Information, Contract Design, and Unsecured Credit Supply: Evidence from Credit Card Mailings

Finance and Economics Discussion Series Divisions of Research & Statistics and Monetary Affairs Federal Reserve Board, Washington, D.C. Information, Contract Design, and Unsecured Credit Supply: Evidence from Credit Card Mailings Song Han, Benjamin Keys, and Geng Li 2015-103 Please cite this paper as: Han, Song, Benjamin Keys, and Geng Li (2015). “Information, Contract Design, and Unsecured Credit Supply: Evidence from Credit Card Mailings,” Finance and Economics Discussion Series 2015-103. Washington: Board of Governors of the Federal Reserve System, http://dx.doi.org/10.17016/FEDS.2015.103. 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.

Information, Contract Design, and Unsecured Credit Supply: Evidence from Credit Card Mailings∗ Song Han Benjamin J. Keys Geng Li Federal Reserve Board University of Chicago Federal Reserve Board September 2015 Abstract How do lenders of unsecured credit use screening and contract design to mitigate the risksofinformationasymmetryandlimitedcommitmentintheabsenceofcollateral? To address this question, we take advantage of a unique dataset of over 200,000 credit card mailsolicitationstoarepresentativesampleofhouseholdsovertherecentcreditcycle—a period that includes the implementation of the CARD Act. We find that while lenders use credit scores as a prominent screening device, they also take into account a wide array of other information on borrowers’ credit histories and financial and demographic characteristics. For instance, the likelihood of receiving an offer is sensitive to the exact timing of a prior bankruptcy filing. We also find that credit market conditions affect the marginal information used in lenders’ offer decisions, as lenders sharply reduced credit supplied to subprime borrowers during the crisis and in response to the CARD Act. Finally, we document that lenders extend multiple distinct offers to the same consumers over a relatively short period, likely designed such that consumers reveal private information in their choice of contract. JEL Classifications: D82, E51, G21, K35 Keywords: credit supply, information asymmetry, credit cards, mail solicitation, personal bankruptcy, CARD Act, household finance ∗TheviewsexpressedhereinarethoseoftheauthorsanddonotnecessarilyreflectthoseoftheFederalReserve Board or its staff. For their helpful comments, we thank Burcu Duygan-Bump, Felicia Ionescu, Michael Palumbo, Min Qi, Karen Pence, Ken Singleton, Annette Vissing-Jørgensen, Jialan Wang and seminar participants at the Centre for Financial Analysis and Policy Annual Conference, the Laboratory of Aggregate Economics and Finance Conference on Credit, Default and Bankruptcy, the 2010 Society of Economic Dynamics Annual Meetings, the Federal Reserve Board, the Federal Reserve Bank of Richmond, Treasury Department, the Office of Currency Comptroller, Bank of Canada, Consumer Financial Protection Bureau, UIC, the 4th FederalReserveSystemASSADay-AheadConference,andthe47thAnnualConferenceonBankStructureand Competition. This paper supersedes an earlier version circulated under the title “Credit Supply to Bankrupt Consumers: Evidence from Credit Card Mailings.” Contact: Song.Han@frb.gov; benkeys@uchicago.edu; Geng.Li@frb.gov.

Information, Contract Design, and Unsecured Credit Supply: Evidence from Credit Card Mailings September 2015 Abstract How do lenders of unsecured credit use screening and contract design to mitigate the risksofinformationasymmetryandlimitedcommitmentintheabsenceofcollateral? To address this question, we take advantage of a unique dataset of over 200,000 credit card mailsolicitationstoarepresentativesampleofhouseholdsovertherecentcreditcycle—a period that includes the implementation of the CARD Act. We find that while lenders use credit scores as a prominent screening device, they also take into account a wide array of other information on borrowers’ credit histories and financial and demographic characteristics. For instance, the likelihood of receiving an offer is sensitive to the exact timing of a prior bankruptcy filing. We also find that credit market conditions affect the marginal information used in lenders’ offer decisions, as lenders sharply reduced credit supplied to subprime borrowers during the crisis and in response to the CARD Act. Finally, we document that lenders extend multiple distinct offers to the same consumers over a relatively short period, likely designed such that consumers reveal private information in their choice of contract. JEL Classifications: D82, E51, G21, K35 Keywords: credit supply, information asymmetry, credit cards, mail solicitation, personal bankruptcy, CARD Act, household finance

1 Introduction Household debt of every type expanded substantially during the credit boom of the 2000s. Previous studies have documented that the growth in secured credit markets, such as mortgagesandautoloans, wasdrivenlargelybyexpandingaccesstoriskyor“subprime”borrowers (Adams, Einav and Levin, 2009; Mian and Sufi, 2009; Keys, Mukherjee, Seru and Vig, 2010). By contrast, there has been considerably less empirical analysis of unsecured consumer credit, where growth was as dramatic as secured credit. For example, credit card debt outstanding grew more than 40 percent in real terms between 1997 and 2008, a period when median household income saw little increase.1 The rapid expansion of unsecured consumer credit is especially remarkable because of the information challenges, incentives, and market environment of unsecured consumer lending. First, it is well-known, beginning with the classic credit rationing literature, that information asymmetry may lead to market failure in the market for unsecured credit (Stiglitz and Weiss, 1981; Riley, 1987; De Meza and Webb, 1987; Hellwig, 1987). Furthermore, even in the absence of information asymmetry, the limited commitment issue remains (Athreya and Janicki, 2006). Because unsecured claims are often wiped out in the event of personal bankruptcy filing, U.S. bankruptcy law does not necessarily provide strong incentives for households to repay their unsecured debt (Fay, Hurst and White, 2002; Agarwal, Liu and Mielnicki, 2003). Indeed, such credit risk has heightened substantially over the past three decades as personal bankruptcyfilingsincreasedfivefold, fromfewerthan300,000filingsin1980toover1.5million filings in 2010.2 Last but not least, the profitability of unsecured lending, like most financial transactions, may be constrained by evolving regulation and is not immune to the fluctuations of credit and business cycles. In this paper, we use a unique sample of over 200,000 credit card mail solicitations linked 1Takingalongerperspective,thegrowthofcreditcarddebthasbeenevenmorespectacular: Whilesecured consumer debtgrew by130 percent (in real terms)between 1980 and2010, creditcard debtgrew 475 percent over this same period. Source: Federal Reserve. 2Source: The Administrative Office of the U.S. Courts. 1

to offer recipients’ credit records to provide novel quantitative evidence on what information credit card lenders use in screening borrowers and designing credit contracts to address information and limited commitment issues. Specifically, guided by existing theory, we focus our empirical analysis on the following four questions: First, what information do lenders use to screen potential borrowers and to design credit card offer terms? Second, how does unsecured credit supply change over credit cycles and respond to regulatory reforms?3 Third, how do lenders react to conspicuous indicators of default risk such as bankruptcy flags? And finally, do credit card lenders adopt strategies to facilitate separating equilibria, where borrowers with different credit risks are sorted into different credit contracts? Our main results are summarized as follows: First, as expected, credit scores play a prominent role in unsecured lenders’ screening of borrowers. However, lenders appear to take a large array of other information, such as the precise timing of personal bankruptcy filing, into account beyond the extent to which this information affects consumers’ credit scores. Thus, the credit score is not a “sufficient statistic” when analyzing the determinants of unsecured credit supply. Further, exploring the geographic heterogeneity of unsecured credit supply, we find that lenders also factor local economic conditions (such as unemployment) into their determination of credit supply, a result that is broadly consistent with the findings of Hsu, Matsa and Melzer (2014), who document regional heterogeneity in credit offers and equilibrium credit outcomes. Second, we find credit supply changed substantially over the past credit cycle. Consumers with “nonprime” credit scores received more credit card offers than prime consumers during the boom, while after the financial crisis these consumers experienced the largest reduction in access to credit, a contraction that persists six years later. This finding corroborates the view that lenders are especially concerned about the endogeneity of consumer responses (in the form of negative unobservable selection) when credit risk is elevated (see Agarwal, Chomsisengphet, Mahoney and Stroebel (2015a)). 3InSection3below,wediscusswhethercreditoffersthemselvescanbetakenasadirectmeasureof“credit supply,” and argue that they are a reasonable (though imperfect) proxy. 2

Furthermore, we find that the Credit Card Accountability Reliability and Disclosure (CARD) Act of 2009 has further limited the supply of unsecured credit to borrowers with greater credit risks. Note that because the implementation of the CARD Act was followed by a stabilization and recovery of the broad economy and credit markets, the net effects of the Act on the supply of credit can be hard to discern. That said, we find that consumers with the lowest credit scores—those who the Act was arguably most intended to protect— have not received any increase in credit card offers during the prolonged recovery. Moreover, reflecting the broad credit market recovery, the Mintel data show that offers of auto loans (not targeted by the CARD Act) increased evenly across the credit score distribution in the post-crisis period. This contrasting experience across types of credit suggests that the CARD Act may have differentially reduced the supply of unsecured credit to consumers with the greatest credit risks. Our results on credit card offers, at both extensive and intensive margins of credit supply, provide a new angle on the effects of the CARD Act, complementing Agarwal, Chomsisengphet, Mahoney and Stroebel (2015b), who examine a large set of existing credit card accounts and argue that the CARD Act effectively reduced fees levied by credit card lenders. The two studies jointly suggest that in the midst of a substantial reduction in credit card mail volume, borrowers became more selective in accepting credit card offers. This selectiveness is broadly consistent with the well-documented phenomenon of household balance sheet deleveraging in the wake of the housing bust (see, e.g. Mian, Rao and Sufi (2013)). The two studies suggest that while the CARD Act may have achieved some of its intended goals to restrict fees, it also may have had the unintended consequence of reducing the supply of credit, particularly to risky consumers.4 Third,theimpactofthemostconspicuousindicatorofaconsumer’screditrisk,abankruptcy flag, on lenders’ credit supply decisions is surprisingly ambiguous and depends crucially on when the bankruptcy was filed. Consumers who filed for bankruptcy fewer than two years 4See also Debbaut, Ghent and Kudlyak (2014) on the impact of the CARD Act on young borrowers, and Ronen and Pinheiro (2014) for a theoretical exploration of the CARD Act’s limitations on re-pricing. 3

earlier are just as likely to receive an offer as comparable nonfilers. In contrast, those who filed for bankruptcy more than five years earlier receive significantly fewer offers, suggesting that lenders are wary of re-filing risk. However, despite relatively small differences in the probability of receiving a credit card offer, offers to filers are more restrictive, more expensive, and provide fewer take-up incentives than offers to their nonfiler counterparts. This sophisticated approach to lending to bankruptcy filers underscores our overall findings that unsecured lenders take into consideration extensive and dynamic information in their decision-making. Finally, we present novel evidence that lenders actively use offers with different terms as a device to facilitate achieving separating equilibria. In particular, lenders not only send different offers to consumers of different credit risk characteristics, but also send offers of different terms to the same consumer over a short period of time. Such a strategy is consistent with search models (as in Butters (1977)) and helps explain the dispersion of borrowing costs among individuals of similar credit attributes as documented in Stango and Zinman (2013). This separating strategy is more commonly used to target higher credit-score individuals, whose credit score alone may provide less information on their “type,” or who may be more valuable customers to lenders in light of the CARD Act’s restrictions on re-pricing and fees. The key innovation of our approach is that we directly identify the supply of credit using a unique dataset of credit card offers. The proprietary survey data not only are linked to subjects’ credit records (both offer recipients and non-recipients), but also contain extensive demographic information, some of which is not allowed by fair credit reporting laws to be used in credit transactions. Existing studies largely focus on borrowers’ behaviors such as their incentives to default (Fay et al., 2002), the equilibrium quantity and price of credit followingmajorcreditevents(HanandLi,2011;Musto,2004;Cohen-Cole, Duygan-Bumpand Montoriol-Garriga,2009), orhowequilibriumloanpricingreflectscreditrisk(Edelberg,2006). Due partly to data limitations, these studies are not able to identify how credit supply per se changes with lenders’ information sets or relevant regulatory and economic conditions.5 5OnenotableexceptionisGrossandSouleles(2002a),whoanalyzeapanelofindividualcreditcardaccounts and are thus able to infer the intensive (but not extensive) margin of credit supply. 4

Importantly, our data cover the time period that overlaps with the latest credit cycle and the implementation of the CARD Act, allowing us to study how unsecured credit supply has evolved over the credit cycle and reacted to tightening regulation. Our paper contributes to a growing literature on information asymmetry and credit contract design amid changing economic and regulatory environments. Recent studies have underscored the critical importance of credit scores in overcoming information asymmetry, particularly in unsecured credit markets (Athreya, Brown, Tam and Young, 2013; Chatterjee, Corbae and Rios-Rull, 2011). Our findings suggest that credit scores alone are not “sufficient statistics”inthecreditcardmarket, aslendersuseanextensivesetofconsumercharacteristics to differentiate between good and bad risks. These results support the view that innovation in information technology has played a crucial role in the expansion of unsecured credit to risky borrowers (Narajabad, 2012; Sanchez, 2010). Finally, our findings in the credit card market enrich our understanding of information asymmetries in consumer credit markets, complementing studies on related issues in auto lending, microcredit, and payday loans, by Einav, Jenkins and Levin (2012), Karlan and Zinman (2009), and Dobbie and Skiba (2013), respectively, as well as in the credit card market by Agarwal, Chomsisengphet and Liu (2010). The rest of the paper is organized as follows. Section 2 briefly reviews the legislative and institutional background related to the U.S. credit card market, then discusses the related theoretical literature and how our results inform this broad research agenda. Section 3 describes the data and presents summary statistics. Sections 4–7 address each of our four research questions mentioned earlier, and Section 8 concludes. 2 Legislative Background and Conceptual Framework 2.1 Legal Background Three areas of regulation—the U.S. bankruptcy code, the Fair Credit Reporting Act (FCRA), and the CARD Act—are most relevant to the unsecured credit market. As discussed before, 5

a key feature of unsecured credit is that the debt owed can be substantially reduced or even discharged outright through a bankruptcy filing. A debtor can file under Chapter 7 of the bankruptcy code to obtain a discharge of unsecured debts.6 Alternatively, the debtor can file under Chapter 13, thereby obtaining a debt discharge after paying off a portion of the debt through a three- to five-year debt repayment plan. The bankruptcy code also affects the post-bankruptcy supply of credit through its restriction on repeated filing. Specifically, a debtor is prohibited from obtaining another bankruptcy discharge (Chapter 7) until eight years after a previous debt discharge.7 The empirical implications of such re-filing restrictions on credit supply are explored in detail later in the paper. The supply of unsecured credit is also affected by how credit scores are derived and how credit information is reported, particularly for borrowers with damaged credit histories. To comply with the FCRA (and the Equal Credit Opportunity Act, or ECOA), credit scoring algorithms may use only the information on one’s credit history and cannot use information such as race, gender, and income. In addition, the FCRA permits a bankruptcy record to stay on credit reports furnished by the credit bureaus for at most 10 years after the date of relief or the date of adjudication (FCRA 605 (a)(1)), and all other non-bankruptcy defaults for up to seven years (FCRA 605 (a)(5)). If credit reports are lenders’ only source for borrowers’ default histories, then lenders cannot distinguish between the consumers who filed more than ten years before from those who never filed for bankruptcy. Indeed, Musto (2004) finds that filers’ credit scores increase appreciably after their bankruptcy flags are removed, inducing greater access to and subsequent use of credit. The CARDAct was enacted in May2009 and took effect in February 2010. Thelegislation strengthensconsumerprotectionforcreditcardcontractsandimposesvariousnewrestrictions on credit card lending. For example, the CARD Act limits the fees that can be charged on some cards, most notably types of over-limit fees and “subprime” fees. In particular, the Act 6Some debts, such as student loans and unpaid tax liabilities, are deemed not dischargeable. See, for example, Administrative Office of the United States Courts (2006). 7TherestrictiononrepeatedChapter7filingwassixyearspriortothe2005BankruptcyAbusePrevention and Consumer Protection Act. 6

limits non-penalty fees to 25 percent of the total amount of the card’s credit line. In addition, the law bans most rate increases on existing balances (such as in the event of a late payment) and requires introductory or promotional interest rates to last at least six months, thereby largely limiting lenders’ ability to quickly re-price contracts based on risk. On balance, the CARD Act makes credit card lending to risky consumers more restrictive, which in turn may lead lenders to reduce the supply of unsecured credit to such consumers, a hypothesis we explore below.8 2.2 Theoretical Motivation and Empirical Hypotheses We now present a stylized model of credit card offering (along the lines of Einav et al. (2012)) astheconceptualframeworkthatguidesourempiricalanalysis. Recentresearchhasgenerated a great deal of insights on how lenders use contract design and modern technology, including credit scoring, to mitigate information asymmetry and limited commitment issues. We will use this framework to develop several specific hypotheses to be tested using the credit card offer data. 2.2.1 Credit Card Offers with Information Asymmetry Consider a two-period model, t = 0,1. At t = 1, after learning his earnings y , a consumer 1 with credit card balance L and rate of interest R decides whether to pay off this debt. He has only limited commitment to the debt repayment because the legal environment gives borrowers options to default on their debt obligations by either filing for bankruptcy or choosing “informal” bankruptcy (Rea, 1984; Fay et al., 2002; Athreya and Janicki, 2006).9 Denote his utility by v(y ) at t = 1 if he defaults, where v reflects costs of default, including the possibil- 1 ity of being excluded from the credit market, social stigma, efforts of avoiding collection, or 8For more details, see Consumer Financial Protection Bureau (2013). 9The term “informal bankruptcy” refers to a scenario where a borrower chooses to default but not to file forformalbankruptcy(AusubelandDawsey,2004). Insuchascenario,statelawsgoverntheextentofdebtor liability and creditor collection rights, typically allowing creditors to pursue repayment more aggressively using methods such as wage garnishment. 7

legal expenses related to court filing (Athreya, Tam and Young, 2009). Thus, the consumer will pay off the debt if and only if u(y −L(1+R)) ≥ v(y ). (1) 1 1 Under fairly general conditions, this simple setup implies that the borrower is more likely to default on loans with a higher interest rate or a larger balance. This result is an important insight by the classic credit rationing literature: credit terms affect the risk of default (Stiglitz and Weiss, 1981). At t = 0, if the consumer receives a credit card offer, φ, consisting of credit limit Lu and interest rate R, he determines the optimal balance L∗ according to: W(L∗;φ,Z) = max [u(y +L)+βE(max(u(y −L(1+R)),v(y )|Z)], (2) 0 1 1 0<L≤Lu where the distribution of y depends on household and economic conditions, Z. 1 In this context, information asymmetry occurs if the lender observes only a subset of variables Z. Let X indicate the set of information known to both the consumer and the lender, such as payment history, debt balance, and other information in credit reports, and θ information private to the borrower, such as time preference or default stigma. That is, Z = (X,θ). For a given credit offer φ, the set of borrowers who would take it up is given by: Θ(cid:98)(φ,X) = {θ : W(L∗;φ,X,θ) ≥ W}, (3) where W = u(y )+βEu(y ), utility in the financial autarky state. 0 1 It is clear from (3) that the offer terms affect the riskiness of the borrower pool, that is, the distribution of Θ(cid:98). Following Einav et al. (2012), we consider a pooling equilibrium to illustrate how this endogenous risk may affect a lender’s offer decision. That is, given observable characteristics X, the lender chooses φ from the permissible contract set, Φ, to 8

maximize her expected profits: (cid:90) Λ(X;Φ,r) = maxP(Θ(cid:98);φ,X)· π(L∗,φ,X,θ,r)dF(θ)−C(φ,X). (4) φ∈Φ θ∈Θ(φ,X) The above problem illustrates how contract design and information asymmetry may interact to affect a lender’s expected profits. In particular, the lender chooses optimally the offer terms by taking into account the loan demand schedule L∗ and the effect of φ on the takeup pool Θ(cid:98). First, φ affects the borrower pool Θ(cid:98) and the probability of consumers with X taking up the credit offer, P(Θ(cid:98);φ,X). Second, φ affects the lender’s expected profit, net of funding cost r, denoted by π(L∗,φ,X,θ,r) on the loan to a consumer with (X,θ).10 Third, the nature of information asymmetry affects how the riskiness of the pool Θ(cid:98), and in (cid:82) turn, expected profits πdF(θ), changes with φ. Finally, the term C captures the cost of Θ(cid:98) processing information and designing and mailing the offers—which may depend in part on the complexity of information and contract terms. With all of these considerations, the lender mails an offer to the set of consumers with observable characteristics X such that Λ(X,r) ≥ 0. 2.2.2 Credit Scores, Bankruptcy Flags, and Other Borrower Information Withintheaboveframework,aclassicresultisthat,undercertainconditions,adverseselection may occur because unfavorable terms may drive out “good” prospective borrowers, leaving a gradually riskier pool of applicants. When the adverse selection risk is severe, the credit market equilibrium may exhibit quantity rationing. Lenders may address the information issue by more accurately measuring borrower quality to better classify risk types and price credit risk. Indeed, the practice of using credit scores as a summary statistic of certain borrower characteristics (such as the length of credit history, lines of credit capacity, and prior use of credit) has become lenders’ most important tool in credit underwriting and pricing (Pagano and Jappelli, 1993; Edelberg, 2006). In the context 10As an example, denote a default by D = 1 and 0 otherwise, the charge-off rate upon default by g, and the lender’s funding costs by r. Then, π(L∗,φ,X,θ)= (cid:0) (1−ID)(1+R)+ID(1−g)−(1+r) (cid:1) L∗. 9

of the stylized model introduced above, credit scores can be interpreted as a sufficient statistic for a subset of observable consumer characteristics, X. Recent studies suggest that the use of credit scores reduces the level of asymmetric information in the unsecured lending market, leading to an increase in the amount of credit provided, and greater dispersion in loan terms (Athreya et al., 2013; Chatterjee et al., 2011; Einav, Jenkins and Levin, 2013). Importantly, various types of information are not used in credit scoring. For example, by law, credit scores cannot use information on race, national origin, sex, and marital status. Further, credit scores may not use age, assets, or employment history. Thus, certain information beyond a credit score may be important as well for underwriting unsecured credit (Sanchez, 2010; Livshits, MacGee and Tertilt, 2010; Chatterjee et al., 2011; Narajabad, 2012).11 Empirically, we test the importance of credit scores and other borrower characteristics on both extensive (offer likelihood) and intensive (offer terms) margins of lenders’ credit card offer decisions. While credit scores are designed to predict the likelihood of default over the subsequent period across a range of credit markets, some direct indicators of creditworthiness may be valuable to unsecured lenders beyond their contribution to the credit score. For instance, theorysuggeststhatmakingbankruptcyanddebtdischargehistoryinformationbroadlyavailable to the credit market may mitigate both adverse selection and moral hazard issues (Pagano and Jappelli, 1993; Padilla and Pagano, 2000; Ordonez, Perez-Reyna and Yogo, 2014). The main premise is that bankruptcy filers can be penalized by being denied for future access to credit markets for a long period of time or forever, increasing the costs of filing for personal bankruptcy. 11Narajabad (2012) and Sanchez (2010) focus on the impact of improvements in information technology on thequalityofsignalsreceivedbylenders. Intuitively,whencreditratingtechnologyisweak,themarketcannot distinguishacrossrisktypes,andapoolingequilibriumarises. Incontrast,ifratingtechnologyimproves,then this information provides lenders with enough guidance to separate types. Narajabad (2012) shows that this improvement in screening leads to a large expansion of credit to low-risk borrowers, and a relatively smaller decreaseofcredittohigher-riskborrowers. Alternatively,Livshitsetal.(2010)arguethatfinancialinnovation lowered the fixed cost of offering new contracts and helped spur growth in the market for unsecured credit. These improvements in credit rating technologies have a bigger impact on the extensive margin, as new lendingcontractscantargetriskierborrowers. Theexpansionontheextensivemarginleadstobothincreased borrowing and increased default. 10

From a lender’s perspective, however, a bankrupt consumer presents both a risk and an opportunity. On the one hand, bankruptcy records generally send a negative signal to lenders regarding consumers’ risk and time preferences, their ability to manage debt, and the uncertainty of their income.12 On the other hand, because bankruptcy allows for the discharge of most unsecured consumer debt, filers emerge from their bankruptcy proceedings with cleaner balance sheets than prior to filing. Moreover, the law’s refiling restriction described earlier effectively prevents recent filers from repeatedly filing. Thus, access to credit for borrowers with a default history presents an interesting subject for empirical analysis. Our estimates below of the dynamic impact of bankruptcy filing on credit card access can enhance the calibration of sophisticated models of unsecured credit and default (e.g. Chatterjee, Corbae, Nakajima and Rios-Rull (2007) and Livshits, MacGee and Tertilt (2007)). More broadly, understanding the information set used by lenders allows for more accurate modeling of the partition between X and θ in the above framework, and establishes the scope for information asymmetry in light of the rapidly expanding use of “big data” in credit risk analysis. 2.2.3 Credit Cycles and the CARD Act Notably, how the quantity and terms of credit vary with borrowers’ characteristics may also evolve over the credit cycle. Within the theoretical framework discussed above, credit cycles may manifest themselves through systematic changes in lenders’ funding costs, r (see, e.g., Arnold and Riley (2009)) and the permissible contract sets, Φ, in that the bust of a credit cycle leads to increased cost of funds, and in turn reduces the set of non-negative NPV contracts. Accordingly, we hypothesize that the offer likelihood is more sensitive to marginal consumers during the bust of the credit cycle. Higher funding costs affect the intensive margin of credit contracts as well. For example, as shown in Arnold and Riley (2009) and Han (2004), 12Bankruptcy not only signals borrowers’ default risks, it may also alter borrowers’ demand for credit, to which the terms of credit supply may react. Ex ante, information sharing affects debtors’ default incentive in that v(y ) in eq. (1) depends on bankruptcy filing status. Ex post, like other consumers, bankruptcy filers 1 need credit for smoothing consumption and facilitating transactions. 11

equilibrium loan amounts tend to decrease and loan rates increase with the costs of funding. Therefore, we hypothesize that during the bust we should observe lower credit limits Lu and higher interest rate spreads R. Regulation of the credit card industry is another key factor that influences the supply of credit. For example, regulation that limits the feasible contract space may reduce lenders’ ability to design contracts with the nuance to separate types in the presence of information asymmetry. Certain fees or interest charges may be specifically designed to compensate for taking on heightened default risk or greater funding costs. In this context, we study the impact of the CARD Act on the supply of unsecured credit through changes in credit card offers. The CARD Act’s impact on credit card lending has attractedagreatdealofattentionfrombothpolicymakersandacademics. Ouruniquedataset oncreditoffershelpsachievebetteridentificationofthesupplyeffectoftheActwithoutrelying on an instrumental variables approach to separate supply and demand. Because the CARD Act was implemented at a time when both the macroeconomy and unsecured credit market began to recover from the aftermath of the financial crisis, discerning the restrictive effects of the Act can be challenging. We therefore focus on a “difference-in-differences” thought experiment to compare the changes in credit card offers received by consumers across the credit score distribution and ask whether the CARD Act had a disproportionate effect on consumers with different levels of creditworthiness. We also use data on auto loan offers to provide a contrast (or “triple-difference” thought experiment) for the CARD Act’s impact on the unsecured lending market, as the Act did not affect the regulatory environment of the auto lending industry. 2.2.4 Lenders’ Strategies for Facilitating Separating Equilibria Arguably, even with the most advanced information collection and screening technologies, unsecured lenders still face challenges related to information asymmetry and unobservable consumer risk. The pooling equilibrium discussed above assumes that lenders treat borrowers 12

withthesameobservablesequally. However,lendersmayfurthermitigatetheadverseselection problem through more sophisticated contract designs. The key insight from the literature is that there has to be some costly and credible signaling device that low-risk types of borrowers arewillingtopay, buthigh-risktypesarenot, resultinginendogenousseparationbyrisktypes in the equilibrium. Importantly, some commonly-used screening devices, such as collateral (Bester,1985), arenotavailableforunsecuredlending. Instead, unsecuredlendershavetorely on alternative mechanisms to solve the information and limited commitment issues. These mechanisms may include various forms of exclusion on defaulted borrowers (e.g., Athreya and Janicki (2006) and Ordonez et al. (2014)), social stigma attached to loss of creditworthiness (Athreya, 2004; Livshits et al., 2010), and implicit collateral.13 Our empirical analysis focuses on a largely unexplored aspect of credit card lenders’ contracting effort for achieving separating equilibria. Specifically, we exploit the panel structure of a subset of our data (discussed below) and examine the common strategy of sequentially sending distinct offers to the same consumer over time. If consumers only engage in a limited search for the best offers, such offer strategies will lead to a dispersed interest rate distribution even among consumers with similar characteristics, a pattern documented by Stango and Zinman (2013). One interpretation is that the extensive set of contract features in the permissible contract set Φ allows for sufficient dimensionality to allow borrowers with different types or different private information to self-select into their preferred contracts. Alternatively, credit card contracts are notoriously complex, and most consumers do not take the time to read the entirety of the credit card offers they accept, instead focusing on the most salient aspects of the contract.14 Our panel data on card offers from the same lender to the same consumer provide an unprecedented look into how banks conduct screening to identify profitable borrower-contract matches in a market with information asymmetries. 13For instance, most student loans are non-dischargeable in bankruptcy filing, effectively converting human capital into collateral in student lending. 14See Agarwal, Chomsisengphet, Liu and Souleles (2006), Agarwal, Driscoll, Gabaix and Laibson (2011), Gabaix and Laibson (2006), and DellaVigna and Malmendier (2004). 13

3 Data and Summary Statistics 3.1 Data Description Our main data source is Mintel Comperemedia’s (henceforth “Mintel”) proprietary survey on credit offers to U.S. consumers.15 Our data span from January 2007 to June 2014, covering three distinct phases of the most recent credit cycle. The period between January 2007 to March 2008 largely covers the final episode of the credit boom.16 The period between April 2008 and February 2010 covers the credit bust and early recovery prior to the full implementation of the CARD Act. Finally, the remainder of our sample (March 2010 to June 2014) covers the recovery period under the CARD Act. On average, about 2,500 households participate in the Mintel survey each month by forwardingallincomingmarketingmail, includingoffersforcreditcardsandautoloans, toMintel and completing an extensive demographic questionnaire. After processing the forwarded mail offers, Mintel sends the database to TransUnion, one of the three major credit reporting agencies, where credit history information, including credit scores, of the individual consumers of the participating households is merged in. Thus, our data provide a unique combination of detailed information about credit card offer terms, credit history, and demographic and socioeconomic characteristics, which is rarely available in other data sources. Prior to July 2011, the data consisted of a cross-section of households surveyed in each month. After that date, a longitudinal sample of 600 consumers on average replaced a portion ofthecross-sectionalsample. Unlikethecross-sectionalsample, however, wehaveonlylimited demographic and socioeconomic information—age, income, and homeownership—as of the monthjoiningthepanelfortheconsumersinthelongitudinalsample. Therefore,thesummary statistics and baseline analysis we present use only the cross-sectional sample. Meanwhile, 15Mintel is a consumer and marketing research company headquartered in the U.K. The data we use are compiled by the company’s American subsidiary, Comperemedia. 16We choose March 2008 as the end of the boom because Bear Stearns was bailed out that month and the decline of credit card mail volume accelerated after that month. Our results are robust to alternative choices of the end of the credit boom. 14

we use the longitudinal sample to examine lenders’ offer strategies, particularly contract dispersion, in offers made to the same borrower from the same lender. To obtain a consistent sample, we restrict the cross-sectional sample to individuals who have a valid credit history, whose household income was between $10,000 and $200,000 (trimming the top and bottom 2.5 percent of the sample household income distribution), and who are between 20 and 60 years old. The final cross-sectional sample contains about 219,700 credit card offers and about 170,000 individuals in more than 105,000 households. Columns (1) and (2) of Table 1 compare the demographics of the Mintel cross-sectional sample with those of the households in the 2007, 2010, and 2013 waves of the Survey of Consumer Finances (SCF) who meet the same income and age restrictions.17 We find that Mintel sample households are broadly comparable with the SCF respondents, with the Mintel sample being, on average, somewhat older, having higher educational attainment and income, andmorelikelytobewhite, married, andhomeowners. Thesedifferencesareduepartlytoour sample restriction to only those Mintel respondents who have a valid credit history. Because the credit records are merged using survey participants’ names and addresses, homeowners, who tend to have more stable addresses, are more likely to have a successful merge. While no personally identifiable information (PII) was provided to the researchers by TransUnion as part of this study, the data merged by TransUnion contain rich anonymized information about Mintel respondents’ debt balances and credit histories. Columns (3) and (4) of Table 1 compare selected attributes of the credit history in the Mintel data with those in the FRBNY CCP/Equifax data (a five percent random sample of U.S. consumers who have a valid credit history) between 2007 and 2013.18 As the table shows, the liability and credit historycharacteristicsoftheconsumersintheMintelsamplearebroadlyconsistentwiththose in the FRBNY CCP/Equifax data, with the former having somewhat lower amounts of debt but more lines of revolving credit. The two samples are similar regarding the frequency of 17All statistics are estimated using the weights provided by Mintel and the SCF, respectively. 18See, for example, Lee and van der Klaauw (2010) for a more detailed description of the FRBNY CCP/Equifax data and a discussion of the data’s statistical properties. 15

personal bankruptcy, serious delinquencies, and other derogatory records. Thus, on balance, the Mintel sample of participating consumers is fairly representative of U.S. consumers in terms of demographic and credit characteristics. Finally, the TransUnion credit history data contain a credit score measure, the VantageScore 2.0 credit score.19 Over the last 30 years, credit scores have become an increasingly prominent factor in consumer lending (see, for example, Federal Reserve Board (2007)). This credit score product, which ranges from 501 to 990, is developed by the three major consumer credit reporting agencies. Consumers whose credit scores are greater than 700 are often labelled as prime or superprime consumers, while those with a credit score below this level are typically referred to as nonprime or subprime consumers. The credit record information that is provided through the Mintel-TransUnion merge is similar to what a lender would receive through a “soft pull” of the credit record, a close approximation of the lender’s information set in the absence of an existing relationship with a consumer. 3.2 Mail Offers as a Measure of Credit Supply Credit card offers in the Mintel data represent not only lenders’ marketing efforts, but also their desired supply of credit given the information available and the economic conditions at the time of mailing. First, because it is costly to design and send an offer, it would be inefficientforalendernottoprovidecreditunlesstheapplicationrevealsnewinformationthat lowers the expected profits below the lender’s break-even point. Indeed, our conversations withlendersrevealedthattheytypicallyconductacomplicated, multi-stagescreeningprocess, verysimilartocreditunderwriting, inselectingcreditcardofferrecipients. Thiscostlyprocess implies that, ex-ante, lenders treat offers as their committed supply of credit. See also our discussion in Section 2.2 on the lender’s decision to extend a credit offer. In addition, both the level and the change in the aggregate volume of mail offers are highly correlated with other indicators of the aggregate supply of unsecured credit. Specif- 19Throughout the paper, our statistical analyses involving credit scores all use the VantageScore 2.0 credit score. 16

ically, total credit card mail volume (as estimated by Mintel) is highly correlated with the aggregate number of credit card accounts opened (estimated using data from the FRBNY CCP/Equifax). As shown in the upper panel of Figure 1, the two time series track each other very closely over the last ten years, with a correlation coefficient of about 0.9. Moreover, as indicated in the lower panel, the quarter-to-quarter change in credit card offer mail volume and the net-easing of credit card lending conditions reported in the Federal Reserve’s Senior Loan Officer Opinion Survey on Bank Lending Practices (SLOOS)—a widely-used gauge of aggregate credit supply—are also positively correlated.20 That said, we caution that credit card offers are not exactly equivalent to credit supply in various aspects. Offers generally contain clauses to allow lenders to react to any new information provided by applicants or changing economic conditions. Thus, lenders have the option of not approving an application responding to an outstanding offer, despite some offers being so-called “pre-approved” offers. For an approved offer, the credit limit ultimately extended is not necessarily the same as the amount specified in the offer. In addition, discussions with various major credit card lenders suggest that the volume of mail offers is also affected by lenders’ marketing strategies and budget limitations, which may not always reflect changes in willingness-to-lend.21 The Mintel database records essentially all information on the forwarded mail offers, allowing us to study not only whether a consumer receives any credit card offers, but also the full set of terms of the contracts offered. For our baseline analysis, we examine five key parameters of an offer—interest rates, credit limits, annual fees, promotional interest rates, and reward programs. Depending on the presence of annual fees or reward programs, a credit card can be categorized into one of the four types of card according to industry practice: plain vanilla (no fee, no reward), credit building (fee, no reward), general market (no fee, reward), 20Note that the metric available in the SLOOS is the share of banks reporting a tightening or loosening of lending standards, which only roughly maps into direct measures of changes in mail volume. The two series have a statistically significant correlation coefficient of 0.45. 21For example, the sharp decline of mail volume in the first quarter of 2012 was primarily due to the reduction in solicitations sent by two major credit card lenders that market participants attributed to shifts in marketing channels at these lenders. 17

and premium rewards (fee, reward). To measure the price of credit, we focus on the so-called “go-to” interest rate—the regular non-promotionalinterestrateforpurchases.22 Regardingcreditlimits, thedatarevealarecent change in industry practice. Historically, credit card offers have usually specified a maximum credit limit. However, since 2006, an increasing share of credit card offers have specified a minimum credit limit, and by early 2009, the vast majority of credit card offers only specified a minimum credit limit.23 Our analysis will thus focus on the minimum credit limit. Table 2 summarizes these characteristics of credit card offers in the Mintel sample. As column (1) of the table indicates, on average, 50 percent of consumers receive at least one credit card offer in a given month.24 Moreover, nearly 40 percent of consumers receive at least one general market offer while only 7.5 percent of consumers receive a credit building offer. The offers on average have a regular purchase interest rate of 13.7 percent and a minimum credit limit of $1,158. In addition, roughly two-thirds of these solicitations offer introductory interest rates and rewards programs, whereas nearly 20 percent of the offers require an annual fee. Columns (2) and (3) of Table 2 compare the offers received by prime (credit score≥ 700) and nonprime (credit score< 700) consumers. Prime consumers are more likely to receive an offer and the offers they receive are more likely to be general market and premium rewards offers, whereas nonprime consumers are much more likely to receive a credit building offer. Furthermore, offers received by prime consumers on average have lower interest rates, higher credit limits, and are more likely to offer introductory rates and rewards programs, but less likely to charge an annual fee. This simple split of the data suggests that credit score alone is a strong predictor of both the frequency and characteristics of credit offers. In the next 22Mintelalsorecordsotherinterestratesspecifiedintheofferssuchastheinterestratesonbalancetransfers and cash advances. Broadly speaking, these offered interest rates exhibit similar contrasts between filers and nonfilers and dynamics over the credit cycle. For more on interest rate pricing, see Ausubel (1991), Stango (2000), and Knittel and Stango (2003). 23Examining offers sent in 2007 that specified both a minimum and maximum credit limit, we find the two limits are positively correlated, with a Spearman rank-order correlation coefficient above 0.5. 24Not shown in the table, among those consumers who received offers, the average number of offers is 2.5 per month. 18

section we explore these relationships in greater detail. 4 Screening in the Unsecured Credit Market 4.1 Baseline Model of Offer Likelihood We begin our analysis with an exploratory model that examines what characteristics influence a consumer’s chances of receiving an offer in a given month and the terms contained in the offers received. The stylized framework in Section 2.2 suggests that lenders’ offer decisions depend on consumers’ characteristics (including their credit scores), macroeconomic conditions, and the regulatory and legal environment. We estimate the following model for the likelihood of receiving an offer (using a probit specification) and for the offer terms conditional on offer receipt (using OLS and probit where applicable) for consumer i in state j in month t: (cid:18) (cid:19) Y ijt = f VS it ,Flag it ,Attr it ,Demo it ,Law jt ,Econ jt ,δ t +(cid:15) i (5) To allow for nonlinear effects of credit scores on the supply of unsecured credit, we specify the effects of credit scores VS nonparametrically by including dummy variables for 50-point it bins in credit score. A prominent question is if lenders use credit scores alone in screening borrowersoralsotakeintoaccountotherfactorsbothinandoutofaconsumer’screditreport. To keep our model parsimonious, we include a relatively small set of key credit attributes in the model. Of note, such variables are also likely used for estimating credit scores. Because we control for the effects of credit scores in a highly flexible way, we argue that the effects estimated reflect the additional weight lenders put on these credit history variables on top of their impact credit scores. Specifically, we include a set of dummies, Flag , indicating flags of adverse credit events: it personal bankruptcy, severe derogatory records (e.g. debt collection or foreclosure), deep 19

debt delinquencies (90 days and longer), and recent debt delinquencies within the previous 24 months. In addition, we include a set of credit attributes, Attr : the total debt-to-income it ratio to reflect general indebtedness (Johnson and Li, 2010), a dummy for having credit cards, a dummy for high credit card utilization (the ratio between outstanding balances and credit limits over 80 percent), and the number of credit inquiries over the past six months. Including the number of credit inquiries, usually associated with a loan application, helps shed light on whether lenders’ actions respond to variation in credit demand. Furthermore, we explore whether consumers’ demographic and financial characteristics, Demo , affectlenders’decisionsbyincludingage, maritalstatus, familysize, race, educational it attainment,homeownerstatus,andincome.25 Wearealsointerestedinwhetherlenders’credit card mailing decisions are influenced by the legal conditions (such whether state law is more favorable to borrowers in the event of default), Law , and economic environment, Econ , of jt jt the consumer’s state of residence. Specifically, we include state-level property and homestead bankruptcy exemptions, and state unemployment rates. Finally, we include year and month fixed effects, δ , to control for aggregate macroeconomic and credit market conditions and t potential seasonal effects. 4.2 Baseline Model Results In column (1) of Table 3, we report the estimated marginal effects of the probit model of offer likelihood for the full sample period (2007–2014). First, as expected, consumers with higher credit scores are, on balance, more likely to receive a credit card offer in a given month. Consumers in the highest credit score bin (> 950) are 26 percentage points more likely to receive an offer than consumers in the lowest score bin (< 550). Also, we note that the relationship between credit scores and the likelihood of receiving an offer is nonlinear and non-monotonic, as consumers with credit scores between 750 and 850 are most likely to 25ThesecharacteristicsarecollectedandmadeavailabletousbyMintel. Someofthesecharacteristics,such as race and gender, are prohibited by law (ECOA) from being used by lenders in credit transactions. The estimated effects of these characteristics may reveal the degree to which lenders use these variables directly, or legitimate variables that are correlated with them in determining credit card mailing strategies. 20

receive an offer. Despite the fact that we allow for a very flexible specification of the effect of credit scores on offer probabilities, other credit history variables nonetheless have a significant impact on the likelihood of receiving an offer. In particular, consumers with personal bankruptcy flags or other severe derogatory records are about six percentage points less likely to receive an offer (Table 3, column 1). In addition, consumers with severely delinquent accounts are three percentage points less likely to received an offer, while having a recent delinquent account also has a small but statistically significant negative effect. On net, lenders are more likely to extend offers to consumers who have existing credit card accounts but less likely to extend offers to those who have high utilization rates (higher than 80 percent) on existing cards. Notably, recent credit inquiries have no independent effect on offer likelihood, suggesting that lenders do not simply target those consumers who are actively seeking credit. Interestingly, several socioeconomic characteristics also appear to influence lenders’ offer decisions, even after controlling for credit histories. For instance, homeownership, college education, and higher household income all boost the likelihood of receiving an offer. Notably, white consumers are almost three percentage points more likely to receive an offer than otherwise identical nonwhite consumers. Indeed, our finding motivated Firestone (2014) to explore possible explanations for this disparity, including, but not limited to, the existence of omittedvariables,modelmisspecification,ordisparateimpactinlenders’marketingstrategies. Therefore, while the use of some socioeconomic variables, such as race, is prohibited, the driving factors behind these results warrant further research. Finally, lenders’ offer decisions depend on state laws and local economic conditions. For example, our estimates indicate that a one percentage point increase in a state’s unemployment rate implies a 1.6 percentage point reduction in the likelihood of receiving an offer. This macroeconomic effect is broadly consistent with Hsu et al. (2014) who find a higher volume of credit card mailings in states with more generous unemployment benefits. In addition, we find that a $100,000 increase in homestead and property exemptions in personal bankruptcy 21

filings imply a 0.3 and 3.5 percentage points reduction in the chances of receiving an offer, respectively. These estimates provide new support to an extensive literature on the impact of the bankruptcy option on credit availability (see, e.g. Gropp, Scholz and White (1997)). 4.3 Results on Offer Terms We now turn our analysis to the factors lenders potentially take into account when deciding the terms in their credit card offers. We examine how the same set of credit, demographic, and financial characteristics, as well as state-level legislative and economic conditions, affect the quantity and price of credit. Specifically, we consider the following terms in the offer—minimum credit limits, interest rate spreads, and whether the contract includes an introductory teaser interest rate, an annual fee, or a rewards program.26 The minimum credit limits and interest rate spreads models are estimated using OLS regressions, while models of whether having teaser rates, annual fees, and rewards programs are estimated using probit regressions.27 As shown in columns (2) through (6) in Table 3, offer terms are generally improving as credit scores increase. Although the relationships are not monotone, in part because of the presence of “premium rewards” cards that charge an annual fee, the pattern over the score distribution is consistent across offer terms. The estimated coefficients of other variables are broadly consistent with those in column (1). For example, conditional on credit scores, offers to consumers with bankruptcy flags and derogatory public records have lower credit limits, while offers to consumers who are white, have higher educational attainment or higher income have higher credit limits.28 The results for pricing variables reveal a similar pattern. Offers to 26As discussed in Section 3, the vast majority of credit card offers mailed during our sample period specify only a minimum credit limit. Moreover, we consider interest rate spreads (relative to the two-year Treasury yield) instead of interest rate levels to take into account variation in risk-free rates. 27Because the distribution of minimum credit limit and interest rate spread is bounded at zero, we also estimated these specifications using Tobit models, which yielded results similar to the OLS estimates (not shown). 28The notable exception is the coefficient of personal property exemption level in bankruptcy filings. Our results indicate that offers to consumers living in states with higher exemption levels are less likely to receive anoffer, butconditionalonreceivinganoffer, theseconsumerstendtoreceiveofferswithhighercreditlimits. 22

consumers with credit scores between 550 and 600 seem to have the least favorable terms— higher interest rate spreads, greater likelihood of having an annual fee, but lower likelihood of having a teaser rate or rewards program. Notably, despite the extensive set of explanatory variables and flexible specifications, our models explain only a relatively small portion of the overall variation in contract terms. For instance, the largest R-squared we obtain is for the interest rate spread, where we can explain only 35% of the variation using observable borrower and geographic characteristics (and time fixed effects). Our low R-squareds are remarkable because the amount of information we use is similar to what a lender would have at its disposal in screening a consumer without a prior business relationship. This large amount of unexplained cross-sectional variation in offer prices may reflect the practice of lenders varying contract terms to induce borrowers to reveal private information through self-sorting, a hypothesis that we test later using our longitudinal sample. 4.4 Evidence from the Longitudinal Sample So far we have demonstrated that credit scores are one of the most important indicators that lenders use in determining credit card offers for a cross-section of borrowers. Do lenders also monitor a borrower’s credit score over time and dynamically adjust their offers accordingly? In answering this question, we explore the longitudinal sample of the Mintel data, containing 13,800 offers to more than 600 households over a three-year period (July 2011 to June 2014). Compared to the cross-sectional sample, which we used in Table 3 above, the longitudinal sample has limited socioeconomic information. However, the advantage of the panel data is that we can include consumer fixed effects in our specifications, which negates the need for time-invariant household characteristics. Accordingly, we estimate a panel regression model of how changes in a consumer’s credit score affects his likelihood of receiving credit card offers and the terms therein, controlling for consumerfixedeffects. TheresultsarereportedinTable4. Ourestimatesindicatethatlenders 23

monitor borrowers’ credit scores and use the new information for credit card offer decisions. In particular, for the same consumer, a 100 point increase in credit score is associated with a 4 percentage point greater likelihood of receiving an offer, a $240 increase in the minimum credit limit, and a 100 basis point reduction in interest rates. In addition, improved credit scores also boost the chances of receiving credit card offers with reward programs. However, changes in credit scores do not appear to influence other offer terms such as introductory interest rate promotions and annual fees. Overall, these results suggest that lenders respond to short-run changes in consumers’ credit positions, and indeed adjust their offers accordingly. In Section 7, we further explore the longitudinal patterns in lender screening by examining multiple distinct offers made by the same lender to the same borrower. 5 Lenders’ Responses to the Credit Cycle and the CARD Act We now examine how the supply of unsecured credit responds to credit cycles and the CARD Act by taking advantage of three distinct periods covered by our Mintel data. Figure 2 presents the likelihood of receiving a credit card offer in a given month by credit score bin over the three periods (boom, crisis/pre-CARD Act, post-CARD Act). The figure shows that thecreditscoregradientinthelikelihoodduringtheboomperiodofJanuary2007–March2008 (thesolidblueline)wasremarkablyflat. Indeed, 40percentofconsumerswiththeworstcredit scores(creditscorebelow550)receivedacreditofferinagivenmonth, comparedto60percent of consumers with the best credit scores (credit score above 950). If anything, consumers in the subprime and near-prime range of the credit score distribution, between 600 and 750, were more likely to receive an offer than any other part of the credit distribution. This pattern highlights the dramatic expansion of unsecured credit to less creditworthy consumers during the credit boom, a trend previously shown in other credit markets (Adams et al., 2009; Mian and Sufi, 2009). 24

In the wake of the crisis, April 2008–February 2010, access to unsecured credit dropped precipitously, as lenders cut existing lines and significantly curtailed credit card mail offers. The overall likelihood of a consumer receiving an offer in a given month fell from 60 to 35 percent, but this decrease was not felt evenly over the credit score distribution. As shown by the orange line in the figure, the credit score gradient steepened sharply during this time period, with consumers at the top of the credit score distribution becoming about five times more likely to receive an offer than those at the bottom. We note that this dramatic reduction in offers to risky borrowers supports the view that lenders were concerned about the disproportionate impact of the Great Recession on subprime borrowers, as well as the potential for endogenous selection concerns (discussed in Section 2.2) to intensify in that part of the creditworthiness distribution (Agarwal et al., 2015a). The overall volume of credit card mail offers has steadily recovered since early 2010, but the recovery has been uneven across the credit score distribution. As indicated by the purple line, following the implementation of the CARD Act (March 2010–June 2014), the likelihood of receiving an offer increased for consumers with credit scores above 650. However, those located in the bottom of the credit score distribution did not see any improvement in their odds of receiving an offer. In addition, the improvement for consumers with credit scores between 650 and 750 is more subdued relative to those with higher credit scores. On balance, comparedtotheorangeline,thepurplelinerepresentsasteepercreditscoregradient,implying a wider gap in the likelihood of receiving an offer between the most and the least creditworthy consumers. The trends shown in this figure provide new evidence of the uneven patterns of access to unsecured credit during the boom and bust. Notably, the lack of growth in credit card offers to consumers with greater credit risks after the CARD Act suggests that the Act may have reduced the supply of credit. Such an inference may be complicated by the concurrent broad credit market recovery. As a parsimonious approach to “control” for the broad market effect, we use our Mintel data to examine offers of auto loans, a type of consumer credit that is not covered by the CARD Act, 25

but may move with improvements across all credit markets over this period. As shown in Figure 3, the likelihood of receiving an auto loan offer improved after 2010 across almost the entire credit score spectrum. Moreover, consumers with subprime credit scores (between 600 and 650) consistently have the best chance of receiving an auto loan offer through the entire credit cycle. This contrast with auto lending highlights the isolated effects of the CARD Act on the credit card industry. Not only the odds of receiving an offer but also the price of offered credit changed substantially across the credit score distribution over the credit cycle. Figure 4 presents the spreads of offered regular purchase interest rates over the yield on two-year Treasury securities. We find that spreads widened significantly during the financial crisis and the early phase of recovery before the CARD Act, with larger increases for consumers with lower credit scores. Such a shift in offered interest rates likely reflects the tighter credit supply during this period, in particular for consumers with less-than-pristine credit records. Furthermore, despite the ensuing general improvements in financing conditions, spreads widened even more after the implementation of the CARD Act over the entire credit score distribution, with the increases being particularly pronounced for consumers with lower credit scores. This result is largely consistent with the hypothesis that lenders raised interest rates on new credit card contracts in part responding to provisions of the Act that made such lending more restrictive and unsecured debt more difficult to subsequently re-price. In related work, Agarwal et al. (2015b) examine a large set of credit card accounts and do not find that interest rates increased following the implementation of the CARD Act, despite the reduction in lenders’ fee revenues. Hence, the two complementary studies jointly suggest that in the midst of a substantial reduction in credit card mail volume, borrowers becamemoreselectiveinacceptingcreditcardoffers. Aheightenedselectivenessbyconsumers during this time period is in turn broadly consistent with the well-documented phenomenon of household balance sheet deleveraging (Mian et al., 2013). Our findings underscore that while the CARD Act may have achieved some of its intended goals, it may have also had the 26

unintended consequence of reducing the supply of credit, particularly to risky consumers. Apart from wider interest rate spreads, other terms in credit card offers extended after the financial crisis, before or after the CARD Act, did not appear to be significantly more stringent than those in the offers sent during the era of credit boom. As shown in the upper left panel of Figure 5, the average credit limit in the offers sent during the credit bust and recovery period before the CARD Act (the red curve) was almost identical to that in the offers sent during the credit boom (the blue curve) except for a moderate reduction among consumers with credit scores higher than 850. Furthermore, credit limits increased across the credit score distribution after the CARD Act. In addition, the other panels of Figure 5 show no evidence that the offers sent after the credit boom ended had worse terms, including introductory interest rates, annual fees, or rewards programs, either before or after the implementation of the CARD Act. For example, the lower left panel indicates that the share of credit card offers with an annual fee did not increase during our sample period. Consistent with Agarwal et al. (2015b), this share declined significantly for consumers with credit scores below 650 after the implementation of the CARD Act. Furthermore, post-CARD Act offers were more likely to contain introductory teaser interest rates and rewards programs than pre-CARD Act offers, with the changes being more pronounced for consumers with lower credit scores. In sum, our results suggest that lenders’ responses to the credit crunch appear to primarily focus on whether to extend an offer and the regular purchase interest rate. In contrast, credit offers mailed during the bust became somewhat more favorable regarding credit limits and other salient terms. This finding is consistent with the notion that during this period, lenders’mainconcernwasborrowers’elevateddefaultrisk, whichledlenderstosharplyreduce the volume of credit card offers. However, for consumers deemed as creditworthy based on observables, lenders appeared to have sent them more attractive offers, potentially reflecting both heightened competition for low-risk customers and responses to the requirements of the CARD Act. 27

6 Conspicuous Credit History: The Case of Bankruptcy Flags Our credit card offer data reveal that lenders take into account borrowers’ credit history information in a sophisticated way when making offer decisions. We further illustrate this practice by examining how consumers’ personal bankruptcy flags affect the offers they receive. To the extent that bankruptcy flags are generally interpreted as signaling heightened default risks, examiningcreditcardofferstopersonalbankruptcyfilerspresentsauniqueperspectiveforunderstanding the supply of unsecured credit, in particular to high-risk borrowers. Furthermore, such an analysis sheds light on the economic costs of filing for personal bankruptcy, which should include not only the expenses related to the filing itself but also the costs associated with possible post-filing limitations to accessing credit markets. In this regard, our results inform the literature that examines the household bankruptcy decision and provides crucial statistics for calibrating recent dynamic general equilibrium models for studying unsecured credit markets.29 We first document that consumers are not excluded outright from the unsecured credit market after filing for bankruptcy, even in the aftermath of the most severe financial crisis in recent history. The Mintel data suggest that, on average, nearly 40 percent of consumers with a history of personal bankruptcy receive at least one credit card offer in a given month. We findbothanecdotalandstatisticalevidencethatofferstoconsumerswithabankruptcyhistory are not the result of a non-discriminatory “blanket campaign.” Rather, some lenders design theiroffersspecificallytosuchconsumers, furtherdemonstratingthatlenderstakeinformation on borrowers’ credit history besides their credit scores into account.30 As the estimates in Table 3 suggest, the likelihood of a filer receiving an offer is, on average, only moderately (six 29See,e.g.,Fayetal.(2002),GrossandSouleles(2002b),Keys(2010),DickandLehnert(2010)forempirical work on personal bankruptcy, and Li and Sarte (2006), Chatterjee et al. (2007), and Livshits et al. (2007) on theoretical advances in this area. 30For example, the header of one mail offer from a top credit card lender states “You deserve some credit for getting through bankruptcy.” 28

percentagepoints)lowerthananon-filerwithcomparableobservablecharacteristics, including credit scores. That said, despite relatively small differences in the probability of receiving a credit card offer, offers to filers tend to have lower credit limits, higher interest rates, and fewer take-up incentives than offers to their non-filer counterparts.31 Lenders’useofbankruptcyfilingstatusinscreeningconsumersnotonlyfocusesonwhether a consumer has a bankruptcy history, but also the time elapsed since the previous filing. As discussed in Section 2, U.S. bankruptcy law prohibits repeated debt discharge within eight years after the previous Chapter 7 filing. Therefore, while in general a filer’s credit score gradually recovers after bankruptcy filing, such a borrower potentially represents a greater default risk over time as he approaches discharge re-eligibility. To explore how time since previous filing influences lenders’ credit card offer decisions, we modify equation (5), replacing thefilerdummywithavectorofthreedummies—“recent”filer(filedwithinthelasttwoyears), “seasoned” filers (filed two to five years ago), and “remote” filers (filed more than five years ago). The estimated coefficients of these three dummies regarding each of the six outcome variables are reported in Table 5. The results regarding the likelihood of receiving an offer (column 1) show that, consistent with the implications of the restrictions on repeated debt discharge, lenders appear to view recent filers as carrying lower credit risk. These recent filers are no less likely than similar nonfilers to receive an offer in a given month, while remote filers are 12 percentage points less likely to receive an offer. Nonetheless, conditional on receiving an offer, the offers to remote filers are more favorable than those to recent filers (though they are mostly worse than offers to nonfilers). For example, comparing the estimated marginal effects of a recent bankruptcy filer with those of a remote filer in columns (2)–(5) of Table 5, we find that offers to remote filers have interest rate spreads that are, on average, about 130 basis points lower, are more than 10 percentage points less likely to impose an annual fee, and are about 20 percentage 31Notethatthetotaleffectofbankruptcyalsoincludesadramaticreductionincreditscore. Theseestimates controlforcreditscoresandthuscomparebankruptcyfilerswithnon-filerswhogenerallyhavequitelowcredit scores for reasons related to non-bankruptcy delinquency. 29

points more likely to include a rewards program. We further contrast how the filer-nonfiler disparity evolved over the credit cycle by estimating the above model separately for the three distinct episodes of time in our sample. We present the results of the likelihood of receiving a credit card offer in Table 6. Column (1) repeats the same result shown in Table 5 as a reference point. As shown in columns (2)–(4), recent filers were 8 percentage points more likely to receive an offer than otherwise comparable nonfilers during the credit boom, whereas remote filers were 5 percentage points less likely. The favorable treatment of recent filers largely persisted even during the depths of the credit crunch (April 2008–February 2010). However, the pattern appears to have changed substantially during the post-CARD Act era, when recent filers lost their edge in receiving credit card offers relative to nonfilers, while the gap between remote filers’ likelihood of receiving an offer and that of comparable nonfilers became ever larger. This result suggests that the CARD Act made lenders more concerned about the credit risks of bankruptcy filers, even those who filed very recently and are not eligible for debt discharge in the near future, possibly due to limitations on re-pricing debt. These findings highlight one example of how information on a conspicuous credit risk may influence lenders’ offer decisions, and support the view that credit underwriting in the unsecured market tightened sharply during the Great Recession.32 7 Lenders’ Strategy for Separating Borrowers Although lenders observe detailed credit histories of potential borrowers (as in our data), and thereforecaninferagreatdealaboutborrowers’likelydemandforanduseofunsecuredcredit, the unobserved heterogeneity among potential borrowers remains substantial. Our analysis indicates that lenders commonly send the same consumer credit offers that contain distinct terms so that consumers with unobserved characteristics may sort themselves into different 32In results not shown, we find that after the implementation of the CARD Act, credit card offers to filers have interest rates spreads nearly 300 basis points higher than nonfilers with similar credit profiles, and are 20 percentage points more likely to require an annual fee. 30

contracts. Our results suggest that lenders may use this approach as a strategy to facilitate achieving separating equilibria. We first focus on the monthly cross-sectional sample of consumers, which we used above in our baseline analysis. We define two offers as “distinct” if they have different specifications regarding at least one of the following five contractual terms: regular purchase interest rate, minimum credit limit, annual fee, introductory interest rate, or rewards program. As shown in columns (1) and (2) in Table 7, 22 percent of consumers receive distinct offers in a given month. Notably, 11 percent of consumers received distinct offers extended by the same lender in the same month. The difference in credit terms across offers to the same consumer are substantial, corroborating the results of Stango and Zinman (2013). For example, among all distinct offers a consumer receives in a given month, the average max-min difference is 370 basis points for regular purchase interest rates, and nearly $1,200 for minimum credit limits. Even among the distinct offers sent from the same lender, such differentials remain significant, at 210 basis points and $1,126 respectively. Moreover, 30 percent of consumers received offers sent by the same lender that contain different terms regarding introductory interest rates, 19 percent with different annual fees (including $0), and 45 percent with differences in the availability of reward programs. Exploiting Mintel’s longitudinal sample of consumers, we further study how the practice of sending distinct offers to the same consumer plays out over a longer time period. As shown in column (3) of Table 7, more than half of consumers in the longitudinal sample received distinct offers from the same lender within six months. The differences in credit terms widens among offers received over this longer period, with the max-min differences of regular purchase interest rates and minimum credit limits reaching 540 basis points and $1,400 on average, respectively. Notably, almost none (< 1%) of the increase in the dispersion of credit terms can be accounted for by within-borrower variation in credit scores over the six-month time period, suggesting that lenders are sequentially experimenting with distinct 31

offers to search for potential borrowers’ revelation of private information. Finally, we examine how the likelihood of receiving distinct offers extended by the same lender and the offered term dispersions vary with borrower characteristics—their credit scores in particular. Specifically, for each consumer in the longitudinal sample, we calculate his sixmonth average credit score. We then estimate models that associate the odds of receiving distinct offers and term dispersions with average credit scores. The results are reported in Table 8. We find that consumers with higher credit scores are more likely to be the target of distinct offers. As shown in column (1), a probit regression indicates that a one-standard deviation increase in credit scores implies a nearly 70 percent higher chance of receiving distinct offers sent by the same lender within six months. That said, conditional on receiving distinct offers, the dispersion of credit terms, on balance, is smaller among offers extended to consumers with higher credit scores. For example, as shown in column (2), the max-min difference in offered regular purchase interest rates narrows by 1.75 percent for consumers with credit scores 100 points higher. For discrete offer term variables (such as whether the offer includes a promotional introductory rate, annual fee, or rewards program), we construct a Herfindahl index to measure offer term dispersion. Specifically, if λ is the fraction of the offers sent by the same lender containing a certain term, then the Herfindahl index is calculated as λ2 +(1−λ)2. A higher Herfindahl index indicates more similar offers regarding this particular term.33 As shown in columns (5) and (6), higher credit scores are associated with more similar offers regarding annual fees and reward programs, with the effect being statistically significant. The only exceptionisthatconsumerswithhighercreditscoresappeartoreceiveofferswithmoredispersed minimum credit limits (column 2). On balance, our results are consistent with the notion that consumers with higher credit scores may have more unobserved heterogeneity with respect to profitability that lenders are 33For example, if a consumer received five offers from Citibank, two of which offer introductory interest rates, the Herfindahl index is equal to (0.4)2+(1−0.4)2 = 0.52. This set of offers represent more dispersed terms regarding introductory interest rates relative to, for example, a set of offers that all have such terms, which has a Herfindahl index equal to 1. 32

trying to separate using more dispersed terms in their offered credit card contracts. The restrictions of the CARD Act may also have led issuers to concentrate their screening efforts on these lower-risk borrowers. These findings provide an unprecedented look into how lenders conduct sophisticated screening to identify profitable borrower-contract matches in a market with information asymmetries. By varying contract features, sometimes substantially, over a relatively short period of time, lenders effectively create a “menu” of contracts into which consumers self-select.34 8 Conclusion Lenders of unsecured credit face the challenges of information asymmetry and limited commitment without the luxury of collateral as a screening device or protection against default. Characterizing how credit card lenders use screening and contract design to mitigate these challenges, particularly in a changing economic and regulatory environment, advances our understandingofcreditmarketsandpotentiallyprovidespolicyguidanceforregulatingthecredit card industry. In this paper, we take advantage of a unique new dataset of over 200,000 credit card mail solicitations to directly observe credit card access and offer terms. The administrative linkage of these credit offers to borrower credit histories allows us to examine effectively the same set of creditworthiness characteristics that a lender (without a pre-existing relationship) would obtain on a “soft pull” of a consumer’s credit record. Although it is generally quite difficult to identify credit supply per se from observed variation in equilibrium prices and quantities, this dataset provides an unprecedented proxy for credit supply in the credit card market. While we confirm the conventional wisdom that credit card lenders use credit scores as their central screening device, we also provide new evidence that lenders also take into ac- 34An alternative view is that consumers may not be especially attentive to the details of the contract, and lenders thus alter contract features in order to adjust the salient aspects of the contract, while “shrouding” other costly contract features (Gabaix and Laibson, 2006; Bordalo, Gennaioli and Shleifer, 2013). In related empirical work, Ru and Schoar (2015) also use the Mintel data to explore the design of back-loaded credit card contracts. 33

count a substantial array of other information on borrowers’ credit histories and financial and demographic characteristics beyond their effects on the calculation of the credit score. For instance, lenders use a dynamic measure of the recency of bankruptcy filing to inform the contract offers they mail to consumers. We also find that lenders extend multiple, distinct offers to the same consumers over a relatively short period, consistent with the predictions of search theory or endogenous separation in a sorting equilibrium with information asymmetry. The recent credit cycle had an enormous impact on the volume of credit card offers, which peaked at nearly 2 billion per quarter in 2007 and fell by a factor of four by mid-2008. We find that subprime offers were prevalent during the peak years of credit expansion, but that this segment of the market contracted most sharply during the downturn. The CARD Act appears to have only further exacerbated market tightening for risky households. Thus the balance-sheet recession and need for deleveraging has not been felt evenly across the credit score distribution. Despite this contraction, however, even the riskiest households maintained some access to new credit offers in the midst of the Great Recession. Finally, we note that this study focuses solely on how unsecured credit supply varied with borrowercharacteristicsoverthelastcreditcycle. Weremainagnosticaboutthedeterminants of the cycle itself: Many factors, such as capital regulations, willingness of lenders to increase leverage, and access to securitization markets likely affected the supply of unsecured credit to increasingly risky borrowers. Understanding the impacts of these specific drivers of cyclical dynamics remains a promising area of future research. 34

References Adams,William,LiranEinavandJonathanLevin(2009),‘Liquidityconstraintsandimperfect information in subprime lending’, American Economic Review 99(1), 49–84. Administrative Office of the United States Courts (2006), ‘Bankruptcy basics’, www.uscourts.gov/bankruptcycourts/bankruptcybasics.html. Agarwal, Sumit, Chunlin Liu and Lawrence Mielnicki (2003), ‘Exemption laws and consumer delinquency and bankruptcy behavior: An empirical analysis of credit card data’, The Quarterly Review of Economics and Finance 43(2), 273–289. Agarwal, Sumit, John C. Driscoll, Xavier Gabaix and David Laibson (2011), ‘Learning in the credit card market’, Working paper. Agarwal, Sumit, Souphala Chomsisengphet and Chunlin Liu (2010), ‘The importance of adverse selection in the credit card market: Evidence from randomized trials of credit card solicitations’, Journal of Money, Credit, and Banking 42(4), 743–754. Agarwal, Sumit, Souphala Chomsisengphet, Chunlin Liu and Nicholas Souleles (2006), ‘Do consumers choose the right credit contracts?’, Working paper. Agarwal, Sumit, Souphala Chomsisengphet, Neale Mahoney and Johannes Stroebel (2015a), ‘Dobankspassthroughcreditexpansions? themarginalprofitabilityofconsumerlending during the great recession’, Working paper. Agarwal, Sumit, Souphala Chomsisengphet, Neale Mahoney and Johannes Stroebel (2015b), ‘Regulating consumer financial products: Evidence from credit cards’, Quarterly Journal of Economics 130(1), 111–164. Arnold, Lutz G. and John G. Riley (2009), ‘On the possibility of credit rationing in the stiglitz-weiss model’, American Economic Review 99(5), 2012–21. Athreya, Kartik B. (2004), ‘Shame as it ever was: Stigma and personal bankruptcy’, Federal Reserve Bank of Richmond Economic Quarterly 90(2), 1–19. Athreya, Kartik B. and Hubert P. Janicki (2006), ‘Credit exclusion in quantitative models of bankruptcy: Does it matter?’, Federal Reserve Bank of Richmond Economic Quarterly 92(1), 17–49. Athreya, Kartik B., Xuan S. Tam and Eric R. Young (2009), ‘Are harsh penalties for default really better?’, Federal Reserve Bank of Richmond Working Paper. No. 09-11. Athreya, Kartik, Meta Brown, Xuan S. Tam and Eric R. Young (2013), ‘A quantitative theory of credit scoring’, FRB-Richmond Working Paper. Ausubel, Lawrence (1991), ‘The failure of competition in the credit card market’, American Economic Review 81(1), 50–81. 35

Ausubel, Lawrence and Amanda Dawsey (2004), ‘Informal bankruptcy’, University of Maryland Working Paper. Bester, Helmut (1985), ‘Screening vs. rationing in credit markets with imperfect information’, American Economic Review 75(4), 850–55. Bordalo, Pedro, Nicola Gennaioli and Andrei Shleifer (2013), ‘Salience and consumer choice’, Journal of Political Economy 121(5), 803–843. Butters, Gerald R. (1977), ‘Equilibrium distributions of sales and advertising prices’, Review of Economic Studies 44(3), 465–491. Chatterjee, Satyajit, Dean Corbae and Jose-Victor Rios-Rull (2011), ‘A theory of credit scoring and competitive pricing of default risk’, Working Paper. Chatterjee, Satyajit, Dean Corbae, Makoto Nakajima and Jose-Victor Rios-Rull (2007), ‘A quantitative theory of unsecured consumer credit with risk of default’, Econometrica 75(6), 1525–1589. Cohen-Cole, Ethan, Burcu Duygan-Bump and Judit Montoriol-Garriga (2009), ‘Forgive and forget: Who gets credit after bankruptcy and why?’, Federal Reserve Bank of Boston. Consumer Financial Protection Bureau (2013), ‘Card act report’. De Meza, David and David C. Webb (1987), ‘Too much investment: A problem of asymmetric information’, Quarterly Journal of Economics 102(2), 281–92. Debbaut, Peter, Andra C Ghent and Marianna Kudlyak (2014), ‘Are young borrowers bad borrowers? evidence from the credit card act of 2009’. DellaVigna, Stefano and Ulrike Malmendier (2004), ‘Contract design and self-control: Theory and evidence’, Quarterly Journal of Economics 119(2), 353–402. Dick, Astrid and Andreas Lehnert (2010), ‘Personal bankruptcy and credit market competition’, Journal of Finance 65(2), 655–686. Dobbie, Will and Paige Marta Skiba (2013), ‘Information asymmetries in consumer credit markets: Evidence from payday lending’, American Economic Journal: Applied Economics 5(4), 256–282. Edelberg, Wendy M. (2006), ‘Risk-based pricing of interest rates for consumer loans’, Journal of Monetary Economics 53(8), 2283–98. Einav, Liran, Mark Jenkins and Jonathan Levin (2012), ‘Contract pricing in consumer credit markets’, Econometrica 80(4), 1387–432. Einav, Liran, Mark Jenkins and Jonathan Levin (2013), ‘The impact of credit scoring on consumer lending’, RAND Journal of Economics 44(2), 249–74. 36

Fay, Scott, Erik Hurst and Michele J. White (2002), ‘The household bankruptcy decision’, American Economic Review 92(3), 706–18. Federal Reserve Board (2007), ‘Report to congress on credit scoring and its effects on the availability and affordability of credit’. Firestone,Simon(2014),‘Race,ethnicity,andcreditcardmarketing’,Journal of Money Credit and Banking 46(46), 1205–1024. Gabaix, Xavier and David Laibson (2006), ‘Shrouded attributes, consumer myopia, and information suppression in competitive markets’, Quarterly Journal of Economics 121(2), 505–540. Gropp, Reint, Karl J. Scholz and Michele J. White (1997), ‘Personal bankruptcy and credit supply and demand’, Quarterly Journal of Economics 112, 217–51. Gross, David B. and Nicholas S. Souleles (2002a), ‘Do liquidity constraints and interest rates matter for consumer behavior? evidence from credit card data’, Quarterly Journal of Economics 117(1), 149–185. Gross, David B. and Nicholas S. Souleles (2002b), ‘An empirical analysis of personal bankruptcy and delinquency’, Review of Financial Studies 15(1), 319–47. Han, Song (2004), ‘Discrimination in lending: Theory and evidence’, Journal of Real Estate Finance and Economics 29(1), 5–46. Han, Song and Geng Li (2011), ‘Household borrowing after personal bankruptcy’, Journal of Money, Credit, and Banking 43(2-3), 491–517. Hellwig, Martin F. (1987), ‘Some recent developments in the theory of competition in markets with adverse selection’, European Economic Review 31(1-2), 319–25. Hsu, Joanne W., David A. Matsa and Brian T. Melzer (2014), ‘Positive externalities of social insurance: Unemployment insurance and consumer credit’, Working Paper. Johnson, Kathleen and Geng Li (2010), ‘The debt-payment-to-income ratio as an indicator of borrowing constraints: Evidence from two household surveys’, Journal of Money, Credit and Banking 42(7), 1391–417. Karlan, Dean and Jonathan Zinman (2009), ‘Observing unobservables: Identifying information asymmetries with a consumer credit field experiment’, Econometrica 77(6), 1993– 2008. Keys, Benjamin J. (2010), ‘The credit market consequences of job displacement’, Finance and Economics Discussion Series, 2010-24. Federal Reserve Board. Keys, BenjaminJ., TanmoyMukherjee, AmitSeruandVikrantVig(2010), ‘Didsecuritization lead to lax screening? evidence from subprime loans’, Quarterly Journal of Economics 125(1), 307–362. 37

Knittel, Chris and Victor Stango (2003), ‘Price ceilings, focal points, and tacit collusion: Evidence from credit cards’, American Economic Review 93(5), 1703–1729. Lee, Donghoon and Wilbert van der Klaauw (2010), ‘An introduction to the frbny consumer credit panel’, Federal Reserve Bank of New York Staff Reports . Li, Wenli and Pierre-Daniel Sarte (2006), ‘U.S. consumer bankruptcy choice: The importance of general equilibrium effects’, Journal of Monetary Economics 53(3), 613–31. Livshits, Igor, James MacGee and Michele Tertilt (2007), ‘Accounting for the rise in consumer bankruptcies’, NBER Working Paper 13363. Livshits, Igor, James MacGee and Michele Tertilt (2010), ‘Accounting for the rise in consumer bankruptcies’, American Economic Journal: Macroeconomics 2(2), 165–93. Mian, Atif and Amir Sufi (2009), ‘The consequences of mortgage credit expansion: Evidence from the u.s. mortgage default crisis’, Quarterly Journal of Economics 124(4), 1449–96. Mian, Atif, Kamalesh Rao and Amir Sufi (2013), ‘Household balance sheets, consumption, and the economic slump*.’, Quarterly Journal of Economics 128(4). Musto, David K. (2004), ‘What happens when information leaves a market? evidence from postbankruptcy consumers’, Journal of Business 77(4), 725–748. Narajabad, Borghan (2012), ‘Information technology and the rise of household bankruptcy’, Review of Economic Dynamics 15(4), 526–550. Ordonez, Guillermo, David Perez-Reyna and Motohiro Yogo (2014), University of Pennsylvania Working Paper. Padilla, A. and M. Pagano (2000), ‘Sharing default information as a borrower discipline device’, European Economic Review 44, 1951–1980. Pagano, Marco and Tullio Jappelli (1993), ‘Information sharing in credit markets’, Journal of Finance 48(5), 1693–1718. Rea, Samuel A. Jr (1984), ‘Arm-breaking, consumer credit and personal bankruptcy’, Economic Inquiry 22(2). Riley, John G. (1987), ‘Credit rationing: A further remark’, American Economic Review 77(1), 224–27. Ronen, Joshua and Tiago Pinheiro (2014), ‘The unintended consequences of the credit card act’, Journal of Law, Finance, and Accounting . Ru, Hong and Antoinette Schoar (2015), ‘Do credit card companies screen for behavioral biases?’, MIT working paper. Sanchez, Juan M. (2010), ‘The it revolution and the unsecured credit market’, Working Paper 2010-022, Federal Reserve Bank of St. Louis. 38

Stango, Victor (2000), ‘Competition and pricing in the credit card market’, Review of Economics and Statistics 82. Stango, Victor and Jonathan Zinman (2013), ‘Borrowing high vs. borrowing higher: Sources and consequences of dispersion in individual borrowing costs’, NBER Working Paper 19069. Stiglitz, Joseph E. and Andrew Weiss (1981), ‘Credit Rationing in Markets with Imperfect Information’, American Economic Review 71. 39

Figure 1: Credit Card Solicitation Volume, New Accounts Opened, and Lending Standards The figure shows how credit card solicitation volume tracks other measures of credit access. The top panel shows the time series of total credit card solicitation mail volume in the U.S. from 2001:Q1 through 2014:Q2 using the Mintel data and the number of new credit card accounts opened estimated using the FRB/NY Equifax Consumer Credit Panel. The two time-series show a strong common pattern, with a correlation coefficient of 0.9. The bottom panel shows the quarter-over-quarter change in credit card solicitation volume and bank-reported changes in lending standards from the Senior Loan Officer Opinion Survey (SLOOS). The two series are again highly correlated. Credit Card Mail Volume and New Accounts Millions of New Accounts Millions of Offers 2000 100 Mail Volume 1800 (Right Axis) 1600 80 1400 1200 60 New Accounts 1000 (Left Axis) 800 40 600 20 400 200 2001 2003 2005 2007 2009 2011 2013 Credit Card Mail Volume and Bank-Reported Lending Standards Changes Percent 100 80 60 Mail Volume Growth 40 20 0 -20 -40 Lending Standards -60 Changes -80 -100 2001 2003 2005 2007 2009 2011 2013 40

Figure 2: Likelihood of Receiving an Credit Card Offer by Credit Score over the Recent Credit Cycle The figure presents the relationship between the likelihood of receiving a credit card offer (in a given month) and VantageScore 2.0 credit scores, separately for three time periods. The blue line shows the relationship during the boom period (January 2007–March 2008), which was flat over the distribution of credit scores, andifanythingpeakedamongsubprimeborrowers. Theyellowlineshowsthe“offercurve”forthepost-crisis periodbutbeforetheCARDActwasimplemented,April2008–February2010. Inthisperiod,thelikelihoodof receivinganofferbecameincreasinglycorrelatedwithcreditscore. Inthefinalpost-CARDActperiod(March 2010–June 2014), the association between creditworthiness and credit access as measured by the likelihood of receiving a credit card offer only strengthened. Consumers with the lowest credit scores were even less likely to receive an offer in this most recent period. Shaded bands represent 95 percent confidence intervals. Likelihood of Receiving a Credit Card Offer by Credot Score over the Recent Credit Cycle PPeerrcceenntt 110000 8800 Pre-Crisis 6600 Post-CARD Act 4400 Post-Crisis, Pre-CARD Act 2200 00 << 555500 555500--660000 660000--665500 665500--770000 770000--775500 775500--880000 880000--885500 885500--990000 990000--995500 >> 995500 CCrreeddiitt SSccoorree 41

Figure 3: Likelihood of Receiving an Auto Loan Offer by Credit Score over the Recent Credit Cycle The figure presents the relationship between the likelihood of receiving an auto loan offer (in a given month) and VantageScore 2.0 credit scores, separately for three time periods. The three time periods represent the credit boom (January 2007–March 2008), the credit crunch period prior to the implementation of the CARD Act (April 2008–February 2010), and the post-CARD Act period (March 2010–June 2014). In contrast to the unsecured credit market, auto loans are more commonly targeted at subprime borrowers, as lenders can rely on repossession of the collateral backing the loan. The pattern in the figure suggests that the auto loan market improved for all consumers between the depths of the crisis and the recovery period. Likelihood of Receiving an Auto Loan Offer by Credit Score over the Recent Credit Cycle PPeerrcceenntt 1100 Pre-Crisis 88 Post-CARD Act 66 44 Post-Crisis, Pre-CARD Act 22 00 << 555500 555500--660000 660000--665500 665500--770000 770000--775500 775500--880000 880000--885500 885500--990000 990000--995500 >> 995500 CCrreeddiitt SSccoorree 42

Figure 4: Spreads of Interest Rates Offered by Credit Score over the Recent Credit Cycle The figure presents the relationship between interest rate spreads on credit card offers and VantageScore 2.0 credit score, separately for three time periods. The three time periods represent the credit boom (January 2007–March 2008), the credit crunch period prior to the implementation of the CARD Act (April 2008– February 2010), and the post-CARD Act period (March 2010–June 2014). The figure shows that interest rates have risen for all types of consumers in the post-CARD Act period, but that rates have risen more sharply for subprime consumers relative to prime consumers. Spreads of Interest Rates Offered by Credit Score over the Recent Credit Cycle PPeerrcceenntt 2200 1188 Post-CARD Act 1166 1144 Post-Crisis, Pre-CARD Act 1122 1100 Pre-Crisis 88 66 44 22 00 << 555500 555500--660000 660000--665500 665500--770000 770000--775500 775500--880000 880000--885500 885500--990000 990000--995500 >> 995500 CCrreeddiitt SSccoorree 43

Figure 5: Other Terms Offered by Credit Score over the Recent Credit Cycle The figure presents the relationship between other terms on credit card offers and VantageScore 2.0 credit score, separately for three time periods. The three time periods represent the credit boom (January 2007– March 2008), the credit crunch period prior to the implementation of the CARD Act (April 2008–February 2010), and the post-CARD Act period (March 2010–June 2014). The top left panel shows the pattern for minimum credit limits, which increased for consumers of all creditworthiness levels in the post-CARD Act period. Thetoprightpanelshowstheprevalenceofpromotionalintroductoryteaserrates,whichhavebecome increasingcommonforlesscreditworthyconsumers. Thebottomleftpanelshowstheuseofannualfees,which have fallen among subprime consumers after the implementation of the CARD Act. The bottom right panel suggests an increase in the use of rewards programs as a contract feature among subprime consumers. Minimum Credit Limit by Credit Score Likelihood of Having Intro Teaser Rate by Credit Score over the Recent Credit Cycle DDoollllaarr VVaalluuee over the Recent Credit Cycle PPeerrcceenntt 22550000 110000 Pre-Crisis Post-Crisis, Pre-CARD Act Post-CARD Act 22000000 8800 11550000 6600 11000000 4400 550000 2200 00 00 << 555500 660000--665500 770000--775500 880000--885500 990000--995500 << 555500 660000--665500 770000--775500 880000--885500 990000--995500 CCrreeddiitt SSccoorree CCrreeddiitt SSccoorree Likelihood of Having an Annual Fee by Credit Score Likelihood of Having a Rewards Program by Credit Score over the Recent Credit Cycle PPeerrcceenntt over the Recent Credit Cycle PPeerrcceenntt 110000 110000 8800 8800 6600 6600 4400 4400 2200 2200 00 00 << 555500 660000--665500 770000--775500 880000--885500 990000--995500 << 555500 660000--665500 770000--775500 880000--885500 990000--995500 CCrreeddiitt SSccoorree CCrreeddiitt SSccoorree 44

Figure 6: Likelihood of Receiving Multiple Distinct Offers Extended by the Same Lender The figure presents the relationship between the likelihood of receiving multiple distinct offers from the same lenderoverasix-monthperiodandcreditscore. Theshadedbandrepresentsa95percentconfidenceinterval. This figure uses the longitudinal sample collected by Mintel from July 2011-June 2014. The figure shows that while relatively few subprime consumers receive multiple offers from the same lender, there is a strong positive relationship up to a credit score of roughly 700, where the relationship flattens out. Over 70% of consumers with credit scores above 700 receive multiple distinct offers from the same lender over a six-month period, suggesting that lenders are intensively searching for low-risk consumers to reveal private information by offering different contracts. Likelihood of Distinct Offers Extended by the Same Lender by Credit Score over the Recent Credit Cycle PPeerrcceenntt 110000 8800 6600 4400 2200 00 << 555500 555500--660000 660000--665500 665500--770000 770000--775500 775500--880000 880000--885500 885500--990000 990000--995500 >> 995500 CCrreeddiitt SSccoorree 45

Table 1: Demographic and Credit Characteristics of the Mintel Cross-Sectional Sample The table presents summary statistics of key demographic and socioeconomic characteristics for the heads of households in the Mintel monthly cross-sectional sample and compares our sample to two representative samples of U.S. consumers. We restrict Mintel the sample to be the households whose heads aged between 20 and 60 and household annual income between $10,000 and $200,000. The final cross-sectional Mintel sample contains about 222,900 credit card offers that were sent to about 173,300 individuals in more than 107,700 households. For comparison, we also include corresponding statistics estimated using the 2007, 2010, and 2013 SCF sample, subject to the same criteria and weighted accordingly, in column (2), and corresponding statistics from the FRBNY/Equifax Consumer Credit Panel in column (4). Demographics Liability and Credit History Mintel SCF Mintel Equifax (1) (2) (3) (4) Mean age 45.3 41.7 Total debt (2013$) 100,065 108,101 Mean household size 2.7 2.9 Rev. debt (2013$) 10,680 12,556 High School (%) 29.8 31.9 Rev. credit limit ($) 42,620 36,375 Some College (%) 23.3 20.4 Utilization rate 0.25 0.33 College (%) 40.6 37.6 Have credit cards 0.75 0.64 Homeowner (%) 70.5 61.1 Number of credit cards 1.88 1.85 White (%) 85.0 80.2 Bankruptcy 0.08 0.07 Married (%) 55.4 50.2 Other derog. 0.07 0.10 Income (2013$) 72,139 66,383 Deep delinquency 0.06 0.07 46

Table 2: Credit Card Offers by Credit Status The table presents summary statistics of credit card offers by credit status (prime vs. nonprime) for the full Mintel cross-sectional sample. Prime consumers refer to those with a VantageScore 2.0 credit score greater than 700, while and nonprime consumers refer to those with credit scores below this threshold. The average numberofoffersiscalculatedconditionalonreceivingatleastonecreditcardoffer. Meanvaluesarereported, withmedianvaluesshowninbracketsbelow(whereapplicable). Allstatisticsarecomputedusingtheweights provided by Mintel. Plain vanilla, credit building, general market, and premium rewards are four types of credit card offers that, in this order, charge no annual fee and carry no rewards program, charge an annual fee and carry no rewards program, charge no annual fee and carry rewards programs, and charge an annual fee and carry rewards programs, respectively. All Prime Nonprime (1) (2) (3) Number of consumers 169,692 127,197 42,495 Received at least one offer (%) 50.4 55.8 36.1 Received at least one credit building offer (%) 7.5 3.2 18.7 plain vanilla offer (%) 20.7 22.8 15.2 general market offer (%) 38.8 47.7 15.3 premium reward offer (%) 14.6 18.0 5.5 Among consumers receiving offers Avg. num. of offers received (monthly) 2.5 2.6 2.2 [2] [2] [1] Number of offers 219,707 185,962 33,745 Avg. interest rate (%) 13.7 13.2 16.1 [13.0] [13.0] [15.0] Avg. min credit limit ($) 1,158 1,313 518 [500] [500] [300] Have introductory rate (%) 68.3 71.3 53.9 Have annual fee (%) 18.9 13.9 43.4 Have rewards program (%) 67.0 74.3 31.2 47

Table 3: Determinants of Credit Card Offers Having an offer Credit limit Spread Intro rate Annual fees Rewards Credit score bins 550-600 0.100*** -208.749*** 1.229*** 0.034** 0.051*** -0.169*** (0.009) (34.383) (0.181) (0.016) (0.013) (0.016) 600-650 0.155*** -251.378*** 2.249*** 0.069*** 0.052*** -0.234*** (0.010) (34.182) (0.186) (0.019) (0.013) (0.016) 650-700 0.223*** -208.157*** 1.693*** 0.146*** -0.070*** -0.143*** (0.011) (32.610) (0.203) (0.016) (0.009) (0.016) 700-750 0.252*** -100.237*** 0.436** 0.187*** -0.136*** -0.014 (0.011) (33.416) (0.188) (0.016) (0.006) (0.014) 750-800 0.285*** 102.084*** -0.367* 0.214*** -0.162*** 0.064*** (0.011) (31.987) (0.191) (0.014) (0.006) (0.013) 800-850 0.292*** 326.984*** -0.871*** 0.220*** -0.164*** 0.119*** (0.012) (35.524) (0.190) (0.014) (0.006) (0.012) 850-900 0.264*** 577.903*** -0.969*** 0.197*** -0.150*** 0.147*** (0.012) (31.360) (0.185) (0.017) (0.007) (0.011) 900-950 0.254*** 714.265*** -0.865*** 0.168*** -0.124*** 0.169*** (0.011) (33.499) (0.192) (0.017) (0.007) (0.010) > 950 0.267*** 809.787*** -0.770*** 0.141*** -0.106*** 0.176*** (0.010) (40.664) (0.178) (0.016) (0.007) (0.009) Credit hist. attr. Bankruptcy filer -0.068*** -238.100*** 0.834*** -0.128*** 0.107*** -0.296*** (0.008) (14.106) (0.111) (0.009) (0.010) (0.010) Other derog rec. -0.065*** -130.100*** 0.153 -0.071*** 0.060*** -0.082*** (0.009) (13.166) (0.106) (0.009) (0.009) (0.008) Deep del. -0.027*** -26.073 0.091 0.065*** -0.008 0.007 (0.010) (19.841) (0.113) (0.013) (0.008) (0.014) Recent del. -0.004*** -31.710*** 0.124*** -0.004*** 0.015*** -0.023*** (0.001) (2.115) (0.015) (0.001) (0.001) (0.002) Num inquiries 0.001 -20.461*** 0.011 -0.008*** 0.008*** -0.011*** (0.001) (2.610) (0.011) (0.001) (0.001) (0.001) Debt-income ratio -0.001 -9.194*** -0.030*** -0.002*** -0.000 -0.001 (0.001) (2.715) (0.007) (0.001) (0.001) (0.001) Have credit card 0.065*** 47.819*** -0.536*** -0.019*** -0.033*** 0.032*** (0.004) (12.860) (0.044) (0.003) (0.004) (0.004) High util -0.022*** -13.200 0.454*** 0.008 0.012*** -0.010 (0.006) (16.002) (0.043) (0.006) (0.004) (0.007) Household char. Head age -0.001 5.551 -0.018 -0.002 -0.000 0.002** (0.001) (4.609) (0.012) (0.002) (0.001) (0.001) Head age2/100 0.001 -5.231 0.019 0.002 -0.000 -0.002** Continued on next page 48

Table 3 – Continued Having an offer Credit limit Spread Intro rate Annual fees Rewards (0.002) (5.291) (0.014) (0.002) (0.002) (0.001) Married 0.007 33.262*** -0.077 0.013*** -0.007* -0.005 (0.005) (10.691) (0.046) (0.004) (0.004) (0.004) Household size 0.007*** -12.111*** 0.049*** 0.011*** -0.009*** -0.006*** (0.001) (3.818) (0.011) (0.001) (0.001) (0.001) White 0.034*** 19.181 -0.049 0.022*** -0.025*** 0.021*** (0.007) (15.554) (0.040) (0.006) (0.006) (0.004) High school 0.012* -8.502 -0.135 0.006 -0.021*** 0.021** (0.007) (16.729) (0.084) (0.006) (0.007) (0.009) Some college -0.006 26.578 -0.192** -0.026*** 0.002 0.030*** (0.008) (23.713) (0.080) (0.007) (0.008) (0.007) College 0.011 137.328*** -0.128 -0.067*** 0.041*** 0.059*** (0.008) (19.567) (0.076) (0.006) (0.007) (0.008) Homeowner 0.014*** -1.546 -0.177*** 0.030*** -0.036*** -0.009** (0.005) (13.039) (0.058) (0.007) (0.004) (0.004) Log(income) 0.042*** 96.004*** -0.027 -0.051*** 0.039*** 0.048*** (0.005) (8.667) (0.030) (0.003) (0.003) (0.003) Legal & econ. cond. Unemp -0.016*** -1.300 0.213*** 0.004 0.010*** 0.003*** (0.002) (4.943) (0.037) (0.003) (0.003) (0.001) Homestead exempt -0.003*** -2.104 0.026** -0.001 0.002*** -0.001* (0.001) (2.759) (0.012) (0.001) (0.001) (0.001) Property exempt -0.035** 72.618 0.273 -0.053* 0.075*** 0.002 (0.018) (50.175) (0.292) (0.028) (0.014) (0.014) Yearly fixed-effects yes yes yes yes yes yes Monthly fixed-effects yes yes yes yes yes yes R2 / pseudo R2 0.061 0.115 0.341 0.095 0.115 0.149 N 169,692 148,656 217,920 219,707 219,707 219,707 Note: The table presents estimates of probit (offer, has intro rate, has annual fee, has reward program) and OLS (credit limit, spread) regressions to explore the determinants of credit card offers and their features. Reported coefficients are probit marginal effects or OLS coefficients. Standard errors in parentheses are clustered by state. *, **, and *** indicate that the estimated coefficients are statistically significant at the 90-, 95-, and 99-percent level, respectively. 49

Table 4: Do Lenders Monitor Credit Scores of an Individual over Time? The table presents estimates of probit (offer, has intro rate, has annual fee, has reward program) and OLS (credit limit, spread) regressions to explore the relationship between individual credit scores and credit card access. The regressions are fixed-effects panel regressions that control for individual consumers, thus the identification comes solely from within-consumer variation in credit scores over time. Reported coefficients are probit marginal effects or OLS coefficients. These specifications use the longitudinal sample of Mintel respondents, covering July 2011-June 2014. Heteroskedasticity-robust standard errors in parentheses. *, **, and *** indicate that the estimated coefficients are statistically significant at the 90-, 95-, and 99-percent level, respectively. Having an offer Credit limit Spread Intro rate Annual fees Rewards Credit score/100 0.061*** 236*** -1.056*** -0.008 0.001 0.072*** (0.015) (41.0) (0.010) (0.011) (0.010) (0.013) R2 / pseudo R2 0.06 0.06 0.17 0.01 0.00 0.09 Number of consumers 631 493 527 533 533 533 Number of observations 10,149 10,259 13,251 13,786 13,786 13,786 50

Table 5: Effect of Bankruptcy on Credit Access The table presents estimates of probit (offer, has intro rate, has annual fee, has reward program) and OLS (credit limit, spread) regressions of the relationship between the recency of bankruptcy filing and credit card offers and their features. The specifications are identical to those presented in Table 3, except the dummy variableforbankruptcyflaghasbeenreplacedwiththreemeasuresofthetimesincefiling. “Recent”filersare those who have filed for personal bankruptcy within the last two years, “seasoned” filers are those who filed between three and five years prior, and “remote” filers are those who filed more than five years earlier and are approaching re-filing eligibility (after 8 years). The table shows that recent filers are as likely as similar nonfilers to obtain a credit card offer, but their offers have decidedly less favorable terms. Remote filers are substantiallylesslikelytoreceiveoffers,reflectiveoftheirre-filingrisk. Thus,asimplebankruptcyflagmasks considerable heterogeneity in how lenders treat this information. Reported coefficients are probit marginal effects or OLS coefficients. Standard errors in parentheses are clustered by state. *, **, and *** indicate that the estimated coefficients are statistically significant at the 90-, 95-, and 99-percent level, respectively. Having an offer Credit limit Spread Intro rate Annual fees Rewards Recent -0.004 -266.151*** 1.790*** -0.066*** 0.152*** -0.403*** (0.015) (21.820) (0.350) (0.015) (0.016) (0.017) Seasoned -0.069*** -238.458*** 1.530*** -0.138*** 0.143*** -0.372*** (0.015) (21.836) (0.153) (0.014) (0.016) (0.017) Remote -0.117*** -282.646*** 0.459*** -0.132*** 0.057*** -0.203*** (0.010) (15.691) (0.112) (0.014) (0.012) (0.016) Controlled for Credit score bins yes yes yes yes yes yes Credit history attributes yes yes yes yes yes yes Household characteristics yes yes yes yes yes yes State fixed-effects yes yes yes yes yes yes Yearly fixed-effects yes yes yes yes yes yes Monthly fixed-effects yes yes yes yes yes yes R2 / pseudo R2 0.065 0.116 0.407 0.105 0.118 0.150 N 169,692 148,656 217,920 219,707 219,707 219,707 51

Table 6: How Did Bankruptcy Flags’ Impact Evolve over the Credit Cycle? The table presents estimates of probit (offer, has intro rate, has annual fee, has reward program) and OLS (credit limit, spread) regressions to relationship between bankruptcy filing status and credit card offers and theirfeaturesoverthreetimeperiods. Thethreetimeperiodsrepresentthecreditboom(January2007–March 2008), the credit crunch period prior to the implementation of the CARD Act (April 2008–February 2010), andthepost-CARDActperiod(March2010–June2014). Column(1)repeatsthesameresultshowninTable 5 as a reference point. As shown in columns (2)–(4), recent filers were 8 percentage points more likely to receive an offer than otherwise comparable nonfilers during the credit boom, whereas remote filers were 5 percentage points less likely. The favorable treatment of recent filers largely persisted even during the depths of the credit crunch. However, the pattern appears to have changed substantially during the post-CARD Act era,whererecentfilerslosttheiredgeinreceivingcreditcardoffersrelativetononfilers,whilethegapbetween remote filers’ likelihood of receiving an offer and that of comparable nonfilers became ever larger. Reported coefficients are probit marginal effects or OLS coefficients. Standard errors in parentheses are clustered by state. *, **, and *** indicate that the estimated coefficients are statistically significant at the 90-, 95-, and 99-percent level, respectively. (1) (2) (3) (4) Full sample Pre-crisis Post-crisis pre-CARD Act Post-CARD Act Recent -0.004 0.046* 0.064** -0.031* (0.015) (0.025) (0.032) (0.016) Seasoned -0.069*** -0.044** -0.046*** -0.100*** (0.015) (0.022) (0.017) (0.017) Remote -0.117*** -0.056*** -0.088*** -0.152*** (0.010) (0.018) (0.016) (0.013) Controlled for Credit score bins yes yes yes yes Credit history attributes yes yes yes yes Household characteristics yes yes yes yes State fixed-effects yes yes yes yes Yearly fixed-effects yes yes yes yes Monthly fixed-effects yes yes yes yes R2 / pseudo R2 0.065 0.022 0.075 0.059 N 169,692 29,701 44,520 95,471 52

Table 7: Multiple Offers to the Same Consumer The table presents results based on analyzing distinct credit offers sent to the same consumer. The first two columns use data from the Mintel cross-sectional sample, while the last column uses six-month intervals from the Mintel longitudinal sample. The summary statistics here suggest that the majority of consumers receive multiple offers from the same lender over a six-month period, with dramatic dispersion in contract characteristics. The average difference in interest rates between the minimum and maximum credit offer over the six-month period from the same lender is 310 basis points, and the average difference in minimum credit limits is $1,401. Cross-sectional sample Longitudinal sample (In a given month) (6-month intervals) (1) (2) (3) All lenders Same lender Same lender % received distinct offers 26.1 13.5 51.1 Number of distinct offers (conditional on receiving) 3.1 2.3 3.1 Max-Min Differences in Interest rates (percent) 3.9 2.2 5.6 Minimum credit limits ($) 1,143 1,044 1,386 Share of consumers received offers with different terms regarding (%) Whether having introductory rates 46.7 30.9 47.8 Whether having annual fees 30.7 19.0 41.4 Whether offering rewards 46.8 45.1 55.0 53

Table 8: Multiple Offers and Credit Scores ThetableusestheMintellongitudinalsampletoexploretherelationshipbetweencreditscores,theprevalence of multiple offers, and the dispersion in those offers. The specifications are collapsed to the level of the respondentinthelongitudinalsample. Dispersioninhavingteaserrate, annualfees, andrewardsismeasured using the Herfindahl index, see text for details. The table shows that consumers with higher credit scores are more likely to receive multiple distinct offers extended by the same lender, with less dispersion in their interest rates but more dispersion in credit limits and other features (annual fee and rewards program). Heteroskedasticity-robust standard errors in parentheses, odds ratio in brackets. *, **, and *** indicate that the estimated coefficients are statistically significant at the 90-, 95-, and 99-percent level, respectively. Received>1offer DispersioninContractTerms Max-Min HerfindahlIndex (1) (2) (3) (4) (5) (6) Reg. int. rate Min. limits Haveintro. rate Haveannualfee Rewards Creditscore/100 0.445*** -1.661*** 477*** -0.004 0.012** 0.033*** (0.044) (0.146) (60) (0.005) (0.005) (0.004) [1.690] N 1,682 938 694 480 431 563 54