular listing.This process allows me to observe the current state of the listing as each lender saw itbefore bidding.

Table 1 shows the descriptive statistics of the borrowers’ characteristics for all listings.Most of the listings request around $6,000–$9,000. Intuitively, the mean requested amountincreases as the credit grade improves, as more credit-worthy borrowers have the ability tosupport larger loans. The median bid amount, regardless of credit grade, is the minimumbid ($50). This result is in line with previous Prosper research that most lenders tend todiversify across listings by pledging small amounts in any one particular listing. Oneimmediate question that might arise, given the size of an individual bid, is why a lenderwould invest in this unsecured market. It has been observed that most lenders commit toinvest around $50–$100 across dozens of loans. Aggregated, a portfolio of an individuallender on a P2P lending site resembles a new asset class, that if diversified correctly, canoffer lenders returns that do not directly follow the motions of stocks and bonds (Lieber,2011).

The nature of the auction mechanism does not allow observing the bid interest ratesfor winning bids. The current standing interest rate is always shown, which is either theborrower’s maximum rate or the interest rate of the first losing bid. The final interest ratesets an upper bound on what the actual bid interest rates may be for the winning bids.Unless otherwise noted, following the literature, I assume that winning bids equal the

6

final interest rate. Not surprisingly, the mean and median bid interest rate increase as thecredit grade worsens. Additionally, the amount of variation in bid interest rates appears toincrease as the credit grade worsens (the correlation between credit grade and standarddeviation of bid interest rate is 0.833).

Table 2 displays that breakdown of the listings and completed loans by credit grade;over half of all of the listing requests are in the bottom credit grades. The simple fundingrate decreases strictly monotonically as the credit grade worsens. I also collect data onthe ex-post loan outcomes for the listings that are actually originated. However, I donot observe how much of the loan’s principle was paid back. I only know if any kindof default occurred. The loan outcome data comes from a different data release, whichcontains only outcomes of the loans that were settled by early September 2011.12 The lastcolumn of Table 2 displays the simple default rate by credit grade for this sample. Whilethese default rates seem rather large, it is important to recall that this is a measure of anykind of default. These magnitudes are well in line with national residential mortgage anddelinquency rates for the same time period.13

4. Model

The central proposition of this paper is that nonzero search cost leads to lenders whoare more informed about local projects. Thus, local lenders are better able to correctlyevaluate the underlying risk in the listing when they bid and are more comfortable biddingearlier, acting mostly on their own information. As pointed out by Devenow and Welch(1996), social learning involves an informational externality where a lender may gain usefulinformation from observing previous lenders’ decisions. Due to the existing informationalasymmetry, nonlocal lenders can learn from observing the actions of others, particularlylocal lenders. This behavior leads to listings with more early local bidding to have morerevealed private information available and thus attract more bidding activity overall.

To motivate my analysis, I abstract away from the auction environment and developa social learning model with heterogeneous agents in the spirit of Banerjee (1992) andBikhchandani, Hirshleifer, and Welch (1992). There is a population of identical risk neutrallenders, who are maximizing their expected utility from investing.14 An L share of thelenders are local and a (1− L) share are nonlocal.15 Each lender is presented with the

12Out of the 9,624 completed listings in my sample, I am able to match 9,099 of them to their final outcomesin 2011.

13Refer to the Richmond Fed Delinquency and Foreclosure Rates Report.14If lenders’ risk tolerances are identical, risk preference has no effect on the outcome of the game.15Smith and Sørensen (2000) also study herding with heterogeneous agents, but the types vary along

preferences, not the information quality.

7

same collection of listings indexed by i ∈ [0, 1]. The monetary return from investing in theith listings is the same for all lenders. Only one unique listing will yield strictly positivereturns, while all other listings will produce returns of zero. This assumption is the sameas if there were a single listing which has excess returns that are strictly greater than allthe other listings. Therefore, the optimal ex-post outcome for all lenders is to invest in thatlisting. For simplicity, each lender’s decision is restricted to simply choosing which listingto invest in.16

At the start of the game, all lenders have the same ex-ante uniform priors about thelistings. However, lenders with probability α ∈ (0, 1) will receive a noisy signal, Si ∈ [0, 1],about which listing to invest in. Local lenders are assumed to be more informed thannonlocal lenders, so their signals will be more informative. The signal of a local lenderis correct with probability β ∈ (0.5, 1), and with probability (1− β) the signal is strictlynoise drawn randomly from U[0, 1]. The signal structure for nonlocal lenders is the sameas for local lenders, except it is less informative. Formally, a nonlocal signal is correct withprobability γ ∈ (0.5, 1), with β > γ.

Lenders do not ex-ante know the local status of the preceding lenders. If their behaviordoes not reveal their status, the current lender will assume that the preceding lender is asinformed as the average lender in the population, θ = Lβ + (1− L)γ. To ensure that thisgame is solvable in a general setting without a complicated dependence on the primitivesof the model in addition to the observed history, I impose some mild structure on howmuch more informed a local lender is. Since lenders move sequentially, in an exogenouslyset randomly order, decisions cannot to be strategically delayed. Each lender’s informationset includes the actions of the preceding bidders, the lender’s local status, and the lender’sprivate information if the lender receives a signal.

The timing of the game is as follows. An individual lender observes the history,potentially receives a private signal, and then computes via Bayes’s rule his private beliefabout which listing is most likely to have positive returns. The lender then chooses toinvest in that listing. The rest of the game proceeds in this way, with each new lenderusing the observed history and his own signal, if he has one, to make his choice. Due tothe symmetry between lenders, some actions, after a particular history, will reveal thata preceding lender must be local. This revealing behavior occurs when a lender choosesto act in a way that only a local lender with a signal would act given the decisions madeby the preceding lenders. I refer to these lenders as revealed local lenders. All followinglenders will rationally update their beliefs when optimizing to account for this fact. After

16It is possible that a lender’s demand level for a particular listing may be a transmission channel forinformation about private signal. However, the median bid, regardless of local status, is $50, so most bidscontain no additional information.

8

all the lenders have chosen a listing, all the uncertainty is resolved and one listing pays outits return to those that invested in it. Rationality and the structure of the game are commonknowledge to all lenders. Consequently, each lender’s decision rule is based solely onthe history and his signal. When indifferent, lenders use the three classical tie-breakingassumptions in Banerjee (1992).

Because each lender’s payoff function is independent from the subsequent choices,there are no dynamic strategic elements in this game. Implying that the game can besolved by forward induction, the uniqueness of the solution is automatically guaranteed.It is obvious that the decisions made by lenders will depend on whether they receive asignal and on their local status. Formally solving the unique Bayesian Nash equilibriumfor this game is outside the scope of this paper; however, I will briefly explain the optimaldecisions for the first three lenders to motivate why early local bidding might have anextra influence on the behavior of lenders later in the auction.

The first lender has no previous history to observe, so the only source of informationavailable is his own signal, if he received one, regardless of his local status. Clearly, the firstlender will follow his signal if he has one, or he will choose the default zero listing if nosignal is received. As shown by Banerjee (1992), this choice will minimize misinformation,since the first lender will only choose the zero listing if (a) he did not receive a signal,or (b) the signal received is the zero listing. Getting a signal for the zero listing is a zeroprobability event, so case (b) can be safely ignored. Seeing the first lender choose the zerolisting does not provide any information to the subsequent lenders.

The second lender’s information set includes her local status, her signal, and the firstlender’s choice. She will follow her own signal, given that she receives one, if the firstlender chooses the zero listing or she is local. Otherwise, the second lender will follow thechoice of the first lender. If the first lender chooses a non-zero listings, then the secondlender knows that the first lender received a signal. Recall that lenders do not exogenouslyknow the local status of the preceding lenders; the second lender will assume that thesignal accuracy of the first lender is that of an average lender. As a result, a nonlocalsecond lender will ignore her private signal, if she has one, and follow the choice of thefirst lender of unknown status, θ > γ since β > γ.

Starting with the third lender, the results of this social learning model deviate fromthe Banerjee (1992) model. The third lender will follow his signal, if he has one, if: (a) allpreceding lenders choose the zero listing, (b) he is a local lender, or (c) his signal matchesa choice made by a preceding lender. Otherwise, the third lender will follow the choiceof the second lender. It is important to point out here that whenever a lender’s signalmatches the choice made by one of the preceding lenders, he should always follow hissignal. This result follows from the fact that the probability of two lenders receiving the

9

same signal and both of them being wrong, regardless of local status, is zero.The effects of the informational asymmetry in this model can clearly be illustrated with

the third lender when both previous lenders choose the same non-zero listing. Neitherof them reveals their local status by their actions. Therefore, a local third lender wouldstill follow his own private signal, while a nonlocal lender would follow the precedingtwo lenders. This is an example of local lenders being more trusting of their privateinformation than their nonlocal counterparts.

The basic intuition from this example gives rise to more general results. The proofs ofthe following lemmas can be found in the appendix.

Lemma 1: If the first three lenders all choose the same non-zero listing, then all subsequentlenders, regardless of their local status, will follow them and choose that listing.

Lemma 2: If the first lender chooses i 6= 0, and the second, third, and fourth lenders all choosethe same non-zero listing that is different from the one chosen by the first lender, j 6= i, then allsubsequent lenders, regardless of their local status, will choose listing j, unless they receive a signalthat matches the choice of the first lender, in which case, they will choose listing i.

These two lemmas are generalizable to the case where all the chosen non-zero listingshave the same number and types of lenders, and one of them receives additional bidsfrom either three unknown status lenders or one revealed local and one unknown-statuslender. Then all subsequent lenders whose signals do not match an already chosen listing,regardless of local status, will choose the listing with the most lenders. After this eventoccurs, a herd starts and no more information will be revealed from lenders’ choice of thislisting, resulting in the relative information structure remaining unchanged. Thus, themodel implies that the speed of herding will accelerate as more local lenders are involved.

Lemma 3: As the most-chosen non-zero listing has been chosen by more local lenders in theearly stages, that particular listings becomes more likely to be the ex-post listing that has positivereturns.

Given that local lenders are more likely to rationally follow their own private informa-tion, having more of them choose the same listings in the early period, before informationtransmission stops, increases the likelihood that the correct listing is being chosen. Thisresult arises from the fact that the signal of local lenders is more accurate than that of anonlocal lender and is less likely to be persuaded by herding instinct.

Although simple, the optimal decision rules that arise from this model provide clear

10

intuition and motivation for how lenders behave on Prosper. In line with the empiricalfindings of Zhang and Lui (2012), this model predicts rational herding from observationallearning. However, this model clarifies that the strength and speed of herding fromlearning depends on the actions of better-informed agents. Translating these theoreticalresults to Prosper, this model produces the following testable hypotheses: a listing withmore early local bidding will (a) attract more lenders; (b) have a higher probability ofbeing funded; (c) conditional on being funded, have a lower final interest rate; and (d)have a larger probability of not defaulting.

5. Results

Intuitively, the Internet makes it cheaper and easier to connect, so online platformsshould reduce at least some informational frictions. Additionally, several features of thisparticular market make the presence of geography-related frictions less plausible: loans areunsecured, and lenders have little legal recourse other than the standard collection processand reporting the loan default to credit reporting bureaus. These constraints minimizethe lenders’ ability to individually monitor and enforce the contract. Therefore, physicalproximity should be less important in online markets. Moreover, given this is an onlinemarket, participants have to possess at least a minimum level of computer and financialcompetence.

Considering the above, it has yet to be shown if there still exists a meaningful differencein P2P based on geography. However, if there are differences based on the location of thelender relative to the borrower, it might be caused by three different mechanisms. Onechannel is an informational asymmetry story where distance-related frictions still matterdue to search costs. The second channel is an offshoot of the information story, but itcenters on friends and family being the primary driver of the asymmetry. The last channelis a preference story where local lenders are not any better informed, but they simplyprefer local projects.

Following the literature on online markets (Wolf, 2000; Hillberry and Hummels, 2003;Hortaçsu et al., 2009), I define “localness" as the condition in which the lender and theborrower reside in the same state. Admittedly, a smaller unit of measure is preferred;however, Prosper not only does not require individuals to publicly post their city, but thesite actually discourages it, to prevent borrowers from personally identifying themselves.Less than 7% of borrowers publicly list their cities, and most of them reside in big cities.If the actual effect of information asymmetry is limited to a smaller physical proximity,then my definition of localness is counting a significant amount of nonlocal lenders aslocal. Therefore, I am making the two groups of lenders more similar and weakening the

11

potential differences that I measure. Thus this data limitation means that my estimates areactually a lower bound on the true effect of localness. To examine the sensitivity of usingstate as the threshold of localness, I perform multiple robustness checks using differentdefinitions of localness. The results from these robustness checks are qualitatively thesame, supporting the claim that the state analysis is a lower bound on the true effect.

I start this analysis by first documenting the fact that local lenders do indeed behavedifferently than nonlocal lenders. I find strong evidence that non-“friends and family"information frictions are the major contributor to the behavioral differences observed inlenders. Next, I examine the prediction of the model about the influence of informationasymmetry on social learning in the P2P loan market. All regressions include Borrowerand Lender State (where applicable), Credit Grade, Category of use, Quarter, Month, andDay of the Week fixed effects.17 The variable Loan Amount is measured in thousands ofdollars, Borrower Max Rate and Current Rate are measured in percentage points (1=1%),and Total Competition and Credit Grade Competition are measured in a single listing.18

5.1. Individual Lender Analysis

The left half of Table 3 displays the summary statistics of bid amount by local status.The mean and 75th percentile local bid amounts are around $10 dollar larger than thenonlocal bids; these differences are significant at the 1% level. I ran a Tobit regression oflog bid amount with left censoring at log($50) while including an indicator for local lender.Standard errors are clustered at the lender level to control for any unobserved correlationbetween the bidding activity of the same lender. As the left half of Table 4 shows, I findthat local lenders bid roughly 7% larger amounts than nonlocal lenders. This effect isof a similar magnitude to the auction’s current standing interest rate being roughly 5.3percentage points higher. It is worth noting that Pseudo R2 is a rather low value (0.159).This is not surprising given that I only have information about lenders’ behavior on the site,and not on the lenders themselves. It is well documented that demographic characteristicssuch as age, educational attainment, marital status, gender, race, and risk tolerance allaffect asset holdings. More information about the heterogeneity of lenders would likelyimprove the fit of this regression.

Knowing that local lenders have a larger demand for local loans indicates nothingabout the channel that is driving this behavior. This result might be consistent withthe findings of local preference in other online markets (Hortaçsu et al., 2009; Lin andViswanathan, 2014); however, local lenders might also be better informed and thus only

17I also explicitly controlled for observable borrower state characteristics in place of the borrower state FEand found no noticeable difference in the results.

18All results from this section are available from the author upon request.

12

bid on local listings that turn out to be good projects with higher returns. In this situation,I would also expect local lenders to have larger demand for better projects. To examinethis, I restrict my focus to listings that become loans by dividing bids into sub-samplesbased on the loans’ ex-post outcome (default or paid back). Regardless of the ex-postoutcome, lenders bid larger amounts on local projects; however, the magnitude of thedifference is significantly larger for loans that pay back in full. As can be seen in the righthalf of Table 4, this behavior suggests the existence of a small preference for local loans.However, given my coarse default data, local lenders might still be choosing loans thatdefault farther into the process and are paying back more of the principle prior to default. Ihave a limited subsample of 100 defaulted loans where I know the date of default; for loanswith a large proportion of early local bidding, the default occurs nine months later, onaverage. While certainly not definite, this result suggests that an appropriate interpretationof lender demand behavior is that local lenders are better informed and thus demandlarger amounts of the expected profitable loans.

If local lenders are actually more informed about the underlying riskiness of theborrower and the general market conditions, they should be better able to evaluate therisk of the listing. To examine this, I restrict my focus to only losing bids where I knowthe actual bid interest rate. The mean and median interest rates bid by local lenders, ascompared to those of nonlocal lenders, are larger for loans that ex-post default, but aresmaller for loans that ex-post pay back in full. These two observations strongly suggestthat local lenders seem to price more accurately the underlying risk of the listings thannonlocal lenders.19

I perform a two-sample Mann-Whitney U-test and two-sample Kolmogorov-Smirnovtest to examine the equality of the local and nonlocal bid interest rate distributions.The M-W and K-S tests both reject the null that the local and nonlocal distributionsare equal at the 1% significance level. Additionally, for loans that default, the localdistribution stochastically dominates the nonlocal distribution, and for loans that pay back,the nonlocal distribution stochastically dominates the local distribution. To further explorethe difference between the local and nonlocal bid interest rate distributions, I perform at-test on the distributions’ mean. Table 5 presents the t-statistics and p-values for testingthe differences between the means, (µNonlocal − µLocal), by the ex-post loan outcome. Forloans that default, the differences are negative, and the difference for credit grade C issignificant at the 10% level, while the other differences are significant at the 5% level. Forloans that pay back, the differences are all positive. These differences are significant atthe 5% or smaller level for all credit grades except A and B. Lenders appear to be better

19Similar results are observed if I include all bids by assuming the interest rate of winning bids is equal tothe final interest rate.

13

evaluators of local projects.Additionally, I run a Type II Tobit regression with left censoring at each listing’s

winning interest rate, since that is the smallest rate that I observe for each listing. Thestandard errors are clustered at the lender level. The regression contains indicators forlocal lender interacted with credit grade, an indicator for if the loan defaulted, and thelocal lender indicator interacted with credit grade and the loan’s default status. The resultsdisplayed in Table 6 are consistent with the previous results, showing that local lendersact differentially based on ex-post loan outcomes. The coefficients on the local indicatorinteracted with credit grade are all negative and significant at the 5% level. The magnitudeof the difference between local and nonlocal bids for loans that do not default variessignificantly across credit grades; B listings have the lowest differential, at 0.032 percentagepoints, while E listings have the largest, at 0.907 percentage points. The reduction thatlocal lenders give to loans that pay back ex-post in their ax-ante bids, generally increasesas the credit grade worsens. The correlation between credit grade and the rate reduction is-0.702. This implies that local lenders are more willing to accept a lower rate from ex-antepotentially riskier borrowers who ex-post turn out to be lower risk than their financialinformation indicates. This result makes sense in the context that the better informationpossessed by the local lenders should have the biggest effect on the listing in the worstcredit grades, since these listings have the largest potential difference between true andperceived risk.

The coefficients for the local indicator interacted with credit grades and default are allpositive and significant. The net effect of being local on the bid interest rate is positiveif the loan ex-post defaults. The premium that local lenders, relative to nonlocal lenders,give to loans that ex-post default ranges from 0.056 to 0.28 percentage points. Reducingthe final interest rate by 1% would lower the total loan payment by a few hundred dollars;this effect would be quite significant when aggregated across a lender’s portfolio. Thisdifferential lender behavior supports the idea that lenders are better able to evaluate theunderlying risk of local projects, with the obvious rationale being that local lenders arerelatively more informed than nonlocal lenders. Theory predicts that this informationalasymmetry arises from the fact that the cost of becoming more informed about the marketconditions and quality of local borrowers is significantly cheaper for local lenders due totheir proximity.

If local lenders are better informed about the quality of listings posted by local borro-wers, then local lenders should be more willing to bid earlier in the auction when the onlyinformation that has been revealed is the original public information and their privatesignal. The right part of Table 3 shows the 25th percentile, mean, and median of bidtimes. Bid times are normalized so that zero is the beginning and one is the end of the

14

auction. The data shows that the average local bid is placed a roughly 2.5 hours and the25th percentile local bid is placed about 5.5 hours earlier than their nonlocal counterparts.A two-sample M-W test on bid times, results in a p-value of less than 0.001, implying thatthe distribution of bid times for local lenders tends to be significantly smaller than thedistribution of bid times for nonlocal lenders.20 To more formally evaluate the timing ofbids, I estimate an OLS and Cox proportional hazard model to investigate the lenders’propensity to bid. Table 8 displays the coefficients from these regressions. In the firstanalysis, the bid are divided into pre- and post-fully funded bids. Within those periods,the bid times are renormalized so that zero refers to the start of each period. Local lendersare estimated to bid about 3 hours earlier than nonlocal lenders in the pre-fully fundedperiod, and 1.5 hours earlier in the post-fully funded period. The hazard model estimatesthat at any time t, being local increases the relative probability of bidding for a lender by3.49%. The results concur with the predictions of the learning model; local lenders bidearlier, with the effect dampening in the post-fully funded phase. This occurs becausemore information about the listing is conveyed from the observed bidding, which closesthe informational gap between local and nonlocal lenders.

One might be concerned that it is not just timing that matters in determining if a bidis actually early or not; therefore, Table 7 displays the summary statistics for the numberof bids placed and money pledged in a listing immediately before a lender bids, by localstatus. As can be clearly seen, local lenders bid earlier in an auction not just chronologically.Local lenders have demonstrated that they are more comfortable bidding in periods wherelimited information is being revealed by other lenders. A two-sample K-S distributionalequality test was performed for both measures of previous bidding activity. The test resultsshow find that the two distributions are statistically different from each other. In addition,the results of a M-W test show that that the nonlocal distribution stochastically dominatesthe local distribution.21 Similar analysis was performed on the number of bids placed andthe amount of funds pledged, as the measure of auction duration. The findings are in linewith the previously shown results that local lenders bid earlier.

Thus far, I have documented that lenders appear to be better informed about localprojects. However, given that I do not observe information about the market participants’offline activity or social networks, it is necessary to check whether the effect of being localis driven by a borrower’s friends and family simply using the site as a way to formalizetheir lending. This is especially true given that offline connections are a big component forthe observed geographic difference in behavior on other crowdfunding platforms.

20The results of a two-sample K-S distributional equality test are consistent with the findings of theMann-Whitney test presented here.

21The p-values for all four of the distributional tests are basically zero.

15

While friends and family would be better informed about the borrower’s characteristics,it is unlikely that they would be motivated by profit. If friends and family are a majorpart of the local bidding, it is reasonable that they will join the site around the time thatthe borrower creates his or her listing. It is unlikely to imagine that a majority of theborrowers creating listings will have friends and family who are previously active onProsper. The average lender has been active for more than 210 days and has placed about50 bids before submitting the first local bid. Less than 2% of lenders place their first bidon a local project and less than 8% of local bids are placed during the first three days afterthe lender joins the site. Furthermore, less than 7% of lenders place more than 25% of thebids in local listings; therefore, a vast majority of lenders are significantly more active innonlocal listings.

Furthermore, Agrawal, Catalini, and Goldfarb (2015) and Lin and Viswanathan (2014)propose empirical definitions of friends and family for crowdfunding platforms basedsolely on observed online behavior. The definition of Agrawal et al. (2015) translated inthe Prosper setting requires that: (a) the first bid be invested by a friend or family member,(b) the largest bid be made by a friend or family member, and (c) the friends and familylender bids in no more than 25 other listings.22 Only 3,534 bids (0.17% of all bids) qualifyas a friend and family bid under this definition, even when allowing for ties in condition(b). The definition of Lin and Viswanathan (2014) states that friends and family are highlyunlikely to invest in other borrowers, as they have signed up to primarily lend to theirfriend or family member. Therefore, a friend and family bid is defined as a bid by a lenderwho only bids on one listing regardless of local status. A total of 4,465 bids (0.22% of allbids) qualify as friend and family bids under this definition. The following results arerobust to relaxing these definitions.

As seen in the top half of Table 9, regardless of the definition used, friends and familybid more on average than a standard local lender. However, local non-friends and familylenders still bid significantly more than nonlocal lenders. Interestingly, as displayed inTable 10, friends and family lenders uniformly bid lower interest rates (0.5 to 2 percentagepoints less, depending on credit grade and definition) regardless of the ex-post loanoutcome than non-friends and family. There is no ex-ante risk adjustment, while localnon-friends and family still adjust for ex-post outcome. This suggests that friends andfamily lenders are motivated primarily by altruism. Even after excluding all the friendsand family bids, the local non-friends and family lenders behave qualitatively similar tobefore, but now have a larger increase for loans that ex-post default.

Prosper allows members to network on site via Prosper groups. Although the Prosper

22Their actual definition restricts friends and family to investing in no more than three other projects;however, their setting is a smaller platform with a lower transaction volume.

16

groups were never widely used, it is also worth evaluating whether these on-site connecti-ons might be driving the local results, similar to the friends and family explanation. Table9 (bottom half) and 11 present the bid amount and interest rate comparison, and showsimilar results to the friends and family analysis. Within-group lenders tend to bid largeramounts and uniformly smaller interest rates, regardless of the ex-post loan outcome.However, when excluding the within-group bids, the local not within-group lenders stillbid larger amounts and adjust their interest rate to better account for ex-post outcomes.Moreover, I reevaluate the individual level Tobit regressions while dropping the differentlydefined friends and family bids, bids from lenders with 75% or more of bids being local,23

and the strictly within-group bids. I also rerun the regressions while adding an indicatorfor friends and family bid. These new results show no noticeable effect on the influence oflocal bids on the previously reported results. However, similar to Zhang and Lui (2012), Ifind mild evidence that friends and family bidding on a listing weakens the herding effect.This can be ascribed to lenders attributing more altruistic motives to friends and familybidding than informational revelation. From all this, I can conclude that although socialconnections matter, they are not the primary driver of the observed lender behavior. Someother mechanism is causing lenders to be better informed about local projects.

Given these results, the other main explanation, that lenders have a strict preferencefor local projects, can also be discredited. Local lenders bid different interest rates thannonlocal lenders depending on the ex-post outcome of the loan. This empirical evidencestrongly suggests that, while altruism might play a small part, the dominant channelexplaining this observed behavior is that local lenders are better informed than otherlenders, which fits well with theory. Additionally, there is no statistically significantdifference in behavior (bid amount, interest rate, timing of bid) for nonlocal listings thatare equally far away from the lender, even though the borrowers who posted these listingsreside in different cultural regions of the country. For example, Ohio lenders demonstratesimilar lending patterns when investing in listings from Nebraska as from Mississippi orNew Hampshire; if the preference story was true, one would expect more local behavior forstates with similar cultures.24 The geography-based asymmetry creates a situation wheresocial learning can occur as local lenders, through their bidding behavior, reveal theirprivate information to nonlocal lenders. Thus as a listing accumulates more bids, a generalsense of quality is signaled that may lead to rational herding behavior in crowdfundingmarkets (Agrawal et al., 2011; Zhang and Lui, 2012; Burtch, Ghose, and Wattal, 2013).

23This is a relaxed version of the Lin and Viswanathan (2014) definition.24Results are available from the author by request.

17

5.2. Listing Level Analysis

The ability of borrowers to signal to lenders on crowdfunding platforms is well studied(Ahlers, Cumming, Günther, and Schweizer, 2015; Nowak, Ross, and Yencha, 2015), inthis paper I examine the less studied ability of lenders to signal to other lenders. Asshown previously, local lenders tend to bid earlier and appear to be more informed aboutgeneral market conditions and underlying riskiness. This behavior allows local lenders topartially reveal their private information. However, the model’s predictions do not requirethat lenders ex-ante know the local status of other lenders, merely that such informationis obtainable and can be revealed.25 These predictions suggest that in this market, theamount of early local bidding is influential and should be a strong predictor of a listing’ssuccess in (a) attracting bidding activity, (b) becoming fully funded, (c) getting a lowerfinal interest rate if funded, and (d) after becoming loan, not defaulting. In this section, Iset out to empirically evaluate these predictions for the P2P lending market.

I include variables measuring the total number of early bids (Total Early Count), theamount of money pledged by all lenders (Total Early Amount), and the interaction of thetwo (Total Early Count x Amount). Additionally, I include the bidding by local lenders,(Local Early Count, Local Early Amount, Local Early Count x Amount). Having the total earlybids and money pledged in the specification controls for the effect that more early biddinghas in general, regardless of local status, on influencing the decisions of other lenders (Kimand Hann, 2014). Using a specification with both variables for total early and local earlybidding allows me to separately identify the additional effect of early local bidding froman generic early bid. I examine four definitions of early bidding, pre-fully funded, firsttwo hours, first hour, and first 30 minutes of the auction. My results are robust to changesof the definition of the early period. For the sake of brevity, only results from the first twohours are reported in this paper.

To test whether more early local bidding attracts more activity to an auction, I runcount regressions using a Poisson specification. Table 12 presents the marginal effectsof early local bidding displays on total and total-nonlocal bid count. The coefficients fortotal early and local early bidding are all positive and significant. This result confirmsthe model’s prediction that more early local bidding attracts, not only more total biddingbut also nonlocal bidding activity to a listing. The interaction terms are negative andstatistically significant, but small enough to be economically insignificant. A single localearly bid has a larger effect on bidding than the borrower being a homeowner. To put themagnitude in context, an extra four $50 local early bids will yield an additional three bids.While that might seem small, the average listing only receives 47 bids and 90% of listings

25If all lenders knew the local status of all other lenders, the local effect would be significantly stronger.

18

receive 150 or fewer bids. Even when focusing just on total-nonlocal bids, one early localbid has roughly the same increase as two nonlocal early bids, given that all the bids arethe same amount, in attracting more lenders to the listing.

One might think that early local bidding only has power in attracting lender activitybefore the listing is fully funded. In the pre-fully funded period of the auction, bids arecomplementary. Since there are no losing bids, no information about the actual interestrates of the bids in the auction is revealed. An additional bid in the pre-fully fundedphase moves the listing closer to completion and demonstrates that the lender’s privateinformation is strong enough to induce the lender to demand positive shares of the listingat an interest rate that is as least as large as the borrower’s maximum rate. After thelisting is fully funded, bids become substitutes and real competition starts between lenders.In this period of the auction, future bidding could cause older bids to become losing;therefore, the interest rate of those bids will start to be shown. If a new bid does not lowerthe standing interest rate, it will reduce the amount of the loan that a previous lender willbe funding.

Given the difference in the nature of the auction and information available across thesedifferent phases of the auction, lenders might no longer consider the previous biddinghistory and only focus on the current standing interest rate. To examine if the strength ofthe attraction of local early bidding varies across this informational shift in the auctionmechanism, I restrict attention to bidding in just the pre-fully funded and post-fullyfunded phases of the auction. Note that only listings that are fully funded will have apost-fully funded phase. Thus, I use a subsample of listings that are completed. Table 13presents the marginal effects of early local bidding on total pre- and post-fully funded bidcount. The coefficients for total early and local early bidding are all positive and significant.These results are similar to the results of the total bid count regression shown previously.The effect of local early bidding is dampened in the post-fully funded period; an extrafour $50 local early bids will yield an additional 5.5 bids pre-fully funded, while the sameamount only results in an additional 2.5 bids post-fully funded. If all listings, not justcompleted listings, are included in the regression, qualitatively similar results are found;however, the magnitude of the early bidding coefficients is reduced. This is unsurprisinggiven that if all listings are included, the average amount of bidding is decreased by theaddition of uncompleted listings.

Local early bidding attracts more lender activity, but it also affects the rate of fundingfor listings. The left part of Table 14 presents the marginal effects, evaluated at the means,of local early bidding on the probability of funding from a logit regression. The resultsconcur with those on individual lender activity. While the coefficient on local early bidamount is negative, the net effect is positive and significant. This result implies that for the

19

same amount of local money pledged, having more local lenders bid smaller individualamounts will have a larger impact than having fewer local lenders bidding larger amounts.An additional four local early bids of $50 will increase a listing’s funding probability by0.028 percentage points above the effect of getting four nonlocal early bids. This effectmay initially seem small, but comparing it to the effect of decreasing the DTI by oneunit (-0.0030), increasing the borrower’s maximum rate by a percentage point (0.0069),or requesting an additionally thousand dollars (-0.0271), the relative effect of early localbidding is significant.

A related issue to the probability of funding is the effect that local early bidding has ona listing’s final interest. The effect of early bidding is not immediately obvious, since it hasbeen shown that local lenders appear to price listings more appropriately than nonlocallenders. If the listing possesses some extra risk that is hidden to uninformed lenders,then the local lenders will bid larger interest rates. In this scenario, the positive effect ofattracting more bidding activity might be negated if there are enough local bids in theauction.

The right part of Table 14 shows that if the effect of early local bidding is negativeand significant, the interaction term is positive but insignificant. A single additional $50local early bid will, on average, decrease the expected final interest rate of a completedlisting by about half a percentage point in excess of the effect of a listing just receivinga single nonlocal early bid. For comparison, this effect is similar in absolute magnitudeto the effect of requesting an additional $3,000. The effect of local early bidding on theprobability of funding and the final interest rate is consistent with the predictions of thesocial learning model. Listings that have more early bidding, especially local activity, willattract more bidding, resulting in these listings being funded at higher rates and enjoyinglower final interests, if funded.

Lastly, the social learning model presented in Section 4 suggests that local early bidspredict the ex-post outcome of a loan. Given that local lenders are better informed thanother lenders, listings with more local early bidding should have a higher probability ofnot defaulting. Table 15 displays the marginal effects, evaluated at the means, for a logitregression predicting whether a completed listing will default using pre-fully funded localbidding as well as local bidding in the first two hours. The effect of early local biddingis negative; however, the coefficients are, in general, statistically insignificant. Moredetailed information, about when the default occurred and how much of the principle wasrepaid, would allow for a finer measure of loan performance. Admittedly, this result issuggestive at best, but the direction of the effect is consistent with social learning. Thebetter information of the local lenders accumulates as more of them bid early in the auctionand this information is transmitted to other lenders, affecting how they bid on the available

20

listings.To further evaluate the friends and family, and the within-group bidding, additional

specifications were performed with the social connection bids removed and with newvariables that explicitly controlled for their presence. However, since the number ofoffending bids is so minor relative to the total of regular local bids, the resulting outputshows no meaningful differences from the reported results. Combining all of these findings,it can be inferred that geography, acting mainly through the channel of informationalfrictions, has a strong effect on P2P lending markets. This research provides more supportfor the claim that although the Internet has made gathering and processing informationcheaper for more people, location of the investor relative to the investment opportunity isstill important due to nonzero search costs.

5.3. Sensitivity Analysis of Localness

Although empirical work on other e-commerce sites also uses state level, it is reasona-ble to question the sensitivity of the "state" definition of localness to actually generateinformational frictions, especially given the vast size difference between states and thepotential spillover from neighboring areas in different states. To examine the robustness ofmy analysis, I perform multiple specification checks: defining local to be within Censusregion, as well as using a restricted subsample of individuals who list their city to definelocal as within city. I also investigate the effect of state size and actual distance betweenstates, and identify lenders from neighboring and near, but not adjacent states from therest of the nonlocal lenders.26

The results from the individual-lender and listing-level analysis with the redefined localdefinition are qualitatively the same as the reported specification, with the estimated localeffect being larger when using city as the local definition. The informational differencebetween local and nonlocal behavior is more pronounced as the definition of local becomesnarrower and focused. Additionally, to see whether the observed localness is being drivendisproportionately by a single state (i.e., "California Effect"), I sequentially drop individualstates from my regressions one at a time. To check the other direction of this effect, Irestrict the sample to just listings from a single state and perform my analysis for eachstate individually. The results are qualitatively the same for both groups of specifications.For state size, I add a small-state indicator (population less than 3 million) and interactit with the local lender. Congruent with the asymmetric information story, the effect ofbeing local is stronger for smaller states. Along the same lines, lenders in neighboringstates, behave similarly to local lenders; however, the adjustments are weaker. Furthermore,

26The results from this section are available from the author upon request.

21

lenders from near but not adjacent states also demonstrate the same pattern but withfurther dampened responses. I include indicators for all three types of physical proximityin the regressions. The estimated effects monotonically decrease as the lender is furtheraway from the borrower. Similar results are also found when using physical distancebetween lender and borrower.27 Lenders further than 700 miles away are estimated tobe the true nonlocal lenders since their behavior is the most different from that of locallenders and does not change as the distance increases past that threshold. The differencefrom local behavior decreases relatively smoothly as the lender is located closer to theborrower within that 700 mile threshold.

Another potential alternative narrative to consider is a story about different levels ofexperience between lenders being the true main driver for the observing signaling. Toexamine this, I include an indicator for a lender having three and six months’ worth ofexperience on the site in the analysis. The coefficients on the variables of interest arenot meaningfully different. Additionally, the distribution of experiences for local andnonlocal lenders is basically the same, even when grouping the lenders based on when inthe auction they bid and what types of listings they bid on.

Moreover, these results provide further evidence supporting the previous findings thatfriends and family are not the primary cause of the observed geographic difference inlender behavior. In particular, the small-state and physical distance analysis are consistentwith the conclusion that information about the borrower and the project is more costly toobtain the further away the lender is, and as such lenders rationally get more informedabout local projects.

6. Discussion of Results

The previously shown results illustrate that geographic frictions still exist the in onlinelending market, and that the primary cause is an informational asymmetry. The existenceof asymmetric information is not new. Zhang and Lui (2012) use a loan-level analysisto find evidences of rational herding in P2P loan markets. However, the cause of thisasymmetry is an ongoing debate. The Zhang and Lui (2012) model, like most studiesin this area, has some lenders receiving private signals about borrower quality throughan unspecified information acquisition process. My work builds on this foundation toprovide a clear mechanism to explain why certain lenders possess better information aboutspecific borrowers. Additionally, I have empirical evidence to collaborate the existenceof my mechanism to generate informational asymmetry. The better information arises

27Distance between lender and borrower is measured using straight-line distance between the population-weighted center of the lender’s and borrower’s state.

22

from the nonzero search costs and the heterogeneity in lenders’ closeness to the differentprojects; thus rational lenders get better informed about local projects. This informationgap between local and nonlocal lenders contributes to the observed rational herding foundin these crowdfunding markets.

While the Internet certainly has made all individuals more connected, localness stillembodies some incidentally obtained initial informational advantage, which causes anatural disparity in lenders’ exogenous, starting information set. This advantage arisesfrom the limited transferability of some local-specific information; this can be thoughtalternatively as information that is especially costly to obtain for nonlocal individuals. Thisbarrier arises because certain information about a local environment is difficult to collect,centralize, digitize, and share. Examples include (1) learning local culture and appetitefor products and services, adoption rates of new styles and trends, and risk tolerances;28

(2) changes to housing and labor markets that are only known in the aggregate withsome delay and are only mentioned by the local or regional news services;29 and (3)compatibility issues across different platforms and site-specific skills needed to operate thedifferent databases used by local, county, state governments, and non-profit organizationsthat might contain valuable information about the potential borrower and his or her area.30

These factors are informative of the performance of a borrower, but are not easily oraccurately converted into a numerical value and, in some cases, difficult to compare acrossborrowers from different locales. The clear effect of this limited transferability can be seenin the state size analysis. As the state gets smaller, the lender’s knowledge about his ownlocal market is more relevant and relatable to borrowers in the rest of the state. A lenderin a small state is even better informed about local conditions than local lenders in a largerstate.

Another way to empirically test the information story is to create subsamples of loantypes where the value of additional information available to local lenders is relatively high(Debt Consolidation, Home Improvement, and Business) and low (Auto and Other). Whilelocal lenders in the both subsamples behave qualitatively the same as the benchmark case,

28For instance, Central Ohio has a growing fashion industry with an expanding presence, and a strongdemand for food trucks (Columbus hosts the Midwestern’s largest, and the country’s second largest foodtruck festival). This is information that is not readily found strictly looking for information about the regiononline.

29Mass layoffs at large companies are reported widely in the news, but for most companies, relocation orpersonnel realignments are unreported nationally. Lenders have a better ability to track a local borrower’soccupation or proposed business ventures in the context of the general local market conditions for thatoccupation.

30Municipalities within the same the state generally use similar technology and terminology making findinginformation easily for lenders with experience in the environment(i.e., county auditor, secretory of state, andhealth department websites).

23

the information adjustment is stronger in the high-information group. Both the increase inbid amount for loans that paid back and the absolute differential in bid interest rates byex-post outcome between local and nonlocal lenders are larger for the high-informationgroup than for the low-information group.31 As suggested by theory, in markets were theinitial information gap is larger or the information is more costly to obtain, the asymmetrybetween local and nonlocal lenders will be more pronounced.

Additionally, it is important to emphasize that borrowers requesting funds for diffe-rent types of projects are vastly different across states, so there is heterogeneity in theinformation available to different lenders. Figure 1 illustrates this point by presenting thedefault rates for different loan categories.32 For example, knowledge about a business loanin California would not be that insightful about an auto loan in Ohio. There exists a largeamount of heterogeneity in the likelihood of default across different states; this illustratesthe actual extent of the limited transferability of local information to foreign markets.

In an environment where all information is obtainable with some nonzero cost, ini-tial information asymmetries, no matter how slight, will lead to persistent informationasymmetries in equilibrium. Van Nieuwerburgh and Veldkamp (2009) theoretically provethat the optimal strategy for investors with some informational advantage is to specializetheir extensive information gathering in their area of comparative advantage. This resultarises because if local investors attempted to fully mitigate their information asymmetryin nonlocal markets, they would forfeit the excess returns that could possibly be gainedlocally. This occurs because when local and nonlocal investors both become fully informedabout the same listings, they will both have large demand for the high-return assets,regardless of localness, bidding down the interest rate and lowering the expected return. Ifinstead, investors learn more about local listings, they have knowledge about high-returnlocal listings, that others do not have, thus the price of that asset will not fully reflectthis information. Thus learning more about local projects increases the investor’s averagereturn. Therefore, investors should seek to make their information set as different fromthat of the average investor as possible. Ivkovic and Weisbenner (2005) show that the morediversification that exists in lenders’ information, the larger the expected return from localinvestments will be. Rational lenders seeking to maximize expected profits will optimallyget better informed about markets where they already have an initial advantage and lesscostly access to information.

31The results from this analysis are available for the author upon request.32A similar result is seen when looking at the default rate for borrower credit grades, range of borrower’s

maximum rate, and ranges of loan amount by state.

24

7. Conclusion

Over the past few years, major commercial organizations have noticed the vast potentialof crowdfunding and have begun to acquire equity stakes in P2P lending firms, whilemainstream financial companies are beginning to offer similar services, mimicking thisonline marketplace (Corkery, 2015). As P2P lending sites are starting to open and expandworldwide, it is clear that P2P lending, and crowdfunding in general, as an alternative aswell as a complement to the traditional finance methods, is a growing trend. Therefore,it is paramount that we better understand the factors that influence the behavior of thisonline marketplace’s participants. Using transaction data from Prosper.com, I documentevidence that the adoption and use of P2P lending has not fully overcome the traditionalinformation asymmetry in investing. I find support for the conclusion that informationalfrictions (outside of friends and family) exist and have strong effects on the behavior inP2P markets. In particular, I observe that lenders bid earlier, bid larger amounts, andappear to be better informed about local projects. This informational asymmetry leads tothe rationally herding found by Zhang and Lui (2012). As shown by my model, this is trueeven if lenders do not ex-ante know other lenders’ local status, the model only requiresmerely that this information is obtainable.

While the empirical literature is split on whether economic or behavioral motives matter,my results are consistent with the informational cost theory of Grinblatt and Keloharju(2001). It appears that local lenders have easier and cheaper access to information and,thus, rationally get better informed about local listings than their nonlocal counterparts.If lenders were only acting altruistically towards local borrowers, then one would expectthat lenders should generally be bidding lower interest rates on local listings regardless ofthe ex-post loan outcome. However, the data shows clearly that local lenders are biddingdifferentially based on the inferred quality of the borrower. Local lenders appear to beaccounting for riskiness that nonlocal lenders are not, implying that local lenders can betterevaluate the probability of default of local projects. That being said, it is evident from thefact that local lenders have a larger demand for local listings, regardless of the ex-post loanoutcome, that lenders have a limited preference for local listings. However, this seemsto be mostly due to friends and family activity. While this coincides with previous workon online markets (Hortaçsu et al., 2009; Lin and Viswanathan, 2014), strict preferenceis not sufficient to fully explain the observed behavior. Thus, it is safe to conclude thatinformation-based frictions are the primary factor explaining the behavioral differencebetween local and nonlocal lenders in P2P lending markets.

Although the cost of information gathering and processing is lower now than beforethe widespread adoption of the Internet, my results suggest that crowdfunding sites

25

suffer from the similar informational asymmetry that exists in the offline credit markets.This strongly suggests that the current online platforms are unlikely to create a moreunified, national credit market. The structure of P2P lending markets is designed toprotect personal privacy and prevent the mob-style baseball bat enforcement; however,the design actually inhibits the full power of online connections to improve informationalefficiencies. What appears to matter, is not simply more sharing generically, but the sharingof more insightful and critical information about potential borrowers. Since loans obtainedthrough Prospers are unsecured and borrowers cannot fully expose all their meaningfulinformation, these loans can alternatively be thought of as an example of the classical“character loan" given out by smaller community banks based on relationship lendingwhere softer factors are more relevant. Allowing borrowers to link their P2P accountsto their social media presence or to more finely specify their place of residence wouldprovide lenders a fuller and richer profile of the potential borrowers.

After the limited personal information that borrowers are allowed to disclose viaProsper, lenders are still left with uncertainty about the hidden characteristics of theborrowers, hence the disparity in behavior between local and nonlocal lenders. Platformslike eBay and Amazon have robust seller reputation and buyer feedback mechanismsto facilitate information sharing to overcome this obstacle. However, on Prosper, theprevious activity between borrowers and lenders is limited and in most cases non-existent,(Greiner and Wang, 2010). Most individuals simply do not request loans very frequently.Furthermore, the lack of a borrower’s history on the site amplifies the importance of thelenders’ ability to extract valuable information from a borrower’s listing.33 This is furtherexacerbated by Prosper’s prohibition on linking a borrower’s listings to his or her socialmedia accounts.

This finding that the reduction in informational cost has yet to be of a sufficientmagnitude to render geography irrelevant in P2P lending markets is not an anomaly inthe empirical literature. Similar results have been found across online markets; manystudies find that the Internet does not universally lower market prices and lead to lessprice dispersion.34 Additionally, Ghose, Goldfarb, and Han (2013) even find that thetype of device used to access information online affects search costs. However, my workcontributes to the growing crowdfunding literature by empirically showing that localpreference and friends and family cannot explain the observed P2P lender behavior; whilea model of costly search with initial information asymmetries gives rise to the observedsituation where lenders are better informed about local borrowers and nonlocal lenders

33Prystav (2016) shows that this is especially important as the borrower’s credit score worsens.34See Brown and Goolsbee (2002); Janssen, Moraga-González, and Wildenbeest (2005); Blevins and Senney

(2014).

26

rationally herd after partial information is revealed by lenders’ behavior.Further research is needed, but placing my results in context of the wider literature

implies that the presence of informational asymmetry and nonzero search costs suggeststhat the current tools and platforms are unlikely to fully remove barriers and vastlyimprove the efficiency of the online credit market. A better understanding of investors’behavior and decision-making on online crowdfunding sites is important, not just formarket participants, but also for the financial market in its entirety. As more researchpushes the frontier on this issue, it can be hoped that better designed mechanisms andpolicies will be able to remove informational barriers, and that online lending can beleveraged to improve the efficiency of the credit market.

References

Adams, C. and I. Ramos (2010). The past, present and future of social networking andoutsourcing: Impact on theory and practice. In AMCIS 2010 Proceedings. UK Academyfor Information Systems Conference Proceedings 2010. [4]

Agrawal, A., C. Catalini, and A. Goldfarb (2011). The geography of crowdfunding. WorkingPaper. [3, 17]

Agrawal, A., C. Catalini, and A. Goldfarb (2015). Crowdfunding: Geography, socialnetworks, and the timing of investment decisions. Journal of Economy and ManagementStrategy 24(2), 253–277. [16, 37, 38]

Ahlers, G. K., D. Cumming, C. Günther, and D. Schweizer (2015). Signaling in EquityCrowdfunding. Entrepreneurship Theory and Practice 39(4), 955–980. [18]

Arthur, B. W. (1988). Urban systems and historical path dependence. In J. Ausubeland R. Herman (Eds.), Cities and their Vital Systems, pp. 85–97. Washington: NationalAcademy Press. [3]

Arthur, B. W. (1990). Silicon Valley locational clusters: When do increasing returns implymonopoly? Mathematical Social Sciences 19, 235–251. [3]

Bakos, J. (1997). Reducing buyer search costs: Implications for electronic marketplaces.Management Science 43(12), 1676–1692. [1]

Banerjee, A. (1992). A simple model of herd behavior. The Quarterly Journal of Econo-mics 107(3), 797–817. [7, 9]

Baye, M., R. Gatii, P. Kattuman, and M. John (2009). Clicks, discontinuities, and firmdemand online. Journal of Economics & Management Strategy 18(4), 935–975. [1]

27

Belleflamme, P., T. Lambert, and A. Schwienbacher (2010, June). Crowdfunding: Anindustrial organization perspective. Working Paper. [2]

Bikhchandani, S., D. Hirshleifer, and I. Welch (1992). A theory of fads, fashion, custom, andcultural change as informational cascades. Journal of Political Economy 100(5), 992–1026.[7]

Blevins, J. R. and G. T. Senney (2014). Dynamic selection and distributional bounds onsearch costs in dynamic unit-demand models. Working Paper 14-02, The Ohio StateUniversity. [26]

Blum, B. S. and A. Goldfarb (2006). Does the internet defy the law of gravity? Journal ofInternational Economics 70(2), 384–405. [3]

Brown, J. R. and A. Goolsbee (2002). Does the internet make markets more competitive?evidence from the life insurance industry. Journal of Political Economy 110(3), 481–507.[26]

Brynjolfsson, E., Y. Hu, and M. Rahman (2009). Battle of retail channels: How productselection and geography drive cross-channel competition. Management Science 55(11),1755–1765. [3]

Burtch, G., A. Ghose, and S. Wattal (2013). An empirical examination of the antecedents andconsequences of contribution patterns in crowd-funded markets. Journal of InformationSystems Research 24(3), 499–519. [17]

Choi, J. and D. Bell (2011). Preference minorities and the internet. Journal of MarketingResearch 48(4), 670–682. [1]

Corkery, M. (2015, July 16). Treasury to examine online lending industry. New York Times,B2. [25]

Cowley, S. (2015, April 8). Online lenders offer a faster lifeline for small businesses. NewYork Times, B4. [2]

David, P. and J. Rosenbloom (1990). Marshallian factor market externalities and thedynamics of industrial localization. Journal of Urban Economics 28, 349–370. [3]

Devenow, A. and I. Welch (1996). Rational herding in financial economics. EuropeanEconomic Review 40(3–5), 603–615. [7]

Duarte, J., S. Siegel, and L. Young (2012). Trust and credit: The role of appearance inPeer-to-Peer lending. Review of Financial Studies 25(8), 2455–2483. [2]

28

Florida, R. and D. Smith (1988). Venture capital formation, investment, and regionalindustrialization. Annuals of the Assiociation of American Geographers 83(3), 434–435. [3]

Forman, C., A. Ghose, and A. Goldfarb (2009). Competition between local and electronicmarkets: How the benefit of buying online depends on where you live. ManagementScience 55(1), 47–57. [1]

Ghose, A., A. Goldfarb, and S. P. Han (2013). How is the mobile internet different? searchcosts and local activities. Information Systems Research 24(3), 613–631. [26]

Goldfarb, A. and C. Tucker (2011). Advertising bans and the substitutability of online andoffline advertising. Journal of Marketing Research 48(2), 207–227. [3]

Greiner, M. and H. Wang (2010). Building consumers-to-consumer trust in e-financemarketplaces: An empirical analysis. International Journal of Electronic Commerce 15(2),105–136. [26]

Grinblatt, M. and M. Keloharju (2001). How distance, language, and culture influencestockholdings and trades. The Journal of Finance 56(3), 1053–1073. [25]

Harrison, R., C. Mason, and P. Robson (2010). Determinants of long-distance investing bybusiness angels in the UK. Entrepreneurship & Regional Development 22(2), 113 – 137. [3]

Herzenstein, M., R. Andrews, U. Dholakia, and E. Lyandres (2008). The democratizationof personal consumer loans? determinants of success in online Peer-to-Peer lendingcommunities. Working Paper. [2]

Hillberry, R. and D. Hummels (2003). Intranational home bias: Some explanations. Reviewof Economics and Statistics 85(4), 1089–1092. [11]

Hortaçsu, A., F. A. Martínez-Jerez, and J. Douglas (2009). The geography of trade inonline transactions: Evidence from eBay and MercadoLibre. American Economic Journal:Microeconomics 1(1), 53–74. [3, 11, 12, 25]

Ivkovic, Z. and S. Weisbenner (2005). Local does as local is: Information content of thegeography of individual investors’ common stock investments. Journal of Finance 60(1),267–306. [24]

Iyer, R., A. Khwaja, E. Luttmer, and K. Shue (2009). Screening in new credit markets: Canindividual lenders infer borrower creditworthness in Peer-to-Peer lending? WorkingPaper. [2]

29

Janssen, M., J. L. Moraga-González, and M. Wildenbeest (2005). Consumer search andpricing behavior in internet markets. Working Paper. [26]

Jaworski, M. (2007, Sept 5). Prosper announces major changes to group concept. Technicalreport, P2P Lending News. [6]

Kim, K. and I.-H. Hann (2014). Crowdfunding and the democratization of access to capital:A geographical analysis. Working Paper. [18]

Klafft, M. (2008a). Online peer-to-peer lending: A lenders’ perspective. In H. R. Arabniaand A. Bahrami (Eds.), Proceedings of the International Conference on E-Learning, E-Business,Enterprise Information Systems, and E-Government, pp. 371–375. CSREA Press. [2]

Klafft, M. (2008b). Peer to peer lending: Auctioning microcredits over the internet.In A. Agarwal and R. Khurana (Eds.), International Conference on Information Systems,Technology and Management, Dubai. [2, 5]

Lendle, A., M. Olarreaga, S. Schropp, and P.-L. Vézina (2013). There goes gravity: HoweBay reduces trade costs. Discussion Papers 9094. [3]

Lieber, R. (2011, Februar 5). The gamble of lending peer to peer. New York Times, B1. [6]

Lin, M. and S. Viswanathan (2014). Home bias in online investment: An empirical studyof an online crowdfunding market. Working Paper. [3, 12, 16, 17, 25, 37, 38]

Mason, C. (2007). Venture capital: A geographic perspectice. In Landström (Ed.), Handbookof Research on Venture Capital, pp. 86–112. Edward Elgar. [3]

Massolution.com (2015). The crowdfunding industry report. Technical report, CrowdSourcing. [4]

Mollick, E. R. (2014). The dynamics of crowdfunding. Journal of Business Venturing 29, 1–16.[2]

Nowak, A., A. Ross, and C. Yencha (2015, July). Small Business Borrowing and Peer-to-Peer Lending: Evidence from Lending Club. Working Papers 15–28, Department ofEconomics, West Virginia University. [18]

NSR Invest (2016). Accumulative Monthly Funds. Technical report, Nickel Streamroller,LLC and Lend Academy Investment, LLC. Data obtained from Prosper and LendingClub. [4]

Pope, D. and J. Sydnor (2011). What’s in a picture?: Evidence of discrimination fromProsper.com. Journal of Human Resources 46(1), 53–92. [2]

30

Powell, W., K. Koput, J. Bowie, and L. Smith-Doerr (2002). The spatial clustering ofscience and capital: Accounting for biotech firm-venture capital relationships. RegionalStudies 36(3), 291–305. [3]

Prystav, F. (2016). Personal information in peer-to-peer loan applications: Is less more?Journal of Behavioral and Experimental Finance 9, 6–19. [26]

Ratchford, B., X. Pan, and V. Shankar (2003). On the efficiency of internet markets forconsumer goods. Journal of Public Policy & Marketing 22(1), 4–16. [3]

Ravina, E. (2012). Love and loans: The effect of beauty and personal characteristics incredit markets. Working Paper. [2]

Smith, L. and P. N. Sørensen (2000). Pathological outcomes of observational learning.Econometrica 68(2), 371–398. [7]

Sohl, J. (1999). The early stage equity markets in the USA. Venture Capital: An InternationalJournal of Entrpreneurial Finance 1, 101–120. [3]

Sorenson, O. and T. Stuart (2001). Syndication networks and social distribution of venturecapital. American Journal of Sociology 106(6), 1546–1588. [3]

Stiglitz, J. and A. Weiss (1981). Credit rationing in markets with imperfect information.American Economic Review 71(3), 393–410. [1]

Van Nieuwerburgh, S. and L. Veldkamp (2009). Information immobility and home biaspuzzle. Journal of Finance 64(3), 1187–1215. [24]

Wolf, H. C. (2000). A intranational home bias in trade. Review of Economics and Statis-tics 82(4), 555–563. [11]

Wong, A., M. Bhatia, and Z. Freeman (2009). Angel finance: The other venture capital.Strategic Change 18(7–8), 221–230. [3]

Zhang, J. and P. Lui (2012). Rational herding in microloan markets. Managment Science 58(5),555–563. [2, 11, 17, 22, 25]

31

Tables & Figures

Table 1: Summary Statistics for Loan Request Listings

All Listings N=42,657

Max Mean Median Min Std Dev

Loan Amount 25,000 8,349.23 6,000 1,000 6,558.28Bid Count 1,124 47.4 5 1 102Borrower Maximum Rate 0.360 0.232 0.224 0.010 0.092DTI 10.10 1.90 0.31 0 3.60Homeowner 1 0.49 0 0 0.49Duration 10 7.10 7 3 1.33Total Competition 3,401 2,119.1 2,158 484 428.9Credit Grade Competition 508 206.6 201 1 93.6

Table 2: Listing and Loan Breakdown by Credit Grade

All Listings Completion Rate All Loans Default Rate

All 42,657 0.2256 9,099 0.3013AA 3,471 0.4676 1,533 0.1768A 4,272 0.3373 1,349 0.2491B 6,118 0.3213 1,862 0.3287C 9,462 0.2313 2,070 0.3193D 11,082 0.1543 1,633 0.3588E 8,252 0.0842 652 0.4233

32

Table 3: Bid Amount and Bid Timing by Local Status

Bid Amount Bid Time

75% Mean Median Mean Median 25% N

All Bids 72.30 78.65 50.00 0.634 0.791 0.304 2,022,910Local Bids 79.00 88.02 50.00 0.619 0.769 0.272 118,798Nonlocal Bids 70.68 78.07 50.00 0.635 0.792 0.304 1,904,112

Table 4: Effect of Local Status on Bid Amount, Tobit with censoring at log($50)

log(Bid Amount) Coef. Std. Err. Coef. Std. Err.

log(Loan Amount) -0.031∗∗∗ 0.002 -0.045∗∗∗ 0.002In Group 0.082∗∗∗ 0.002 0.078∗∗∗ 0.003Local Lender 0.069∗∗∗ 0.013 0.070∗∗∗ 0.006Default -0.008∗∗ 0.002Local Lender_Default -0.027∗∗ 0.009Borrower Max Rate 0.011∗∗∗ 3.96E-04 0.008∗∗∗ 4.17E-04DTI 0.007∗∗∗ 3.83E-04 0.007∗∗∗ 4.65E-04Homeowner 0.043∗∗∗ 0.002 0.046∗∗∗ 0.002Total Competition -4.09E− 06† 2.15E-06 -7.59E-06∗∗ 2.41E-06Credit Grade Competition 1.44E-04∗∗∗ 2.30E-05 2.86E− 04∗∗∗ 2.58E-05Fully Funded 0.083∗∗∗ 0.002 0.070∗∗∗ 0.003Current Bid Count -2.99E-04∗∗∗ 1.13E-05 -1.319E-04∗∗∗ 1.26E-05Current Rate 0.013∗∗∗ 4.02E-04 0.017∗∗∗ 4.36E-04

N 2,022,910 1,598,786Pseudo R2 0.159 0.157

Significance: 0.10†, 0.05∗, 0.01∗∗, and 0.001∗∗∗.

FE: state, category, credit grade, quarter, month, and day of the week.

Standard errors are clustered at the lender level, intercept not shown.

33

Table 5: Test for Difference Between Nonlocal and Local Bid Interest Rate Distributions

t-Test for Difference in Mean

Default=1 Default=0

AA -3.14 (0.001) 4.47 (0.000)A -2.97 (0.002) 1.23 (0.215)B -6.00 (0.000) 0.83 (0.406)C -1.70 (0.088) 6.93 (0.000)D -4.30 (0.000) 3.79 (0.000)E -2.53 (0.011) 2.40 (0.016)

Note: Test statistics are for (Nonlocal − Local)

p-values are in parentheses.

34

Table 6: Effect of Local Status on Bid Interest Rate, Type II Tobit censored at Final Rate

Bid Interest Rate Coef. Std. Err.

Loan Amount -0.037∗∗∗ 0.002In Group 0.222∗∗∗ 0.016LocalAA -0.142∗∗∗ 0.033LocalA -0.071∗∗∗ 0.011LocalB -0.032∗ 0.013LocalC -0.051∗ 0.017LocalD -0.241∗∗ 0.098LocalE -0.907∗∗∗ 0.311Defaulted 0.041∗∗∗ 0.010LocalAA_Defaulted 0.282∗∗∗ 0.063LocalA_Defaulted 0.1536† 0.086LocalB_Defaulted 0.122† 0.067LocalC_Defaulted 0.332∗∗∗ 0.104LocalD_Defaulted 0.298∗ 0.148LocalE_Defaulted 0.992∗ 0.414Borrower Max Rate 0.072∗∗∗ 0.003DTI 0.012∗∗ 0.004Homeowner 0.323∗ 0.012Total Competition -8.40E-5∗∗∗ 8.29E-6Credit Grade Competition 8.44E-3∗∗∗ 2.27E-4Current Bid Count 1.13E-3∗∗∗ 1.04E-4Current Rate 0.698∗∗∗ 0.005

N 1,598,786Prob > χ2 0

Significance: 0.10†, 0.05∗, 0.01∗∗, and 0.001∗∗∗.

FE: state, category, credit grade, quarter, month, and day of the week.

Standard errors are clustered at the lender level, intercept not shown.

35

Table 7: Bid Count and Amount Pledged in an Auction Before a Bidder Bids by LocalStatus

Number of Previous Bids

75% Mean Median 25% N

Local Bidders 182 118.7 70 23 118,798Nonlocal Bidders 186 132.1 84 28 1,904,112

Amount Pledged

75% Mean Median 25% N

Local Bidders 14,093.20 9,918.64 5,562.97 2,178.97 118,798Nonlocal Bidders 14,393.88 10,558.23 6,358.73 2,252 1,904,112

Table 8: Lender Propensity to Bid

OLS Cox Proportional HazardCoef. Std. Err. Coef. Std. Err.

Loan Amount 1.60E-03∗∗∗ 6.57E-05 1.217E-3 1.436E-3In Group -0.026∗∗∗ 9.01E-04 0.127∗∗∗ 2.189E-3Local Lender -0.019∗∗∗ 0.001 0.033∗∗∗ 3.011E-3Post Fully Funded -0.371∗∗∗ 7.07E-04 0.033∗∗∗ 3.011E-3Local x Post Fully Funded 0.011∗∗∗ 0.002 0.033∗∗∗ 3.011E-3Borrower Max Rate 1.081∗∗∗ 0.012 0.133∗∗∗ 0.031DTI 0.012∗∗∗ 1.18E-04 -0.056∗∗∗ 2.921E-4Homeowner 0.015∗∗∗ 6.49E-04 -0.080∗∗∗ 1.562E-3Total Competition 4.06E-06∗∗∗ 6.64E-07 -1.280E-5∗∗∗ 1.620E-6Credit Grade Competition 4.97E-04∗∗∗ 5.42E-06 -3.532E-3∗∗∗ 1.440E-5Current Bid Count 2.37E-04∗∗∗ 3.47E-06 -1.481E-3∗∗∗ 8.500E-6Current Bid Rate 0.198∗∗∗ 0.01 -0.204∗∗∗ 0.032

N 2,022,910 2,000,483Prob > F 0 0

Significance: 0.10†, 0.05∗, 0.01∗∗, and 0.001∗∗∗.

FE: state, category, credit grade, quarter, month, and day of the week.

36

Table 9: Influence of Social Connections on Bid Amount

Effect of Friends & Family on Bid Amount

75% Mean Median N

All Bids 72.30 78.65 50.00 2,022,910Local Bids 79.00 88.02 50.00 118,798F&F(Agrawal et al.) Bids 100.00 150.24 50.00 3,534Local and Non-F&F(Agrawal et al.) Bids 76.29 86.12 50.00 115,264F&F(Lin and Viswanathan) Bids 100.00 132.79 50.00 4,465Local and Non-F&F(Lin and Viswanathan) Bids 77.90 86.76 50.00 118,106

Effect of Being In Group on Bid Amount

75% Mean Median N

In Group 75.00 84.51 50.00 243,906Not Group 70.00 77.85 50.00 1,779,004Local and In Group 100.00 98.27 50.00 16,020Nonlocal and In Group 75.00 83.55 50.00 227,886Local and Not In Group 75.00 86.43 50.00 102,778Nonlocal and Not In Group 70.00 77.32 50.00 1,676,226

37

Table 10: t-Tests for Difference in Mean for Friends and Family Bid Interest Rate

Agrawal et al. Definition

Default=1 Default=0

AA 4.32 (0.000) 2.60 (0.009)A 2.52 (0.011) 3.18 (0.001)B 3.95 (0.000) 0.83 (0.408)C 1.50 (0.133) 4.51 (0.000)D 2.82 (0.004) 4.96 (0.000)E 1.68 (0.093) 3.48 (0.000)

Lin and Viswanathan Definition

Default=1 Default=0

AA 8.55 (0.000) 9.11 (0.000)A 5.61 (0.000) 8.80 (0.004)B 7.87 (0.000) 7.33 (0.178)C 4.82 (0.000) 5.48 (0.000)D 4.54 (0.000) 5.54 (0.000)E 0.59 (0.555) 0.57 (0.566)

Note: Test statistics are for (Nonlocal − Local)

p-values are in parentheses.

Table 11: t-Tests for Difference in Mean for In Group Bid Interest Rate

Default=1 Default=0

AA 3.98 (0.000) 10.03 (0.000)A 20.14 (0.000) 5.87 (0.000)B 13.53 (0.000) 25.67 (0.000)C 5.63 (0.000) 0.81 (0.419)D 21.36 (0.000) 19.19 (0.000)E 8.01 (0.000) 9.96 (0.000)

Note: Test statistics are for (Nonlocal − Local)

p-values are in parentheses.

38

Table 12: Marginal Effect of Early Local Bids on Bid Count

Total Bid Count Total Nonlocal Bid Count

dy/dx Std. Err. dy/dx Std. Err.

Total Early Count 0.283∗∗∗ 0.003 0.074∗∗∗ 0.002Total Early Amount 1.51E-03∗∗∗ 4.11E-05 1.46E-03∗∗∗ 2.18E-05Total Early Count x Amount -1.82E-05∗∗∗ 1.53E-07 -8.54E-06∗∗∗ 1.03E-07Local Early Count 0.164∗∗∗ 0.019 0.055∗∗∗ 0.012Local Early Amount 1.80E-03∗∗∗ 1.31E-04 1.25E-03∗∗∗ 7.55E-05Local Early Count x Amount -8.20E-06† 4.83E-06 -2.21E-05∗∗∗ 3.04E-06In Group 3.463∗∗∗ 0.106 2.039∗∗∗ 0.068Loan Amount 2.529∗∗∗ 0.006 0.911∗∗∗ 0.004Borrower Max Rate 1.744∗∗∗ 0.007 101.127∗∗∗ 0.497DTI -2.491∗∗∗ 0.014 -0.857∗∗∗ 0.010Homeowner 0.220∗∗ 0.075 0.990∗∗∗ 0.051Total Competition -7.35E-04∗∗∗ 7.64E-05 -5.23E-04∗∗∗ 5.12E-05Credit Grade Competition 0.101∗∗∗ 0.001 0.100∗∗∗ 0.001

N 42,657 42,657Pseudo R2 0.830 0.858

Significance: 0.10†, 0.05∗, 0.01∗∗, and 0.001∗∗∗.

FE: state, category, credit grade, quarter, month, and day of the week, intercept not shown.

39

Table 13: Marginal Effect of Early Local Bids on Pre & Post-Fully Funded Total Bid Count

Total Pre-FF Bid Count Total Post-FF Bid Count

dy/dx Std. Err. dy/dx Std. Err.

Total Early Count 1.036∗∗∗ 0.011 0.416∗∗∗ 0.002Total Early Amount 7.25E-03∗∗∗ 1.49E-04 5.62E-03∗∗∗ 1.89E-05Total Early Count x Amount -2.45E-05∗∗∗ 3.32E-07 -4.88E-05∗∗∗ 1.74E-07Local Early Count 0.101∗ 0.051 0.194∗∗∗ 0.012Local Early Amount 4.07E-03∗∗∗ 4.24E-04 1.45E-04∗ 6.37E-05Local Early Count x Amount 5.08E-05∗∗∗ 1.16E-05 -3.19E-05∗∗∗ 4.34E-06In Group -5.164∗∗∗ 0.306 5.633∗∗∗ 0.068Loan Amount 7.743∗∗∗ 0.02 -0.505∗∗∗ 0.004Borrower Max Rate -0.587∗∗∗ 0.021 1.893∗∗∗ 0.005DTI -1.272∗∗∗ 0.042 -2.545∗∗∗ 0.01Homeowner 0.790∗∗∗ 0.213 0.153∗∗∗ 0.049Total Competition 6.33E-04∗∗ 2.14E-04 -1.12E-03∗∗∗ 5.05E-05Credit Grade Competition 0.047∗∗∗ 0.002 0.064∗∗∗ 0.001

N 9,624 9,624PseudoR2 0.723 0.538

Significance: 0.10†, 0.05∗, 0.01∗∗, and 0.001∗∗∗.

FE: state, category, credit grade, quarter, month, and day of the week, intercept not shown.

40

Table 14: Marginal Effect of Early Local Bidding on Probability of Funding and FinalInterest Rate

Probability of Funding Final Interest Rate

dy/dx Std. Err. Coef. Std. Err.

Total Early Count 0.028∗∗∗ 0.001 -0.025∗∗∗ 0.004Total Early Amount 1.33E-04∗∗∗ 1.05E-05 -4.21E-04∗∗∗ 5.73E-05Total Early Count x Amount -1.43E-06∗∗∗ 2.85E-08 2.56E-06∗∗∗ 1.81E-07Local Early Count 0.009∗ 0.004 -0.038† 0.022Local Early Amount -3.39E-05∗ 1.52E-05 -9.33E-03∗∗∗ 1.38E-04Local Early Count x Amount -5.49E-06∗∗∗ 1.01E-06 6.68E-08 6.14E-06In Group 0.033∗∗∗ 0.004 -0.346∗∗ 0.108Loan Amount -0.027∗∗∗ 0.000 0.163∗∗∗ 0.008Borrower Max Rate 6.91E-03∗∗∗ 2.77E-04 0.509∗∗∗ 0.008DTI -0.003∗∗∗ 0.000 0.037∗∗ 0.014Homeowner -0.003 0.003 0.224∗∗ 0.078Total Competition 2.72E-06 2.96E-06 1.27E-04∗∗ 8.09E-05Credit Grade Competition -3.50E-04∗∗∗ 2.75E-05 -7.06E-03∗∗∗ 6.41E-04

N 42,656 9,624Adj R2 0.554 0.759

Significance: 0.10†, 0.05∗, 0.01∗∗, and 0.001∗∗∗.

FE: state, category, credit grade, quarter, month, and day of the week, intercept not shown.

41

Table 15: Marginal Effect of Early Local Bidding on Probability of Default

Pre-Fully Funded First 2 hr

Default dy/dx Std. Err. dy/dx Std. Err.

Total Early Count -0.091∗∗∗ 0.023 -0.035 0.058Total Early Amount 3.77E-05 7.38E-04 -9.33E-04 7.57E-04Total Early Count x Amount -3.71E-06∗∗∗ 9.18E-07 1.34E-06 2.32E-06Local Early Count -0.173† 0.171 -0.250 0.309Local Early Amount -6.66E-04 7.10E-04 -2.68E-04 1.80E-03Local Early Count x Amount -2.78E-05 2.36E-05 -1.20E-04 9.92E-05In Group -4.774∗∗ 1.414 -4.883∗∗ 1.419Loan Amount 0.664 0.786 0.874∗∗∗ 0.108Borrower Max Rate 0.707∗∗∗ 0.117 0.679∗∗∗ 0.117Loan Rate 0.973∗∗∗ 0.122 0.859∗∗∗ 0.124DTI 0.414∗ 0.187 0.349∗∗∗ 0.186Homeowner 7.282∗∗∗ 1.008 7.195∗∗∗ 1.008Total Competition 3.55E-04 1.05E-03 4.44E-04 1.05E-03Credit Grade Competition -6.31E-03 8.05E-03 -6.86E-03 8.19E-03

N 9,091 9,091Pseudo R2 0.089 0.087

Significance: 0.10†, 0.05∗, 0.01∗∗, and 0.001∗∗∗.

FE: state, category, credit grade, quarter, month, and day of the week, intercept not shown.

42

Figure 1: Default Rate by Category and State

43