«SCREENING PEERS SOFTLY: INFERRING THE QUALITY OF SMALL BORROWERS Rajkamal Iyer Asim Ijaz Khwaja Erzo F.P. Luttmer Kelly Shue Working Paper 15242 ...»
NBER WORKING PAPER SERIES
SCREENING PEERS SOFTLY:
INFERRING THE QUALITY OF SMALL BORROWERS
Asim Ijaz Khwaja
Erzo F.P. Luttmer
Working Paper 15242
NATIONAL BUREAU OF ECONOMIC RESEARCH1050 Massachusetts Avenue Cambridge, MA 02138 August 2009 This paper previously circulated as "Screening in New Credit Markets: Can Individual Lenders Infer Borrower Creditworthiness in Peer-to-Peer Lending?." We are extremely grateful to Prosper.com for being so generous with their time in answering our queries. We thank Liran Einav, Raymond Fisman, Matthew Gentzkow, Lawrence Katz, Atif Mian, Enrichetta Ravina, David Scharfstein, Jesse Shapiro, Jeremy Stein, three anonymous referees and seminar participants at Harvard, the NBER finance meetings and UCLA Anderson for helpful comments. We thank David Robinson for excellent research assistance.
The views expressed in this paper are solely our own. The views expressed in this paper are solely our own and do not necessarily reflect the views of the National Bureau of Economic Research.
NBER working papers are circulated for discussion and comment purposes. They have not been peer- reviewed or been subject to the review by the NBER Board of Directors that accompanies official NBER publications.
© 2009 by Rajkamal Iyer, Asim Ijaz Khwaja, Erzo F.P. Luttmer, and Kelly Shue. All rights reserved.
Short sections of text, not to exceed two paragraphs, may be quoted without explicit permission provided that full credit, including © notice, is given to the source.
Screening Peers Softly: Inferring the Quality of Small Borrowers Rajkamal Iyer, Asim Ijaz Khwaja, Erzo F.P. Luttmer, and Kelly Shue NBER Working Paper No. 15242 August 2009, Revised April 2013 JEL No. D53,D8,G21,L81
ABSTRACTThe recent banking crisis highlights the challenges faced in credit intermediation. New online peer-to-peer lending markets offer opportunities to examine lending models that primarily cater to small borrowers and that generate more types of information on which to screen. This paper evaluates screening in a peer-to-peer market where lenders observe both standard financial information and soft, or nonstandard, information about borrower quality. Our methodology takes advantage of the fact that while lenders do not observe a borrower’s exact credit score, we do. We find that lenders are able to predict default with 45% greater accuracy than what is achievable based on just the borrower’scredit score, the traditional measure of creditworthiness used by banks. We further find that lenders effectively use nonstandard or soft information and that such information is relatively more important when screening borrowers of lower credit quality. In addition to estimating the overall inference of creditworthiness, we also find that lenders infer a third of the variation in the dimension of creditworthiness that is captured by the credit score. This credit-score inference relies primarily upon standard hard information, but still draws relatively more from softer or less standard information when screening lower-quality borrowers.
Our results highlight the importance of screening mechanisms that rely on soft information, especially in settings targeted at smaller borrowers.
Rajkamal Iyer Erzo F.P. Luttmer MIT 6106 Rockefeller Center, Room 305 riyer@MIT.EDU Department of Economics Dartmouth College Asim Ijaz Khwaja Hanover, NH 03755 Kennedy School of Government and NBER Harvard University Erzo.FP.Luttmer@Dartmouth.Edu 79 JFK Street Cambridge, MA 02138 Kelly Shue and NBER University of Chicago email@example.com Booth School of Business 5807 S. Woodlawn Avenue Chicago, IL 60637 firstname.lastname@example.org I. Introduction An important function of credit markets is to screen borrowers and allocate credit efficiently. Traditionally, the credit score provided by rating agencies has been the main tool banks use to screen smaller borrowers. These scores are compiled using sophisticated models based on the payment history of the borrower along with other verified information, such as the number of credit lines and the outstanding balances. However, the recent banking crisis highlights some of the challenges of traditional credit screening in predicting default. In addition, one of the difficulties faced in allocating credit to smaller borrowers is that the credit score is primarily based on historical repayment history, and is therefore very susceptible to small shocks to borrowers’ financial conditions. This often creates difficulties for smaller borrowers in accessing credit.
A variety of new lending models offer potentially valuable insights on how to best address shortcomings of traditional screening mechanisms. Peer-to-peer online lending platforms provide a non-hierarchical, market-based mechanism that facilitates screening by aggregating information on borrower creditworthiness over multiple individual lenders. One distinguishing feature of these markets is the availability of nonstandard or “soft” information, which may provide valuable information about borrower creditworthiness (Petersen, 2004; Berger et al., 2002). In this paper, we evaluate how these market-based screening mechanisms compare to screening based on credit score and traditional methods. Is the screening by these markets better than the screening achievable based on just the credit score? Does the market screen high-quality borrowers more effectively than low-quality ones? How important is the role of soft/nonstandard information in screening borrowers, and does this role depend on borrower quality?
This paper uses a setting that provides an ideal laboratory to quantify collective inference by individual, non-expert, market participants. It also allows us to estimate the extent and nature of inference arising from different sources of information, such as hard and soft information. The setting used is that of an online peer-to-peer lending market, Prosper.com, where borrowers post loan listings and where multiple individual lenders bid to fund a portion of the loan at a desired interest rate. Lenders have access to standard hard financial information commonly used by banks, such as the borrower’s income and number of past delinquencies. In addition, lenders can view nonstandard information, such as the maximum interest rate the borrower is willing to pay as well as softer and less quantifiable information, such as the borrower’s picture and a textual description of his/her reasons for the loan application. The interest rate for a funded loan is determined through 1 sequential bidding and reflects the lenders’ collective perception of the quality and, hence, the creditworthiness of the borrower.1 We exploit a unique feature of the Prosper marketplace in our proprietary data: while lenders only see the borrower’s aggregate credit category, we as econometricians observe a borrower’s exact credit score – a much finer measure of the borrower’s underlying creditworthiness. We first examine the efficiency of screening in these markets by comparing the power of the interest rate (proxying lenders’ inference) set by market participants in predicting default against the default predictability obtained by using the exact credit score of the borrower. In theory, the credit score should be the best available aggregator of the standard financial variables in terms of predicting default because it is based upon a sophisticated prediction model, estimated using the same (and possibly more extensive) type of standard hard financial data. However, individual Prosper lenders may be able to improve upon the predictive power of the credit score because they make use of nonstandard and softer borrower information in addition to standard financial variables. At the same time, the interest rate set by lenders could be less predictive of default than the credit score because lenders tend to be non-expert individuals, may be driven by their personal biases, and lack access to a larger pool of data on which a credit score is based.
We find that the lenders in these markets are able to substantially outperform the credit score in terms of predicting default. We first show that the market interest rate on loans explains more variation in ex-post default than the credit score can explain. We then present a more formal comparison using tools from signal detection that are common in credit scoring. Specifically, we construct “Receiver Operator Curves” (ROCs) and show that the “area under the curve” (AUC) - a simple metric used to judge the screening power of a screening score – is both large in an absolute sense and also significantly higher for the market interest rate than for the borrower’s exact credit score. In particular, the interest rate set by lenders predicts default 45% more accurately than the borrower’s credit score. In addition, we compare lender inference to the best possible predictor that an econometrician could construct using all available codeable data (hard and soft). We find that the market interest rate (our sufficient statistic for lenders’ inference) also screens favourably relative to this more demanding benchmark as it exceeds 80% of the AUC of the econometrician.
To address a possible concern that Prosper lenders directly use a noisy proxy for the exact credit score in forming their inference (lenders observe seven aggregate credit score categories), we 1 The loan is funded only if the total amount bid equals or exceeds the amount requested by the borrower, and the final interest rate is determined by the highest reservation interest rate among the set of lenders that bid successfully.
2 examine AUC curves corresponding to interest rates set within each observed credit category. We find that even within credit categories, the AUC for the interest rate remains high and significantly outperforms the predictive power of the credit score.
Next, we examine whether the extent of lender inference differs based on borrower quality.
We find that inference is greater in the higher credit categories (better borrowers) than in the lower ones. However, the interest rate set by lenders is a better predictor of default than the credit score across all credit categories. We also explore how lenders weight standard financial versus nonstandard/soft information in forming their predictions of default. We find that both sources of information are important in screening, but that inference from soft/nonstandard information is relatively more important when assessing worse borrowers.
To the extent that a higher interest rate leads directly to increased default (as in Stiglitz and Weiss, 1981), our results may not only reflect lender inference but also reverse causality. We test whether the interest rate has a causal effect on default using credit-category borders as instruments for exogenous changes in the interest rate. The intuition for the instrument is that, at the exogenously defined borders, there is a sharp jump in interest rates even though borrower quality is continuous. We do not find any evidence of reverse causality. Similarly, the interpretation of our estimates as lender inference would be threatened if lenders directly learn borrowers’ exact credit scores from self-reported borrower information in the listing text or through public and private communication via Prosper’s “questions-and-answers” feature. While this channel is unlikely because Prosper strongly discourages borrowers from revealing detailed personal information and a text search through all listing text does not reveal any self-reported credit scores, we further examined this possible channel by restricting our sample to the period before the introduction of the question-and-answer feature, and find similar results. Our results also hold under other sample restrictions/splits that account for periods where Prosper introduced policy/information changes.
The results above highlight the ability of lenders to infer borrower creditworthiness along dimensions not captured by the credit score. Inference beyond the credit score is important because the credit score is primarily based upon hard information (e.g., past repayment history) and will miss other predictors of borrower quality. However, we are also interested in how well lenders can infer the information content of the credit score itself. In the remainder of the paper, we present a complementary analysis of how well lenders infer creditworthiness along the dimension that is directly captured by the credit score. An advantage of doing so is that it allows us to develop a 3 methodology to obtain precise magnitudes of inference arising from different sources of information.
We find that, within a given credit category (spanning 40 points in the credit score), lenders are able to infer a third of the difference in creditworthiness that is captured by a borrower’s exact credit score. This effect is economically significant because such a degree of inference allows lenders to offer a rate that is 140 basis points lower for borrowers at the top of a typical credit category than for borrowers at the bottom of that category. Given that the credit score is computed based on proprietary formulas developed by credit bureaus and not all variables that go into the computation are available to lenders, it is by no means obvious that lenders can piece together the information provided in the listing and infer a third of the true credit score.2 However, we estimate that lenders infer as much as 69% of what they could have potentially extracted from the information provided on the Prosper website.