Risk Overhang and Loan Portfolio Decisions: Small Business Loan Supply Before and During the Financial Crisis

Transcription

1 Rsk Overhang and Loan Portfolo Decsons: Small Busness Loan Supply Before and Durng the Fnancal Crss Robert DeYoung, Unversty of Kansas* Anne Gron, NERA Economc Consultng Gokhan Torna, Unversty of Kansas Andrew Wnton, Unversty of Mnnesota Ths draft: August 1, 2012 Short Abstract: We buld a model of bank lendng wth captal market mperfectons (asset llqudty, costly equty captal) and rsk-averse decson-makers, and estmate the model usng data from portfolo-lendng U.S. commercal banks between 1990 and The evdence s consstent wth our theory. Pror to the fnancal crss, new busness lendng declned wth the llqudty of pre-exstng loans (loan overhang) and ncreased wth stores of equty captal (rsk averson). As predcted, these phenomena grew stronger durng the crss as asset llqudty, credt rsk, and captal costs all ncreased. We also fnd evdence of busness loan credt ratonng durng the crss. Portons of ths paper are based on an unpublshed manuscrpt ttled Rsk Overhang and Loan Portfolo Decsons (DeYoung, Gron and Wnton 2006). The opnons expressed n ths paper do not necessarly reflect the vews of NERA Economc Consultng. We thank Allen Berger, Lamont Black, Paolo Fulgher, Ted Juhl, Greg Udell and semnar partcpants at Bangor Unversty, the Bank of Canada, the Federal Depost Insurance Corporaton, the Federal Reserve Bank of Chcago, the Unversty of Gronngen, the Unversty of Kansas and the Unversty of Lmoges for ther nsghtful comments and suggestons. * Correspondng author: Robert DeYoung, Unversty of Kansas School of Busness, 1300 Sunnysde Avenue, Lawrence, KS 66045, rdeyoung@kuk.edu

2 1. Introducton Corporate fnance theory suggests that when external fnance s costly, value-maxmzng frms make nvestment decsons n a rsk-averse manner: they base these decsons not only on the expected returns from the nvestment opportunty n queston, but also on ther avalable captal and that nvestment s return covarance wth the rest of ther busness. Such behavor ncreases a frm s expected profts by reducng the probablty that the frm wll forego a valuable future nvestment opportunty due to a lack of nternal captal when the return on the nvestment does not justfy the costs of rasng addtonal external captal. In ths paper, we adapt ths stran of fnance theory to descrbe the nvestment decsons of bank lenders, and then we test whether the predctons of the theory are consstent wth actual busness lendng behavor of U.S. commercal banks. Our emprcal nvestgaton focuses on small commercal banks, whch face the types of captal market mperfectons assumed n the theory: ther portfolos of llqud, nformatonally opaque small busness loans lock-up nternal equty captal and make external equty captal fnance expensve. Thus, n addton to provdng a strong test of corporate fnance theory, we are able to test whether the nvestment behavors descrbed by that theory caused a reducton and/or ratonng of credt to U.S. small busnesses durng the global fnancal crss. Commercal banks are an attractve subject for ths research queston for several reasons. Because banks act as delegated montors, they have prvate nformaton about ther loans that can lead to lemons problems f they attempt to sell the loans. Ths llqudty leads to testable predctons: f old, llqud loans create rsk exposures that cannot be cheaply sold off, then banks should take these exposures nto account when decdng what addtonal loans should be made. The degree of preexstng rsk exposure whch we refer to here as rsk overhang or loan 1

3 overhang wll vary across dfferent types of loans due to dfferences n ther lqudty, and also across tme as economc crcumstances vary. For example, durng economc downturns both the rsk and llqudty of loans should ncrease, and ths wll amplfy any rsk overhang effects assocated wth preexstng portfolo loans. The recent fnancal crss s a natural envronment for testng these phenomena, a perod durng whch fnancal markets were llqud, credt rsk was hgh, and bank equty captal was scarce and hence expensve. To examne these conjectures, we derve a theory model of loan supply and then test the model s predctons usng panel data from the fnancal statements of U.S. commercal banks. Our model, whch s based on Froot and Sten (1998), generates emprcally tractable predctons about the effects of loan overhang, expected loan returns, and competng lendng opportuntes on banks supply of new loans, gven banks current portfolo composton and captal. We test these loan supply predctons for banks wth assets less than $2 bllon (2010 dollars) operatng n metropoltan and urban markets between 1990 and These banks make loans n three man sectors (busness, real estate and consumer) that dffer n terms of loan lqudty, credt rsk, and performance co-varaton, and we use two-stage least squares estmaton to control for the smultanety of banks lendng decsons across these dfferent sectors. We pay specal attenton to the crss perod data years n whch macro-economc condtons were far more severe than the mld downturns of the early 1990s and early 2000s to test whether the busness loan supply predctons of our model are pro-cyclcal. Small, so-called communty banks are a crtcal source of fundng for small busnesses, chefly due to the nformatonal advantages nherent n the local focus of these banks. 1 Because 1 Hstorcally, small banks have generated dsproportonate amounts of small busness loans and have tended to use relatonshp lendng technques to generate ths credt (Petersen and Rajan 1994, Berger, Saunders, Scalse, and Udell 1997, Berger, Mller, Petersen, Rajan and Sten 2005). More recently, large banks have begun to provde 2

4 of ths, small banks can be unquely mportant for macro-economc growth: Berger, Hasan, and Klapper (2004) fnd a postve emprcal lnk between a large, healthy small bankng sector and macro-economc growth across 49 developed and developng natons. Crucally, data from small banks provde an especally clean test of our theory. Because communty banks lend to small, prvately-held frms that are opaque to publc captal markets, the rsk overhang assocated wth ther loans can be substantal, as these loans have no resale market and can be large relatve to the overall loan portfolo. Most communty banks lack access to publc fundng markets; ths ncreases ther cost of external fnancng, whch n turn magnfes the consequences of all new lendng decsons. These banks are unlkely to use credt dervatves (CDS do not exst for small busness loans, and usng exstng CDS to hedge these loans would ental extreme bass rsk), so they must manage the rsk of ther loan portfolos by adjustng on-balance sheet loan concentratons. And because bank managers are often placng ther own or ther famly s captal at rsk when makng lendng decsons (communty banks are often owner-managed), rsk-averse lendng behavor should be relatvely free of potentally confoundng prncpal-agent effects (DeYoung, Spong and Sullvan 2001; Spong and Sullvan 2007). Overall, our results ndcate these banks make small busness loan supply decsons consstent wth rsk-averse value-maxmzng behavor. Frst, we fnd strong evdence of loan overhang effects. All else equal, banks make fewer new busness loans when ther portfolos contan large amounts of preexstng busness loans, and make more new busness loans when ther portfolos contan large amounts of loans wth expected returns that co-vary negatvely wth busness loans. Preexstng stocks of busness loans (whch tend to be less lqud and more default-prone than consumer or real estate loans) generate the strongest rsk overhang effects. In ncreasng amounts of small busness loans (Berger, Rosen, and Udell 2001, Petersen and Rajan 2002) although the cyclcal mplcatons of ths for small busness loan supply are not yet fully understood. 3

5 general, loan overhang effects grew stronger durng the fnancal crss, consstent wth reductons n loan lqudty and lender rsk tolerance durng a downturn. Second, our results confrm that we are estmatng a supply relatonshp. Throughout most of our twenty year sample perod, new busness lendng ncreased wth the expected return from these loans. However, we are unable to fnd ths postve relatonshp durng the fnancal crss, whch mples that new busness loan supply grew nelastc durng the crss and suggests that banks may have been ratonng credt to small busnesses. Thrd, the sze of equty captal cushons nfluences both new loan supply decsons and the degree to whch loan overhang nfluences these decsons. Durng normal tmes, reductons n bank captal ratos are assocated wth a shft n new loan supply away from busness loans and toward loan categores wth hstorcally lower levels of default rsk, a pattern that mples ncreased rsk averson n lendng by banks. Consstent wth ths, we fnd that reductons n captal are also assocated wth stronger loan overhang effects. Durng the crss perod, however, the mpled lnk between equty captal and rsk-averse lendng behavor dsappears for low-captal banks, consstent wth the long establshed lterature (e.g., Merton 1977, Marcus 1984) on rsk-seekng behavor at poorly captalzed banks. That these fndngs are generated usng data from U.S. banks s especally mportant. In the U.S., about two-ffths of all small busnesses obtan some form of credt from a commercal bank and, as prevously noted, these loans come dsproportonately from small banks. 2 But the lon s share of extant knowledge on the behavor of small busness lenders durng economc downturns comes from European markets where loan-level data are more plentful. Whle our 2 Based on data from the Federal Reserve Survey of Consumer Fnances reported n DeYoung, Hunter, and Udell (2004), and the Federal Reserve Survey of Small Busness Fnances reported n Btler, Robb, and Wolken (2001), respectvely. 4

6 results confrm the central fndng of those studes.e., that supply-sde phenomena are the predomnant drvers of reduced small busness lendng actvty durng recessons we also extend ths body of knowledge by emprcally dentfyng the meta-drvers of small busness lender behavor (asset llqudty, rsk overhang, rsk averson) and showng that these forces vary n strength across the busness cycle. Understandng that portfolo lenders allocate ther captal n a rsk-averse fashon has mportant ramfcatons for small busnesses. In the long run, a rsk-averse lender s more lkely to be around to provde fundng and other fnancal servces, thus makng bank-borrower relatonshps possble. But some borrowers wll face tghter credt supply durng short-run perods of heghtened bank rsk averson, when lender balance sheets exhbt an unusually hgh amount of rsk overhang and/or when banks experence nternal or external pressure to ncrease ther equty captal. More generally, our fndngs are consstent wth models of pro-cyclcal bank lendng drven by nternal bank behavor (e.g., Rajan 1994; Berger and Udell 2004; Ruckes 2004). Durng an economc expanson, demand for bank credt s hgh and busness proftablty s good, resultng n proftable loans, ncreasng bank captal, and an expandng credt envronment n whch banks lend more at lower rates as they compete for busness. But as the expanson nevtably ends, busness proftablty wll declne, resultng n delnquent loan payments or outrght default, declnng bank captal, and a tghter credt envronment as banks make fewer loans at hgher rates. We fnd that rsk overhang effects are themselves pro-cyclcal at banks, workng to decrease small busness loan supply durng economc downturns by even more than would be mpled by recessonary reductons n bank captal alone. As loan securtzaton markets broke down durng the fnancal crss, banks were less able to sell ther outstandng stocks of real estate and consumer loans; ths ncrease n asset llqudty created cross-sector 5

7 loan overhang effects, neutralzng equty captal that could otherwse have been used to back new small busness lendng. As stock market declnes made ssung new equty captal more expensve, banks became more crcumspect n ther allocaton of exstng rsk captal; ths ncrease n rsk averson exacerbated the overhang effects of exstng portfolo loans, makng banks less lkely to make the margnal busness loan. Thus, our study provdes a mcro-theoretc framework that helps explan recent emprcal fndngs that small busness lendng declned n Europe durng the fnancal crss (e.g., Popov and Udell 2010; Jmenez, Ogena, Peydro and Saurna 2012; Cotugno, Monferra and Sampagnaro 2012). More to the pont, we perform a structural estmaton of our theoretcal loan supply functon usng lendng data from U.S. banks, and fnd parameter estmates that are strongly consstent wth the predctons of our theory across the busness cycle. The rest of the paper proceeds as follows. Secton 2 provdes a bref overvew of the bank lendng lterature most relevant to ths study, frst the theoretcal then the emprcal. Secton 3 presents our theory model of loan supply wth captal market mperfectons, whch lnks bank loan portfolo management to preexstng (.e., overhangng) loan stocks, expected loan proftablty, current lendng opportuntes, loan performance covarances, and effectve rsk averson. Secton 4 operatonalzes the model for emprcal estmaton and lays out our man hypotheses to be tested. Secton 5 descrbes our detaled bank-level data set and defnes the varables used n our regresson tests. Secton 6 presents the results for our basc model of busness loan supply, some extensons of the basc model, and a smultaneous verson of the basc model that also ncludes loan supply functons for real estate loans and consumer loans. Secton 7 summarzes our man fndngs and dscusses mplcatons for polcy. 6

8 2. Related lterature Our work s rooted n the theoretcal lterature that models fnancal nsttuton portfolo management when external fnancng s costly due to captal market mperfectons. These theores apply partcularly to banks wth enough equty so that moral hazard va rsk shftng does not become an ssue. 3 Froot, Scharfsten and Sten (1993) show that frms facng costly external fnance, stochastc net worth, and attractve future nvestment opportuntes wll behave n a rsk-averse manner. Froot and Sten (1998) extend ths model to nclude the nfluence of preexstng portfolos of nvestments on fnancal nsttutons new nvestment decsons. These authors show that the amount the nsttuton wll want to nvest n a new opportunty wll depend upon ts level of captal, the covarance of that nvestment s cash flows wth the cash flows of the frm s stock of llqud (or non-tradable) asset exposures, and the covarance of the non-tradable cash flows of any other new nvestments the frm s consderng. Froot (2007) extends the framework further n a model of nsurance companes, ntroducng product market mperfectons and allowng some of the rsks faced by nsurers to be hedged. Several emprcal applcatons of ths framework exst. Froot and O Connell (1997) apply ths model to prce determnaton n the catastrophe rensurance market. They show that such fnancng mperfectons can lead to costly rensurer captal and also to rensurer market power, and estmate the correspondng supply and demand curves. Gron and Wnton (2001) coned the term rsk overhang to descrbe how outstandng and llqud rsk exposure from long-term nsurance polces can affect the current supply of new nsurance polces. In extreme 3 It s well-known that banks wth very low captal levels may engage n moral hazard va rsk-shftng, possbly by overly aggressve lendng, as n Marcus (1984). Ths s more lkely f depost nsurance s prced at a flat rate. By contrast, f captal levels are not very low, banks may become more conservatve n ther lendng when captal levels fall, as n Besanko and Kanatas (1996), Thakor (1996), Holmstrom and Trole (1997), Damond and Rajan (2000) and Perott, Ratnovsk and Vlahu (2011). 7

9 cases, ncreases n rsk overhang may lead frms to reduce ther total exposure to the underlyng rsk by cancelng exstng polces. A large number of emprcal studes on bank captal and lendng nvestgate whether mplementaton of the Basel I captal requrements caused a credt crunch n the U.S. In general, these studes relate loan growth to captal measures and other controls. 4 Although ths lterature does not generate a consensus on the relatonshp between bank captal and loan supply, Sharpe (1995) dentfes two robust results across the studes: bank proftablty has a postve effect on loan growth, and loan losses have the opposte effect. Snce profts (loan losses) tend to ncrease (decrease) bank captal, these fndngs are consstent wth a postve assocaton between bank captal and loan growth. In more recent work, Beatty and Gron (2001) fnd that banks wth stronger captal growth have greater loan growth, wth the most sgnfcant effects comng from the most captal-constraned banks. The global fnancal crss has motvated a new stream of studes on bank captal and bank loan supply. Perott, Ratnovsk and Vlahu (2011) derve a non-monotonc theoretcal relatonshp between bank captal and bank rsk-takng. When banks are operatng near ther regulatory captal mnmums, addtonal captal results n fewer tal rsk projects (consstent wth a reducton n the value of the depost put opton, e.g., Merton 1977, Marcus 1984). However, when captal s so hgh that banks have no worry of breachng ther regulatory captal mnmums, addtonal captal results n more tal rsk projects. Emprcal studes by Black and Hazelwood (2011), Duchn and Sosyura (2011) and L (2011) all fnd at least some evdence of ncreased lendng (.e., greater rsk-takng) at banks that receved government captal njectons. The 4 Examples nclude Bernanke and Lown (1991), Hall (1993), Haubrch and Wachtel (1993), Berger and Udell (1994), Hancock and Wlcox (1993, 1994a, 1994b), Berger and Udell (1994), Brnkman and Horvtz (1995), and Peek and Rosengren (1995). 8

10 fndngs of these studes are mportant; however, because they focus narrowly on bank lendng behavor n response to artfcal (non-market) captal njectons durng a perod of severe fnancal stress, they provde an ncomplete treatment of the bank captal-loan supply relatonshp. Much of our current knowledge about the mpact of the fnancal crss on small busness loan markets comes from European economes, where credt regstres provde researchers wth hghly detaled data on loans and loan applcatons. Jmenez, Ogena, Peydro and Saurna (2012) fnd that reductons n busness lendng n Span durng the fnancal crss were predomnantly caused by supply-sde effects due to weak bank balance sheets, rather than demand-sde forces. Popov and Udell (2010) fnd that both supply-sde and demand-sde factors led to reduced small and medum szed enterprse (SME) lendng n 14 European countres: banks experencng stress to ther assets and equty values extended less credt, and hgh-rsk SMEs wth fewer tangble assets receved less credt, durng the early stages of the fnancal crss. Cotugno, Monferra and Sampagnaro (2012) fnd that SMEs n Italy experenced reduced credt supply durng the fnancal crss, but that credt ratonng was substantally mtgated for loan applcants wth exclusve borrowng relatonshps wth ther banks. Research on U.S. bank lendng durng ths perod tends to use data on large busness lendng. Ivashna and Scharfsten (2010a, 2010b) show that shocks to bank lqudty (e.g., depost wthdrawals, credt lne draw downs) were assocated wth reduced lendng to large corporate customers durng the crss. Montoral- Garrga and Wang (2012) derve a model of bank loan prcng wth endogenous credt ratonng, and estmate t usng a sample of U.S. bank loans durng the 2000s; the authors conclude that large busness borrowers were less lkely than small frms to be ratoned out of the bank loan market durng the fnancal crss. 9

11 Our study dffers from the prevous lterature n several respects. Frst, most prevous studes focused on large banks, chefly because regulatory captal constrants are more lkely to be bndng for large banks and because large banks produce the lon s share of the aggregate loan supply. We focus on small banks because, for the reasons we outlned above, these banks provde a more natural laboratory for testng the concepts developed n the theoretcal corporate fnance lterature on nvestment decsons when assets are llqud and external captal s costly. Second, prevous studes estmated reduced-form regresson models, whereas we estmate a structural model that ncludes other loan supply decsons. Ths framework provdes a more complete test of rsk-management practces at lendng nsttutons and the effects of rsk overhang on loan supply. Thrd, most prevous studes used annual data over a lmted perod of tme, whereas we observe detaled changes n portfolo composton and loan supply at quarterly ntervals over 20 years. Fnally, wthn the small set of studes that have tested for small busness credt ratonng durng the fnancal crss, we are one of the very few to take ths queston to U.S. data. 3. Loan Supply wth Captal Market Imperfectons: Theory In ths secton we develop a portfolo model of bank loan supply. We begn wth a representatve bank whch has lendng opportuntes n several sectors. Loans can be funded out of net nternal captal W or external funds F, where external funds are assumed to be more costly than nternal funds. Ths addtonal cost reflects nformaton asymmetres between the frm and outsde nvestors (e.g., Myers and Majluf (1984), Sten (1998), and DeMarzo and Duffe (1999)), as well as other transacton costs n accessng publc markets. In addton to current perod loans, the bank may be able to make proftable loans n future perods. As shown by Froot, Scharfsten 10

12 and Sten (1993), proftable future nvestment opportuntes combned wth costly external funds and stochastc nternal funds cause the frm's objectve functon to be ncreasng and generally concave n the stock of nternal funds. Intutvely, more nternal funds lessen the extent to whch a bank must rely on costly external funds, but ths beneft s generally decreasng because, at the margn, there are fewer proftable uses for these funds. Denotng the ndrect form of the bank's objectve functon as P(W), we have P W > 0 and P WW < 0 where the subscrpt denotes the partal dervatve. The bank begns perod t wth W t-1 n net nternal funds ( captal ), L t-1, n outstandng loans n each sector, and net external ( debt ) fnance of F t-1 = (L t-1, ) -W t-1. Wthout loss of generalty, we assume that F t-1 s postve, as s the case for most banks; we also assume that all external fnance takes the form of debt. 5 For the moment, assume that all of the bank s outstandng loans are llqud and cannot be sold due to the bank s prvate nformaton on loan qualty. Snce the bank must bear the rsk of L t-1, loans n each sector regardless of ts subsequent decsons n perod t, L t-1, s the bank s rsk overhang n sector n perod t. Durng perod t the bank can make new loans NL t, 0 to each sector, resultng n endof-perod outstandng debt of F t = (L t-1, + NL t, ) - W t-1. The gross per dollar cost of debt fundng s 1+r t, whch ncludes any costs of accessng external markets rather than usng nternal ~ captal. Durng perod t, the bank realzes the gross per dollar return of Rt, / t 1 on loans to sector ~ that were orgnated n perod t-1. Rt, / t 1 equals 1+r t +p t-1, -η ~ t,, where p t-1, s the per dollar credt 5 We make ths assumpton for smplcty alone, as t s well known that ssung new equty also nvolves sgnfcant transacton and nformatonal costs. Whle banks can ssue long-maturty, federally nsured retal deposts that are less lkely to be affected by such nformaton concerns, banks also regularly ssue costly non-depost debt nstruments such as subordnated debt, trust preferred stock, and Federal Home Loan Bank advances. Moreover, nsured deposts are not a perfect, costless substtute for unnsured debt. Bllett et al. (1998) fnd that large banks ncrease ther use of nsured deposts followng downgrades of ther publcly traded debt, but also fnd that total debt fnance (nsured plus unnsured labltes) declnes, consstent wth ncreased external costs of debt fnance. Further 11

13 spread or markup charged on sector loans that orgnated n perod t-1, and η ~ t s the random per dollar loan losses on sector loans n perod t. Smlarly, the bank realzes the gross per dollar return ~ R t, / t = 1+r t +p t, -η ~ t, on the new loans to sector orgnated n perod t, where p t, s the per dollar credt spread on these loans. For smplcty, we assume that all losses on loans to sector n perod t are perfectly correlated, regardless of when the loan was made. Current perod loan losses are assumed to be normally dstrbuted: ~ η t, ~ N( µ t,, σ t, ) where both µ t, and σ t, depend on the sector s economc outlook at the start of that perod. 6 Both µ t, and σ t, are decreasng n the sector's economc outlook: when borrowng frms have better prospects, both ex ante credt rsk and ex post realzed loan losses are lower because the borrowng frms chances of default are reduced. Gven these assumptons, t follows that the bank s net captal at the end of perod t s ~ W = t n = 1 [ L ~ R = W (1 + r ) + 0 t 1, t t, / t 1 n = 1 + NL [ L t 1, t, ~ R ( p t, / t t 1, ] F (1 + r ) t ~ η ) + NL ( p η~ )] ~ ~ where we have made use of the defntons of R t, / t 1, R t, / t, and F t. t, t t, t, t, (1) The bank chooses new loan amounts NL t, that maxmze expected proft E[P( W ~ )], gven the fnancng constrants. Ths leads to the frst order condton for each sector ~ Wt 0 = E[ P ] [ ( ~ )] [ ]( ) (, ~ W = E PW pt, ηt, = E PW pt, µ t, Cov PW ηt, ), (2) NL t, where we have made use of (1) and the dentty E(xy) = E(x)E(y) + Cov(x,y). Snce loan losses t support that external fundng s costly for banks comes from Jayaratne and Morgan (2000), who fnd that banks fnance an unusually large porton of ther assets wth nternal funds. 6 In realty, loan losses are skewed to the rght: they cannot be less than zero, there s a hgh probablty that they won t be too large, and a low probablty of very large losses (see Carey, 1998, and Wnton, 2000). The assumpton of normalty allows us to gve a smple, tractable analytc soluton to the bank s portfolo choce problem. 12

14 ~η t, and the level of nternal funds W ~ t are both normally dstrbuted, we can apply Sten s Lemma and the defnton of covarance to derve the bank s supply of new loans NL, to sector 7 S t NL S t, σ = S j NLt, j Lt 1, L j j t 1, j σ σ j 1 pt, µ t, +. (3) σ G σ where for convenence we have suppressed the tme subscrpt on the loan performance varance and covarance terms. In (3), σ s the varance of loan losses n sector over tme; σ j s the covarance of loan losses across sectors and j over tme; and E[ PWW ] G = measures the E[ P ] W bank s effectve rsk averson (and we shall refer to ts recprocal 1/G as the bank s rsk tolerance) nduced by the costs of external fnance. The bank s supply of new loans to sector s determned by several factors on the rghthand sde of equaton (3). The frst term s the effect of covarance-adjusted lendng opportuntes n other sectors j at tme t. The second term s the preexstng portfolo exposure n sector, that s, the overhang of outstandng loans n sector at tme t. The thrd term s the effect of the covarance-adjusted loan overhangs n other sectors j. The fnal term s the bank s tolerance 1/G multpled by the rsk-adjusted proft rato (p t, -µ t, )/σ. It s straghtforward to verfy that equaton (3) has the features of a supply curve. The supply of new loans to sector s ncreasng n the current credt spread (or markup ) p t, and decreasng n expected loan losses (or costs) µ,t. Assumng that p t, exceeds µ t,, new loan supply s also decreasng n the bank s effectve rsk averson G. Further, the supply of new loans to sector s decreasng n the overhang of outstandng loans n that sector, L t-1,. Fnally, f the covarance between sector and 7 Sten s lemma mples Cov(P W, ~η t, ) = E[P WW ]Cov( W ~ t, ~η t, ). We also use Cov( W ~ t, ~η t, ) = j (L t 1, j + NL t, j )σ, j. 13

15 sector j s postve, then the supply of new loans n sector s decreasng n both the overhang of outstandng loans n sector j and the supply of new loans n sector j; by contrast, f the covarance s negatve, then the supply of new loans n sector s ncreasng n loans to sector j. 4. Loan Supply wth Captal Market Imperfectons: Issues for Emprcal Specfcaton Equaton (3) forms the bass for our emprcal analyss. Before we proceed to the data and estmaton, however, we must ncorporate two features of the data that run counter to our assumptons above. The frst s that banks hold lqud as well as llqud loan stocks. The second s that we do not drectly observe new loan supply, only the change n loan stock. We then present our estmaton equaton and predcted outcomes for the regresson parameters. 4.a. Banks hold lqud and llqud loan stocks Durng a gven accountng perod, some loans wll mature and be repad. The remanng loan stocks exhbt varyng degrees of lqudty. As shown by Froot and Sten (1998), under optmal portfolo allocaton wth mperfect captal markets, t s optmal for banks to shed all loans that can be sold at far value. However, the market prces for loan sales may be below the banks expected values due to nformaton asymmetres or transacton costs of sellng loans, resultng n llqud loans whch are held rather than sold. To nclude the effects of llqud loan stocks, let δ t-1, (0,1) be the llqud porton of the outstandng loans at the begnnng of perod t (end of perod t-1). The remanng loans are assumed to be lqud and wll be sold off at no cost, or wll run off naturally (e.g., not rolled over), to make room for new loans. Snce only llqud loan stocks wll affect new lendng, we can rewrte equaton (3) as NL S t, S σ j σj 1 pt, µ t, = j NLt, j δt 1, Lt 1, j δt 1, jlt 1, j + σ σ G σ. (3') 14

16 Whle equaton (3') s predcted by theory, the avalable data do not allow us to observe the portons δ t-1, and δ t-1,j of preexstng loan stocks that are llqud. Instead, we must use the total (lqud and llqud) outstandng stocks of loans L t-1, and L t-1,j n our estmaton equatons n place of llqud outstandng loans δ t-1, L t-1, and δ t-1,j L t-1,j. Thus, although equaton (3') predcts that the coeffcent on outstandng same-sector loan stocks (δ t-1, L t-1, ) wll be exactly -1, the estmated coeffcent n our regressons wll reflect the effect of loan stock lqudty on the supply of new same-sector loans. 8 The degree to whch outstandng loans are lqud or llqud s not fxed but can change wth exogenous condtons. A recesson, or a downturn n a specfc sector, wll reduce the lqudty of outstandng loans for two reasons. Frst, borrowers are n worse shape so they are more lkely to try to roll over ther maturng loans. Second, the bank faces greater adverse selecton problems when tryng to sell or securtze these ncreasngly rsky loans. Addtonally, a recesson or a sector downturn may have a captal effect: banks wll expect ncreased future losses on outstandng loans, whch wll reduce expected equty captal and make the bank effectvely more rsk averse. We test for these effects by estmatng our model separately durng ( ) and pror to ( ) the fnancal crss. 4.b. New loans are unobservable A second concern s that new loan supply NL S s not drectly observable n the data. Instead, we use the net perod-to-perod change n the stock of loans, whch we refer to as the net lendng change, or NLC. Note that the stock of outstandng sector loans L t, at the end of perod 8 Gven ths substtuton of L t-1, for δ t-1, L t-1, n our estmatons, a strct nterpretaton of the model (3') s that the estmated coeffcent on L t-1, wll reveal the average share of outstandng sector loans that are llqud. However, we make several addtonal adjustments to our estmaton equaton (e.g., we estmate the equaton n loan shares rather than loan levels, to avod sze effects) whch nvaldate ths strct predcton. 15

17 t s the sum of three tems: the llqud porton of the perod t-1 loan stock, any retaned lqud porton of the perod t-1 loan stock, and the new perod t loans. Lettng τ t, (0,1) represent the fracton of outstandng lqud sector loans from perod t-1 that the bank retans at the end of perod t, t follows that L t, equals (δ t, + τ t, (1-δ t, ))L t-1, + NL S t,. Thus, we have NLC t, = L t, = NL = NL S t, S t, L t 1, + [ δ t, + τ (1 δ )] L t, t, t, [(1 τ )(1 δ )] L t, t 1, t 1, L t 1, whch shows that net lendng change equals the actual supply of new loans less the porton of lqud loan stocks that are actually sold. In practce, banks wll sell some lqud loans f they can do so at far prces, wll hold some lqud loans for strategc purposes, and wll hedge some of these held lqud loans f the prce of nsurance s favorable. Regardless, banks wll tend to draw down or sell off a larger porton of ther lqud outstandng loans when ther captal falls (due to ncreased rsk averson) or f the portfolo rsk assocated wth ther lqud loans ncreases (.e., ncreased correlatons wth other loans). If the share of lqud loans (1 δ )L t-1, s small relatve to the flow of new loans or f the bank retans a large porton of ts lqud loans both condtons are more characterstc of small banks than of large banks then NLC wll be hghly correlated wth new loan supply NL S. 4.c. Specfcaton Equaton (4) presents our estmaton equaton for busness loan supply, whch devates somewhat from the theoretcal equaton (3'). We use net lendng change NLC t as a proxy for the unobservable new loan supply NL S t ; the L t-1 measure total preexstng loans (not just the llqud porton δ t-1 L t-1 ); the bank rsk tolerance G -1 and rsk-adjusted loan return (p t, -µ t, )/σ measures are specfed separately (rather than multplcatvely) n order to estmate the ndependent effects of 16

18 these measures; 9 the covarance-varance ratos σ j /σ are suppressed; and the regresson coeffcents φ, β, ρ, χ and ξ are parameters to be estmated: NLC p µ φ β ρ χ ξ (4) t,1 t,1 1 t,1 = NLCt, 1Lt 1,1 Lt 1, Gt = 2,3 = 2,3 σ11 The subscrpt ndexes each of the three loan sectors n our data (busness = 1; real estate = 2; consumer = 3) and t ndexes tme. The coeffcents φ and ρ absorb the effects of the suppressed covarance-varance ratos σ j /σ whle the coeffcents β and ρ absorb the unobserved lqudty effects δ t-1 dscussed above. In our estmatons, we addtonally control for fxed bank effects, seasonal effects, and economc condtons n banks local markets. Snce banks make new busness loan supply decsons smultaneously wth new real estate and consumer loan supply decsons, the rghthand sde NLC t, terms are endogenous, and we account for ths by estmatng equaton (4) usng two-stage nstrumental varables technques. Full detals of our estmaton methods appear below. 4.d. Predcted sgns for estmated coeffcents Based on the dscusson above, we can make the followng predctons about the estmated coeffcents of equaton (4): Same-sector loan overhang: Wthn the busness loan sector, net lendng change wll be negatvely related to overhang (β 1 <0). Ths effect wll be stronger when the sector s less lqud. Cross-sector loan overhang: If the portfolo model s the prmary determnant of net lendng 9 Estmatng the model usng the combned term yelds only trval dfferences n the other coeffcents. 17

19 changes, then the mpact of cross-sector loan overhang on net lendng change (ρ j ) wll be ncreasngly negatve (or less postve) as the covarance between loan losses n sectors and j ncreases. Holdng covarance constant (and not equal to zero), the magntude of ρ j wll be larger the more llqud s loan stock j. Cross-sector net lendng change: If our model holds strctly, the estmated effect of net lendng change n sector j on net busness lendng change (φ j ) should be the same sgn as the estmated effect of sector j loan stocks on net busness lendng change (ρ j ). The coeffcents wll be exactly the same (φ j =ρ j ) only f the loan stocks and net lendng change have the same degree of lqudty and f loan losses for each have the same correlaton wth loan losses for the net busness lendng change. Rsk-adjusted loan return: Wthn the busness loan sector, net lendng change wll ncrease wth the rsk-adjusted return rato (χ>0). Effectvely, ths coeffcent captures the rskadjusted slope of the busness loan supply functon. Rsk tolerance: Wthn the busness loan sector, net lendng change wll ncrease wth the bank s rsk tolerance (ξ>0). 5. Data and Varables We estmate equaton (4) usng quarterly fnancal statement data for small U.S. commercal banks. These data are taken from the Federal Reserve s Report of Condton and Income (call reports) from the frst quarter of 1990 (1990:Q1) through the fourth quarter of 2010 (2010:Q4). We lmt the data to nclude only so-called communty banks wth less than $2 18

20 bllon n assets n real 2010 dollars. 10 We exclude banks wth less than $25 mllon n assets n current dollars because the call reports contan very lttle detal for these tny banks. We also exclude banks located n rural areas, as rural banks face a dfferent set of lendng opportuntes than urban banks, whch results n dfferent exposures to loan overhang and dfferent ncentves to deal wth ths rsk. 11 Snce our goal s to examne banks loan portfolo decsons, we only consder banks that make non-trval amounts of loans n all three major categores of loans reported n the call reports: busness loans, real estate loans, and consumer loans. We defne these non-specalst lenders each perod as follows: the dollar value of ther sector loans must be no more than ten tmes, and no less than one-tenth, of the dollar value of ether of ther sector j loans ( j). These upper and lower boundary restrctons elmnate around one-thrd of the bank-quarter observatons, and the restrctons become more bndng over tme. As shown n Fgure 1, the asset share of real estate loans for the average non-specalst bank approxmately doubled durng our sample perod before decreasng durng the fnancal crss, whle the asset share of busness loans remaned relatvely stable and the asset share of consumer loans declned by about half. 12 Over tme, as real estate loans provded a larger porton of small bank lendng, and as consumer 10 For decades, both bank regulators and bank researchers used $1 bllon as a convenent upper sze threshold to defne the U.S. communty bank sector (DeYoung, Hunter, and Udell 2004). Our $2 bllon threshold s smlarly convenent, but recognzes several decades of nflaton. 11 Rural banks typcally have local market power; wth greater rents at stake, ther ablty and wllngness to absorb rsk overhang may dffer markedly from those of urban banks. The extreme localness, or ruralness, of these banks nfluences the manner n whch they underwrte loans and results n lower levels of credt rsk (DeYoung, Glennon, Ngro and Spong 2011). Rural banks hold relatvely low levels of total loans, hgh levels of marketable securtes, and hgh levels of equty compared to smlarly szed urban banks (DeYoung, Hunter, and Udell, 2004), consstent wth a less sophstcated approach to rsk management. And because the agrcultural economy permeates the performance of all lendng sectors at rural banks (e.g., busness loans are domnated by agrcultural producton loans and loans to farm-related busness concerns, and real estate loans nclude large amounts of farm mortgages and farm resdental mortgages), the loan performance covarances wll dffer from those observed n urban markets. 12 The sum of these three loan-to-asset shares ncreases over tme. Ths mrrors the secular ncrease n total loan-toasset ratos at small U.S. banks durng the post-deregulaton era, durng whch ncreased competton and ndustry consoldaton removed neffcent banks that loaned out only a small porton of ther assets (DeYoung, Hunter and Udell 2004, Tables A1 and A2). 19

21 loan shares became less mportant at small banks, fewer banks qualfed as non-specalst lenders; hence, the number of observatons n our tests unavodably declnes over tme. We make several addtonal data adjustments to avod the effects of data errors, mergng banks, or banks wth an abrupt change n lendng strategy. We delete bank-quarter observatons when the rato of nonperformng loans to begnnng-of-perod loans, the rato of net lendng change to begnnng-of-perod assets, the quarterly change n assets, or the quarterly change n equty captal are over the 99 th percentle or below the 1 st percentle of the sample dstrbutons. Smlarly, we delete bank-quarter observatons when the expected proft varable n any of the three loan sectors s less than the 0.5 th or greater than the 99.5 th percentle of the sample dstrbuton. We also delete bank-quarter observatons n whch the assets of another bank are acqured, bank-quarters when banks are less than 5 years old, all observatons for banks that never lend out more than 25% of ther assets, and all observatons for banks that were not present n the data for at least fve consecutve quarters. Ths sample perod ncludes data from before and durng the global fnancal crss. We defne the begnnng and the end of the crss based on the self-reported small busness lendng behavor of U.S. banks, as measured by the Federal Reserve s Senor Loan Offcer Opnon Survey on Bank Lendng Practces (SLOS). The SLOS s admnstered four tmes each year to a relatvely stable set of around 55 large and medum szed U.S. commercal banks. Among other questons, the survey asks each bank whether ts credt standards for approvng small busness loan applcatons have eased, remaned unchanged, or tghtened over the past three months. Not surprsngly, banks reported that they tghtened lendng standards early n the crss, and reported that they eased lendng standards as the crss waned. The net percentage of banks tghtenng ther small busness lendng standards exceeded 10 percent for 20

22 the frst tme n the January 2008 SLOS, so we mark 2007:Q4 as the begnnng of the crss. The net percentage of banks easng ther small busness lendng standards exceeded 10 percent for the frst tme n the Aprl 2011 SLOS, so we mark 2011:Q1 as the end of the crss (.e., 2010:Q4 s the fnal quarter of the crss). 13 Hence, we refer to the 17 years of data from 1990:Q1 though 2007:Q3 as the pre-crss perod and the 16 quarters of data from 2007:Q4 through 2010:Q4 as the crss perod The small bank lendng envronment The lmted lendng capacty of small banks precludes them from makng or partcpatng n busness loans to large publcly traded frms; nstead, small banks specalze n busness loans to small, prvately-held busnesses. These loans typcally rely on relatonshps between a small bank s loan offcers and ts busness borrowers that allow the bank to observe soft (.e., not quantfable) nformaton about the borrower that can be used to evaluate the borrower s credtworthness (Sten 2002). Because the supportng nformaton for these relatonshp loans cannot be credbly conveyed to outsde nvestors, these loans should be less lqud than loans based upon quantfable nformaton; Berger et al. (2005) fnd evdence consstent wth ths predcton. Real estate loans and consumer loans made by a gven small bank may or may not be less lqud than those made by larger banks. Large banks orgnate and securtze, or orgnate wth the ntent to securtze, large portons of ther real estate loans (e.g., resdental mortgages, home equty lnes of credt) and consumer loans (e.g., auto loans, student loans, credt card 13 In the January 2008 SLOS, 17 banks tghtened standards, 39 dd not change ther standards, and 0 eased ther standards. Thus, the net percentage of banks that tghtened standards = (17 0)/56 = 30.4%, up from just 9.6% n the prevous survey. In the Aprl 2011 SLOS, 0 banks tghtened standards, 45 dd not change ther standards, and 7 eased ther standards. Thus, the net percentage of banks that eased standards = (7 0)/52 = 13.5%, up from just 1.9% n the prevous survey. 21

23 recevables). The orgnate-and-securtze producton process generates addtonal costs not present n portfolo lendng (e.g., legal and credt ratng agency fees, overhead for performng statstcal analyss, establshng a reputaton n the asset-backed securtes market, provdng credt enhancements to the buyers of the asset-backed securtes), but the scale economes and fee ncome assocated wth ths process offset these costs for large lendng operatons. Because hgh volumes of loan orgnaton are necessary to run ths process effcently, and because sellng off rather than holdng loans s antthetcal to close bank-borrower relatonshps, small lenders may choose to securtze a smaller porton of the real estate and consumer loans they orgnate, and hold a larger porton as portfolo nvestments. Moreover, because small banks have less ncentve to make loans that conform to the sze, documentaton, and credt score standards necessary for securtzaton, the real estate and consumer loans made by small banks are more lkely to be dosyncratc and hence less lqud. The prncple excepton to ths s home mortgage loans sold for securtzaton through government-sponsored enterprses such as Fanne Mae, Fredde Mac and Gnne Mae. Small banks have several other attractve features for our study. Small banks operate wthn a smaller geographc area than large banks and hence are less well dversfed; ths makes small banks more senstve to fluctuatons n local busness condtons that can shft ther optmal loan portfolo composton away from ther current (perhaps llqud) loan portfolo composton. Small banks lack the scale and expertse to produce many nontradtonal, off-balance-sheet bankng products (e.g., nsurance and securtes underwrtng, securtes brokerage, loan securtzaton), so ther strategc focus remans on tradtonal portfolo lendng. Ths not only makes small banks a relatvely homogeneous populaton for statstcal analyss, but also means that lendng portfolo concerns such as loan overhang should loom larger for small banks than 22

24 for large banks. Durng much of our perod of nvestgaton, small banks were less lkely to be nvolved n the knds of mergers that sgnfcantly altered ther busness strateges. Because most small banks lack the expertse to hedge credt and other rsks wth off-balance sheet dervatve securtes, balance sheet-based measures are a more accurate measure of a small bank s capacty for bearng rsk than they would be for a larger bank. And because a large porton of small U.S. banks are famly owned and managed, any lendng decsons drven by rsk averson should be relatvely free of potentally confoundng prncpal-agent effects Regresson varables The defntons of the varables used to specfy the regresson equatons are presented n Table 1, and descrptve statstcs for these varables are dsplayed n Table 2, Panel A. We defne three categores of loans: busness loans L BUS, real estate loans L RE, and consumer loans L CON. Each of these three categores aggregates loans wth dfferent characterstcs; whle n some cases ths hgh level of aggregaton s undesrable, ths s unavodable due to the structure of the data n the call reports. 14 L BUS ncludes all commercal and ndustral loans. L RE ncludes all loans secured by a len on real estate: commercal and development loans, frst and second mortgages on sngle famly and mult-famly resdental propertes, and mortgages on commercal propertes. L CON ncludes all revolvng, nstallment, or sngle payment loans to ndvduals (e.g., auto loans, student loans, personal lnes of credt), wth the excepton of credt card loans whch we exclude because they are relatvely unmportant for small banks. 15 For all 14 Whle the call reports do dsaggregate the portfolo balances for BUS, RE and CON loans nto a varety of subcategores, they do not dsaggregate loan nterest revenue. Ths precludes us from calculatng rsk-adjusted loan returns (RAR) for loan sub-categores, and as such we are lmted to usng only the three hghly aggregated loan categores n our tests. 15 Small banks exted credt card lendng wth the development of loan producton processes (.e., credt scorng and loan securtzaton) that exhbted huge scale economes. For the banks n our data, credt card loans never exceeded 1% of bank assets on average durng our sample perod. Loans to government enttes, loans to other fnancal nsttutons, loans to fnance agrcultural producton, and loans to fnance the purchase of farm land also comprse a neglgble porton of the loan portfolos of the small, urban banks n our sample. 23

25 three of these loan categores, we measure NLC t (the net lendng change n sector lendng n quarter t) as the end-of-quarter t loan stock mnus the begnnng-of-quarter t (end-of-quarter t-1) loan stock, plus net loan charge-offs (loans charged off mnus loans recovered) durng the quarter. In order to reduce the effect of sze-nduced dfferences between banks, we normalze all loan stock and net lendng change varables by dvdng them by begnnng-of-quarter t assets. 16 In our dscussons above, we have characterzed busness loans as beng less lqud, and exhbtng greater credt rsk, than consumer and real estate loans. The statstcs dsplayed n Table 2, Panel B provde confrmaton. Credt rsk data are dsplayed n tem 1. Busness loans had the largest average quarterly loan charge-off rato (0.62%), followed by consumer loans (0.49%) and then real estate loans (0.09%). Ths rankng s unchanged when specalst lenders are ncluded n the averages. Whle real estate loans defaulted at hgh rates durng the fnancal crss, they have hstorcally exhbted a relatvely low level of credt rsk. Loan lqudty data are dsplayed n tem 2. Unfortunately, the call reports do not contan complete or unform data on loan lqudty across loan types or across tme. We use the sum of the best varables avalable Outstandng prncpal balances of assets sold and securtzed by the reportng banks wth servng retaned or wth recourse or other seller-provded credt enhancements plus Assets sold wth recourse of other seller-provded credt enhancements and not securtzed by the reportng bank to construct loan lqudty ratos for the second half of our sample. Busness loans are the least lqud, wth ncreasng lqudty for consumer loans and the hghest amount of lqudty for real estate loans (whch ncludes data on resdental mortgages and home equty loans, but not commercal real estate loans). Agan, the rankng s unchanged when specalst 16 In the theory model we assume that loans are perfectly llqud and once made never leave the balance sheet. Hence, NL s non-negatve. In contrast, NLC (our emprcal proxy for NL) s often negatve because actual bank 24

26 lenders are ncluded n the averages. The small magntudes of the lqudty ratos understate the extent of loan lqudty for two reasons. Frst, small banks do not sell loans contnuously throughout the year; hence, n any gven quarter, the average ratos contan lots of zeros. Second, these data report only loans for whch the sellng bank s stll exposed to recourse or other credt guarantees, whch often expre wth a year after the loan has been sold. 17 We measure the rsk-adjusted return rato (p t,1 -µ t,1 )/σ 11 for busness loans wth the followng rato: the bank-specfc expected returns on busness loans n perod t dvded by the market-specfc varance of these returns over the entre sample perod. The numerator n ths rato s the expected percent return (the bank s nterest and fee ncome from busness loans durng perod t dvded by ts stock of accrung busness loans at the end of perod t) multpled by the expected performance of busness loans (the hstorcal percentage of accrung busness loans) mnus the average depost rate pad by the bank (the nterest pad on deposts durng perod t dvded by the average deposts n the current and pror perod). 18 The denomnator n ths rato s the varance of the quarterly change n expected proft from busness loans for the whole sample perod, calculated separately for banks n each state. 19 We measure bank-specfc rsk tolerance G -1 as the bank s total equty captal dvded by ts total assets at the begnnng of quarter t, denoted as EQ t. 20 Intutvely, banks wth lower loans are only mperfectly llqud, and can leave the balance sheet va sales, maturty, or charge-offs. 17 Not surprsngly, the specalst lenders exhbt hgher overall levels of both credt rsk and loan lqudty. By specalzng rather than dversfyng, these banks (a) are sgnalng that they are wllng to operate wth hgher levels of credt rsk and (b) must rely more on loan sales to manage ther rsk profles. 18 Hstorcal nonaccrung loans are calculated as the four-quarter laggng average of nonperformng loans to begnnng of perod loan stock when avalable. When the four-quarter average s not avalable but a three-quarter average s, the three-quarter average s used. 19 Whle the theory model (3) does not constrant the loan performance varances and covarances to be constant over tme, the quarterly call report data are smply too nfrequent to construct good tme-varyng measures of these varables. We estmate the proft varance at the state level rather than at the bank level n order to ensure exogenety. Usng the varance of nonperformng loans nstead of proft varance has no qualtatve effect on the results. 20 We construct EQ usng the book values of equty and assets. The component parts of the Basel I rsk-adjusted captal ratos are not avalable for our entre sample perod. 25