Payday Lending Regulation

Finance and Economics Discussion Series Divisions of Research & Statistics and Monetary Affairs Federal Reserve Board, Washington, D.C. Payday Lending Regulation Alex Kaufman 2013-62 NOTE: Staff working papers in the Finance and Economics Discussion Series (FEDS) are preliminary materials circulated to stimulate discussion and critical comment. The analysis and conclusions set forth are those of the authors and do not indicate concurrence by other members of the research staff or the Board of Governors. References in publications to the Finance and Economics Discussion Series (other than acknowledgement) should be cleared with the author(s) to protect the tentative character of these papers.

PAYDAY LENDING REGULATION ALEXKAUFMAN* ABSTRACT. Todatethedebateoverpaydaylendinghasfocusedonwhetheraccesstosuchlending is on net beneficial or harmful to consumer welfare. However, payday loans are not one product but many, and different forms of lending may have different welfare implications. The current diversityinpaydaylendingstemsfromthediversewaysinwhichstateshaveregulatedtheindustry. Thispaperattemptstoquantifytheeffectsthatvariousregulatoryapproacheshavehadonlending terms and usage. Using a novel institutional dataset of over 56 million payday loans, covering 26 states fornearly 6 years, I find thatprice caps tend tobe strictly binding, size caps tendto be less binding,andprohibitionsonsimultaneousborrowingappeartohavelittleeffectonthetotalamount borrowed. Minimum loan terms affect loan length while maximum loan terms do not. Repeat borrowingappearstobenegativelyrelatedtorolloverprohibitionsandcooling-offperiods,aswell as to higher price caps. Several states have used law changes to sharply cut their rate of repeat borrowing. However,thisprocesshasbeendisruptive,leadingtolowerlendingvolumesand,inat leastonecase,higherdelinquency. JELClassifications: D14,D18,G21,G28. Keywords: AlternativeFinancialServices,FinancialRegulation,PaydayLending. Date:August15,2013. I thank many helpful employees of the anonymous payday lender that provided the data for this paper. I also thank Eleanor Blume, Josh Gallin, Alex Horowitz, Michael Palumbo, Steve Sharpe, and Paul Smith for their thoughtful commentsanddiscussions. MattHoopsprovidedexcellentresearchassistance. Theviewsexpressedinthispaperare myown,anddonotindicateconcurrencebymembersoftheBoardofGovernorsoftheFederalReserveSystemorits staff. * Board of Governors of the Federal Reserve System, Division of Research and Statistics. Email: alex.kaufman@gmail.com. 1

1. INTRODUCTION Over two decades since its emergence, payday lending remains a divisive topic for economists and policymakers. No conscensus has been reached on whether access to these high-cost, shortterm balloon loans makes consumers better off or worse. Advocates point to cases where payday loans appear to be a customer’s best option. For instance, if unexpected medical expenses leave a familyshortonmoneytopayutilities,apaydayloanmaybepreferabletoanelectricityshutoffand eventual reconnect fee. Alternative sources of funds may be unavailable in the case of emergency (forinstance,creditcardsmaybemaxedout)ormoreexpensivethanpaydayloans(asareoverdraft fees at many banks). Research such as Morgan and Strain (2008), Elliehausen (2009), Fusaro and Cirillo(2011),andMorse(2011)hassupportedthenotionthataccesstopaydaylendingiswelfareenhancing. However, opponents of payday lending point out that customers rarely report borrowing in response to such emergency situations. Pew Charitable Trusts (2012) finds that only 16% of payday customers took out their initial loan in response to an unexpected expense, while 69% reported borrowing to cover a recurring expense such as rent or groceries. In addition, though they are marketed as short-term loans designed to deal with transitory shocks, a significant fraction of customers use payday loans repeatedly.1 Such repeat borrowing fuels the claim that payday loans can trap borrowers in cycles of debt. Research such as Parrish and King (2009), Melzer (2011), and Carrell and Zinman (2013) suggests that the damage caused by such debt cycles outweighs the benefitsofaccess. Given the continued debate over its merits and the long history of high-cost, short-term loans aimed at credit-compromised customers (Caskey, 1996) it seems likely that payday lending, or something similar to it, will remain a feature of the credit landscape for the forseeable future. For 1The exact fraction of payday lending that should be considered repeat borrowing is a contentious subject. The distribution of borrowing is heavily skewed, with occasional borrowers making up the bulk of the customers but repeatborrowersmakingupthebulkoftheloans. Thiscausesstatisticstovarydrasticallyaccordingtowhetherthey are person-weighted or loan-weighted, and whether the mean or median is considered. In addition, statistics vary according to whether repeat borrowing is defined as an unbroken string of loans or as the number of loans within a fixed time period. Rather than report simple summary statistics, Table 1 presents a range of percentiles in order to morecomprehensivelycharacterizethedistributionofborrowing. 2

this reason it may be productive to ask not whether payday lending is good or bad on net, but insteadwhichtypeofpaydaylendingwouldbebest. Both sides of the debate tend to treat “payday lending” as a monolithic entity, but in practice it is a pastiche of practices shaped by a diverse set of state laws. States have approached payday lendingwithavarietyofregulatorystrategiesincludingpricecaps,sizecaps,prohibitionsonrepeat borrowing, prohibitions on simultaneous borrowing, “cooling-off” periods, mandates to provide amortizing alternatives, and many combinations thereof. Some of these forms of regulation may create payday loans that lead to better outcomes than others. Though a few papers, notably Avery and Samolyk (2011), have attempted to compare regulations of differing strengths (in the case of Avery and Samolyk (2011), higher price caps versus lower ones), efforts to distinguish among regulatorystrategieshavesofarbeenlimited. This paper breaks down the monolith of payday lending in order to judge the relative merits of lending under different regulatory regimes. It uses a novel institutional dataset covering all loans originated by a single large payday lender between January 2007 and August 2012, in 26 of the 36 states in which payday lending is allowed—a total of over 56 million loans. Unlike previous payday datasets, the depth and breadth of these data span a variety of regulatory environments, makingitpossibletoestimateoftheeffectsofavarietyofregulatoryapproaches. However, the data are also limited in some ways. Most importantly, customer activity outside of payday borrowing is unobserved, making it impossible to estimate effects on overall financial health. Second,becausethedatacomefromasinglelenderonecannotcrediblyestimatetheeffect ofstatelawsontotallendingvolume. Forthesereasonsthispaperfocusesonloantermsandusagebasedoutcomes. Inparticular,itfocusesoncustomers’propensitytoborrowrepeatedly. Whatever theirotherviews,paydaylending’ssupportersanddetractorsoftentendtoagreethatverypersistent indebtedness is undersirable and indicative of counterproductive use, making repeat borrowing a usefulobjectofstudy. I find that payday loan price caps tend to be strictly binding on prices, while size caps are much less binding on loan size. Prohibitions on simultaneous borrowing appear to have little effect on total amount borrowed. Minimum term limits affect loan length, but maximum term 3

limitsdonot. Sourcesofdelinquencyaredifficulttoidentify,thoughdelinquencyseemspositively related to higher price caps. Repeat borrowing appears negatively related to rollover prohibitions and cooling-off periods, as well as to higher price caps. Extended repayment options have little identifiable effect, though that may be due in part to idiosyncracies of the dataset. Looking at individual states that changed their laws, South Carolina, Virginia, and Washington all enacted changesthatsignificantlycuttheirratesofrepeatborrowing. Thesechangeswereaccompaniedby significantupheavals,particularlyinVirginiaandWashingtonwhereloanvolumeplummetedand, inthecaseofVirginia,delinquencyspiked. Section 2 provides background on the payday lending industry and the state regulations that affect it. Section 3 describes the data, the sources of regulatory variation, and the econometric specifications. Section4presentsresultsusingcross-statepooledregressionsandwithin-statelawchangeregressions. Section5concludes. 2. PAYDAY LENDING AND STATE REGULATION Payday lending is widespread. FDIC (2013) estimates that 4.7% of all U.S. households have at sometimeusedpaydaylending,whilePewCharitableTrusts(2012)putsthefigureat5.5%ofU.S. adults. In 2005, payday storefronts outnumbered McDonald’s and Starbucks locations combined (Graves and Peterson, 2008). Lenders extended $40 billion in payday credit in 2010, generating revenuesof$7.4billion(StephensInc.,2011). To date the federal government has not directly regulated payday lending (save via general statutes such as the Truth in Lending Act and the Military Lending Act), though this may change now that the Consumer Financial Protection Bureau (CFPB) has been given rulemaking authority over the industry. Traditionally, payday lending regulation has been left to the states. Prior to the mid-2000s, states’ ability to regulate payday lending was undermined by the so-called “renta-bank” model, wherein a local lender would partner with a federally-chartered bank not subject to that lender’s state laws, thereby importing exemption from those laws (Mann and Hawkins, 2007; Stegman, 2007). In March 2005 the Federal Deposit Insurance Corporation (FDIC) issued guidanceeffectivelyprohibitingbanksfromusingthismodel,givingstatelawsmorebite. 4

The advent of online payday lending offers a potential alternative model for skirting state law. However, initial evidence suggests only very limited substitution between storefront and online payday products. Online payday customers tend to be younger, richer, and more educated than storefront customers, and states that ban storefront payday have virtually identical rates of online borrowingas statesthat allowstorefrontpayday (PewCharitable Trusts,2012). This suggeststhat customers have not responded to more stringent state regulations by substituting toward online paydayinappreciablenumbers. 2.1. The payday lending model. A payday loan is structured as a short-term advance on a paycheck. The borrower provides proof of employment (usually via pay stubs) and writes a check for theprincipaloftheloanplusthefee,post-datedforafterthenextpayday. Forinstance,aborrower might write a check for $345 and walk out with $300 in cash. Once the payday arrives the lender cashesthecheckwrittenbytheborrower. Though payday loans are technically uncollateralized, the lender’s possession of the post-dated check (or, increasingly often, the permission to directly debit the borrower’s checking account) plays a collateral-like role. By taking the repayment decision out of the borrower’s hands, payday lenders effectively ensure they are repaid ahead of the borrower’s other debts and expenses. Though default is still possible, loss rates of around 3.5% of loan volume (Stephens Inc., 2011) are very low given borrower creditworthiness.2 The high price of payday loans reflects their high overhead cost more than it does high losses from default. Stephens Inc. (2011) estimates that in 2010lossescomprisedonly21%oftotalcost.3 Becausepaydayloansaretypicallydueontheborrower’snextpayday,termsof14daysarecommon. Given prices around $15 per $100 borrowed, APRs are often in the range of 300%–500%. 2Bhutta,Skiba,andTobacman(2012)findsthatpaydayapplicantshaveanaverageEquifaxcreditscoreof513. 3FlanneryandSamolyk(2007)arguesthatthepaydayindustry’shighoverheadisduetolowbarrierstoentry. Each loanisprofitablebutdemandisrelativelyfixedandprice-insensitive,sostorefrontsenterthemarketandcompeteover scarcelendingopportunitiesuntileachonejustcoversitsoverhead. Thisstoryisconsistentwiththelargenumberof stores in operation, and with each store’s relatively low lending volume (an average of 15.3 loans per store per day inmydata). Iftheseargumentsarecorrecttheyimplythatstricterpricecapswillreducethenumberofstorefronts, causingeachstoretooperateatmoreefficientscale,withoutactuallyinhibitinglending. AveryandSamolyk(2011) offersevidenceinsupportofthisconclusion. Inmydatathereisacorrelationof-0.54betweenpricecapsandloan volumeperstore. Likewise,regressionestimatesimplythatdroppingthecapby$10per$300borrowedraisesloans perstoreby10.9%,equivalentto1.7loansperday. 5

Ontheduedatethewholeamountoftheloanisdueinasingleballoonpayment. Borrowerswishing to renew their loan can theoretically recreate the structure of an amortizing loan by borrowing slightly less each time. In practice, it is much more common for customers to borrow the same amountwitheachrenewaluntilsuchtimeastheloancanberetired. 2.2. Strategies to regulate payday lending. States concerned about payday lending within their bordershavepassedavarietyoflawstoregulateit. Thefollowinglistdetailsthemostwidely-used regulatorystrategies. 2.2.1. Price caps. A very common form of payday lending regulation is price caps. States that “prohibit” payday lending usually do so by setting APR caps that are too low for the payday business model to profitably operate, effectively driving lenders from the state. Caps of 36% APR are used by many states for this purpose. States with caps high enough to allow payday lending alsomayuseAPRlimits,butmorecommonlythecapsarestatedasadollarlimitperamountlent. A cap of $15 per $100 is typical. Some states use tiered schedules of price caps: for instance, Indiana limitsfees to 15% ofthe first $250lent, 13% of thenext $251-$400, and 10%of anything abovethat. 2.2.2. Size caps. Many states limit the maximum size of a payday loan. The modal size limit is $500. Some states don’t use a fixed size limit but instead set the limit as a percentage of the borrower’smonthlyincome. Sizelimitsaremeanttolimitaborrower’sabilitytobecomeindebted, though they can potentially be circumvented in states that allow borrowers to take multiple loans atatime. 2.2.3. Loan term limits. Maximum term limits put an upper cap on the length of a payday loan. Minimum term limits potentially directly address one of the alleged problems with payday loans: short maturity that leaves borrowers scrambling to repay by the due date. By requiring longer minimumterms,statesmightgivecustomersthetimenecessarytosortouttheirfinancesbeforethe loan is due. However, if the main source of repayment difficulty is that the loan doesn’t amortize, a slightly longer balloon loan may be no easier to retire than a slightly shorter one. Some states 6

don’tuseafixedminimumloanterm,butinsteadvarytheminimumaccordingtothelengthofthe borrower’spayperiod. 2.2.4. Limits on simultaneous borrowing. Some states set limits on the absolute number of loans a customer can borrow at a given time, while others set limits on the number of loans a customer can borrow from a single lender at a given time. The former type of regulation requires that there be some way for the lender to check the activity of other lenders; the latter type does not. For this reason, limits on the absolute number of simultaneous loans are often enacted along with legislationestablishingastatewideloandatabase. 2.2.5. Rolloverprohibitions. Prohibitionsonrenewing(“rollingover”)loansareextremelypopular, though their efficacy is debated. Superficially, rollover bans seem like a good tool to address the problem of repeat borrowing. In practice, these laws may at times be circumvented by paying off the first loan and then immediately taking out a second loan, which is technically not the same loan as the first. States vary according to how a rollover is defined and in the number of rollovers, ifany,thattheypermit. Somestatespermitrolloversonlyifaportionoftheprincipalispaiddown. 2.2.6. Cooling-offperiods. Afteraperiodofrepeatborrowingsomestatesrequirea“cooling-off” period, which is a length of time during which borrowing is not allowed. Cooling-off periods vary in length, though 1 to 10 days is common, and may be triggered according to the number of consecutiveloansorbythetotalnumberofloansintheyear. Likerolloverprohibitions,cooling-off periodsareanattempttodirectlyprohibitrepeatborrowing. 2.2.7. Extended repayment options. A number of states require that under certain circumstances lenders make available an extended, amortizing loan option in addition to their basic payday loan option. Extended repayment loans may be made available after a certain number of rollovers, or may be always available. There is a huge degree of variation among states in the form that the extended repayment options take. Most states only require that the option be made available; they do not require that the option be used.4 Variation between states in extended repayment options 4Coloradohasenactedauniquelawthatentirelyreplacespaydayloanswithanextendedrepaymentoption. Unfortunately,Coloradoisnotincludedinthisdataset. 7

may be somewhat muted in this dataset because the lender that provided the data, unlike many lenders,makesextendedrepaymentoptionsavailableeveninstateswheretheyarenotrequired. 3. THE DATA The data in this paper were provided by a large, anonymous payday lender and consist of all loans made by this lender in 26 states between January 2007 and August 2012. Figure 1 maps thestatesincludedinthedata. Thedatacontainnodemographicinformationaboutborrowers,but loans made to the same borrower can be linked across time and location. The street address of the storefront at which the loan was made is known. The data include all dimensions of the loan contract,aswellasitsrepaymenthistory. Thelendermakesnodirectonlineloans,thoughitrefers customers to online lending affiliates through its website. The dataset contains only directly made storefrontloans. The data consist of 56,143,566 loans made at 2,906 different stores to 3,428,271 distinct customers. Oncesimultaneousloansarecombinedandconsideredassingleloans(asexplainedbelow) this number drops to 54,119,468, for an average of 15.8 loans per customer. However, the median number of loans per customer is 7, reflecting the skewness of the distribution. Table 1 presents distributionsformanyvariablesinthedata. 3.1. VariableDefinitions. Becausepaydayloansvaryinsize,price,andlengthofterm,anycomparisons should be robust to relabeling. For instance, two simultaneous loans of $250 should be considered equivalent to a single loan of $500—it would be problematic to conclude that in the former case “twice as much” payday lending had occurred as in the latter, since all that must be done to convert one scenario to the other is relabel. Similarly, a customer who takes out twelve 1-week loans in a row, paying $20 each time, and a customer who takes out two 6-week loans at a cost of $120 each, should be treated similarly. Though superficially the former had 11 rollovers whilethelatterhadonlyone,ineachcasethecustomerspentexactly12consecutiveweeksindebt andpaid$240. 8

In order to construct outcome variables that are agnostic to labeling I depart slightly from standard practice. Rather than count sequences of consecutive loans, my main repeat borrowing measure is a binary variable measuring whether, exactly 90 days after origination of the current loan, the customer again has an active loan.5 This definition is agnostic about patterns of borrowing in the interim. For instance, it makes no difference if a customer takes many short loans or fewer longer loans, or whether a customer takes consecutive 2-week loans, or 1-week loans on alternating weeks. All that matters is that indebtedness 90 days later is a positive indication of propensity tostayindebt. Additionally, all simultaneous loans are combined and considered as single loans. This is done in order to facilitate comparisons in both the volume and average size of loans across regulatory regimesthatallowanddon’tallowsimultaneousborrowing. Consistently coding state regulations themselves presents another challenge. For analytical tractibility, complex regulations must necessarily be simplified and regularized. The challenge is to do this in such a way as to capture the important details and distinctions of the laws, while elidinglessrelevantdetails. Tables2and3presentasimplifiedmatrixofstatepaydayregulations. Explanations of how regulations were interpreted to create the variables in this matrix, as well as how the information in the matrix was further coded in order to perform regression analyses, are providedindetailinAppendixA. 3.2. Regulatory Variation in the Data. The data contain regulatory variation both across states and across time. Of the two forms of variation, regulatory variation across time may be econometrically cleaner. States differ from one another in many ways unrelated to their payday lending regulations (for instance, in their other consumer protections) and these differences may impact borrowing outcomes directly. In addition, state regulation itself is likely influenced by previous borrowing outcomes. For instance suppose that, for unrelated reasons, customers in State A have greater problems with repeat borrowing than customers in State B. This may cause lawmakers in 5Ninetydaysislongerthananyindividualloaninthedata,butshorterthanmanyspellsofrepeatborrowing. Results arerobusttousingalternatefollow-upperiods. 9

StateAtoenactstricterlawsthanlawmakersinStateB.Theselawsmaythemselveshavesomeeffectonoutcomes,butitwouldbeincorrecttoattributetheentiredifferenceinborrowingoutcomes between the states to the difference in laws. The inclusion of macroeconomic covariates such as thelocalunemploymentratemayhelpamelioratethisproblem,butonlypartially. In contrast, variation within state over time is likely to be less problematic. Though states that enact law changes may differ systematically from states that do not, it is likely the case that within-state before-and-after comparisons, particularly if they are focused tightly around the time of the law change, reflect the actual effects of the change in regulatory regime. Though there may be differences in usage across time for reasons unrelated to the law change, these changes a) are unlikelytobesharpdiscontinuities,andb)canbeidentifiedbyexaminingtrendsovertimeinstates without law changes. Econometrically we can apply a regression discontinuity design to look for sharp changes in outcomes, and a difference-in-difference design in order to difference out trends thatarecommontoallstates. However, such a design can only identify the effect of whatever bundle of laws each state altered—there is no easy way to separate out the effect of a price cap from, say, the effect of a cooling-offperiodrequirementifastateimplementedbothofthesethingsatonce. Inordertoseparatelyidentifytheeffectsofcomponentsofregulation,onewouldideallyhavemanydifferentlaw changes and run a pooled regression with both state and time fixed effects. However, of the states inthedata,onlysixamendedtheirpaydaylendinglawsinsomefashionduringthesampleperiod: Ohio, Rhode Island, South Carolina, Tennessee, Virginia, and Washington.6 Unfortunately, this is too few law changes to allow for a regression containing state fixed effects. Instead, to attempt to separately identify the impact of different components of the law we run pooled regressions with time fixed effects and macroeconomic convariates. This regression relies partially on cross-state regulatoryvariation. Though without a doubt regulations are not randomly assigned to states, it is also the case that they do not follow obvious patterns. For instance, Figure 2 presents a map of the states, divided 6Aseventhstate,Mississippi,amendeditslawsinJulyof2012,whichwhiletechnicallyfallingwithinthetimeframe ofthedataoccurredtooclosetotheendofthesampletoallowforanalysisofthepost-period. 10

according to the strigency of their price caps. High and low caps are well-distributed across the map, rather than clustering in particular regions. Figure 3 shows an equivalent map for rollover prohibitions. Lawdistributionssuchasthesegiveonesomereassurancethatregressionsemploying cross-stateregulatoryvariationarenothopelesslycontaminatedbyomittedvariablesbias. Though neither of these approaches (cross-state variation with time fixed effects, within-state variation due to law changes) is perfect, each corrects some of the shortcomings of the other. Cross-state regressions allow us to break apart bundles of laws, and make use a wide range of regulatory variation. Within-state law changes allow us to better control for state-specific factors andmoreconvincinglyidentifytheeffectsofthelawsthemselves.7 3.3. Econometric Specifications. In order to take advantage of cross-state law variation we use thefollowingspecification: (1) Y = α +α fee300 +α maxsize +α minterm +α maxterm i 0 1 ts 2 ts 3 ts 4 ts +α nosimult +α nosimultlender +α norollover +α cooling +α extended 5 ts 6 ts 7 ts 8 ts 8 ts +α M +α T +ν +(cid:15) 9 ti 10 t s i where Y is an outcome of interest such as amount borrowed, fee300 and maxsize are in i ts ts dollars, minterm and maxterm are in days, and the other five law variables are binary. Because ts ts themainsourceofvariationisdifferencesinlawsacrossstateswecannotaddstatefixedeffects,but we can at least partially account for cross-state differences with M , a vector of macroeconomic ti variables including monthly unemployment at the state level provided by the Bureau of Labor StatisticsandmonthlyhousepricesatthezipcodelevelprovidedbyCoreLogic. T isasetoftime t 7An earlier version of this paper employed a third empirical strategy: comparisons across state borders. Assuming thatmacroeconomicvariablesdonotcaptureallrelevantlocalvariation,aborrowerwithin,say,25milesofaborder ononesidemaymakeagoodcontrolforaborrowerwithin25milesontheotherside. However,thisworksbestfor omittedvariablesthatarelikelytovarysmoothlyoverspace;omittedvariablessuchasotherstatelawswillalsovary sharply at the border and so will not be controlled for with this method. Furthermore, there must also be a critical densityofbranchesonbothsidesoftheborder. Intheend,theborderregressionsweredroppedduetoconcernsthat state-specificidiosyncracies,ratherthandifferencesinpaydayregulations,weredrivingtheresults. Theyareavailable fromtheauthorbyrequest. 11

dummies for every month in the data, ν is a state-specific error term, and (cid:15) is the idiosyncratic s i errorterm. For regressions in which Y is delinquency or repeat borrowing, both of which are binary, the i regression is estimated as a probit with marginal effects reported. In all other cases it is estimated as ordinary least squares. All standard errors are clustered at the state level. For regressions in which Y is indebtedness three months later, the relevant law is the law in force three months later. i For this reason, whenever this dependent variable is used the laws are coded to reflect the law in force at the time of the outcome, rather than the time of origination. Because in many cases the transition from one legal regime to another disrupts loans made very close to the time of the change,makingthematypicalofloanseitherbeforeorafter,allregressionsareestimatedremoving loansmadewithin30daysofthechangeitself. Thewithin-statelawchangeanalysesuseregressionsofthefollowingform: (2) Y = β +β A ∗S +β S +β A +β t+β t∗A +β M +β Q +ν +(cid:15) i 0 1 t s 2 s 3 t 4 5 t 6 ti 7 t s i where A is a dummy variable equal to 1 if the loan was originated after the law change, S is t s a dummy variable equal to 1 if the loan was originated in the state that changed its law, t is the time running variable, and Q is a set of month dummies meant to capture seasonal factors. Y, t i M , ν , and (cid:15) are the same as before. In this setting the coefficient β captures the discontinuous ti s i 1 jumpatthetimeofthelawchangeinthestatethatchangedthelaw,withβ andβ capturinglinear 4 5 trends on either side of the discontinuity and β capturing jumps that happen in other states at the 3 time ofthe change. Again, whenY is delinquencyor repeatborrowing theregression isestimated i as a probit, and when Y is repeat borrowing the laws are coded to correspond to the time of the i outcomeratherthanthetimeoforigination. SouthCarolinaprovidesaninterestingcasebecauseithadnotonelawchangebuttwo. Thestate amended its law on June 16, 2009, raising the maximum loan size to $550, creating an extended repaymentoption,institutinga1-daycooling-offperiodbetweenloans(2-dayaftertheeighthloan 12

in the calendar year) and prohibiting customers from taking more than one loan at a time. However,in ordertoallow timefortheestablishment ofastatewidedatabase thesimultaneouslending and cooling-off provisions did not take effect until February 1, 2010. This delay of part of the law makes it potentially possible to separate the effects of the simultaneous lending prohibition and cooling-off period from the effects of the size limit and extended repayment option, and necessitatesaslightlydifferentspecification: (3) Y = β +β A1 ∗S +γ A2 ∗S +β S +β A1 +γ A2 +β t+β t∗A1 +γ t∗A2 i 0 1 t s 1 t s 2 s 3 t 3 t 4 5 t 5 t +β M +β Q +ν +(cid:15) 6 ti 7 t s i where A1 is a binary variable equal to 1 after the first law change, and A2 is a binary variable t t equal to 1 after the second law change. Now β and γ capture the effects of the first and second 1 1 lawschanges,respectively. 4. RESULTS 4.1. Using Cross-State Variation. Table 4 presents the results of regressions employing crossstate regulatory variation. Each column corresponds to a separate regression of the form given in Equation (1). These regressions help us understand the contributions of various regulatory components. The first column uses fees per $100 as the dependent variable. Only two coefficients are significant: the price cap on a $300 loan, and the maximum loan size. It is easy to imagine why the price cap would matter for the price, and the coefficient of 0.25 implies that for each $1 the price cap increases, the actual price goes up 75 cents.8 It is more difficult to see why the size cap would matter for the price. A likely explanation is that this is due to the functional form used to express the price cap in the regressions. Price caps are not single numbers; instead they tend to be price schedules, and those schedules tend to be concave in the size of the loan. In other words, 8Notethatonefigureisstatedintermsofthefeeona$300loanwhiletheotherisstatedasanaverageper$100. 13

in many states as loans get larger the per-dollar price cap drops. Using one number for the price cap effectively assumes that all price schedules are linear. It may be that maxsize picks up the ts non-linearityofactualpricecapschedules. It’salsonotablethattheestimatedeffectisverysmall: anincreaseof30centsper$100increaseinthesizecap. The next column’s dependent variable is total loan size. Unsuprisingly, maximum size caps matter, with an estimated increase of $41 per $100 increase in the size cap. However, this is well below the one-to-one correspondence we would see if size caps are binding. Maximum loan term androlloverprohibitionsalsocomeinassignificant,thoughtheconnectionislessclear. Onlyonevariablesignificantlyaffectsloanterm,andthatisminimumloanterm. Thecoefficient just misses the 5% significance mark (p = 0.052) and implies a 10-day increase in the minimum will raise lengths by 2.6 days on average. This effect is likely non-linear and concentrated among states with longer minimum loan terms. Notably, the estimate for maximum term is insignificant andeconomicallysmall,suggestingitrarelyifeverbinds. Price caps and size caps are the only types of regulation that are significantly predictive of delinquency, with coefficients implying that a $10 increase in the cap on a $300 loan increases delinquencyby0.6percentagepoints,anda$100increaseinthesizecapincreasesdelinquencyby 0.4 percentage points. These effects are moderate relative to an overall delinquency rate of 4.3%, andthemechanismbywhichtheymightaffecttherateisnotcertain. Onepossibilityisthatlarger andmoreexpensiveloansaresimplymoredifficulttopayoff,leadingtodelinquency. Four types of regulation appear predictive of repeat borrowing: price caps, maximum term limits,rolloverprohibitions,andcooling-offperiods. Itiseasytoseewhytheremightbeaconnection between rollover prohibitions and cooling-off periods—both are specifically designed to limit repeat borrowing, and indeed both coefficients are significant and negative. Though much of the debate over rollover prohibitions focuses on the ability of lenders and borrowers to circumvent them, it is possible that on the margin such prohibitions still make rollovers a bit less convenient, withconsequencesforoverallrepeatborrowing. Itislessstraightforwardtoseethelinkbetweenpricecapsandrepeatborrowing. Thecoefficient impliesasignificant3percentagepointdecreaseintherepeatborrowingrateforeach$10increase 14

in the cap. One possibility is that this is a simple price effect: cheaper loans are more attractive to prospective customers and so they choose to use them more often. Another possibility is that, assuming higher price caps lead to greater delinquency, delinquent borrowers are less likely to be allowedtoborrowinthefuture,leadingtolessrepeatborrowing. However,theestimatedeffectof price caps on repeat borrowing is larger than the estimated effect on delinquency, suggesting this cannotbethesolemechanism. Lastly, maximum loan term is negatively associated with repeat borrowing. Given that this form of regulation appears to have no effect on loan term itself, its putative target, it is difficult to imagineachannelbywhichitwouldaffectrepeatborrowing. 4.2. Using Variation from Law Changes. Next we examine states that changed their laws in order to see whether the results obtained from the pooled regressions of the previous section are supportedorcontradictedinasettingwithfewerconfoundingfactors. Table5presentsanalysesof the six states in the data with law changes. Each cell of the table represents a separate regression using the specification in Equation (2), except for the South Carolina cells which use the specification in Equation (3). For reference, Figures 4, 5, 6, 7, 8, and 9 present raw means over time for fees, amount borrowed, loan term, lending volume, delinquency, and repeat borrowing for each statewhoselawschanged.9 Thepooledregressionssuggestedafairlytightconnectionbetweenpricecapsandprice,andthis relationshipappearsatleastasstronginthelaw-changeregressions. Asnotedinthelawmatrixin Tables 2 and 3, price caps went up in Ohio and Rhode Island, while Tennessee and Virginia both loosened theirs. All four states saw price changes in the direction of the price cap changes, and the sizes of the price changes closely track the size of the cap changes: $1.03, 96 cents, 56 cents, and$1.16changesper$1changeinthecap,respectively. Theremainingstatesdidnotadjusttheir price caps, and their prices did not change. These results support the conclusion that actual prices adherecloselytopricecaps. 9Thefiguresrevealsignificantseasonalcyclesforsomevariables,notablydelinquencyandrepeatborrowing,though theseasonalfactorsintheregressionsensurethesecyclesdon’tcontaminatethelawchangeestimates.Thisseasonality stems largely from tax returns: many people use their returns as a lump sum payment to retire their loans, causing temporarydecreasesinratesofdelinquencyandrepeatborrowing. 15

Theconnectionbetweenloansizelimitsandloansizeappearsweakerinthelaw-changeregressions than it did in the pooled regressions. Ohio’s limit increased but its loan size did not, while Tennessee’s limit and loan size actually went in opposite directions. South Carolina’s loan size mayhaveincreasedslightlywhenitraiseditslimit,onlytodecreaseagainwhenitaddeditssimultaneous loan prohibition (Table 5 shows a marginally-significant $27 increase, though there is no observable jump in Figure 6). The lack of connection between legal limit and amount borrowed maybebecause,unlikepricecaps,sizecapsareoftennotlowenoughtobebindingonlenders. Thepooledregressionsfoundnorelationshipbetweensimultaneousborrowingprohibitionsand total amount borrowed even though amount borrowed, as contructed, merged simultaneous loans together. Thelaw-changeregressionssupportasimilarconclusion. Ohioremoveditssimultaneous borrowing limit, while Virginia instituted a new limit, neither of which appears to have affected total amount borrowed. The result is particularly notable for South Carolina, which prior to its changes had a single-loan size limit of $300. Approximately 71.5% of all its loans were made simultaneouslywithatleastoneotherloan,foranaverageborrowingamountofabout$420. After thefirstlawchangethesingle-loanlimitincreasedto$500butsimultaneousloanswerestilllegal, effectively making it easier to borrow much larger amounts. However, the total amount borrowed roseonlyslightly. Afterthesecondchangesimultaneousloansbecameillegal,anddroppedtoonly 2.4% of loan volume. Average single-loan size increased, leaving total amount borrowed largely unchanged. Overall, it appears that customers were able to borrow the desired amount no matter whetherthelimitwasstructuredasasizecaporasimultaneousborrowingban. Thissuggeststhat unlessstatesenactmuchmorebindinglimitsonthemaximumamountborroweditmaynotmatter whetherornottheyalsohavelimitsonsimultaneousborrowing. The pooled regressions found that minimum loan terms affect loan length, and the law-change results support that. Only one state changed its laws regarding minimum or maximum loan term: Virginia raised its minimum loan term from 7 days to two times the length of the borrower’s pay cycle. Assuming a standard pay cycle of two weeks, this raises the effective limit by about 21 days. The third column of Table 5 estimates that loan length in Virginia increased nearly 20 days on average as a result, suggesting that the change was binding. OH and WA both exhibit more 16

modest changes in average loan term, though neither directly changed their loan term regulations andOhio’schangewasnotstatisticallysignificant. All six states saw statistically significant changes in their rates of loan delinquency. The largest change occurred in Virginia, where delinquency rose nearly 7 percentage points over a base rate of about 4%. The law-change evidence shows a connection between price caps and delinquency, consistent with the pooled regressions. Price caps and delinquency alike dropped in Ohio and Rhode Island, while price caps and delinquency rose in Tennessee and Virginia. The connection between size caps and delinquency found in the pooled regressions gets notably less support: the threestatesthatchangedtheirsizecapssawdelinquencymoveinthewrongdirectionornotatall. The rate of repeat borrowing also changed in all six states, though the change was large in only four of them. Ohio’s rate increased about 14 percentage points, while South Carolina, Virginia, andWashingtondecreasedtheirratesby15,26,and33percentagepoints,respectively. Thepooled regressions indicated that repeat borrowing should decrease with the implementation of rollover prohibitions and cooling-off provisions. Unfortunately no state changed its rollover prohibition so the law-change regressions can provide no evidence either way. South Carolina, Virginia, and Washington all instituted cooling-off provisions and all saw large decreases in repeat borrowing, supporting the pooled regressions. South Carolina in particular saw its largest decrease after its second regulatory change, when it instituted its cooling-off provision. Washington implemented a strict 8-loan per year limit on lending, which can be thought of as an unusual form of cooling-off provision,andsawthelargestrepeatborrowingdecreaseofall. The pooled regressions also suggested that higher fee caps lowered repeat borrowing, and this toogetsfurthersupport. Thetwostatesthatraisedtheirfeecaps,TennesseeandVirginia,sawdrops inrepeatborrowingwhilethetwostateswheretheydecreased,OhioandRhodeIsland,sawjumps. Though the pooled regressions showed no relationship, the two states that instituted simultaneous borrowing prohibitions, South Carolina and Virginia, saw big drops in repeat borrowing, while Ohio,whosesimultaneousborrowingbanwasrenderedobsoletewhenlendersbegantolendunder anewstatute,sawabigincreaseinrepeatborrowing. 17

Taking a step back it appears that three states—South Carolina, Virginia, and Washington— enacted changes that had large effects on lending within their borders. For Washington the key provisionmayhavebeenthe8-loanmaximum,andforVirginia,theunusuallylongminimumloan term. SouthCarolinachangedmanysmallerthingsatonce. Allthreestatessawtheirratesofrepeat borrowingplummet. Thechangesweredisruptive: VirginiaandWashington,andtoalesserextent South Carolina, all saw large drops in total lending.10 Besides being an interesting outcome in its ownright,thechangeinlendingvolumesuggeststhatcustomercompositionmayhavechangedas well. Without demographic data it is difficult to assess changes in composition. Table 6 attempts to getahandleonthequestionbyaskinghowoftencustomerswhowererepeatborrowerspriortothe law change appear in the data after the law change. Customers are divided according to whether their pre-period loans led to indebtedness a greater or smaller proportion of the time than was the median for all pre-period borrowers. A borrower is considered to appear in the post-period if he or she takes any loan in the post-period. Naturally, repeat borrowers are more likely to appear in the post-period no matter what the regulatory environment, so similar figures are computed for customers in other states in order to get a baseline. The rightmost column presents odds ratios, withnumbers1indicatingthedegreetowhichpre-periodrepeatborrowersareover-representedin thepost-period. As expected, the data show that repeat borrowers are much more likely to show up than occasional borrowers in the post-period in all states. The odds ratio for Virginia is much lower than forotherstates,suggestingthatinVirginiathelawchangesignificantlyalteredcustomercomposition. In South Carolina and Washington, however, the odds ratios look more normal. Both states were marginally more likely than other states to retain non-repeat borrowers, but the differences 10Loansfromasinglelenderare,ingeneral,illsuitedtoestimatingtheeffectsofregulationontotallendingvolume. Thelendermayhavegreaterpentrationinsomestatesthanothers,andmayexpandorcontractoperationsforreasons unrelatedtothelegalenvironment. However,thereisreasontobelievetheseparticulardropsinvolumeareduetothe law changes themselves. These drops do not correspond to mass branch closings, but instead to decreases in loans perbranch. InbothVirginiaandWashingtonthelenderdideventuallyclosemanybranches,butthisappearstobea consequenceratherthanacauseofthedropinvolume. SouthCarolinaneverhadanymassclosings. 18

aresmall,suggestingthatthesestatesdidnotexperiencenotablecustomerselectionwhenlending volumedropped. Finally, as in the pooled regressions, the law-change results show no evidence that extended repayment options matter. This may be due to the omission of Colorado, the only state where extended repayment is mandatory, not just an option. It may also be due to the fact that the lender providingthedatamakesextendedrepaymentoptionsavailableeveninstatesthatdon’trequireit. As such, these regressions may not capture the impact of extended repayment options on lenders withoutsuchapolicy. 5. CONCLUSIONS Overall, pooled cross-state regressions and within-state regressions examining law changes show a remarkable amount of agreement. Both suggest the following conclusions about payday lending regulation: price caps tend to be strictly binding, size caps tend to be less binding, and prohibitions on simultaneous borrowing appear to have little effect on the total amount borrowed. Minimum term limits affect loan length, but maximum term limits do not. Delinquency seems positively related to higher price caps. Rollover prohibitions and cooling-off periods, as well as to higherpricecaps,appeartoreducethefrequencyofrepeatborrowing. Focusingonstateswithlawchanges,SouthCarolina,Virginia,andWashingtonwereallableto significantly cut their rates of repeat borrowing. These changes were accompanied by significant upheavals, however, particularly in Virginia and Washington where loan volume dropped sharply and, inthe caseof Virginia, delinquencyspiked andcustomer compositionshifted. It seemslikely that Virginia’s changes were connected to its adoption of a 2-pay-period minimum term, which is longer than the minimum term of most states. It will be interesting to follow what happens in Mississippi, which like Virginia recently adopted a long minimum term limit. Washington’s changes seem plausibly related to its adoption of an 8-loan yearly maximum, another form of regulation unusual among states. In South Carolina the decline in repeat borrowing is less readily pinnedonasingleprovision. 19

Thispaperhasattemptedtogetinsidethemonolithofpaydaylendingandexaminehowdifferent regulatory environments affect loan terms and usage. Without a doubt there remains greater detail toexplore—forinstance,bothcooling-offprovisionsandextendedrepaymentoptionsvarygreatly across states. It is possible that particular instances of these regulations, like for instance those adopted by South Carolina, might have effects on delinquency or repeat borrowing that are not captured by the average effect of all laws in that regulatory category. In the face of state-specific idiosyncracies, however, the more fine-grained the question the more challenging it is to move beyondinformedspeculation. Payday lending is not one product but many. The price, size, and duration of payday loans, as well as the manner in which customers use them, varies greatly according to their regulatory environment. As we possibly move toward a regime of federal regulation, it is crucial to better understandhowthesedifferenttypesofregulationwork. REFERENCES AVERY, R., AND K. SAMOLYK (2011): “Payday Loans versus Pawnshops: The Effects of Loan FeeLimitsonHouseholdUse,”Workingpaper. BHUTTA, N., P. SKIBA, AND J. TOBACMAN (2012): “Payday Loan Choices and Consequences,” VanderbiltUniversityLaw&EconomicsWorkingPaperno.12-30. CARRELL, S., AND J. ZINMAN (2013): “In Harm’s Way? Payday Loan Access and Military PersonnelPerformance,”Workingpaper. CASKEY, J. (1996): Fringe Banking: Check-Cashing Outlets, Pawnshops, and the Poor. The RussellSageFoundation. ELLIEHAUSEN, G.(2009): “AnAnalysisofConsumers’UseofPaydayLoans,”FinancialServices ResearchProgramMonograph,no.41. FDIC (2013): “Addendum to the 2011 FDIC National Survey of Unbanked and Underbanked Households: UseofAlternativeFinancialServices,”FederalDepositInsuranceCorporation. FLANNERY, M., AND K. SAMOLYK (2007): “Scale Economies at Payday Loan Stores,” Working paper. 20

FUSARO, M., AND P. CIRILLO (2011): “Do Payday Loans Trap Consumers in a Cycle of Debt?,” Workingpaper. GRAVES, S., AND C. PETERSON (2008): “Usury Law and The Christian Right: Faith-Based Political Power and the Geography of American Payday Loan Regulation,” Catholic University LawReview,57(3). MANN, R., AND J. HAWKINS (2007): “JustUntilPayday,”UCLALawReview,54(4),855–912. MELZER, B. (2011): “The Real Costs of Credit Access: Evidence from the Payday Lending Market,”QuarterlyJournalofEconomics,126,517–555. MORGAN, D., AND M. STRAIN (2008): “Payday Holiday: How Households Fare after Payday CreditBans,”FederalReserveBankofNewYorkStaffReports,no.309. MORSE, A.(2011): “PaydayLenders: HeroesorVillians?,”JournalofFinancialEconomics,102, 28–44. PARRISH, L., AND U. KING (2009): “Phantom Demand: Short-term due date generates need for repeatpaydayloans,accountingfor76%oftotalvolume,”CenterforResponsibleLending. PEW CHARITABLE TRUSTS (2012): “Who Borrows, Where They Borrow, and Why,” Payday LendinginAmerica. STEGMAN, M. (2007): “PaydayLending,”JournalofEconomicPerspectives,21(1),169–190. STEPHENS INC. (2011): “PaydayLoanIndustry,”IndustryReport. 21

APPENDIX A Notesoncodingbytypeofregulation. Price caps. For analytical tractibility this paper collapses complex fee schedules into a single number: the dollar limit on fees for a hypothetical $300 loan. For example, Indiana limits fees to 15% of the first $250 lent, 13% of the next $251-$400, and 10% of anything above that. In this casethefeefora$300loanwouldbe0.15∗250+0.13∗50 = $44. Allcapsareconsideredinclusive of database fees, verification fees, and other add-on fees. States without any price cap are treated as if they had a cap equal to the highest cap of any state in the data, which is the $73.52 cap for VirginiaafterJanuary1,2009. Size caps. States vary according to whether their size cap is stated inclusive of exclusive of fees. For comparability, this paper codes all size caps as if they were exclusive of fees. In other words, if a state limits loan size to $500 inclusive of fees, as for instance Nebraska does, this is coded as an exclusive size limit of $425 because $75 has gone to fees. (Technically a lender in Nebraska could offer a loan with principal higher than $425 if its fees were set below the state statuatory maximum, but in practice lenders tend to charge the maximum allowed.) For states that set their size cap as the minimum of an absolute size limit and a percentage of the borrower’s monthly limit I assume an annual income of $31,000, which is the median annual income of payday loan borrowers in the 2010 Survey of Consumer Finances. Using this income level, monthly income limitsarenotbindingforanystate. Stateswithnosizecapsarecodedashavingacapequaltothe capinthestatewiththehighestcap,whichis$1000forIdaho. Minimum term limits. For states that set the minimum term limit in terms of pay periods rather than days, a standard pay period of 2 weeks is assumed. For instance, Virginia’s limit of 2 pay periodsiscodedas28days. Maximum term limits. States with no maximum term limits are coded as having a limit equal to thestatewiththehighestlegallimit,whichis60daysforKentucky. Limitsonsimultaneousborrowing. Simultaneousborrowinglimitsaredividedintotwovariables: thelimitonabsolutenumberofloans,andthelimitofthenumberofloansperlender. Inregression analysis both of these are collapsed into binary variables. These variables take the value 1 if the state limits customers to one loan at a time, and 0 otherwise. This means that states limiting customers to two or more loans at a time are considered equivalent to states with no limit. This decisionwasmadeinlightofthefactthatinstateswithnolimititisraretoborrowmorethantwo loansatatime;therefore,alimitoftwoloansisunlikelytobebindingonmanycustomers. Rollover prohibitions. For states in which the rollover limit is stated in weeks rather than in the number of renewals, 2 weeks is considered equivalent to 1 renewal. In regression analysis the rollover variable is collapsed into a binary equal to 1 if rollovers are completely prohibited, and 0 if some form of rollover is allowed (even if it requires part of the principle to be paid down). Note that an alternate definition, considering paydown-only rollovers as equivalent to rollover prohibitions,yieldsempiricalresultsverysimilartotheresultspresentedinthepaper. 22

Cooling-off periods. Cooling-off periods are stated in days. Given variability in both the length of cooling-off periods and in the conditions under which they are triggered, in regression analysis they are collapsed into a binary variable equal to 1 if the state employs some type of cooling-off regulation,and0otherwise. Extended repayment options. Extended repayment options are extremely variable both in their formandintheconditionsunderwhichtheyaretriggered. Inregressionanalysistheyarecollapsed intoabinaryvariableequalto1ifthestateemployssometypeofextendedrepaymentoption,and 0otherwise. Notesoncodingbystate. California. CalculatingCalifornia’spricecapper$300isachallengebecausethestatehasa$300 loan size cap that is inclusive of the fee. This means that if a lender were to charge the statuatory maximum of 15% of the face value of the check, or $45, the principal would be limited to $255. Lenders could make a loan with $300 principal, but it would need to have no fee. In order to calculate the per-$300 maximum fee for comparison with other states I calculate the percentage feeallowedon$255thenapplythatpercentageto$300. Thisyields(45/255)∗300 = $52.94. Ohio. The Ohio Short Term Loan Act, meant to govern payday lending, sets an APR cap of 28%, effectively making payday lending impossible. However, lenders have circumvented the Act by lending under either the Ohio Small Loan Act or, more commonly, the Ohio Mortgage Lending Act. Because the Short Term Loan Act is irrelevent to lending in the state, this coding uses values derivedfromtheMortgageLendingAct. Tennessee. Tennessee allows a maximum of two loans simultaneously, and they cannot sum to an amount greater than $500. Given that $500 is also the size limit for a single loan, the dollar limit will bind more strongly that the limit on the number of simultaneous loans, making the effective loan limit 1. Tennessee has a further complication in that it is the only state with a limit on the absolutenumberofloansperborrower,butnodatabasethroughwhichlenderscancheckforother outstanding loans. This lack of an enforcement mechanism effectively renders the absolute loan limit moot. Hence, even though on the books both the absolute and lender-specific limits are 2, in practiceIhavecodedthemas“nolimit”and1,respectively. Washington. Washington uses a form of regulation that is unique among states in the data: an absolute limit of 8 loans per customer per year. This regulation most closely resembles a coolingoff period, in that it could be considered a permanent cooling-off period triggered after the 8th loan. For this reason I’ve coded Washington’s cooling-off variable as 1, though the regulation is differentenoughfromothercooling-offregulationtomeritconsiderationinitsownright. 23

B XIDNEPPA nwohseraatadehtnitoneratubgnidnelyadyapwollatahtsetats,eulbkradninwohseraatadehtnisetatS .1 ERUGIF .eulbthgilninwohseragnidnelyadyaptibihorptahtsetatsdna,eulbmuidemni 24

naht retaerg spac htiw ,neerg krad ni nwohs era 55$ naht retaerg 003$ rep spac eef htiw atad eht ni setatS .2 ERUGIF ni nwohs era 04$ ot lauqe ro naht ssel spac htiw dna ,neerg muidem ni nwohs era 55$ ot lauqe ro naht ssel dna 04$ .neergthgil 25

fo mrof emos wolla taht esoht dna ,elprup krad ni nwohs era srevollor lla tibihorp taht atad eht ni setatS .3 ERUGIF .elprupthgilninwohserasrevollor 26

52 02 51 01 5 sralloD seeF naeM HO 2102 1102 0102 9002 8002 7002 raeY deworroB001$repseeF )a( 005 054 004 053 003 sralloD deworroB naeM HO 2102 1102 0102 9002 8002 7002 raeY )snaoLsuoenatlumiSgnidulcnI(eziSnaoL )b( 120291817161514131211101 9 8 7 syaD htgneL naeM HO 2102 1102 0102 9002 8002 7002 raeY syaDnihtgneL )c( 00002100000100008 00006 00004 00002 0 emuloV emuloV HO 2102 1102 0102 9002 8002 7002 raeY snaoLfoemuloV )d( 51. 521. 1. 570. 50. 520. 0 etaR ycneuqnileD etaR ycneuqnileD HO 2102 1102 0102 9002 8002 7002 raeY ycneuqnileD )e( 8. 7. 6. 5. 4. 3. 2. 1. 0 retaL shtnoM 3 tbeD ni noitroporP gniworroB taepeR HO 2102 1102 0102 9002 8002 7002 raeY gniworroBretfAshtnoM3tbeDninoitroporP )f( stneserper enil der lacitreV .gniworrob taeper dna ,ycneuqniled ,emulov ,mret naol ,ezis naol ,seef :oihO .4 ERUGIF .egnahcwalfognimit 27

52 02 51 01 5 sralloD seeF naeM IR 2102 1102 0102 9002 8002 7002 raeY deworroB001$repseeF )a( 005 054 004 053 003 sralloD deworroB naeM IR 2102 1102 0102 9002 8002 7002 raeY )snaoLsuoenatlumiSgnidulcnI(eziSnaoL )b( 120291817161514131211101 9 8 7 syaD htgneL naeM IR 2102 1102 0102 9002 8002 7002 raeY syaDnihtgneL )c( 0002100001 0008 0006 0004 0002 0 emuloV emuloV IR 2102 1102 0102 9002 8002 7002 raeY snaoLfoemuloV )d( 51. 521. 1. 570. 50. 520. 0 etaR ycneuqnileD etaR ycneuqnileD IR 2102 1102 0102 9002 8002 7002 raeY ycneuqnileD )e( 8. 7. 6. 5. 4. 3. 2. retaL shtnoM 3 tbeD ni noitroporP gniworroB taepeR IR 2102 1102 0102 9002 8002 7002 raeY gniworroBretfAshtnoM3tbeDninoitroporP )f( enil der lacitreV .gniworrob taeper dna ,ycneuqniled ,emulov ,mret naol ,ezis naol ,seef :dnalsI edohR .5 ERUGIF .egnahcwalfognimitstneserper 28

52 02 51 01 5 sralloD seeF naeM CS 2102 1102 0102 9002 8002 7002 raeY deworroB001$repseeF )a( 005 054 004 053 003 sralloD deworroB naeM CS 2102 1102 0102 9002 8002 7002 raeY )snaoLsuoenatlumiSgnidulcnI(eziSnaoL )b( 120291817161514131211101 9 8 7 syaD htgneL naeM CS 2102 1102 0102 9002 8002 7002 raeY syaDnihtgneL )c( 00006 00005 00004 00003 00002 00001 0 emuloV emuloV CS 2102 1102 0102 9002 8002 7002 raeY snaoLfoemuloV )d( 51. 521. 1. 570. 50. 520. 0 etaR ycneuqnileD etaR ycneuqnileD CS 2102 1102 0102 9002 8002 7002 raeY ycneuqnileD )e( 8. 7. 6. 5. 4. 3. 2. retaL shtnoM 3 tbeD ni noitroporP gniworroB taepeR CS 2102 1102 0102 9002 8002 7002 raeY gniworroBretfAshtnoM3tbeDninoitroporP )f( senil der lacitreV .gniworrob taeper dna ,ycneuqniled ,emulov ,mret naol ,ezis naol ,seef :aniloraC htuoS .6 ERUGIF .segnahcwalfognimittneserper 29

52 02 51 01 5 sralloD seeF naeM NT 2102 1102 0102 9002 8002 7002 raeY deworroB001$repseeF )a( 005 054 004 053 003 052 sralloD deworroB naeM NT 2102 1102 0102 9002 8002 7002 raeY )snaoLsuoenatlumiSgnidulcnI(eziSnaoL )b( 120291817161514131211101 9 8 7 syaD htgneL naeM NT 2102 1102 0102 9002 8002 7002 raeY syaDnihtgneL )c( 00005 00004 00003 00002 00001 0 emuloV emuloV NT 2102 1102 0102 9002 8002 7002 raeY snaoLfoemuloV )d( 51. 521. 1. 570. 50. 520. 0 etaR ycneuqnileD etaR ycneuqnileD NT 2102 1102 0102 9002 8002 7002 raeY ycneuqnileD )e( 8. 7. 6. 5. 4. 3. 2. retaL shtnoM 3 tbeD ni noitroporP gniworroB taepeR NT 2102 1102 0102 9002 8002 7002 raeY gniworroBretfAshtnoM3tbeDninoitroporP )f( -erper enil der lacitreV .gniworrob taeper dna ,ycneuqniled ,emulov ,mret naol ,ezis naol ,seef :eessenneT .7 ERUGIF .egnahcwalfognimitstnes 30

52 02 51 01 5 sralloD seeF naeM AV 2102 1102 0102 9002 8002 7002 raeY deworroB001$repseeF )a( 005 054 004 053 003 sralloD deworroB naeM AV 2102 1102 0102 9002 8002 7002 raeY )snaoLsuoenatlumiSgnidulcnI(eziSnaoL )b( 53 82 12 41 7 syaD htgneL naeM AV 2102 1102 0102 9002 8002 7002 raeY syaDnihtgneL )c( 00008 00006 00004 00002 0 emuloV emuloV AV 2102 1102 0102 9002 8002 7002 raeY snaoLfoemuloV )d( 51. 521. 1. 570. 50. 520. 0 etaR ycneuqnileD etaR ycneuqnileD AV 2102 1102 0102 9002 8002 7002 raeY ycneuqnileD )e( 8. 7. 6. 5. 4. 3. 2. retaL shtnoM 3 tbeD ni noitroporP gniworroB taepeR AV 2102 1102 0102 9002 8002 7002 raeY gniworroBretfAshtnoM3tbeDninoitroporP )f( stneserperenilderlacitreV .gniworrobtaeperdna,ycneuqniled,emulov,mretnaol,ezisnaol,seef :ainigriV .8 ERUGIF .egnahcwalfognimit 31

52 02 51 01 5 sralloD seeF naeM AW 2102 1102 0102 9002 8002 7002 raeY deworroB001$repseeF )a( 005 054 004 053 003 sralloD deworroB naeM AW 2102 1102 0102 9002 8002 7002 raeY )snaoLsuoenatlumiSgnidulcnI(eziSnaoL )b( 53 82 12 41 7 syaD htgneL naeM AW 2102 1102 0102 9002 8002 7002 raeY syaDnihtgneL )c( 00004 00003 00002 00001 0 emuloV emuloV AW 2102 1102 0102 9002 8002 7002 raeY snaoLfoemuloV )d( 51. 521. 1. 570. 50. 520. 0 etaR ycneuqnileD etaR ycneuqnileD AW 2102 1102 0102 9002 8002 7002 raeY ycneuqnileD )e( 8. 7. 6. 5. 4. 3. 2. retaL shtnoM 3 tbeD ni noitroporP gniworroB taepeR AW 2102 1102 0102 9002 8002 7002 raeY gniworroBretfAshtnoM3tbeDninoitroporP )f( enil der lacitreV .gniworrob taeper dna ,ycneuqniled ,emulov ,mret naol ,ezis naol ,seef :notgnihsaW .9 ERUGIF .egnahcwalfognimitstneserper 32

TABLE 1. SummaryStatistics Percentiles Mean 10% 25% 50% 75% 90% Totalborrowedatonce $200 $255 $350 $500 $501 $370.50 (N = 54,119,468) Totalfeesatonce $29.07 $38.75 $47.95 $65.45 $85 $53.27 (N = 54,119,468) Loantermindays 11 13 14 18 29 16.7 (N = 54,119,468) APR 185.3 252.6 365 465.25 547.5 376.1 (N = 54,119,468) Isarepeatloan - - - - - 86.0% (N = 54,119,468) Isasimultaneousloan - - - - - 3.6% (N = 54,119,468) Delinquent - - - - - 4.3% (N = 53,677,124) Indebted3monthslater - - - - - 57.2% (N = 54,119,468) Consecutiveloansafteranewloan 0 0 2 7 16 5.9 (N = 7,571,675) Consecutiveloansafteranyloan 0 2 6 15 30 11.5 (N = 54,119,468) Totalloanspercustomerindata 1 2 7 21 43 15.8 (N = 3,428,271) Notes: Simultaneousloansarecombinedandtreatedassingleloans,bringingthesample sizefromover56milliondowntojustover54million. “Repeat/consecutive”loandefined as any loan originated less than 31 days after the previous loan was due. “New” loan definedasanyloanthatisnotarepeatloan. 33

snoitalugeRgnidneLyadyaPetatSfoxirtaM .2 ELBAT setaD dednetxE gnilooC srevolloR# tlumiS# tlumiS# mreTxaM mreTniM eziSxaM 003$repeeF etatS noitpO )syad( dewollA redneLreP snaoL )syad( )syad( )$( )$( 21/8–70/1 sey 1 1 1 1 13 01 005 5.25 LA 21/8–70/1 sey 0 0 1 timilon 13 0 552 49.25 AC 21/8–70/1 on 0 4 timilon timilon 95 0 005 timilon ED 21/8–70/1 sey 1 0 1 1 13 7 005 53 LF 21/8–70/1 sey 1 0 2 timilon 13 0 544 57.83 AI 21/8–70/1 on 0 3 timilon timilon timilon 0 0001 timilon DI 21/8–70/1 sey 7 0 1 1 timilon 41 055 44 NI 21/8–70/1 on 0 0 2 timilon 03 7 005 54 SK 21/8–70/1 on 0 0 1 1 06 41 005 49.35 YK 21/8–70/1 on 0 nwodyapfi timilon timilon 03 0 053 05 AL 21/8–70/1 sey 0 0 1 2 13 0 006 54.24 IM 21/8–70/1 on 0 nwodyapfi6 2 timilon 13 41 005 timilon OM 21/6–70/1 on 0 0 timilon timilon 03 0 89.833 45 SM 21/8–21/7 on 0 0 timilon timilon 03 82 414 58.56 21/8–70/1 on 0 0 2 timilon 43 0 524 9.25 EN 21/8–70/1 sey 0 5 1 timilon 53 0 526 timilon VN 80/9–70/1 on 0 timilon 1 timilon timilon 0 008 54 HO 21/8–80/01 on 0 timilon timilon timilon timilon 0 timilon 35.82 21/8–70/1 sey 2 0 2 2 54 21 005 64.54 KO 01/6–70/1 on 0 1 3 timilon timilon 31 005 54 IR 21/8–01/7 on 0 1 3 timilon timilon 31 005 03 90/51/6–70/1 on 0 0 timilon timilon 13 0 003 54 CS 01/1–90/61/6 sey 0 0 timilon timilon 13 0 055 4.54 21/8–01/2 sey 1 0 1 1 13 0 055 4.54 .AxidneppAeeS :setoN 34

)d’tnoc(snoitalugeRgnidneLyadyaPetatSfoxirtaM .3 ELBAT setaD dednetxE gnilooC srevolloR# tlumiS# tlumiS# mreTxaM mreTniM eziSxaM 003$repeeF etatS noitpO )syad( dewollA redneLreP snaoL )syad( )syad( )$( )$( 21/8–70/1 on 0 nwodyapfi4 timilon timilon timilon 0 005 timilon DS 11/91/5–70/1 on 0 0 1 timilon 13 0 074 03 NT 21/8–11/02/5 on 0 0 1 timilon 13 0 524 49.25 21/8–70/1 sey 0 5 timilon timilon timilon 0 timilon timilon TU 80/21–70/1 on 0 0 timilon timilon timilon 7 005 54 AV 21/8–90/1 sey 1 0 1 1 timilon sdoirepyap2 005 25.37 90/21–70/1 on 0 0 timilon timilon 54 doirepyap1 007 54 AW 21/8–01/1 sey raeyfotser 0 timilon timilon 54 doirepyap1 007 54 21/8–70/1 sey 1 1 1 1 timilon 0 578 timilon IW 21/8–70/1 on 0 0 timilon timilon 13 0 timilon 03 YW .AxidneppAeeS :setoN 35

noitairaVwaLetatS-ssorCgnisUsnoissergeR .4 ELBAT gniworroBtaepeR ycneuqnileD mreTnaoL eziSnaoL 001$repeeF )stniopegatnecrep( )stniopegatnecrep( )syad( )$( )$( ***03.0- ***60.0 30.0 54.0 ***52.0 α 003eef 1 )90.0( )10.0( )50.0( )68.0( )30.0( .e.s 10.0- ***400.0 00.0 ***14.0 **300.0- α ezisxam 2 )10.0( )100.0( )00.0( )70.0( )100.0( .e.s 20.0 10.0 *62.0 *73.2 50.0- α mretnim 3 )20.0( )40.0( )21.0( )13.1( )60.0( .e.s **42.2- 20.0- 40.0 **66.2- 30.0 α mretxam 4 )00.1( )20.0( )40.0( )41.1( )20.0( .e.s 90.6 84.0- 51.1- 21.42 19.0 α tlumison 5 )62.5( )46.0( )22.1( )0.93( )08.0( .e.s 37.1- 11.0- 81.0 8.15- 50.0 α redneltlumison 6 )71.4( )76.0( )64.1( )9.54( )37.0( .e.s ***69.6- 45.0- 20.0- **5.64 74.0 α revolloron 7 )84.2( )63.0( )68.0( )6.81( )87.0( .e.s **24.9- 54.0 38.0- 3.71 76.0- α gnilooc 8 )57.4( )55.0( )37.0( )3.34( )24.0( .e.s 41.0- 62.0 42.1 4.41 20.0- α dednetxe 9 )20.0( )94.0( )68.0( )0.54( )56.0( .e.s 743,354,15 743,354,15 743,354,15 743,354,15 743,354,15 N dna ycneuqnileD .)1( noitauqE ni noitacfiiceps eht gnisu noisserger etarapes a si nmuloc hcaE :setoN slortnoc evah snoissergeR .serauqs tsael yranidro era tser eht ;snoisserger tiborp era gniworroB taepeR etats eht ta deretsulc srorre dradnats evah dna ,htnom yreve rof seimmud dna srotcaf cimonoceorcam rof setoned *** dna ,level %5 eht ta ecnacfiingis setoned ** ,level %01 eht ta ecnacfiingis setoned * .level .level%1ehttaecnacfiingis 36

segnahCwaLgnisUsnoissergeR .5 ELBAT egnahCfoetaD gniworroBtaepeR ycneuqnileD mreTnaoL eziSnaoL 001$repeeF etatS )stniopegatnecrep( )stniopegatnecrep( )syad( )$( )$( 80/1/01 ***39.31 ***11.1- 35.1 84.7- ***66.5- β HO 1 )47.1( )62.0( )89.0( )55.51( )75.0( .e.s 104,715,25 358,599,15 791,834,25 791,834,25 791,834,25 N 01/1/7 ***11.4 ***36.0- 40.0- ***43.53 ***08.4- β IR 1 )33.1( )31.0( )22.0( )45.11( )44.0( .e.s 134,468,25 818,122,25 261,466,25 261,466,25 261,466,25 N 90/61/6 40.2- 82.0 55.0- *57.62 10.0- β CS 1 )09.1( )62.0( )43.0( )34.51( )55.0( .e.s 01/1/2 ***33.51- ***43.1 ***10.2- **96.02- 91.0 γ 1 )85.0( )56.1( )72.0( )16.9( )62.0( .e.s 250,492,15 800,328,05 253,562,15 253,562,15 253,562,15 N 11/02/5 ***67.6- ***37.1 51.0 ***08.44 ***23.4 β NT 1 )94.1( )42.0( )64.0( )94.7( )14.0( .e.s 204,967,25 290,471,25 634,616,25 634,616,25 634,616,25 N 90/1/1 ***38.52- ***78.6 ***57.91 ***02.62- ***00.11 β AV 1 )17.1( )63.0( )73.0( )04.8( )66.0( .e.s 844,593,25 397,470,25 731,715,25 731,715,25 731,715,25 N 01/1/1 ***37.23- ***67.2 ***78.1 ***47.06- 74.0 β AW 1 )36.1( )22.0( )92.0( )95.6( )54.0( .e.s 468,255,25 061,360,25 405,505,25 405,505,25 405,505,25 N noitacfiiceps eht esu AW dna ,AV ,NT ,IR ,HO setatS .noisserger etarapes a si elbat eht fo llec hcaE :setoN ta ytiunitnocsid eht etamitse snoisserger llA .)3( noitauqE ni noitacfiiceps eht sesu CS elihw ,)2( noitauqE ni .egnahc eht retfa dna erofeb sdnert raenil dna slortnoc sa egnahc-wal-non gnisu ,egnahc wal eht fo emit eht deretsulc srorre dradnats evah dna ,srotcaf lanosaes dna srotcaf cimonoceorcam rof slortnoc evah snoissergeR *** dna ,level %5 eht ta ecnacfiingis setoned ** ,level %01 eht ta ecnacfiingis setoned * .level etats eht ta .level%1ehttaecnacfiingissetoned 37

TABLE 6. CustomerSelectioninStateswithLargeDropsinVolume State RepeatBorrower ProbabilityofAppearance OddsRatio inPre-Period? inPost-Period Yes/No SC No 16.2% 4.34 Yes 45.6% OtherStates No 17.2% 5.28 Yes 52.3% VA No 18.6% 2.77 Yes 38.8% OtherStates No 23.5% 5.61 Yes 63.3% WA No 11.4% 4.82 Yes 38.3% OtherStates No 18.0% 5.37 Yes 54.1% Notes: Repeat Borrower defined as any borrower whose pre-period loans led to indebtedness 3 months after origination a greater proportion of the time than was the median for all pre-period borrowers. A borrower is considered toappearinthepost-periodifheorshetakesanyloaninthepost-period. Preandpost-periodsdefinedbystate,withSCusingitssecondlawchange. Odds ratioscalculatedas p1 / p2 . 1−p1 1−p2 38