Rules versus Discretion in Loan Rate Setting

Transcription

1 Rules versus Dscreton n Loan Rate Settng Geraldo Cerquero * CentER - Tlburg Unversty Department of Fnance PO Box 90153, NL 5000 LE Tlburg, The Netherlands Telephone: , Fax: E-mal: g.m.cerquero@uvt.nl Hans Degryse CentER - Tlburg Unversty, TILEC, K.U. Leuven and CESfo Department of Fnance PO Box 90153, NL 5000 LE Tlburg, The Netherlands Telephone: , Fax: E-mal: h.degryse@uvt.nl Steven Ongena CentER - Tlburg Unversty and CEPR Department of Fnance PO Box 90153, NL 5000 LE Tlburg, The Netherlands Telephone: , Fax: E-mal: steven.ongena@uvt.nl Frst Draft: January 27 th, 2007 * Correspondng author. We are grateful to Pedro Bom, Carlos Lourenço, Vorel Roscovan and Bas Werker, as well as semnar partcpants at the NAKE research day 2006 (Amsterdam) for many valuable comments and suggestons. The authors gratefully acknowledge fnancal support from NWO-The Netherlands and FWO-Flanders. Hans Degryse holds the TILEC AFM Char on Fnancal Regulaton.

2 Rules versus Dscreton n Loan Rate Settng Abstract We propose a heteroscedastc lnear regresson model to dentfy the determnants of loan rate dsperson. We nterpret unexplaned devatons as evdence of the banks dscretonary use of nformaton n the loan rate settng process. Dscreton n the loan-prcng process s most mportant, we fnd, f: () loans are small and uncollateralzed; () frms are small, rsky and dffcult to montor; () frms owners are older, (v) banks have weak tes wth ther borrowers; and, (v) the concentraton n the bankng market s hgh. Our results are not sample-specfc or drven by bank heterogenety n loan-prcng technology. Keywords: fnancal ntermedaton, loan rates, prce dscrmnaton, varance analyss. JEL classfcaton: G21, L11 2

3 1. Introducton Technologcal progress has shaped the evoluton of the bankng ndustry over the last few decades. In partcular, new nformaton and communcaton technologes asssted banks n processng and summarzng nformaton n credt scores and n prcng credt wth great statstcal accuracy. Despte ntal concerns that the prolferaton of cost-effectve, automated decson mechansms mght blunt the banks ncentves to engage n relatonshp lendng, recent studes have shown that the adopton of credt scorng models has actually ncreased the banks portfolo shares of small busness loans (Frame, Srnvasan and Woosley (2001)). Consequently, Berger, Frame and Mller (2005) argue that the new technologes complement, rather than substtute, the relatonshp hand-shake technology and that ther adopton by banks may even trgger a net ncrease n lendng to rsky, margnal borrowers. Regardless of the prolferaton of credt scorng technologes based solely on quanttatve nformaton (Akhaven, Frame and Whte (2005)), banks nternal ratng systems are stll generally based on both quanttatve and qualtatve factors. Bankers often rely on ther experence and dstrust the sole use of fnancal nformaton. Consequently, the fnal decsons concernng credt approvals and loan terms are then based on many dfferent attrbutes, from whch the experence and the judgment of the credt staff contnue to play a sgnfcant role (Crouchy, Gala and Mark (2001)). In ths paper, we study how partcular loan, frm, relatonshp and market characterstcs affect a bank s propensty to rely on statstcal methods ( rules ) or subjectve judgments ( dscreton ) n the loan rate settng process. In partcular, we emprcally examne how these relevant characterstcs relate to the explaned as well as unexplaned varance of a lnear loanprcng model. Startng from the trval observaton that loan rates are not set at random rates are rather the outcome of a complex set of nformaton the bank possesses about the borrower we nterpret the ncreasng predctve power of a loan-prcng model as evdence of the greater mportance of rules n the loan rate settng process. Larger unexplaned varance, on the other hand, s then assocated wth the prevalence of dscreton. In a world wthout asymmetres of nformaton and market mperfectons, there should be no room for dscreton. However, n the real world banks do engage n dscretonary loan-prcng (Guttentag (2003)). The extent to what loan rates reflect the prevalence of rules or dscreton depends on partcular macro and mcro determnants. At the macro level, a bank s ablty to engage n dscretonary prcng essentally depends on the severty of nformaton asymmetres frms face n the credt markets. Other reasons nvolve nformatonal search costs (Salop and

4 Stgltz (1977)), mperfectly compettve credt market structures, and the exstence of regulatory constrants such as fxed or capped loan rates. 1 These market mperfectons determne the barganng power banks have vs-à-vs frms, and hence set the boundares wthn whch banks engage n dscretonary loan-prcng practces. Recent organzaton theory has emphaszed the mportance of bank herarchy for the nature and success of the lendng technology beng employed (Berger and Udell (2002), Sten (2002)). Ths lterature suggests that decentralzed, small banks have a comparatve advantage n small busness lendng, an actvty that s often vewed as dosyncratc and relatonshp-based. In practce, of course, loan offcers may act ndependently of the formal herarchcal structure and have the lattude and ncentves towards dscretonary loan-prcng. The desgn of the compensaton scheme frequently renforces these ncentves (Godbllon-Camus and Godlewsk (2005)). 2 Our study s the frst to emprcally analyze the determnants of dscretonary loan-prcng. Loan rate dsperson tself has been wdely documented before, 3 but no study so far (to the best of our knowledge) has dentfed the actual sources of ths documented dsperson. The study closest to ours s that by Berger, Frame and Mller (2005), who analyze the effects of small busness credt scorng on the avalablty, prce, and rsk of credt. They test for dfferences between banks that support ther decsons usng credt scores and those banks that use the scores only as a dscretonary nput n ther decsons. Ther results suggest that rules-banks employ credt scorng as a substtute for exstng lendng technology to reduce costs. In contrast, dscretonbanks use credt scorng as a complement to exstng lendng technology to mprove accuracy. Nevertheless, ther study s slent on the factors that shape a bank s propensty to rely on ether rules or dscreton. Our paper contrbutes to ths lterature by ntroducng varance analyss to endogenously dentfy the determnants of dscretonary behavor n the loan-prcng process. As a result, recent 1 Degryse and Ongena (2005a) for example revew the sources of bank rents. 2 In partcular, barganng sklls are recognzed as a key determnant of the contracted loan prce. Guttentag (2003) for example emphaszes the relevance of these sklls for the charged rates n the U.S. mortgage market: If the loan offcer tabs you as unknowledgeable and tmd, you wll probably pay an "overage" -- a prce above the prce lsted on the loan offcer's prce sheet. The lender and the loan offcer usually share overages. If you are smart and forceful, on the other hand, you mght get an underage - a prce below the lsted prce. There are many more overages than underages. 3 Heffernan (2002) fnds that the margn between the hghest and lowest lendng rates for U.K. mortgages s relatvely small (0.45 percentage ponts), compared wth the market for personal loans, where there s a range of 8.17 percentage ponts. Hassnk and Van Leuvenstejn (2003) fnd that lendng rates n the Dutch mortgage market are hghly dspersed across lenders, even after controllng for borrowers characterstcs and regons (about 1 percentage pont). Degryse and Ongena (2005) analyze data from a large Belgan bank and report substantal varaton n loan rates at the branch level. Other papers analyze how racal dscrmnaton may determne loan-prcng n the U.S. mortgage market (see Courchane and Nckerson (1997), and Black, Boehm and DeGennaro (2001 and 2003)). These papers dsentangle market drven dfferentals (credt rsk and loan characterstcs) from arbtrary dfferentals n loan prces, and subsequently dentfy the potental effects of race on ether the magntude of overages or the lkelhood of beng charged an overage. 4

5 theoretcal developments n the fnancal ntermedaton lterature can be accommodated wthn the same framework. A partcular dstnctve feature of our study s that we concentrate on the nformaton embedded n the contracted loan rates, rather than on the loan approval decson. Both the credt approval and the loan-prcng stages rely, to some extent, on the outcomes of credt scorng models. However, loan offcers are lkely to have more dscreton n the prcng decson (Berger, Frame and Mller (2005)). Our results show that banks prce larger loans and loans granted to larger frms accordng to more objectve crtera, or rules. We also fnd evdence that banks explot for dscretonary purposes avalable publc nformaton about frms, whch we measure by the age of the frm s owner. In contrast, banks seem to avod prce dscrmnaton based on prvate nformaton. We nterpret ths result as a defensve strategy undertaken by banks to sheld ther nformatonal advantage aganst potental compettors. We employ several alternatve proxes for the rsk of the borrower and fnd that they are unlaterally assocated wth a larger unexplaned dsperson of loan rates. We nterpret ths result n lne wth the vew that banks ncentves to create nformatonal monopoles depend on the severty of nformaton asymmetres faced by the borrowers. We further model the condtonal probablty of a frm payng abnormally hgh or low loan rates and fnd that rsker and dstantly located frms are more lkely to pay abnormally low rates, n lne wth Agarwal and Hauswald (2006). The conclusons we draw from the varance analyss are present n the three dfferent datasets we nvestgate. Our prmary dataset s the 1993 Natonal Survey of Small Busness Fnances (NSSBF). We further analyze a dataset provded by an mportant Belgan bank that contans relatonshp and geographcal nformaton. We fnally merge the 1993, the 1998 and the 2003 (N)SSBF nto a pseudo-panel. We fnd no major dspartes n results between the samples, suggestng that our results are not sample-specfc, tme-specfc, nor drven by bank heterogenety n loan-prcng technology. Our panel further seems to ndcate that the use of dscreton per se dd not change substantally durng the last ffteen years, but that shfts n market condtons trggered by the development n nformaton technology explan much of the ncrease n loan rate dsperson. The rest of the paper proceeds as follows. Secton 2 provdes the theoretcal background, whle Secton 3 presents and motvates the methodology. Secton 4 dscusses the prmary dataset and Secton 5 the results. Secton 6 provdes the robustness tests. Secton 7 tests for temporal changes n our results and Secton 8 concludes. 5

6 2. Theoretcal Background There s substantal theoretcal and emprcal lterature addressng how characterstcs of the borrower, the nature of the frm-bank relatonshp and the compettve envronment affect loan rates. In contrast, very lttle research has focused on the banks ncentves to engage n dscretonary prcng. In partcular, explct predctons concernng the determnants of the dsperson of loan rates seem lackng. To overcome ths apparent gap, we resort to nferrng what one could call latent predctons from the extant bankng theory. We beleve that the nature of the problem studed eases ths task. After all, prce dscreton s justfed by the exstence of market mperfectons of the sort the lterature has extensvely analyzed. We organze the theoretcal predctons concernng the determnants of loan rate dsperson nto fve subsectons: frm rsk, frm sze, the nature of the lender-borrower relatonshp, bank organzaton and the structure of the bankng market. The unavodable overlap n content between the dfferent subsectons notwthstandng, the followed categorzaton substantally smplfes the nterpretaton of the results n our subsequent emprcal work Frm Rsk The last decade wtnessed an exponental ncrease n the adopton by banks of credt scorng models to evaluate loan applcatons, especally commercal credts (Akhaven, Frame and Whte (2005)). Scorng models permt banks to determne the lkelhood applcants wll default on ther repayments, employng nformaton avalable at the tme of the applcaton. A common practce n the credt ndustry s to classfy applcants n lne wth the score obtaned nto one of three classes of rsk (good, bad and ndetermnate rsks), and to seek further nformaton on the last group (Hand and Henley (1997)). The nformaton collected about those frms assgned to the grey area, but that are stll granted credt, ncreases the bank s ablty to prce dscrmnate across borrowers. In fact, once that (costly) nformaton s obtaned, we can presume t wll be used throughout the entre loan applcaton process, ncludng the prcng stage. Ths two-stage process suggests that we should expect more dscreton n loan rate settng to rsker, or margnal, borrowers (Berger, Frame and Mller (2005)). In a recent study, Gan and Rddough (2006) demonstrate that banks may strategcally avod engagng n rsk-based prcng to deter entry from compettors. Because the loan prce conveys prvate nformaton about the qualty of the borrower, banks have an ncentve to charge a unform rate to hgher credt-qualty frms, and a rsk-based rate to lower credt-qualty frms. Consequently, also n ths model wll loan rate dsperson be an ncreasng functon of the frm s rskness. 6

7 Studes that focus on the possblty of ntertemporal rsk sharng between frms and banks yeld smlar predctons. For nstance, Petersen and Rajan (1995) show that, n concentrated markets, banks may optmally subsdze young (and hence opaque) frms n early stages. Later on, banks recapture ths loss by chargng hgher loan rates to older, locked-n frms. Km, Krstansen and Vale (2006) propose a theoretcal model that predcts a lfecycle pattern for nterest rate mark ups. Ths lfecycle pattern s shaped by the nteracton between the severty of nformatonal asymmetres and the degree of market competton. Ther model also mples that rsker frms face a more pronounced lfecycle pattern of nterest rates. We nterpret ths predcton as ndcatng that loan rates for rsker frms should exhbt larger unexplaned devatons. 4 In short, all referred studes clearly pont towards a postve correlaton between the rsk of the frm and the mportance of dscreton n the loan rate settng process; the drvng mechansm n all these models s the nformatonal advantage lendng banks have over compettors Frm Sze We expect rules to play a larger role than dscreton n loan rate settng to large frms, for several reasons. Frst, large frms are more lkely to have audted fnancal records, ncreasng the relablty of the numbers provded n the loan applcaton process. 5 Ths suggests that banks face an addtonal source of rsk when analyzng loan applcatons from small frms basng themselves solely on rules.e., the fact that the nformaton provded may be an mprecse sgnal of the true frm s credtworthness. Second, because less publc nformaton s avalable about small frms, they are ntrnscally more opaque than large frms. In addton, large frms are more lkely to engage multple credtors (Machauer and Weber (2000), Ongena and Smth (2000)), assuagng potental holdup problems. Consequently, banks dealng wth large clents are endogenously constraned to prce loans on the bass of objectve rules as the excessve use of dscreton may result n ether unproftable deals (loan rates are too low) or hgh rsk of losng other proftable deals (loan rates are too hgh) Borrower-Lender Relatonshps The banks man source of compettve advantage over the arm s length credt market s ther ablty to reduce nformaton asymmetres through relatonshp lendng. Banks acqure propretary 4 Petersen and Rajan (1995) and Km, Krstansen and Vale (2006) provde emprcal evdence that generally supports these predctons. Machauer and Weber (1998) analyze how nterest rate prema relate to the frms credt ratngs and fnd a postve, though non-monotone, relaton between the level of rsk and the dsperson of the charged prema. 5 Prevous studes demonstrate that the usefulness of fnancal nformaton on the predcton of payment performance of small-busness loans s a drect functon of the sze of the frm. The smaller the frm, the lower s the predctve content of the frm s fnancals (see e.g. Esenbes (1996)). 7

8 nformaton over tme through frequent and personal contacts wth ther clents and through multple transactons. As a result, banks get a more complete pcture about ther clents credtworthness and future prospects. In prncple, frms could beneft from ths relaton n terms lower loan rates (Boot and Thakor (1994)). However, the nformatonal advantage ganed over the course of the relatonshp can generate a lock-n effect, n whch the bank takes advantage of the hgh nformatonal swtchng costs faced by the frm to charge hgher loan rates (Greenbaum, Kanatas and Veneza (1989), Sharpe (1990), Rajan (1992)). A common denomnator n these apparently conflctng vews s that the bank s explotaton of propretary, soft nformaton generates devatons n loan rates that cannot be predcted by an objectve prcng model. Ths argument straghtforwardly mples a postve correlaton between the strength of the bank-frm relatonshp and the extent to whch loan rates are set on a dscretonary bass. In short, we expect rules to domnate n transactonal lendng, whereas dscreton should preval n relatonshp lendng. However, the opposte predcton s proposed by other theores. For nstance, Gan and Rddough (2006) suggest that banks have an ncentve to conceal prvate nformaton to deter entry from potental compettors. Because loan rates convey nformaton about the borrowers, banks strategcally resort to unform prcng wth respect to ther best customers Bank Organzaton Recent organzatonal theores have emphaszed the mportance of bank organzaton on the loan offcers ncentves to produce soft nformaton. Berger and Udell (2002), and Sten (2002) suggest that decentralzed, small banks have a comparatve advantage n the producton of soft nformaton. Consstent wth ths vew, Cole, Goldberg and Whte (2004) fnd that the lendng decsons of large banks are more lkely based on fnancal varables, whereas the lendng decsons of small banks are more lkely a functon of the strength of the frm-bank relatonshp. Berger et al. (2005) also test for potental dfferences n the lendng technology of small and large banks. They fnd that large banks often lend to large, physcally dstant frms wth good accountng records. These studes suggest that small banks should rely more on dscreton n the loan-prcng process. In a related paper, Godbllon-Camus and Godlewsk (2005) recognze soft nformaton as a potental source of moral hazard, by showng that loan offcers may have ncentves to manpulate soft nformaton for dscretonary purposes. Ther results pont out the relevance of other dmensons n a bank s organzatonal structure dsregarded n Berger and Udell (2002) and Sten (2002). Specfcally, the desgn of compensaton schemes as well as the nternal bank rules (that 8

9 defne, among other thngs, the prcng lattude to gve loan offcers) may provde partcularly strong ncentves towards dscretonary loan-prcng. In a recent study, Lbert and Man (2006) show that the senstvty of credt approval to hard nformaton s larger for credt applcatons handled at hgher herarchcal levels,.e. levels that are more dstant from the source of nformaton. In addton, they dentfy the frm and loan szes, as well as the rsk of the borrower as key rule varables n the determnaton of the lowest herarchcal level wthn the bank that must approve the terms of the loan. As a result, we expect rules to domnate dscreton n the loan rate settng process for the banks largest clents and loans, as well as for the rskest applcants Structure of the Bankng Market There s a developng theoretcal lterature analyzng the mpact of the structure of the bankng market on the nature of the frm-bank relatonshps. Gven the close correspondence between the rules versus dscreton and transactonal versus relatonshp lendng dyads, we nfer from these theores predctons about the expected effects of market structure on the dsperson of loan rates. A frst set of theores argues that competton and relatonshps are ncompatble (Mayer (1988), Petersen and Rajan (1995)). These theores suggest that banks only have an ncentve to engage n relatonshp lendng, and hence confer an mportant role to dscreton n ther prcng decsons, n concentrated markets. The opposte predcton s proposed n Boot and Thakor (2000). They argue that banks have an ncentve to engage n relatonshp lendng to sheld ther rents aganst compettors. Dnç (2000) and Yafeh and Yosha (2001) develop models n whch relatonshp lendng s non-monotoncally, n partcular nversely U-shaped, related to the degree of concentraton n bankng markets. Interestngly, Elsas (2005) and Degryse and Ongena (2006) fnd emprcal support for a U-shaped correspondence between competton and the lkelhood of relatonshp lendng. Gven the dversty of predctons, the assocaton between the level of market concentraton and the relatve mportance of rules or dscreton ultmately becomes an emprcal queston. It s also dffcult to buld an unequvocal argument regardng the expected mpact of the frm-bank physcal dstance on loan-prcng practces. Dstance may reduce the cost-effectveness of collectng propretary nformaton about the borrowers. Hauswald and Marquez (2005) for example develop a model where the qualty of a bank s nformaton-generaton process s a decreasng functon of the bank-borrower dstance. In addton, Sussman and Zera (1995) and Almazan (2002) start from the premse that banks montorng costs ncrease n dstance. Consequently, these models predct that when banks practce prce dscrmnaton, loan rates 9

10 decrease n level and ncrease n precson wth the lender-borrower dstance, because dstance leads to more transactonal lendng. However, there s an alternatve, rather contradctory nterpretaton. Whle reducng a bank s ncentve to collect soft nformaton, the lenderborrower dstance can also be nterpreted as a rsk measure, snce t affects the precson of all nformaton avalable to the bank (Scott (2004)). Consstent wth ths clam, DeYoung, Glennon and Ngro (2006) fnd that the borrower-lender dstance ncreases the posteror probablty of a frm defaultng on a loan. Gven the dscusson n secton 2.1, where we provded support for a postve lnk between frm rsk and dscreton, ths second group of studes suggests an ncreasng role for dscreton n frm-bank dstance. 3. Methodology In order to dentfy the determnants of the dsperson of loan rates, we employ the regresson model wth multplcatve heteroscedastcty popularzed by Harvey (1976). The heteroscedastc verson extends the lnear regresson model by also parametrzng the unexplaned varance as functon of exogenous covarates. We may thnk of the heteroscedastc regresson model as comprsng two equatons one to model the mean of the dependent varable, and one for the resdual varance. We wll refer throughout the paper to the mean equaton as the loan-prcng model, and concentrate our analyss on the varance equaton (and ts varables), whch determne the precson of the loan-prcng model. A conventonal formulaton of the heteroscedastc model s gven by: y = X ' β + u, wth the dentfyng assumptons: 2 Log σ = 'γ, Z E [ u X ] = 0, 2 Var [ u Z ] = σ = exp{ Z ' γ }, where y s the dependent varable, X a vector of explanatory varables n the mean equaton, u a dsturbance term, varables n the varance equaton. 2 σ the resdual varance, and Z a vector of explanatory Under the normalty assumpton, the condtonal dstrbuton of y s gven by: 10

11 y X ( X ' β,exp{ Z 'γ }) d, Z N As a result, we obtan Maxmum-Lkelhood estmates n the heteroscedastc regresson model by maxmzng the followng log-lkelhood functon wth respect to β and γ : 6 LogL = n log(2π ) n = 1 Z ' γ 1 2 n = 1 exp( Z ' γ ) ( y X ' β ) 2 An alternatve estmaton procedure s to estmate the parameters n the mean equaton by ordnary least squares, and to use the squared errors as raw estmates of the ndvdual varances. Then, one obtans estmates of the parameters n the varance equaton by regressng the log of squared errors on the set of covarates n the vector Z. Despte beng computatonally smpler, there s a dramatc loss of effcency n ths two-step procedure (Harvey (1976)). For that reason, we obtan our estmates va maxmum lkelhood. The nterpretaton of the parameters of nterest (γ ) s crucal n our analyss. Pck one varable n Z, say Z k, and the respectve parameter, γ k. A postve γ k ndcates that the precson of the loan-prcng model decreases n postve correlaton between the varable settng process. Z k. We nterpret ths result as evdence of a Z k and the weght of dscreton n the loan rate Next, we provde an example that should ntutvely substantate the adequacy of the methodology used n ths study. Suppose we have detaled nformaton on a sample of loan contracts, ncludng observed characterstcs of the loan and borrower. We lne up (from smallest to largest) the exstng observatons accordng to one of these characterstcs, say the loan amount; then, we perform rollng regressons of the loan rate on all avalable loan and borrower characterstcs usng a fxed-wdth wndow. Ths procedure generates a sequence of parameter estmates and measures of ft relatng to the rollng regressons. Our man nterest les n the relaton between the explanatory power (or ts recprocal, the resdual varance) of the loanprcng model and the sortng varable (the loan amount). Fgure 1 llustrates the outcome of the descrbed method, when appled to our dataset. 7 The graph plots the sequence of the resdual 6 From a theoretcal perspectve, the maxmum lkelhood estmators for the parameters n the mean and varance equatons are, n expectaton, uncorrelated (see Harvey (1976)). 7 We delay the descrpton of our data to the next secton. 11

12 standard errors obtaned n the rollng regressons aganst the loan amount (the sortng varable). Smple vsual nspecton dscloses a negatve relaton between the unexplaned varaton of loan rates and the sze of the loan. Ths result ndcates that the loan-prcng model (.e. rules ) gans predctve power as we move towards wndows contanng larger loans. Ths also mples that we should expect n the varance equaton a negatve sgn for the parameter assocated wth the loan amount varable. The methodology employed n ths example has as ts most mportant lmtaton the fact that t allows only for one explanatory varable (the sortng varable). In contrast, we model n the heteroscedastc regresson framework the resdual varance n a multvarate fashon. 4. Sample Descrpton The data used n ths study stem from the 1993 Natonal Survey of Small Busness Fnances (NSSBF), a survey conducted by the Federal Reserve Board and the Small Busness Admnstraton. The 1993 NSSBF collected data for the fscal year 1993 for a natonally representatve sample of 4,637 for proft, non-governmental, non-agrcultural busnesses wth fewer than 500 employees. The dataset provdes a detaled look at these frms ther characterstcs and ther use of credt and other fnancal servces. We are nterested, n partcular, n the frms most recent borrowng experences durng the sample perod ( ), whch nclude data on the characterstcs of the borrower and of the lender, as well as the contracted loan terms. We restrct our analyss to the sample of 1,695 frms that provded nformaton on ther most recent loans (about 36% of the total sample). We dropped 29 observatons pertanng to frms that dd not report all the requred nformaton and 39 observatons related to loans granted by non-fnancal frms. Followng Cole, Goldberg and Whte (2004) we dropped 202 potentally endogenous observatons correspondng to loans granted before 1993, the year n whch frm specfc data s collected. We ended up wth a fnal sample of 1,425 observatons. The NSSBF s deally suted for our purposes, for several reasons. Frst, the sample s focus on small frms ensures that the lendng banks have suffcent barganng power n settng loan contract terms and hence plenty of room to engage n dscretonary prcng. Second, because t s a comprehensve source of nformaton on loan contracts, the NSSBF permts us to exhaustvely control for a sgnfcant share of the varaton n loan rates that s attrbutable to objectve crtera, or rules. Ths pont s partcularly relevant snce we nterpret unexplaned devatons from our loan-prcng model as measurng the banks use of dscreton. Another advantageous attrbute of the NSSBF s that data s also avalable for the subsequently conducted surveys (SSBF 1998 and SSBF 2003). Snce a consstent defnton and a majorty of dentcal questons are used 12

13 across the three surveys, we are able to test whether our results are specfc to the perod / sample analyzed. Table 1 presents the varables used n our study, along wth defntons and descrptve statstcs. We now turn to a detaled descrpton and motvaton for each of these varables Interest Rate Varables The dependent varable s the nterest rate on the frm s most recent loan we label the Loan Rate. On the rght hand sde, we nclude the varable Prme Rate to control for changes n the underlyng cost of captal n the economy. The prme rate s the reference rate when banks grant credt to ther most credtworthy customers. Typcally, most frms pay a premum on top of the prme rate. There are, however, 71 frms n our sample wth loan rates equalng the current prme rate, and 91 frms enjoyng loan rates below the prme rate Loan Characterstcs We employ the dummy varable Floatng to control for dfferences n level between fxed and varable nterest rates. Ln(Loan Amount) s the log of the amount of the loan n dollars. 8 Ln(Loan Maturty) s the log of the loan repayment duraton and proxes for the term rsk. Collateral and Guaranty are dummy varables ndcatng whether the loan s collateralzed or guaranteed by a thrd party, respectvely. These varables represent the potental rsk faced by the lender n the loan contract. On the other hand, Collateral may also sgnal the qualty of the frm (Bester (1985), Besanko and Thakor (1987)) Frm Characterstcs Frm Owner Characterstcs The dummes Propretorshp, Partnershp S-Corporaton and Corporaton control for legal and governance aspects of the frm. We ntroduce the varable Mnorty, whch ndcates whether the frm s owned by members of a mnorty groups (Afrcan-Amercan, Asan or Natve). Our motvaton to nclude ths varable follows from the evdence that mnorty groups tend to be more adverse to lookng around for the best deals and to barganng (Black, Boehm and DeGennaro (2003)). Ln(Owner s Age) s the log of the age of the frm s owner n years as of year-end Petersen and Rajan (1994) fnd that the reputaton of the frm s owner s more mportant than that of the busness n predctng loan rates. We correct Ln(Owner s Age) by the length of the frmbank relatonshp to avod spurous correlaton between these two varables. Followng Petersen 8 We used the logarthmc transformaton n all contnuous varables, except the fnancal ratos, n order to assuage scale problems as a potental source of heteroscedastcty n our regressons. 13

14 and Rajan (1994) and Berger and Udell (1995), we nterpret the age of the frm (owner) as the amount of publc nformaton avalable about the frm Frm Accountng Informaton Ln(Assets) s the log of the book value of the frm s assets. We nclude Ln(Assets) n the model, despte beng hghly correlated wth loan sze, for these varables stand for dfferent rsks. Whle the sze of the loan symbolzes how much the bank has at stake, the frm sze measures to what extent a bank has access to relable fnancal nformaton about the frm (Esenbes (1996)). Lbert and Man (2006) fnd evdence that both frm and loan sze determne the herarchcal level wthn the bank that must approve the terms of the loan. In addton, these varables correlate wth the sze of the bank. As a result, Ln(Assets) and Ln(Loan Amount) should also capture between, as well as wthn dfferences n banks organzatonal structures (Berger and Udell (2002), Sten (2002)). We nclude several accountng fgures to ncrease the precson of our loan-prcng model. Specfcally, we nclude the followng accountng denttes for the fscal year 1993, deflated by total assets: sales, profts, cash holdngs, the total amount of loans and total debt. 10 In relaton to the resultng varables, Sales s a measure of effcency, Profts a measure of performance, Cash a measure of lqudty, and Total Loans and Leverage are conventonal fnancal rsk proxes. In addton, we nclude a varable that captures the frm s relance on trade credt (Trade Credt Use). Petersen and Rajan (1994) suggest that trade credt usage measures the extent to whch frms are ratoned by fnancal nsttutons, hence affectng a bank s percepton of the qualty of the frm Credt Hstory Bankrupt ndcates whether the frm or ts prncpal owner has declared bankruptcy wthn the past seven years, Owner (Frm) Delnquent ndcates whether the owner (frm) has been 60 or more days delnquent on personal (busness) oblgatons wthn the past 3 years, and Judgments s a dummy that equals one f any judgments have been rendered aganst the prncpal owner wthn the past 3 years. Prevous studes (e.g., Esenbes (1996), Berger, Frame and Mller (2005)) found that the hstory of the prncpal s a strong predctor of payment performance of small-busness loans. We summarze the credt hstores of the owner and frm nto the varable Clean Record. Clean Record s a dummy that equals one when all prevous varables (Bankrupt, Owner Delnquent, Frm Delnquent and Judgments) equal zero. 9 We obtan smlar results when we replace the age of the owner by the age of the frm. 10 Cash holdngs nclude the total amount of cash on hand, checkng, savngs and money market accounts, certfcates of depost, and other tme deposts. 14

15 4.4. Relatonshp Characterstcs Ln(Relaton Length) s the log of duraton n years of the relatonshp the frm has had wth the lender. There s ample evdence n the lterature that the duraton of the frm-bank relatonshp affects credt terms (see e.g. Berger and Udell (1995), and Petersen and Rajan (1994, 1995)). Man Bank ndcates whether the lender s the frm s prmary source of fnancal servces, capturng the scope of the relatonshp. Personal s a dummy varable ndcatng whether the frm s most frequent method of conductng busness wth the lender s n person, or face-to-face. The varables Ln(Relaton Length) and Man Bank measure dstnct aspects of the nature of the frm-lender relatonshp. Whle Ln(Relaton Length) represents the amount of nformaton accumulated over the course of the relatonshp, Man Bank corresponds to nformaton accumulated over dfferent transactons (products). We expect Man Bank to negatvely affect loan rates, because cross sellng ncreases the frms barganng power vs-à-vs banks ceters parbus. Boot (2000) argues that relatonshp lendng s a mult-dmensonal concept. Accordngly, we combne the prevously referred dmensons of the bank-frm relatonshp nto a varable we call the Strength of the Relatonshp. Ths newly created varable equals the product of Ln(Relaton Length), Man Bank and Personal,.e. t s the duraton of the frm-bank relatonshp condtonal on the bank beng the prmary provder of fnancal servces to the frm, and on the bank and the frm favorng a personal mode of nteracton Competton/Locaton Measures The varable Concentrated ndcates whether the Herfndahl-Hrschman Index (HHI) n the deposts market of the MSA or county where the frm s headquarters offce s located s greater than The bankng market structure should have both a drect and ndrect effect on the dstrbuton of loan rates. On the one hand, greater market concentraton ncreases the banks barganng power wth respect to the frms. On the other hand, t should also nfluence the nature of the frm-bank relatonshp (Boot and Thakor (2000)), as well as the lkelhood that banks engage n ntertemporal rsk sharng (Petersen and Rajan (1995), Km, Krstansen and Vale (2006)). The varable Ln(Dstance) s the log of the dstance between the frm s man offce and the lendng nsttuton s offce. Degryse and Ongena (2005b) demonstrate that banks engage n spatal prce dscrmnaton. Petersen and Rajan (2002) and Berger et al. (2005) document that the lender-borrower relaton becomes ncreasngly mpersonal as ther physcal dstance grows. Personal controls for the alternatve transmsson channel through whch the frm-bank dstance 11 Unfortunately, ths s the only measure of bankng market concentraton avalable n the 1993 NSSBF. 15

16 mght affect loan rates. Followng a common practce n the lterature, we also nclude the varable MSA, whch ndcates whether the frm s located n a Metropoltan Statstcal Area Other Control Varables We addtonally nclude several sets of dummy varables to control for potental heterogenety n the loan-prcng model. The referred heterogenety may relate to the purpose of the loan (Loan Types), census regon (Regons), frm ndustry (SIC Codes), year of approval (Loan Approval Year) and type of lendng nsttuton (Lender Type) Varable Selecton n the Varance equaton We employ a consderable number of varables n the mean equaton, dsregardng potental collnearty problems, for ts role s smply to predct loan rates. In the varance equaton, on the other hand, we seek to understand how varables assocated to partcular market mperfectons affect the remanng dsperson of loan rates. Although t s techncally possble to employ the same set of covarates n both equatons, the above argument should justfy per se that we perform varable selecton n the varance equaton. An addtonal reason s that collnearty ssues mght nflate our standard errors, preventng us from dentfyng (n a statstcal sense) the prncpal forces affectng the dsperson of loan rates. In fact, collnearty n the varance equaton mght be partcularly problematc n our case due to the lmted sample sze and to the large number of varables employed n the mean equaton. As a result, we carry out the followng varable selecton n the varance equaton, whch s rooted n the theoretcal dscusson n Secton 2. Ln(Loan Amount) and Collateral relate to specfc rsks of the loan contract. The theory suggests a postve correlaton between the rsk of the frm and the mportance of dscreton, whch results from the nformatonal advantage banks have over compettors. Ths nformatonal advantage, whch s greater vs-à-vs more opaque frms, may enhance a bank s ablty to prce dscrmnate, and hence generate unpredctable lfecycle patterns for nterest rates (Petersen and Rajan (1995), Km, Krstansen and Vale (2006)). We nclude the varables Mnorty, Ln(Assets) and Clean Record as proxes for a frm s ntrnsc opacty. In contrast, the varables Strength of Relatonshp and Ln(Dstance) measure how easy t s for the bank to montor the frm,.e. the frm s opacty from the bank s perspectve. We expect the varables Ln(Loan Amount) and Ln(Assets) to be strongly correlated wth bank sze. As a result, they should capture potental dfferences n organzaton structures. Sten (2002) and Berger and Udell (2002) suggest that small banks are better at collectng and handlng soft nformaton than ther larger counterparts. Because these soft elements are unobservable and 16

17 easly manpulated, credt decsons by small banks should reflect a greater weght of dscreton. The varables Ln(Owner s Age) and Strength of Relatonshp capture, respectvely, the amount of publc and prvate nformaton avalable to the bank about the frm. These varables should relate to more dscreton n the loan-prcng process, snce added nformaton sharpens a bank s ablty to prce dscrmnate. In contrast, banks may strategcally conceal ther prvate nformaton by settng loan rates merely on bass of publcly observable sgnals (Gan and Rddough (2006)). Fnally, the varables Concentrated and MSA control for dfferences n geographcal locatons and n bankng market structures. As contended before, the theory offers multple predctons concernng the expected effect of bankng market concentraton on the relatve mportance of dscreton. The channel medatng the referred effect s the nature of the bank-frm relatonshp. For nstance, Mayer (1988) suggests that a hgher level of market concentraton s consstent wth more relatonshp lendng, whle Boot and Thakor (2000) develop a theoretcal argument that supports the opposte result. Consequently, the effect of the bankng market concentraton on the mportance of dscreton s an emprcal queston. 5. Emprcal Results We estmate a heteroscedastc lnear regresson model to analyze the determnants of the dsperson of loan rates. The mean equaton specfes a lnear prcng model that extracts from loan rates that nformaton (or varaton) pertanng to rules. In the varance equaton, we dentfy the factors affectng the resdual varance of loan rates,.e. the dscreton. Whle our nterest s manly n the parameters of the varance equaton, we also report the estmates of the mean equaton. Column I of Table 2 reports the coeffcents of the mean equaton, whle Column II reports the varance equaton. The dependent varable s the nterest rate on the frm s most recent loan. In the mean equaton we employ all varables descrbed n Table 1 wth the excepton of Clean Record and Strength of Relatonshp. 12 In the varance equaton we refne the varable selecton as motvated n the prevous secton. We center the explanatory varables to render a drect nterpretaton of the constants of the model; they represent the expected loan rate (mean equaton) and the resdual varance (varance equaton) of a frm wth sample average characterstcs. 12 We use the composte varables Clean Record and Strength of Relatonshp to model the varance snce we ntend to dentfy unambguous drectons of the effects of what these varables represent (frm rsk and nature of frm-bank relatonshp, respectvely) on the resdual varance. In the mean equaton we employ ther respectve consttuents (Bankrupt, Owner Delnquent, Frm Delnquent, Judgments and Ln(Relaton Length), Man Bank and Personal). 17

18 We turn now to the dscusson of the emprcal results obtaned. We brefly summarze the results of the mean equaton and then present, n detal, those pertanng to the varance equaton Mean Equaton A frm wth average characterstcs n our sample faces an 8% loan rate. Consstent wth the results n Petersen and Rajan (1994 and 1995) we fnd that loan rates are relatvely nsenstve to changes n the cost of captal for banks. We acknowledge, nevertheless, a nearly twofold ncrease n the coeffcent of the varable Prme Rate wth respect to ther studes (from 28 to 55). Larger loans seem to beneft from lower nterest rates. Ths result probably reflects the borrowers efforts to get the best possble deals concernng large loans, as well as the dluton of contractual and operatonal fxed costs. We also observe that, even after controllng for loan and relatonshp characterstcs, larger frms tend to obtan rates that are more favorable. Ths result suggests that banks are able to extract larger rents or that they perceve hgher rsk n smaller, nformatonally opaque frms. All the effects mentoned are statstcally sgnfcant at the 1% level. Consstent wth moral hazard theores we fnd that collateralzed loans pay on average 32 bass ponts (bp) more than unsecured loans. The magntude of ths effect les below the range estmated by Brck and Pagla (2005), who employed the same dataset and proposed a set of nstruments to dentfy the mpact of Collateral on loan rates. 13 The coeffcent of Ln(Owner s Age) s postve but statstcally nsgnfcant. Thus, the data does not substantate a learnng effect the fact that frm owners accumulated experence has probably taught them to prospect the market for the best deals, as well as mproved ther negotaton sklls. Frms wth healther balance sheets and better track records seem to enjoy lower loan rates, although the majorty of the estmated coeffcents are statstcally nsgnfcant at conventonal levels. Concernng accountng nformaton, banks perceve frms wth low sales-to-assets and hgh debt-to-assets ratos as belongng to hgher rsk categores. Regardng credt hstory, we hghlght the fndng that frms whose owners have been delnquent on personal oblgatons pay a premum of about 80 bp. Ths result confrms that the hstory of the owner s a strong predctor of payment performance of small-busness loans (Esenbes (1996)). The estmate obtaned for Mnorty suggests that frms owned by mnorty groups pay a premum of 28 bp, even after controllng for several rsk characterstcs. We cannot determne, however, whether ths result s 13 The omsson of relevant covarates n ther OLS regresson can also explan the nsgnfcant effect of Collateral on Loan Rate documented n Brck and Pagla (2005). Our results are not drectly comparable to thers, for Brck and Pagla (2005) employ only lnes of credt (L/Cs). Nevertheless, we note that our results reman unaltered f we restrct our sample to L/Cs. 18

19 due to dfferences n negotaton sklls, dfferences n the wllngness to search for better loan terms, or whether t smply ndcates the practce of race-based dscrmnaton. 14 Consstent wth the theoretcal predctons n Boot and Thakor (1994) we fnd that loan rates declne over the duraton of the frm-bank relatonshp, though the effect s statstcally nsgnfcant. Contrary to our predctons, the coeffcents of the varables Man Bank and Personal are postve, though the latter s not statstcally sgnfcant. The estmates for the varables under the label Competton Measures are partcularly nosy; not one s statstcally sgnfcant at conventonal levels. We justfy the non-statstcal sgnfcance of many potentally mportant predctors of loan rates wth collnearty problems. As suggested before, the relatvely small sample sze and the large number of covarates n the mean equaton nflate consderably our standard errors. The (unadjusted) R 2 of the loan-prcng model s 24%. Ths value s substantally hgher than those obtaned n other studes that employ data from the NSSBF. 15 We attrbute the superor ft of our model to the greater number of explanatory varables employed n our regressons. The dstnctve aspect n our study s that we recognze the explanatory power of the loan-prcng model to depend on certan characterstcs of the loan contract, the borrower, and the lender. These characterstcs crcumscrbe a bank s ablty to devate from rules, and hence determne how accurately rules predct observed loan rates. In the varance equaton, we explctly nvestgate how the ft of the loan-prcng model depends on these sources of heterogenety Varance Equaton The constant measures the resdual varance of the loan-prcng model for the average frm n the sample. Its estmated value mples an average standard error of 1.55, whch corresponds to a unlateral devaton from the loan-prcng model of about 77 bp. We now nvestgate n whch drecton the varables n the varance equaton shft the predcted devatons. Postve coeffcents n the varance equaton ndcate larger unexplaned devatons. In turn, large unexplaned devatons are consstent wth the banks explotaton of market mperfectons for dscretonary purposes n the loan rate settng process. In partcular, we assocate small devatons to the prevalence of rules and large devatons to the predomnance of dscreton n the loan-prcng process. 14 The Equal Credt Opportunty Act prohbts credtors from dscrmnatng n any aspect of a credt transacton because of an applcant s race. However, prevous studes have shown that unexplaned dfferences n loan rates between Afrcan Amercan- and whte male-owned frms persst (Cavalluzzo and Cavalluzzo (1998) and Cavalluzzo, Cavalluzzo and Wolken (2002)). 15 For nstance, Petersen and Rajan (1994) employed 1,389 observatons from the NSSBF 1988 and obtaned a R 2 of 14.5%. Berger and Udell (1995) and Brck and Pagla (2005) employ a sample of lnes of credt and obtan R 2 of, respectvely, 9.5% and 11.1%. 19

20 The coeffcent of Ln(Loan Amount) s negatve and statstcally sgnfcant at the 1% level. 16 We nterpret the fndng that banks prce larger loans upon more objectve crtera, or rules, n lght of recent organzatonal theores (Berger and Udell (2002), Sten (2002)). On the one hand, applcatons nvolvng larger loans are generally apprased by hgher herarchcal levels wthn the lendng bank, levels that soft nformaton, and hence dscreton may eventually not reach (Lbert and Man (2006)). On the other hand, loan amount typcally correlates wth the sze of the lendng bank. Consequently, the negatve coeffcent Ln(Loan Amount) also ndcates that larger banks rely more on hard nformaton, or automated decsons based on credt scores (Akhaven, Frame and Whte (2005), Frame, Srnvasan and Woosley (2001)). The postve sgn of Ln(Assets) contradcts our expectatons, whch are founded on the evdence that smaller frms are ntrnscally rsker from the bank s perspectve. 17 Ether because the nformaton provded by small frms s less relable (Esenbes (1996), subject to faster deprecaton (Chan, Greenbaum and Thakor (1986)), or due to the lack of publc nformaton avalable about these small frms, banks have ncentves to collect propretary nformaton, and afterward arbtrarly captalze on ther nformatonal monopoly. The coeffcent of Mnorty s postve and statstcally sgnfcant at the 1% level. In addton, the varables Collateral and Clean Record are negatve, the latter beng also statstcally sgnfcant at the 1% level. We pool these varables n our analyss because they all pont n the same drecton,.e. towards the noton that the dsperson of loan rates ncreases wth the rsk of the borrower. 18 We provded two arguments to explan the relaton found. Frst, banks ncentves to ntertemporally cross-subsdze frms ncrease wth the severty of nformatonal problems they face (Petersen and Rajan (1995), Km, Krstansen and Vale (2006)). Second, we already ponted out that rsker frms that obtaned credt probably underwent a more comprehensve screenng process, endowng banks more prvate nformaton about these frms. In sum, both mechansms act as catalysts n the amplfcaton of unpredcted devatons from an objectve loan-prcng model. Alternatvely, the postve relaton between frm rsk and dscreton may also reflect a bank s ncentve to dsguse the true credt qualty of hgh qualty borrowers through pool prcng. Gan and Rddough (2006) show that banks can engage n such practces to deter entry from potental compettors. The negatve estmate obtaned for Strength of Relatonshp ndcates a greater prevalence of rules when there are strong tes between the frm and the lendng bank. Ths seems to contradct 16 We turn to the economc sgnfcance of all varables n subsecton However, f we consder a categorcal measure of frm sze (.e. Corporaton) nstead of Ln(Assets), ts correspondng estmate becomes negatve and statstcally sgnfcant at the 5% level. 18 See Phelps (1972) for a formal argument n respect of the correlaton between credt qualty and race. 20