Credit Scoring and the Availability, Price, and Risk of Small Business CreditAllen N. BergerBoard of Governors of the Federal Reserve SystemWashington, DC 20551 U.S.A.Wharton Financial Institutions CenterPhiladelphia, PA 19104 [email protected]. Scott FrameFederal Reserve Bank of AtlantaAtlanta, GA 30303 [email protected] H. MillerBoard of Governors of the Federal Reserve SystemWashington, DC 20551 [email protected] 2002AbstractWe examine the economic effects of small business credit scoring (SBCS) and find that it is associatedwith expanded quantities, higher average prices, and greater risk levels for small business credits under 100,000. These findings are consistent with a net increase in lending to relatively risky “marginalborrowers” that would otherwise not receive credit, but pay relatively high prices when they are funded.We also find that: 1) bank-specific and industrywide learning curves are important; 2) SBCS effects differfor banks that adhere to “rules” versus “discretion” in using the technology; and 3) SBCS effects differfor slightly larger credits.JEL Classification Numbers: G21, G28, G34, L23Keywords: Banks, Credit Scoring, Small Business, RiskThe views expressed in this paper do not necessarily reflect those of the Federal Reserve Board, FederalReserve Bank of Atlanta, or their staffs. The authors thank Bob Avery, Chris Cornwell, Bob Eisenbeis,Jill Richardson, Greg Udell, Larry Wall, Cordell Weiss, and seminar participants at the University ofGeorgia and the 2001 Credit Scoring and Credit Control meetings in Edinburgh, Scotland for helpfulcomments and suggestions.Please address correspondence to Allen N. Berger, Mail Stop 153, Federal Reserve Board, 20th and CStreets. NW, Washington, DC 20551, call 202-452-2903, fax 202-452-5295, or email [email protected].
Credit Scoring and the Availability, Price, and Risk of Small Business CreditI.IntroductionAdvances in information processing, telecommunications, and financial engineering have beenintegral parts of the transformation of the U.S. commercial banking industry in recent years. This paperfocuses on one rapidly spreading technology that embodies these advances – small business credit scoring(SBCS). Specifically, we examine how the adoption of SBCS affects the availability, price, and risk ofsmall business credit.Our analytical framework provides an intuition for how the adoption of SBCS may change theway banks process information and make credit decisions about small business loan applicants. We showhow SBCS may act as a substitute for or complement to other lending technologies, and may result inreduced lending costs, improved accuracy in evaluating creditworthiness, or both.In addition, weindicate how the adoption of SBCS may result in either expanded or contracted availability of credit,higher or lower average prices for credit, and greater or lesser credit risk for different pools of smallbusiness loan applicants. Importantly, the framework provides a number of testable predictions for ourempirical models.We estimate the effects of SBCS on the availability, price, and risk of credit for the years 19951997 using data from a survey of large U.S. banks on whether and how they use SBCS. We measure theextent to which institutions rely more on “rules” – automated decisions for approval/rejection and pricesbased on externally produced credit scores – versus “discretion” – banks developing their own scoringmodels and using other inputs in credit decisions. The scoring information is combined with data on thesample banks’ individual small business loan prices, risk ratings, and other contract terms from theFederal Reserve’s Survey of Terms of Bank Lending. We also match these data with Call Report andSummary of Deposits data on the sample banks and other banks operating in their local markets.By way of preview, the data suggest that the adoption of credit scoring is associated withexpanded quantities, higher average prices, and greater risk for small business credits under 100,000.These findings are consistent with the hypothesis that a dominant effect of SBCS is a reduction in lendingcosts and/or improved accuracy of credit assessments that allows banks to expand credit to some
relatively risky “marginal borrowers” that would otherwise not receive credit. These borrowers tend topay relatively high prices when they are funded because of higher risks, more informational opacity, orother associated additional costs of processing their loans. Other results include that: 1) both bankspecific and industrywide learning curves are important; 2) the effects of SBCS differ according to theextent to which the bank adheres to “rules” versus “discretion” in using the technology, and 3) the effectsof SBCS differ for slightly larger credits (between 100,000 and 250,000).The findings of this research may have a number of important implications. For example, there isan on-going concern that the consolidation of the banking industry may result in less credit available tosome small businesses. Our results suggest that this effect may be offset to some degree by SBCS, atechnology that appears to allow large banks to expand their small business lending to some “marginalborrowers” that would not otherwise receive credit.The paper proceeds as follows. Section II provides background information on SBCS and therelated literature. Section III outlines our analytical framework. Section IV discusses the data sources.Section V describes the variables, gives some sample statistics, and shows our empirical models. SectionVI presents the empirical results, and Section VII concludes.II.Small Business Credit Scoring and Related LiteratureSBCS is a relatively new technology for lending to informationally opaque small businesses thatinvolves processing data about the firm and its owner using statistical methods.1 The outcome is a score,or summary statistic about the borrower’s expected future loan performance (Feldman 1997, Mester1997). Although credit scores have been used for some time in the underwriting of consumer loans, thistechnology has only recently been applied to small commercial credits, which had been thought to havenonstandardized documentation and to be too heterogeneous (Rutherford 1994/1995). However, creditanalysts ultimately determined that the personal credit history of small business owners is highlypredictive of the loan repayment prospects of the business.2 The personal information used in SBCS1See Hand and Henley (1997) for a review of the statistical methods used for consumer credit scoring. See Avery,Bostic, Calem, and Canner (2000) for a discussion of statistical issues, such as omitted variable bias, that may affectthe accuracy of credit scoring models.2Similar statistical techniques, such as discriminant analysis, are also used in lending to larger businesses but these2
models may include the owner’s monthly income, outstanding debt, financial assets, employment tenure,home ownership, and previous loan defaults or delinquencies (Mester 1997). The personal information isobtained from one or more consumer credit bureaus and may be combined with data from commercialcredit bureaus and basic business-specific data collected by the bank to enter into the prediction model. 3In the long run, the use of SBCS may help in the development of a secondary market for pools of smallbusiness debt, similar to the way in which consumer credit scoring helped in the development ofsecondary markets for consumer debt.While some large banks have developed proprietary SBCS models (e.g., Wells Fargo), most haveturned to models from outside vendors.The largest provider of external models, Fair, Isaac andCompany, introduced its first SBCS model in 1995. The model used a sample of more than 5000 smallbusiness loan applications over five years from 17 large U.S. banks designed to represent a national poolof small business loan applicants.4 Today, there are alternative external vendors that are typically also inthe commercial credit information business (e.g., Dun & Bradstreet, Experian). SBCS products also varyconsiderably in the amount and type of information they require and in the way the lender accesses them.For example, the data entered by the lender may be as little as an SIC code and checking account balanceall the way up to including extensive financial statement information on the business as well. The creditscores may be accessed through desktop software on a personal computer, via mainframe server, or overthe Internet. While these models may be designed for use for credits up to 250,000, many lenders usethem only for amounts less than 100,000. A fee is typically charged for every small business creditscore obtained, as opposed to a flat fee for the service.To our knowledge, the extant evidence on the effects of SBCS on small business credit is limitedto two studies. One study estimated that the use of SBCS increased the portfolio share of small businessloans under 100,000 by 8.4 percentage points for a sample of large commercial banking organizations in1997 (Frame, Srinivasan, and Woosley 2001). The other study found that in 1997 the use of SBCS bylarge banks increased small business loans under 100,000 in low- and moderate-income census tractsare typically not focused on the personal credit history of the business owners (Saunders 2000).3See Eisenbeis (1996) and Mester (1997) for detailed background information about the introduction of SBCS.4This model, which was constructed in cooperation with the Robert Morris Associates, was further refined in 1996using data from 25 banks.3
and that this effect was twice that for higher-income areas (Frame, Padhi, and Woosley 2001).While we employ the same SBCS survey data used in these two previous studies, this papermakes several significant extensions and contributions. First, we examine the effects of SBCS on theprice and risk of small business credits under 100,000, as well as the quantity of this credit, providing amore complete picture of how small business borrowers fare from the adoption of SBCS. Second, we testthe presence and slope of bank-specific learning curves, or the effects on quantity, price, and risk of thelength of time since an individual bank adopted SBCS. Third, we conduct our analysis over a 3-yearperiod and examine the effects of the industrywide learning curve, rather than focusing exclusively on thesingle year of 1997. Fourth, we examine how the effects of SBCS differ with the extent to which banksadhere to “rules” versus “discretion” in using the technology. Finally, we investigate how the effects ofSBCS differ for slightly larger credits between 100,000 and 250,000.Recent research is consistent with a view of SBCS as one of several recent innovations fordelivering financial services driven by technological improvements. Two studies found that SBCS ismore likely to be adopted, or adopted earlier, by larger banking organizations with more branches, but fewerseparately chartered banks – consistent with the benefits of spreading the costs of technology adoption overmore resources, and with potential organizational diseconomies of using this technology within the holdingcompany form (Akhavein, Frame, and White 2001, Frame, Srinivasan, and Woosley 2001). Studies ofother innovations in the banking industry – including Internet banking (e.g., Furst, Lang, and Nolle 2000),automated clearinghouse (e.g., Gowrisankaran and Stavins 1999), and ATMs (e.g., Hannan and McDowell1984) – in some cases had similar and in some cases had dissimilar findings in terms of the effects ofbank size and holding company affiliation on speed of adoption.Other recent research is consistent with a view of SBCS as one part of an on-going movement touse more quantitative methods in bank small business lending. Some studies found that banks increasedthe distances at which they lent to small businesses and the use of impersonal methods of contact prior tothe widespread implementation of SBCS in the mid-1990s (e.g., Kwast, Starr-McCluer, and Wolken1997, Petersen and Rajan 2002).Other research found that banks increased their lending to smallbusinesses outside their own local market in the late 1990s, after SBCS became widespread (Cyrnak andHannan 2000). One study also found that banking organizations have been increasing their control over4
distant affiliates over time, consistent with the use of more quantitative methods such as SBCS, whichmay require less monitoring of loan officers (Berger and DeYoung 2001).Finally, another studysuggested that the latter trend may be directly linked to SBCS – as its use was found to significantlyincrease the probability that a large bank will make loans in a given census tract (Frame, Padhi, andWoosley 2001).The introduction of SBCS and its economic effects are also related to the issue of technologicalchange and productivity growth in the banking industry. Studies of U.S. bank productivity growth duringthe 1990s often found either productivity declines or only very slight improvements using costproductivity or linear programming methods (e.g., Wheelock and Wilson 1999, Stiroh 2000, Berger andMester 2001). However, the latter study also found profit productivity to be increasing even while costproductivity declined, especially for large banks.This finding is consistent with the hypothesis thattechnological progress allowed banks to offer more or better services that may have raised costs, but thatcustomers were willing to pay more for these services, raising revenues by more than the cost increases.A number of technological changes are likely responsible for this finding, possibly including SBCS if itresulted in greater lending to high-cost borrowers that were charged high prices for credit that more thancovered their high costs. As shown below, our empirical results are consistent with this possibility.Finally, the issue of the effect of SBCS on credit availability is intertwined with the issue of theeffect of banking industry consolidation on credit availability. A number of studies found that largebanks tend to devote lower proportions of their assets to small business lending than smaller institutionsand that mergers and acquisitions (M&As) involving large banking organizations generally reduce smallbusiness lending substantially (e.g., Berger, Saunders, Scalise, and Udell 1998). The adoption of SBCSmay offset some or all of the reduction in relationship credit by large consolidated banks.Theconsolidating institutions or other institutions may use the SBCS technology to serve some of theborrowers that might otherwise be dropped, or use this technology to extend credit to other borrowers thatmight otherwise not receive bank funding.III.Analytical FrameworkIn this section, we outline a framework for analyzing the economic effects of the adoption of5
SBCS. We first provide some intuition about how the introduction of SBCS may change the way banksprocess information and make credit decisions. Second, we show how the adoption of SBCS may altersmall business credit availability, price, and risk.Third, we discuss the testable implications of thisframework for our empirical models.A.The Effects of the Adoption of SBCS on a Bank’s Credit ProcessOur framework begins with a bank that adopts SBCS technology. Prior to adoption, the bankused one or more incumbent lending technologies to evaluate small business credits, such as financialstatement lending (based primarily on financial statements), asset-based lending (based primarily oncollateral), or relationship lending (based primarily on “soft” or nonquantitative information about thefirm and its owner gained through contact over time).5 SBCS may substitute for one or more of theseother technologies in gathering information and making decisions for various potential borrowers. Thebank may also use SBCS as a complement to another lending technology by employing the credit scoresas additional information in implementing that technology.The adoption of SBCS may result in the discontinuation of an incumbent lending technology bythe bank. This could occur if SBCS substitutes for this technology for analyzing a substantial fraction ofthe loan applicants, perhaps leaving the incumbent technology at below minimum efficient scale. Atechnology might also be dropped for the purpose of avoiding organizational diseconomies associatedwith employing too many different technologies together. Relationship lending may be the most likelytechnology to be discontinued because it differs from SBCS and other transactions-based lendingtechnologies in the use of information (“soft” nonquantitative information versus “hard” quantitativedata). SBCS may also tend to substitute for relationship lending because both of these technologies relyheavily on the use of personal information about the firm’s owner for informationally opaque firms thatdo not have strong financial statements.We assume that on average, the adoption of SBCS decreases lending costs, increases accuracy inevaluating creditworthiness, or both.Presumably, such cost savings and/or more accurate creditevaluations are the reasons for adopting this technology and the reason that banks adopting SBCS tend tokeep this technology (i.e., it is an absorbing barrier). It seems unlikely that the adoption of SBCS would5See Berger and Udell (2002) for brief descriptions of these lending technologies.6
result in any significant, lasting increase in market power, since many other banks generally have accessto essentially the same credit scores.B.The Effects of SBCS Adoption on Small Business Credit Availability, Price, and RiskThe adoption of SBCS may significantly alter the availability and price of credit to individualsmall business loan applicants, and change the risk composition of the borrowers that receive credit.Here, we briefly outline the range of possible effects.For some small business loan applicants, the bank may continue to employ an incumbenttechnology without the use of credit scores. For these applicants, there should be little or no change in theavailability or price of credit unless the implementation of SBCS creates significant organizationalproblems or draws away significant resources. There may also be loan applicants for which SBCS is usedthat are denied credit, but would have been denied credit under the incumbent technology as well. Again,there is no significant effect on these firms in terms of price or availability of credit due to the adoption ofSBCS.We divide the remaining firms – those that may be affected by the adoption of SBCS into twocategories – 1) “nonmarginal borrowers” or firms that would be approved for credit whether or not SBCSis adopted; and 2) “marginal borrowers,” or firms that gain or lose access to credit as a result of theadoption of SBCS. Under SBCS, some “nonmarginal borrowers” may pay lower prices for credit or faceother contract terms that are less restrictive (e.g., reduced collateral requirements, larger credit lines, etc.)for at least three possible reasons. First, if the overall costs of lending to a firm (including due diligence,on-going monitoring, etc.) decrease, some of these cost savings may be passed on to the firm as lowerinterest rates.Second, if the use of SBCS improves the bank’s accuracy in evaluating thecreditworthiness of some types of firms, then the price or other terms may also become more favorablebecause of reduced moral hazard or adverse selection problems. The improved accuracy may allow thebank to eliminate some of the worst borrowers in their applicant pool (adverse selection) or to reduce theability of borrowers to take on more risks after the loan is granted (moral hazard). Third, irrespective ofany change in costs or accuracy, some firms may pay lower prices because they are evaluated as higherquality when SBCS is used. 6 That is, for some firms, the “new” information set that includes credit6That is, for some firms, the “new” information set that includes a credit score may suggest that these firms are less7
scoring information may suggest that these firms are less risky than the “old” information set.Other “nonmarginal borrowers” may pay higher prices as a result of the adoption of SBCSbecause the converse of these conditions may hold. That is, lending terms may worsen because bankcosts could increase, accuracy could decrease, or the “new” information set may suggest more risk thanthe “old” set for these borrowers.7 Although all of these effects may occur to some degree acrossdifferent banks and different borrowers, it seems most likely that there would be a net reduction inaverage prices for “nonmarginal borrowers” due to the overall improvements in lending costs and/oraccuracy discussed above.Some “marginal borrowers” may gain access to credit for essentially the same reasons givenabove. First, any cost savings from the use of SBCS may transform some loans from negative net presentvalue (NPV) investments into positive NPV investments and let the bank expand its lending reach intosome pools of firms that were previously unprofitable to serve. Second, improved accuracy may allowthe bank to expand its lending into pools of firms that were previously too informationally opaque to becreditworthy. That is, the adoption of SBCS may reduce moral hazard and/or adverse selection problemsto the point where some loans are positive NPV rather than negative NPV investments.Third,irrespective of any change in costs or accuracy, some “marginal borrowers” may gain access to creditusing SBCS because these firms are evaluated as higher quality credits. Loans to these firms may bepositive NPV investments using the “new” information, whereas these loans were negative NPVinvestments using the “old” information.The use of SBCS may cause other “marginal borrowers” that would otherwise be granted creditto lose access to credit because 1) the costs of lending increase, 2) the bank’s accuracy decreases, or 3)the “new” information set is less favorable than the “old” information set. Note that by assumptionabove, either (1) or (2) may be true, but they cannot both hold, since SBCS would not be adopted andkept if both raised costs and was less accurate.Other “marginal borrowers” may also lose access to credit because the adoption of SBCS causesthe bank to discontinue a lending technology under which these borrowers’ loans would be evaluated asrisky than the “old” information set, and the bank may respond with easier credit terms.7Note that all these types of price and contract term changes may also occur if firms are switched from oneincumbent lending technology to another because one of these technologies is discontinued.8
positive NPV investments to a technology under which these loans are negative NPV investments. Note,however, that if “marginal borrowers” lose their access to credit, there may be “external effects,” orgeneral equilibrium effects, in which other banks react and pick up some of these borrowers. Similareffects have been found for the effects of bank consolidation (e.g., Berger, Saunders, Scalise, and Udell1998). For example, if the firm is denied credit because the bank discontinued a technology under whichthe firm’s loan is a positive NPV investment, then another bank that continues to use that technology maypick up the loan.“Marginal borrowers” do not face price increases or decreases from the bank adopting SBCS,since these firms simply either gain or lose access to credit (although they may pay a higher or lowerprice at another bank). Nonetheless, the addition to or subtraction of these firms from the pool of firmsreceiving loan contracts from the adopting bank may have important effects on the average price paid forcredit at that bank. We assume that “marginal borrowers” pay higher loan prices on average than“nonmarginal borrowers” because they generally have higher credit risk, greater informational opacity,and/or are more costly to serve than other borrowers that are not at the margin of acceptance/denial.C.Testable Implications of the Framework for Our Empirical ModelsThe analytical framework provides some testable implications for the effect of SBCS on theavailability, price, and risk of credit. Our empirical models below compare the predicted quantities,prices, and risk ratings of small business credits for banks that did and did not adopt SBCS, controllingfor other exogenous differences among the banks and their markets. The arguments above suggest thatthe adoption of SBCS is likely to have mixed effects on small business loan applicants – some are likelyto face virtually no change in credit availability or terms, whereas some will likely gain or lose access tocredit, and others will likely face higher or lower prices. The credit risks of the borrowers receiving creditmay also change significantly as “marginal borrowers” are added and subtracted.The framework also suggests which of these effects are likely to dominate for different pools ofsmall businesses.Specifically, it suggests that credit prices are likely to be lower on average for“nonmarginal borrowers” at banks that have adopted SBCS relative to banks that have not adopted thistechnology. This is due to reduced bank lending costs and/or improved bank accuracy. The frameworkdoes not predict whether more “marginal borrowers” will gain versus lose access to credit, but it does9
predict that if there is a net gain (loss) of these firms, this will tend to increase (decrease) the averageprices and credit risks at the bank because of the relatively high prices paid by “marginal borrowers” andthe relatively high credit risk of these firms when they receive credit. We note the effects averaged acrossall borrowers, because we cannot identify the “nonmarginal borrowers” from the “marginal borrowers” inthe empirical application.Based on the analytical framework, an empirical finding of higher lending quantities for banksthat have adopted SBCS versus those that have not adopted this technology may be attributed to a netincrease in lending to “marginal borrowers.” That is, higher lending quantities imply that more firmsreceive credit that would otherwise be denied than firms that are denied that would otherwise be accepted.Conversely, a finding of lower lending quantities for SBCS banks may be attributed to a net decrease inlending to “marginal borrowers.” A finding of no significant change in lending quantities would implythat approximately as many “marginal borrowers” are added as are subtracted.The interpretation of any difference found in the average small business loan prices betweenbanks that have and have not adopted SBCS depends in a critical way on the finding for the effects of theadoption of SBCS on lending quantities. If lending quantities are higher on average for SBCS banks, thenan additional finding of higher average small business loan prices for banks that have adopted SBCS maybe attributed to an empirical dominance of 1) the relatively high prices paid by the net additional“marginal borrowers” over 2) the net reduction in average prices for “nonmarginal borrowers.” Thisempirical dominance would be reversed for a finding of higher quantities and lower pric es for SBCSbanks. A finding of higher quantities and no significant differences in average prices of SBCS bankswould suggest that the influences on prices of the increase in high-priced “marginal borrowers” and thereduction in average prices paid by “nonmarginal borrowers” are approximately offsetting.In contrast, if SBCS is associated with lower lending quantities, both of the major effects onprices point toward lower prices. The net reduction in high-priced “marginal borrowers” would tend toreduce the average prices paid for credit, which would complement the lower average prices paid by“nonmarginal borrowers.” A finding of lower quantities and equal or higher prices for SBCS bankswould be inconsistent with the predictions and would lead us to reject the framework.The framework also predicts that a net increase (decrease) in lending to “marginal borrowers”10
should increase (decrease) loan credit risks on average, as “marginal borrowers” are likely to havesignificantly higher credit risks on average than “nonmarginal borrowers.” However, the effect of theadoption of SBCS on the credit risks of “nonmarginal borrowers” cannot be predicted ex ante. Theaverage risks on “nonmarginal borrowers” is likely to decline if accuracy is improved and the bankreduces moral hazard and adverse selection problems significantly. Conversely, the average risk on loansto “nonmarginal borrowers” is likely to increase if accuracy suffers significantly as a result of theadoption of SBCS. The overall effects on credit risk will be the net effect of these different influences.These quantity, price, and risk outcomes may differ at various points along the industrywide orbank-specific learning curves as the banking industry as a whole or individual banks that have adoptedSBCS gain more experience with the technology. As examples, at later points on the learning curves,banks may be able to 1) reduce costs more and/or gain greater improvements in accuracy than at earlierpoints, and/or 2) apply SBCS to expand lending to more “marginal borrowers” than at earlier points. Inour empirical model, we test for these changes by estimating the effects on prices and quantities 1, 2, and3 years after the adoption of SBCS (i.e., bank-specific learning curves), and by evaluating the effects ofSBCS for the three years 1995, 1996, and 1997 (i.e., along the industrywide learning curve).We also test how the outcomes differ with the extent to which banks rely more on “rules” –making automated approval and pricing decisions based on externally produced credit scores – versusexercising more “discretion” – using other inputs in credit decisions and developing their own models.Banks that follow SBCS “rules” may be
While some large banks have developed proprietary SBCS models (e.g., Wells Fargo), most have turned to models from outside vendors. The largest provider of external models, Fair, Isaac and . scores may be accessed through desktop software on a personal computer, via mainframe server, or over the Internet. While these models may be designed .