Given that i do not observe information about the market participants'offline activity or connections, one immediate concern that needs to be addressed is whether my definition of local is picking up borrowers friends and family simply using the site as a way to formalize their loans. While friends and family would be better informed about the borrower's characteristics, it is unlikely that they be motivated by profits. If friends and families are a major part of the local bidding, it is reasonable that they will join the site around the time that borrower creates his or her listing. It is unlikely to imagine that a majority of the borrowers creating listings will have friends and family who are previously active on Prosper. Table 5 presents some descriptive statistics on the amount of previous bidding and length of time on the site of the lenders at the time of their first local bid. The median local lender has been active on the site for over 124 days and has submitted 15 previous bids by the time that they place their first local bid. Additionally, the average lender has been active for over 210 days and has placed about 50 bids before he submits his first local bid. Furthermore less than 2% of lenders place their very first bid on a local project and less than 8% of local bids are placed during the first three days after the lender joins the site. This suggests that it is highly unlikely that a major part of local bidding is coming from friends and family I start this analysis by documenting the fact that local lenders do behave differently than nonlocal lenders in the P2P lending market under study. The evidence I find strongly suggests that information-based frictions are still a major contributor to the observed behavioral differences between local and nonlocal lenders. Thus, i develop a simple model of social-learning to motivate these empirical observations and to assist in explaining how local lenders' actions can transmit their private information about the underlying quality of borrowers to nonlocal lenders who are less informed This social-learning will act like a signal, making a listing with more local lenders more attractive to other lenders All regressions include Borrower and lender State(where applicable), Credit Grade, Category of use, Quarter, Month, and Day of the Week fixed effects. The variable Loan amount is measured in thousands of dollars 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 4. 1. Local lenders Bid Larger amounts The left half of Table 6 displays the mean, median, and 75 percentile of bid amounts by local status. The median bid is the minimum allowable amount of $50. However the average and 75th percentile local bid amounts are around $10 dollar larger than the 15i also explicitly control for observable borrower state characteristics in place of the borrower state FE and found no noticeable difference in the results
Given that I do not observe information about the market participants’ offline activity or connections, one immediate concern that needs to be addressed is whether my definition of local is picking up borrowers’ friends and family simply using the site as a way to formalize their loans. While friends and family would be better informed about the borrower’s characteristics, it is unlikely that they be motivated by profits. If friends and families are a major part of the local bidding, it is reasonable that they will join the site around the time that borrower creates his or her listing. It is unlikely to imagine that a majority of the borrowers creating listings will have friends and family who are previously active on Prosper. Table 5 presents some descriptive statistics on the amount of previous bidding and length of time on the site of the lenders at the time of their first local bid. The median local lender has been active on the site for over 124 days and has submitted 15 previous bids by the time that they place their first local bid. Additionally, the average lender has been active for over 210 days and has placed about 50 bids before he submits his first local bid. Furthermore, less than 2% of lenders place their very first bid on a local project and less than 8% of local bids are placed during the first three days after the lender joins the site. This suggests that it is highly unlikely that a major part of local bidding is coming from friends and family. I start this analysis by documenting the fact that local lenders do behave differently than nonlocal lenders in the P2P lending market under study. The evidence I find strongly suggests that information-based frictions are still a major contributor to the observed behavioral differences between local and nonlocal lenders. Thus, I develop a simple model of social-learning to motivate these empirical observations and to assist in explaining how local lenders’ actions can transmit their private information about the underlying quality of borrowers to nonlocal lenders, who are less informed. This social-learning will act like a signal, making a listing with more local lenders more attractive to other lenders. All regressions include Borrower and lender State (where applicable), Credit Grade, Category of use, Quarter, Month, and Day of the Week fixed effects.15 The variable Loan Amount is measured in thousands of dollars, 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. 4.1. Local lenders Bid Larger Amounts The left half of Table 6 displays the mean, median, and 75th percentile of bid amounts by local status. The median bid is the minimum allowable amount of $50. However, the average and 75th percentile local bid amounts are around $10 dollar larger than the 15I also explicitly control for observable borrower state characteristics in place of the borrower state FE and found no noticeable difference in the results. 11
nonlocal bids; these differences are significant at the 1% level. It is clear that local lenders bid larger amounts than their nonlocal counterparts To formally test the question of whether local lenders bid larger amounts than nonlocal lenders, I run a Tobit regression of log 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 correlation between the bidding activity of the same lender. Table 7 shows the results; I find that even after controlling for all the observable listing characteristics, local lenders bid roughly 7% larger amounts than nonlocal lenders. Thi effect is of the similar magnitude to the increase in bid amount that occurs if lender and borrower are in the same Prosper group, and is roughly equivalent to the current standing interest rate in the auction being 5.3 percentage points higher. This result is consistent with the findings of local preference in other online markets (Hortacsu et al., 2009, Lin and Viswanathan, 2014). 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 lender's behavior on the site, and not on the lenders themselves. It is well documented that demographic characteristics like age, educational attainment, marital status, gender, race, and risk tolerance all affect asset holdings. b More information about the heterogeneity of lenders would likely improve the fit of thi Knowing that local lenders have a larger demand for local loans suggests that lenders behave differently based on locality, but tells nothing about the channel that is driving this behavior. To examine this i restrict my focus to listings that become loans and their outcome is known. Bids are divided into sub-samples based on the loans' ex-post outcome (default or paid back). Regardless of the ex-post outcome, local lenders bid larger amount than their nonlocal counterparts; however the magnitude of the difference is significantly larger for loans that pay back in full. As can be seen in Table 8, while the differential between local and nonlocal lenders is significantly greater than zero regardless of loan outcome, it is 2.7 percentage points larger for loans that ex-post pay back. This behavior suggest that there does exist a preference for local loans, but the local lenders are better informed and thus demand larger amounts of the less risky local loans 4. 2. Local lenders Evaluate the Probability of default Better than Nonlocal lenders If local lenders are merely acting altruistically towards individuals living in their area then local lenders should bid lower interest rates, regardless of the borrower's quality, than their nonlocal counterparts. However, if local lenders are actually more informed about the underlying riskiness of the borrower and the general market conditions, they should Kreinin(1959); Baker and Haslem(1974); Figner and Weber(2011)
nonlocal bids; these differences are significant at the 1% level. It is clear that local lenders bid larger amounts than their nonlocal counterparts. To formally test the question of whether local lenders bid larger amounts than nonlocal lenders, I run a Tobit regression of log 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 correlation between the bidding activity of the same lender. Table 7 shows the results; I find that even after controlling for all the observable listing characteristics, local lenders bid roughly 7% larger amounts than nonlocal lenders. This effect is of the similar magnitude to the increase in bid amount that occurs if lender and borrower are in the same Prosper group, and is roughly equivalent to the current standing interest rate in the auction being 5.3 percentage points higher. This result is consistent with the findings of local preference in other online markets (Hortaçsu et al., 2009; Lin and Viswanathan, 2014). 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 lender’s behavior on the site, and not on the lenders themselves. It is well documented that demographic characteristics like age, educational attainment, marital status, gender, race, and risk tolerance all affect asset holdings.16 More information about the heterogeneity of lenders would likely improve the fit of this regression. Knowing that local lenders have a larger demand for local loans suggests that lenders behave differently based on locality, but tells nothing about the channel that is driving this behavior. To examine this, I restrict my focus to listings that become loans and their outcome is known. Bids are divided into sub-samples based on the loans’ ex-post outcome (default or paid back). Regardless of the ex-post outcome, local lenders bid larger amount than their nonlocal counterparts; however the magnitude of the difference is significantly larger for loans that pay back in full. As can be seen in Table 8, while the differential between local and nonlocal lenders is significantly greater than zero regardless of loan outcome, it is 2.7 percentage points larger for loans that ex-post pay back. This behavior suggest that there does exist a preference for local loans, but the local lenders are better informed and thus demand larger amounts of the less risky local loans. 4.2. Local lenders Evaluate the Probability of Default Better than Nonlocal lenders If local lenders are merely acting altruistically towards individuals living in their area, then local lenders should bid lower interest rates, regardless of the borrower’s quality, than their nonlocal counterparts. However, if local lenders are actually more informed about the underlying riskiness of the borrower and the general market conditions, they should 16Kreinin (1959); Baker and Haslem (1974); Figner and Weber (2011). 12
be better able to evaluate the risk of the listing. For example, if there are two listings that look identical to nonlocal lenders, but one is riskier than another, local lenders should bid larger interest rates on the listing with the hidden extra riskiness. To examine this, I estrict my focus to listings that became loans and their outcomes are known. These bids are divided based on the loans' ex-post outcome. I further focus on only losing bids where I know the actual bid interest rate. Table 9 presents the mean, median, and 75th percentile of bid interest rate for losing bids by credit grade, local status, and loan outcome. For loans that ex-post default, local lenders tend to submit ex-ante larger interest rates for all credit grades. For loans that ex-post pay back in full, the opposite result is seen, with local lenders tending to bid lower ex-ante interest rates. These two observations strongly suggest that local lenders seem to more accurately price the underlying risk of the listings than nonlocal lenders To formally evaluate this claim, I perform two tests on the distributions of bid interest rates by locality: (1)a t-test on whether the difference between the distributions'means is significantly different from zero, and(2)a two-sample Mann-Whitney u-test(Wilcoxon rank-sum test), which is a nonparametric rank-sum test to determine if one sample stochastically dominates the other sample. The top half of Table 10 presents the t-statistics and the p-values for testing the differences between the means, (uNonlocal-FLocal), by ex-post loan outcome. For loans that default, the differences are negative and significant for all credit grades; the difference for credit grade C is significant at the 10%, while the rest are significant at the 5% level. For loans that pay back, the differences are all positive. These differences in means are significant at the 5% or smaller level for all credit grades except A and B. To further explore the difference between the local and nonlocal bid interest rate distributions, the bottom half of Table 10 displays the u-statistics and p-values from the one-sided two-sample Mann-Whitney U-tests. The testing procedure can be interpreted as comparing the medians of the two samples. The results are for a one-sided test to determine if one of the distributions of bid interest rates by credit grade, stochastically dominates the other. Given the large sample size, the test statistics For loans that eventually default, the null hypothesis is that values drawn from the nonlocal distribution tend to be larger than or equal to values drawn from the local distribution( the nonlocal distribution stochastically dominates or is equal in distribution to the local distribution). The u-statistics for this one-side test are negative for all credit grades. Given that all the p-values are less than 0.02, I can reject the null in favor of the alternative hypothesis (i.e. the local distribution stochastically dominates the nonlocal 1For further reference see Wilcoxon(1945); Mann and Whitney(1947): Hettmansperger and McKean(1998); ehmann and D'Abrera(2006)
be better able to evaluate the risk of the listing. For example, if there are two listings that look identical to nonlocal lenders, but one is riskier than another, local lenders should bid larger interest rates on the listing with the hidden extra riskiness. To examine this, I restrict my focus to listings that became loans and their outcomes are known. These bids are divided based on the loans’ ex-post outcome. I further focus on only losing bids where I know the actual bid interest rate. Table 9 presents the mean, median, and 75th percentile of bid interest rate for losing bids by credit grade, local status, and loan outcome. For loans that ex-post default, local lenders tend to submit ex-ante larger interest rates for all credit grades. For loans that ex-post pay back in full, the opposite result is seen, with local lenders tending to bid lower ex-ante interest rates. These two observations strongly suggest that local lenders seem to more accurately price the underlying risk of the listings than nonlocal lenders. To formally evaluate this claim, I perform two tests on the distributions of bid interest rates by locality: (1) a t-test on whether the difference between the distributions’ means is significantly different from zero, and (2) a two-sample Mann-Whitney U-test (Wilcoxon rank-sum test), which is a nonparametric rank-sum test to determine if one sample stochastically dominates the other sample. The top half of Table 10 presents the t-statistics and the p-values for testing the differences between the means, (µNonlocal − µLocal), by ex-post loan outcome. For loans that default, the differences are negative and significant for all credit grades; the difference for credit grade C is significant at the 10%, while the rest are significant at the 5% level. For loans that pay back, the differences are all positive. These differences in means are significant at the 5% or smaller level for all credit grades except A and B. To further explore the difference between the local and nonlocal bid interest rate distributions, the bottom half of Table 10 displays the U-statistics and p-values from the one-sided two-sample Mann-Whitney U-tests. The testing procedure can be interpreted as comparing the medians of the two samples. The results are for a one-sided test to determine if one of the distributions of bid interest rates, by credit grade, stochastically dominates the other. Given the large sample size, the test statistics are approximately normal.17 For loans that eventually default, the null hypothesis is that values drawn from the nonlocal distribution tend to be larger than or equal to values drawn from the local distribution (the nonlocal distribution stochastically dominates or is equal in distribution to the local distribution). The U-statistics for this one-side test are negative for all credit grades. Given that all the p-values are less than 0.02, I can reject the null in favor of the alternative hypothesis (i.e., the local distribution stochastically dominates the nonlocal 17For further reference see Wilcoxon (1945); Mann and Whitney (1947); Hettmansperger and McKean (1998); Lehmann and D’Abrera (2006). 13
distribution). Values drawn from the local distribution tend to be larger than values drawn from the nonlocal distribution. Additionally, for loans that pay back ex-post, the null hypothesis is that the local distribution stochastically dominates or is equal in distribution to the nonlocal distribution. I find that the u-statistics are positive for all credit grade and the p-values are less than 0.05 for all credit grades except B and E. However, the difference for e listings is significant at the 10% level. Accordingly, I reject the null in favor of the alternative (i.e that the values drawn from the local distribution tend to be smaller than the values drawn from the nonlocal distribution) for most of the credit grades. To put these results into context, local lenders bid interest rates seem to better reflect the true revealed riskiness of a local listin To further examine this difference in the bid interest rates, I run a Type II Tobit egression with left censoring at each listings winning interest rate, since that is the smallest rate that i observe for each listing. The standard errors are clustered at the lender level. The regression contains indicators for local lender interacted credit grade, an indicator for if the loan defaulted, and the local lender indicator interacted with credi grade and the default dummy. The results displayed in Table 1l are consistent with the previous results, showing that local lenders act differentially based on ex-post loan outcomes. The coefficients on the local indicator interacted with credit grade are all negative and significant at the 5% level. The magnitude of the difference between local and nonlocal bids for loans that do not default varies significantly across credit grades B listings have the lowest differential at 0.032 percentage points while E listings have the largest at 0.907 percentage points. The reduction that local lenders give to loans that pay back ex-post in their ax-ante bids, generally increases as 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-ante potentially riskier borrowers who ex-post turn out to be lower risk than their financial information indicates. This result makes sense in the context that the better information possessed by the local lenders should have the biggest effect on the listing in the worst credit grades, since these listings have the most potential space for a difference between true risk and perceived risk The coefficients for the local indicator interacted with credit grades and default are all positive and significant. The net effect of being local on the submitted bid interest rate is positive if the loan ex-post defaults. The premium that local lenders give to loans that ex-post default, relative to nonlocal lenders, ranges from 0.056 to 0. 28 percentage points The pattern across credit grades is less clear here than for the reduction seen for loans that do not default. E listings generally get the smallest premium followed very closely by a i also r o-sample Kolmogorov-Smirnov distributional equality tests; the results are consistent with the findings of the Wilcoxon-Mann-Whitney test presented here
distribution). Values drawn from the local distribution tend to be larger than values drawn from the nonlocal distribution. Additionally, for loans that pay back ex-post, the null hypothesis is that the local distribution stochastically dominates or is equal in distribution to the nonlocal distribution. I find that the U-statistics are positive for all credit grade and the p-values are less than 0.05 for all credit grades except B and E. However, the difference for E listings is significant at the 10% level. Accordingly, I reject the null in favor of the alternative (i.e., that the values drawn from the local distribution tend to be smaller than the values drawn from the nonlocal distribution) for most of the credit grades. To put these results into context, local lenders’ bid interest rates seem to better reflect the true revealed riskiness of a local listing.18 To further examine this difference in the bid interest rates, 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. The standard errors are clustered at the lender level. The regression contains indicators for local lender interacted credit grade, an indicator for if the loan defaulted, and the local lender indicator interacted with credit grade and the default dummy. The results displayed in Table 11 are consistent with the previous results, showing that local lenders act differentially based on ex-post loan outcomes. The coefficients on the local indicator interacted with credit grade are all negative and significant at the 5% level. The magnitude of the difference between local and nonlocal bids for loans that do not default varies significantly across credit grades; B listings have the lowest differential at 0.032 percentage points while E listings have the largest at 0.907 percentage points. The reduction that local lenders give to loans that pay back ex-post in their ax-ante bids, generally increases as 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-ante potentially riskier borrowers who ex-post turn out to be lower risk than their financial information indicates. This result makes sense in the context that the better information possessed by the local lenders should have the biggest effect on the listing in the worst credit grades, since these listings have the most potential space for a difference between true risk and perceived risk. The coefficients for the local indicator interacted with credit grades and default are all positive and significant. The net effect of being local on the submitted bid interest rate is positive if the loan ex-post defaults. The premium that local lenders give to loans that ex-post default, relative to nonlocal lenders, ranges from 0.056 to 0.28 percentage points. The pattern across credit grades is less clear here than for the reduction seen for loans that do not default. E listings generally get the smallest premium followed very closely by A, 18I also ran two-sample Kolmogorov-Smirnov distributional equality tests; the results are consistent with the findings of the Wilcoxon-Mann-Whitney test presented here. 14
B, and d listings but these differences are all less than 0. 1 percentage points. However, the local premium is significantly larger for AA and C listings. The total effect of this local behavior on the final outcome for the borrower depends on many factors, but reducing the final interest rate by 1% would lower the total loan payment made by the borrower by a few hundred dollars. Given that there exists a preference for local projects, the total ffect of this behavior would be significant when aggregated across a lenders portfolio This differential lender behavior supports the idea that local lenders seem better able to evaluate a listings underlying risk than nonlocal lenders, with the obvious rationale being that local lenders are relatively more informed. Theory predicts that this informational symmetry arises from the fact that the cost of becoming more informed about the market conditions and quality of local borrowers is significantly cheaper for local lenders due to their proximity. This result supports the idea that informational frictions are a main driver in explaining the behavioral difference between local and nonlocal lenders 4.3, Local lenders bid earlier If local lenders are better informed about the quality of listings posted by local borrowers, they should be more willing to bid earlier in the auction when the only information that has been revealed is the original public information and their private signal. Additionally, if nonlocal lenders are learning from local bids, then local lenders should be bidding earlier in the auction so that nonlocal lenders have time to react and process this newly revealed information before they bid. The right part of Table 6 shows the 25th percent mean,and median of bid times. Bid times are normalized so that zero is the beginning and one is the end of the auction the data shows that local lenders tend to bid earlier during the auction than nonlocal lenders. To put these normalized time differences into perspective, a vast majority of the listings are active for seven days, so 0.00595 is rough qual to a single hour. Consequently, the average local bid is placed a little more than 2.5 hours and the 25 percentile local bid is placed about 5.5 hours earlier than its nonlocal counterparts o formally test if local lenders do bid significantly earlier than nonlocal lenders, I perform a two-sample Mann-Whitney test on bid times. The p-value for test is 0.000, implying that the distribution of bid times for local lenders tends to be significantly smaller than the distribution of bid times for nonlocal lenders. 19 To further illustrate that local lenders bid earlier more often, Figure 2 plots the cumulative distribution functions of the bid time distribution by local status. The nonlocal bid time distribution first order stochastically dominates the local bid time distribution. The distance between the 19A two-sample Kolmogorov-Smirnov distributional equality test is consistent with the findings of the Mann-Whitney test presented here 15
B, and D listings but these differences are all less than 0.1 percentage points. However, the local premium is significantly larger for AA and C listings. The total effect of this local behavior on the final outcome for the borrower depends on many factors, but reducing the final interest rate by 1% would lower the total loan payment made by the borrower by a few hundred dollars. Given that there exists a preference for local projects, the total effect of this behavior would be significant when aggregated across a lender’s portfolio. This differential lender behavior supports the idea that local lenders seem better able to evaluate a listing’s underlying risk than nonlocal lenders, with the obvious rationale being that local lenders are relatively more informed. Theory predicts that this informational symmetry arises from the fact that the cost of becoming more informed about the market conditions and quality of local borrowers is significantly cheaper for local lenders due to their proximity. This result supports the idea that informational frictions are a main driver in explaining the behavioral difference between local and nonlocal lenders. 4.3. Local lenders Bid Earlier If local lenders are better informed about the quality of listings posted by local borrowers, they should be more willing to bid earlier in the auction when the only information that has been revealed is the original public information and their private signal. Additionally, if nonlocal lenders are learning from local bids, then local lenders should be bidding earlier in the auction so that nonlocal lenders have time to react and process this newly revealed information before they bid. The right part of Table 6 shows the 25th percentile, mean, and median of bid times. Bid times are normalized so that zero is the beginning and one is the end of the auction. The data shows that local lenders tend to bid earlier during the auction than nonlocal lenders. To put these normalized time differences into perspective, a vast majority of the listings are active for seven days, so 0.00595 is roughly equal to a single hour. Consequently, the average local bid is placed a little more than 2.5 hours and the 25th percentile local bid is placed about 5.5 hours earlier than its nonlocal counterparts. To formally test if local lenders do bid significantly earlier than nonlocal lenders, I perform a two-sample Mann-Whitney test on bid times. The p-value for test is 0.000, implying that the distribution of bid times for local lenders tends to be significantly smaller than the distribution of bid times for nonlocal lenders.19 To further illustrate that local lenders bid earlier more often, Figure 2 plots the cumulative distribution functions of the bid time distribution by local status. The nonlocal bid time distribution first order stochastically dominates the local bid time distribution. The distance between the 19A two-sample Kolmogorov-Smirnov distributional equality test is consistent with the findings of the Mann-Whitney test presented here. 15