SSRN Id3405897 PDF

Regulatory Interventions in Consumer Financial
Markets: The Case of Credit Cards*

Manolis Galenianos¶ Alessandro Gavazza§
April 2020
Abstract
We build a framework to understand the effects of regulatory interventions in credit

markets, such as caps on interest rates and higher compliance costs for lenders. We
focus on the credit card market, in which we observe U.S. consumers borrowing at
high and very dispersed interest rates, despite receiving many credit card offers. Our
framework includes two main features that may explain these patterns: endogenous
effort of examining offers and product differentiation. Our calibration suggests that these
patterns occur because borrowers do not examine most of the offers that they receive.
The calibrated model implies that interest rate caps reduce credit supply modestly and
curb lenders’ market power significantly, leading to large gains in consumer surplus,
whereas higher compliance costs unambiguously decrease consumer surplus.
* We thank Victor Stango for sharing data with us, Mark Armstrong, Glenn Ellison, John Kennan, Scott
Nelson, Nikita Roketskiy, and Matthijs Wildenbeest for useful comments and suggestions. William Matcham
provided excellent research assistance. Alessandro Gavazza gratefully acknowledges support from the European
Research Council (ERC-Consolidator grant award no. 771004).
¶ Department of Economics, Royal Holloway, University of London. Egham Hill, Egham TW20 0EX, United
Kingdom. Email: [email protected].
§ Department of Economics, London School of Economics. Houghton Street, London WC2A 2AE, United
Kingdom. Email: [email protected].
Electronic copy available at: https://ssrn.com/abstract=3405897

1 Introduction
After the financial crisis, policymakers in several countries began intervening in markets for
consumer financial products more aggressively than before, both through new legislation and
the creation of new regulatory agencies, such as the Consumer Financial Protection Bureau
(CFPB) in the U.S. and the Financial Conduct Authority (FCA) in the U.K. Although
the interventions have taken different forms, two broad, prevalent directions are a break
from the recent past: 1) the imposition of direct constraints on some prices and fees of
financial products, because policymakers viewed them as “predatory,”, that is, as targeting
unsophisticated and poorly-informed households;1 and 2) a tighter supervision of lenders,
which often increased compliance and operating costs for these financial institutions.
An important question is what effects these policies will have on the operation of markets
for consumer financial products. Standard competitive theory predicts that binding price caps
and higher operating costs reduce market efficiency and lower market access, particularly for
marginal borrowers. Policymakers’ primary motivation, however, is that markets for consumer
financial products do not satisfy the conditions of perfectly-competitive markets, because of
informational and other frictions.2 Nevertheless, the theoretical analysis of frictional markets
provides ambiguous predictions on the effects of these policies on consumer surplus and on
aggregate welfare. For example, although price caps have the direct effect of lowering some
prices, they also reduce incentives to become informed, which may increase market power and
lead to higher, rather than lower, average prices (Fershtman and Fishman, 1994; Armstrong,
Vickers, and Zhou, 2009). Therefore, determining the overall effect of these policies is an
empirical/quantitative question.
The objective of this paper is to quantitatively study the effects of the two types of
intervention on the market for consumer financial products. We focus on the U.S. credit card
market, for which we have been able to combine several sources of data. Our data build on
Stango and Zinman (2016) (henceforth SZ) and display two key patterns that appear, at first
sight, somewhat contradictory. First, the interest rates that borrowers pay on their credit cards
1
Specifically, the 2009 U.S. Credit Card Accountability Responsibility and Disclosure Act explicitly
prohibited lenders from charging some fees on credit cards (Agarwal, Chomsisengphet, Mahoney, and Stroebel,
2015). Similarly, in the U.K. the FCA has introduced regulatory caps for several financial products: in
November 2014 it enacted a price structure for payday loans, capping the initial cost of a loan to a maximum
of 0.8 percent per day; in November 2016, it restricted fees for individuals who want to access their pensions
to a maximum of one percent. Furthermore, the financial press reports that the FCA is currently evaluating
limits on fees for other products, such as mutual fund fees (The Financial Times, Funds’ lucrative entry fees
under attack, May 26, 2016) and mortgage origination fees (The Financial Times, Mortgage lenders under
FCA review for masking high fees, December 12, 2016).
2
Indeed, Sirri and Tufano (1998) and Hortaçsu and Syverson (2004) argue that information frictions play
a prominent role in mutual fund markets, and Allen, Clark, and Houde (2019) and Woodward and Hall (2012)
show their relevance in mortgage markets.

are high in comparison to funding costs, and they are very dispersed, even after controlling
for observable borrower characteristics—including their creditworthiness, as captured by their
credit score—or card characteristics (e.g., rewards). Thus, this first pattern suggests that
lenders enjoy a significant amount of market power. Second, the average consumer receives
several pre-approved credit card offers every month and these offers advertise very different
interest rates. This second pattern suggests that lenders face considerable competition,
because borrowers can choose low-interest-rate credit cards among the offers that they receive.
Interpreting these seemingly contradictory patterns and deriving the policy implications of our
interpretation are the main contributions of our paper.
We develop a rich but tractable modeling framework based on search theory that accounts
for these patterns.3 Our framework includes two features to help interpret high and dispersed
interest rates: information frictions, such as costs of examining and evaluating different offers,
and product differentiation. In our model, lenders, which are heterogeneous in their funding
costs, choose whether to enter a particular market characterized by borrower creditworthiness
(i.e., sub-prime, near-prime, prime, and super-prime), choose what interest rate to offer, and
send credit card offers to borrowers. Borrowers, who are heterogeneous in their willingness
to pay for a loan, choose how much effort to exert in examining the offers that they receive
and decide which offer, if any, to accept. The acceptance decision depends on the offer’s
interest rate and on an offer-specific attribute, interpreted as product differentiation. Hence,
a borrower might reject a low-interest-rate credit card because he does not examine it or
because he does not like the offer-specific attribute, thereby leading to a high level of and a
large dispersion in accepted interest rates.
We calibrate the model to match the statistics on the distribution of interest rate offers and
the distributions of accepted offers in each market, as well as the fractions of borrowers and the
charge-off rates in each market, and lenders’ average funding costs. The model fits the data
well. The calibration implies that the high cost of examining offers, leading to low examination
effort, mostly account for borrowers’ outcomes, whereas product differentiation plays a smaller
role. This important finding drives the outcomes of our counterfactual experiments. The
intuition for this finding is that, although both high examination costs and high product
differentiation are potentially consistent with the high level of and the high dispersion in
accepted interest rates, only high examination costs can additionally account for the high
dispersion of offered rates and the moderate fraction of borrowers in the data. By contrast,
high product differentiation leads to a lower dispersion of offered rates and a larger fraction
of borrowers in the population than those observed in the data. In other words, to match all
3
Some recent papers that structurally estimate search models of consumer product markets include
Hortaçsu and Syverson (2004), Hong and Shum (2006), Wildenbeest (2011), Allen, Clark, and Houde (2019),
Gavazza (2016), and Galenianos and Gavazza (2017). Galenianos and Nosal (2016) develop a dynamic search
model of unsecured credit and default.

relevant moments, our model requires high examination costs.
We use our calibrated model to perform counterfactual experiments. First, we consider
the effect of the introduction of a 25-percentage-point interest rate cap. Our calibrated model
predicts that the cap has a relatively small effect on outcomes in the markets for creditworthy
borrowers (prime and super-prime) because it is rarely binding in our data. In markets for
riskier borrowers, that is, sub-prime and near-prime, where the cap is binding for 35 and 25
percent of borrowers in our data, respectively, the effect is more pronounced: the average
accepted interest rate declines by more than 10 percent, while the number of people getting
a loan declines modestly, by less than five percent. The decline in loans is modest because,
although the number of offers falls by more than 10 percent in the sub-prime market, borrowers
respond to the more favorable interest rate distribution by examining a larger share of the
offers that they receive. The overall effect is a large redistribution of surplus from lenders
to borrowers—consumer surplus increases substantially, by more than 10 percent in the sub-
prime and near-prime markets, whereas lender profits decline quite steeply, by more than 45
percent—and a smaller decline in aggregate welfare of at most four percent.
We view the large positive effect of an interest rate cap on the consumer surplus of sub-
and near-prime borrowers to be an interesting result: in a perfectly-competitive market with
complete information such caps will reduce supply precisely toward marginal borrowers and
will, therefore, negatively affect their surplus. Our results are reminiscent of monopolistic
markets, though our model explicitly accounts for the large number of credit card offers that
borrowers receive. Most notably, the presence of informational frictions, that is, the costs
of examining and evaluating credit card offers, rationalizes the appearance of a competitive
market with the reality of high lender market power.
We further use the calibrated model to understand the effect on market outcomes of higher
operating costs, which we model as an increase in lenders’ fixed costs. In this case, we are
interested in the effect of cost-increasing policies on borrower outcomes, admittedly neglecting
some of the potential benefits that motivate the policies, such as greater financial stability
and/or fewer abusive lending practices. We calibrate an increase in entry costs to replicate
the reduction in entry of our price-cap case to facilitate the comparison of market outcomes
between these two cases. We find again that higher operating costs have only minor effects on
the markets for safer borrowers. However, in markets for riskier borrowers, higher operating
costs trigger the exit of lenders with high marginal costs; because competition is lower than in
the baseline case, all surviving lenders increase their interest rates. These changes in supply
lead to a 15-percent drop in the fraction of borrowers with credit card loans and a 20-percent
drop in consumer surplus. Hence, in contrast to the price-cap case, higher operating costs
unambiguously harm borrowers and decrease aggregate welfare.
The paper proceeds as follows. Section 2 reviews the literature, highlighting our

contribution. Section 3 describes the data. Section 4 presents the theoretical model. Section
5 presents our calibration of the model and illustrates its main quantitative implications.
Section 6 performs our counterfactual analyses. Section 7 concludes. The appendices report
further results and collect all proofs.
2 Related Literature
The paper contributes to several strands of the empirical literature. The first is the literature
that studies imperfect competition and frictions in credit card markets. In an important
contribution, Ausubel (1991) showed that interest rates on credit cards are substantially
higher than lenders’ funding costs and display limited intertemporal variability, citing search
frictions as a potential departure from a competitive market. Calem and Mester (1995) present
empirical evidence on consumers’ limited search and switching behavior. Stango (2002) studies
credit card pricing when consumers have switching costs. Grodzicki (2015) analyzes how
credit card companies acquire new customers. We contribute to this literature by building a
framework that allows us to quantify the effects of information frictions and consumer inertia
on lenders’ loan pricing and on consumers’ cost of borrowing.
Second, a vast literature in household finance studies whether consumers behave optimally
in credit markets: among others, Agarwal, Driscoll, Gabaix, and Laibson (2008) and Agarwal,
Chomsisengphet, Liu, and Souleles (2015) analyze consumer mistakes in the credit card
market. Ru and Schoar (2016) study how credit card companies exploit consumers’ mistakes.
In this strand of literature, the most related paper is Woodward and Hall (2012), who study
consumers’ shopping effort in the U.S. mortgage market. We contribute to this literature
by developing and calibrating an equilibrium model in which consumers’ shopping effort is
endogenous, which allows us to analyze how it adjusts after regulatory interventions.
Third, many countries have recently enacted reforms and introduced new regulations in
markets for consumer financial products (Campbell, Jackson, Madrian, and Tufano, 2011a,b).
Several recent contributions provide descriptive analyses of the effects of these reforms. In the
case of U.S. credit card markets, Agarwal, Chomsisengphet, Mahoney, and Stroebel (2015)
and Nelson (2020) analyze how regulatory limits on credit card pricing introduced by the
2009 CARD Act affect borrowing costs exploiting rich administrative data. Similarly, in a
contemporaneous contribution, Cuesta and Sepúlveda (2019) study price regulation in the
Chilean consumer loan market. We complement these papers by analyzing some of these
regulatory interventions in a quantitative model that introduces one additional key feature,
namely, borrowers’ cost of examining offers, and evaluate its importance for market power
and pricing in the credit card market.
Finally, this paper is related to the literature on the structural estimation of consumer

search models. Recent contributions include Hortaçsu and Syverson (2004), Hong and Shum
(2006), Ellison and Ellison (2009), Wildenbeest (2011), Allen, Clark, and Houde (2019),
Galenianos and Gavazza (2017), and Salz (2017). Our theoretical framework innovates on
these previous empirical papers by building on the models of Butters (1977) and Burdett
and Judd (1983), thereby combining search frictions and endogenous shopping effort by
consumers. Fershtman and Fishman (1994), Armstrong, Vickers, and Zhou (2009), Janssen
and Moraga-González (2004) show that these key features could potentially offset the effects of
the regulations that we focus on; thus, our framework that incorporates these features seems
well suited for a quantitative analysis of these policy interventions.4
3 Data
The available data dictate some of the modeling choices of this paper. For this reason, we
describe the data before presenting the model. This description also introduces some of the
identification issues that we discuss in more detail in Section 5.2.
3.1 Data Sources

Our quantitative analysis combines several sources of data. More specifically, we exploit some
of the datasets that SZ use in their descriptive analysis of households’ credit card terms,
supplementing them with some statistics obtained from credit card market reports of the
CFPB and from the Survey of Consumer Finances. We now describe these datasets in more
detail.
The first dataset is an account-level panel that samples individuals and reports the main
terms of their credit card accounts during (at most) 36 consecutive months between January
2006 and December 2008, including their credit limit, the end-of-month balance, the revolving
balance, the annual percentage rate (APR), and the cash advance APR. The dataset also
reports limited demographic characteristics of the cardholders, most notably their FICO credit
score.5
The second dataset reports the terms of all pre-approved credit card offers that a sample
of individuals received in January 2007.6 This second dataset samples different individuals
than those in the first dataset, but allows us to measure the number of offers that individuals
4
Knittel and Stango (2003) show that price ceiling served as a focal point for tacit collusion in the U.S.
credit card market during the 1980. However, they also show that price dispersion was a lot more limited in
that period than in the period of our data.
5
We are grateful to Victor Stango for sharing this dataset with us.
6
In the U.S. market, lenders often send personalized pre-approved credit card offers in the mail committing
to terms with borrowers.

receive in a given month, as well as the dispersion in the interest rates of these offers. As SZ
emphasize, the dispersion in interest rates on all credit cards offered to a given individual in a
given month removes any effect of individual-specific factors on the cross-sectional distribution
of interest rates on credit cards that individuals hold. We should point out that we do not
have access to the individual survey data, and thus we exploit data reported in tables of SZ.
We complement these datasets with some additional statistics: the shares of sub-prime,
near-prime, prime, and super-prime borrowers in the U.S population, calculated using the
distribution of FICO scores reported in Consumer Financial Protection Bureau (2012);7 the
fraction of individuals with credit cards, computed from Campbell, Haughwout, Lee, Scally,
and van der Klauuw (2016) and the 2007 Survey of Consumer Finances; the charge-off rate on
credit card loans in the first quarter of 2007, reported by the Board of Governors of the Federal
Reserve System; the interest rate of the one-year Treasury bill on January 16, 2007, which we
use as the risk-free rate;8 and Standard & Poor’s U.S. Credit Card Quality Index at January
2007, which is a monthly performance index that aggregates information of securitized credit
card receivables, most notably reporting an average cost of funding (i.e., excluding expected
charge-offs) for credit card loans.
3.2 Data Description

We use the first dataset on individuals’ credit card terms to sum up and extend one of the main
results of SZ’s descriptive analysis: a large dispersion of the interest rate distribution persists,
even after taking into account: 1) different default risk across individuals, as measured by
their FICO scores; 2) different card characteristics across borrowers, such as rewards; and 3)
different revolving balances across borrowers.
Specifically, the basic framework for this analysis is the following equation:
Rijt = γX Xit + γZ Zijt + ǫijt , (1)
where the dependent variable Rijt is the APR that individual i pays on credit card j in month
t; Xit are characteristics of individual i in month t, namely his default risk, measured by the
FICO score;9 Zijt are characteristics of individual i’s credit card j in period t, namely the
credit limit, rewards, and the credit balance; ǫijt are residuals.
7
These shares equal 0.215, 0.140, 0.166, and 0.479, respectively.
8
We retrieved the values of the charge-off rate and of the interest rate of the one-year Treasury bill from
FRED, Federal Reserve Bank of St. Louis, series https://fred.stlouisfed.org/series/CORCCT100S and
https://fred.stlouisfed.org/series/DGS1, respectively.
9
The dataset reports household income brackets for approximately 50 percent of the individuals in
the sample. In order to have larger sample sizes, we choose to report results obtained without including
income among the individual characteristics, but we have estimated equation (1) including income among the
individual characteristics as well, and obtained very similar results to those reported in Table 1.

Based on regression equation (1), we calculate the centered interest rate residuals as
′
Rijt = γ̂X X it + γ̂Z Z ijt + ǫ̂ijt , (2)
where γ̂X and γ̂Z are the coefficient estimates, X it and Z ijt are the sample averages of the
covariates of each regression, and ǫ̂ijt are the estimates of the residuals. Hence, (2) removes
the variation in Rijt due to the variation in Xit and in Zijt , while keeping that due to ǫijt .
We perform regression (1) and calculate interest rate residuals according to equation (2)
separately for four different groups of cardholders based on their FICO score: 1) sub-prime
borrowers, with FICO score strictly below 620; 2) near-prime borrowers, with FICO scores
between 620 and 679; 3) prime borrowers, with FICO scores between 680 and 739; and
4) super-prime borrowers, with FICO scores above 740. These different groups constitute
the main classification of borrowers used in the credit card industry (Consumer Financial
Protection Bureau, 2015). Hence, performing separate regressions for each group allows us
to capture in a flexible way the heterogeneity across them, and thus to obtain a reasonably
accurate measure of the dispersion in interest rates within each group of borrowers.
Table 1 reports coefficient estimates of several specifications of equation (1) and the main
percentiles of the resulting distribution of interest rates based on equation (2). Column (1)
uses the raw data over the entire sample period, which exhibit a large dispersion of interest
rates: the difference between the 90th and the 10th percentiles equals 18 percentage points
for subprime borrowers; it decreases for more-creditworthy borrowers, reaching a difference
of 10 percentage points for super-prime borrowers. Column (2) restricts the data to January
2007 (the date of our other data sources), showing the large dispersion of interest rates is
almost identical to that in column (1), for two reasons: a) limited aggregate variation exists
in interest rates over time; and b) limited within-account variation exists in interest rates.
Column (3) further restricts the data to cards without introductory “teaser” rates (i.e., low
initial rates that reset to higher rates after an initial offer period); of course, interest rates
increase relative to those displayed in column (2), but the increase is minimal; for example,
the difference between the 90th and the 10th percentiles slightly decreases to 16 percentage
points for subprime borrowers and 9 percentage points for super-prime borrowers.
The specification of column (4) introduces the main individual characteristic that should
affect pricing, that is, the credit risk of the individual, measured by the FICO score. Within
all groups, higher-risk individuals face higher interest rates. Averaging across all groups, a
10-point increase in the FICO score corresponds approximately to a 30-basis-point decrease
in interest rates, which is almost identical to the magnitude that Nelson (2020) estimates.10
10
Moreover, Nelson (2020) shows that: 1) the interest rate on a credit card changes in response to a change
in the credit card holder’s FICO score over time; and 2) the magnitude of the change in response to a change in
the FICO score over time is almost identical to the cross-sectional difference between individuals with different

However, the corresponding distribution of residual interest rates constructed as in equation
(2) displays a dispersion that is almost identical to that computed from the raw data in column
(3). The specification of column (5) further controls for other card characteristics, such as the
credit limit and an indicator variable that equals one if the card features some rewards (e.g.,
frequent flier miles or cash back), and zero otherwise, as well as the revolving balance. The
specification of column (6) further restricts the sample to cards with a revolving balance (i.e.,
cards used for borrowing beyond the 25-day grace period). Nevertheless, the large dispersion of
residual interest rates persists almost unaffected. Moreover, Appendix A presents robustness
checks that weigh each observation by the size and the persistence of its revolving balance
which leads to qualitatively similar results. This finding bolsters the argument that interest
rate dispersion is a robust feature when we focus our attention on individuals who use their
credit cards to borrow (and, hence, pay interest on their balances) rather than transactors
(who pay off their credit card balance at the end of each month).11
Overall, Table 1 attests to some remarkable features of credit card markets. First, although
more creditworthy borrowers on average pay lower interest rates, the difference in interest
rates within groups is substantially larger than the difference across groups. Notably, the
difference between the 90th and the 10th percentiles equals approximately 16 percentage
points for sub-prime borrowers, near-prime borrowers, and prime borrowers, whereas it equals
approximately 13 percentage points for super-prime borrowers. Hence, because the average
outstanding revolving balance of borrowers equals approximately $4,000 in our sample, moving
from the 90th percentile to the 10th percentile of interest rates would reduce borrowers’ annual
payment by approximately $500-$600. Second, observable credit card characteristics do not
seem to have a major effect on card pricing. A consequence of these two features is that a large
dispersion of interest rates persists once we account for borrower and card characteristics.
Table 2 combines all empirical targets of our quantitative model. Panel A reproduces
the percentiles of the distributions of interest rates derived in Table 1. Panel B reports
statistics on credit card offers that SZ document. Specifically, Section 5.1 of SZ recounts
that approximately 75 percent of individuals received two or more credit card offers during
January 2007; among them, the median and the mean number of offers was three and four,
respectively. For these individuals who received two or more offers, Table 4 of SZ reports
key percentiles of the distribution of the difference between the highest and the lowest offered
interest rates charged after the expiration of any introductory “teaser” period (if any).
FICO scores at credit card origination. These two observations imply that the long-term nature of the credit
card contract does not affect the magnitude of the correlation between the FICO score and the interest rate
in our data.
11
Some people use their credit cards as a means of payment and repay their balance in full at the end of
each month. For such transactors the interest rate is arguably not a salient feature of their credit card, since
they never actually pay interest charges.

Table 1: Dispersion of Interest Rates, by Borrower Group
Subprime Borrowers (1) (2) (3) (4) (5) (6)

FICO Score -0.018 -0.017 -0.018
(0.006) (0.006) (0.006)
Reward Card 0.863 0.405
(0.618) (0.674)
Credit Limit -0.170 -0.158
(0.094) (0.125)
Credit Balance 0.197 0.174
(0.110) (0.140)
R2 0.010 0.015 0.013
Observations 27,024 903 877 877 871 766
10th Percentile 11.90 12.74 14.24 14.30 14.41 14.39
25th Percentile 16.15 16.95 17.24 17.55 17.46 17.58
50th Percentile 20.65 21.20 21.74 21.72 21.65 21.93
75th Percentile 27.49 27.99 27.99 27.52 27.45 27.80
90th Percentile 29.99 30.24 30.24 30.06 30.01 30.16
Near-prime Borrowers
FICO Score -0.046 -0.043 -0.052
(0.011) (0.011) (0.013)
(0.453) (0.565)
(0.046) (0.078)
(0.064) (0.100)
R2 0.019 0.044 0.043
Observations 27,059 944 900 900 885 661
10th Percentile 10.49 11.24 12.99 13.16 13.09 13.20
25th Percentile 14.90 14.99 15.94 16.01 16.06 16.55
50th Percentile 18.24 18.99 19.24 19.09 19.31 20.20
75th Percentile 23.15 23.24 23.30 24.07 23.87 25.72
90th Percentile 28.99 29.24 29.24 29.03 28.75 29.16
Prime Borrowers
FICO Score -0.053 -0.048 -0.051
(0.011) (0.011) (0.015)
Reward Card 0.084 -0.257
(0.374) (0.520)
(0.026) (0.049)
(0.038) (0.058)
R2 0.024 0.039 0.029
Observations 31,115 1,003 953 953 932 605
10th Percentile 9.90 9.90 11.99 11.79 11.85 11.56
25th Percentile 12.99 13.99 14.31 14.65 14.64 14.81
50th Percentile 16.74 17.31 18.24 17.74 17.70 17.93
75th Percentile 19.99 20.24 20.34 21.04 20.97 21.90
90th Percentile 25.99 27.24 28.15 27.32 27.25 28.68
Super-prime Borrowers
FICO Score -0.008 -0.008 -0.024
(0.004) (0.005) (0.010)
(0.220) (0.471)
Credit Limit 0.010 0.028
(0.014) (0.031)
Credit Balance 0.058 -0.040
(0.032) (0.051)
R2 0.002 0.007 0.012
Observations 56,880 1,741 1,645 1,645 1,611 546
10th Percentile 9.90 9.90 11.24 11.33 11.20 10.79
25th Percentile 12.99 13.24 14.15 14.05 13.97 13.82
50th Percentile 15.98 16.74 16.99 17.02 16.91 16.84
75th Percentile 18.24 18.24 18.24 18.48 18.52 19.54
90th Percentile 20.24 20.24 20.31 20.56 20.81 23.98
Notes: This table reports OLS coefficient estimates of equation (1) and the corresponding percentiles of the distribution of
centered interest rates as in equation (2).

Panel C reports auxiliary statistics on credit card markets. We compute the fraction of
credit card revolvers in each group by combining the share of individuals with a credit card
in 2007 reported in Campbell, Haughwout, Lee, Scally, and van der Klauuw (2016) with the
probability of revolving conditional on having a credit card, which we compute directly as the
ratio of observations in column (6) to observations in column (5) in Table 1. Interestingly,
the share of individuals with a credit card is lower for borrowers with lower credit scores,
whereas the probability of revolving conditional on having a credit card is higher, exceeding
80 and 90 percent for near-prime and sub-prime borrowers, respectively. Hence, the fraction of
credit card borrowers is non-monotonic in borrowers’ credit scores; on average, 46 percent of
individuals borrow on their credit card.12 Finally, the aggregate charge-off rate approximately
equals four percentage points and the average funding cost reported by Standard & Poor’s
Credit Card Quality Index is approximately two percentage points above the risk-free rate.
3.3 Implications for Modeling

Table 2 provides an interesting description of the credit card market and informs the model
that we develop in Section 4. We focus on two key data patterns.
First, as we noted above, Panel A shows that the dispersion in the interest rates that similar
borrowers pay on their credit card debt is very large, even after we control for observable
borrower and card characteristics. Second, Panel B points out that many individuals receive
several credit card offers at substantially different interest rates. Hence, public information
about individuals’ repayment probability, as measured by FICO scores, does not account for
the dispersion observed in Panel A and, even more so, in Panel B. Moreover, if all individuals
chose the credit card with the lowest interest rate among all their offers, the level and the
dispersion of the accepted interest rate distributions would be considerably lower than those
reported in Table 2.
These striking patterns motivate some of our key modeling assumptions. We focus on
borrowers (rather than transactors) and we allow two possible explanations for why these
borrowers do not accept the credit card offers with the lowest interest rates: 1) They may not
examine all the offers that they receive; and 2) They may have idiosyncratic preferences for
card attributes that our data do not report. The calibration of Section 5 aims to quantitatively
match the dispersion of interest rates as well as the sizable number of offers that individuals
receive, and it allows us to assess the contributions of the two theoretical explanations for
the empirical patterns. Nevertheless, our calibration admits other (unmodeled) factors which
12
The aggregate share of the population with a credit card and the aggregate share of revolvers, computed
as the weighted averages of the corresponding group shares in our data, equal 0.76 and 0.46, respectively.
These shares closely match the corresponding aggregate statistics in the 2007 Survey of Consumer Finances,
which equal 0.73 and 0.44, respectively.
10

Table 2: Empirical Targets
Panel A: Accepted Offers

10th Percentile Accepted Offer Distribution, Sub-Prime Borrowers 14.39
10th Percentile Accepted Offer Distribution, Near-Prime Borrowers 13.20
10th Percentile Accepted Offer Distribution, Prime Borrowers 11.56
10th Percentile Accepted Offer Distribution, Super-Prime Borrowers 10.79
Panel B: Received Offers
Fraction Receiving 2+ Offers (%) 75.00
Median Number of Offers Received, Conditional on 2+ Offers 3.00
Average Number of Offers Received, Conditional on 2+ Offers 4.00
10th Percentile Distribution of Differences in Offered Rates 0.00
Panel C: Auxiliary Statistics
Fraction with Credit Card Debt, Sub-Prime Borrowers 54.56
Fraction with Credit Card Debt, Near-Prime Borrowers 55.33
Fraction with Credit Card Debt, Prime Borrowers 54.00
Fraction with Credit Card Debt, Super-prime Borrowers 36.02
Charge-Off Rate 4.01
Average Funding Cost 7.02
Notes: This table provides the empirical targets of our calibrated model. Panel A reports statistics
on the interest rates that borrowers pay on their credit cards. Panel B displays statistics on credit
card offers that SZ report. Panel C reports auxiliary statistics.
11

may feed into the large variance of the cross-sectional distribution of accepted rates, such as
adjustment of the interest rate after the offer is accepted, as in Nelson (2020).
Despite all of their advantages, however, we should acknowledge that our data have some
limitations. First, they are mostly cross-sectional, and therefore we do not observe borrowers’
and lenders’ behavior over time. Specifically, we cannot precisely assess how frequently
borrowers switch across credit cards. Hence, in the absence of more-detailed measurement
on borrowers’ switching behavior, we seek to match the cross-sectional distribution through
a static model. Moreover, although the theory can accommodate large multidimensional
heterogeneity of borrowers, our cross-sectional data make identifying such a model difficult.
Hence, our framework allows for flexible heterogeneity across markets (i.e., sub-prime,
near-prime, prime, and super-prime), as well as within-market heterogeneity in borrowers’
willingness to pay for credit and in borrowers’ valuation of non-price card attributes, whereas
some other parameters are common across borrowers within markets. Most notably, we
do not observe individual default and, thus, we abstract from within-market heterogeneity
in repayment risk, as well as from asymmetric information between lenders and borrowers
(though we should note again that differences in observable FICO scores across individuals
do not account for the dispersion in interest rates in Panel A and in Panel B). We discuss the
implications of these data limitations for our results further in Section 7.
4 The Model
The economy consists of J different markets, labeled by j, which are populated by borrowers
and lenders. The different markets operate independently from each other, and each agent
(borrower or lender) participates in a single market. Our calibration of Section 5 will consider
four different markets corresponding to the general classifications of creditworthiness used in
the credit card industry: sub-prime, near-prime, prime, and super-prime.
Each market j has measure 1 of borrowers (a normalization) who have market-specific
default risk ρj , want to take a loan of market-specific size bj and are heterogeneous in their
marginal valuation of a loan, z̃. We abstract from within-market heterogeneity in repayment
probability for reasons we describe in Section 3.3. Furthermore, we abstract from the intensive-
margin decision of how much to borrow for the sake of tractability, following Allen, Clark,
and Houde (2019), Crawford, Pavanini, and Schivardi (2018), and Nelson (2020), among
others. We allow for unobserved heterogeneity in borrowers’ marginal valuation z̃ which is
distributed according to a market-specific discrete distribution M̃j (·) with an Nj -point support
Z̃ = {z̃1 , ..., z̃Nj }, where z̃1 ≤ ... ≤ z̃Nj .13 We define sz̃ to be the share of type-z̃ borrowers,
13
We assume a discrete distribution of borrower types to facilitate some technical derivations. The
interaction between borrowers and lenders does not hinge on that assumption.
12

where z̃ ∈ Z̃.
Each market j has measure Λj of potential lenders who face entry cost χj to enter the
market and are heterogeneous in their marginal cost of providing a loan, k̃. The marginal
cost k̃ is distributed according to a market-specific smooth distribution Γ̃j (·) with connected
support [k̃ j , k̃ j ]. The measure of lenders who choose to enter market j and the distribution of
their marginal costs are Lj and G̃j (·), respectively. Every entering lender can give one loan of
size bj .14
Matching between borrowers and lenders in a market is subject to frictions. Each lender
sends one loan offer with an associated interest rate to a random borrower. Each borrower
chooses his examination effort and then receives a random number of offers that follows a
Poisson distribution, examines every offer with a probability that depends on his effort, and
decides which, if any, offer to accept.15 The effective number of offers (i.e., offers received and
examined) to a borrower who exerts examination effort e in a market with Lj lenders follows a
Poisson distribution with parameter e ∗ Lj . A borrower who exerts examination effort e incurs
cost qj (e, Lj ), where qj (·, ·) is strictly increasing and convex in both arguments and satisfies
qj (0, Lj ) = 0, lime→0 ∂qj (e, Lj )/∂e = 0 and lime→1 qj (e, Lj ) = ∞.
Borrowers consider two components to rank loan offers. The first component is the
net interest rate R, which is chosen by the lender and is drawn from the equilibrium
offer distribution FRj (·). The second component is an idiosyncratic (i.e., borrower-specific)
attribute a, which is stochastic and represents every other aspect of the loan that might
affect the borrower’s valuation. The attribute draw captures the importance of horizontal
product differentiation in this market, whose value may vary across borrowers. Idiosyncratic
attribute a is drawn from an exogenous distribution Faj (·) that is smooth, has zero mean,
and has support in a connected set [a, a] ⊂ (−∞, ∞). In this section and in the calibration of
Section 5, we assume attribute draw a is independent across lenders. Appendix D considers
the case where some lenders’ offers might consistently draw better values of a. We call the
sum c = R + a the (net) cost of a loan to the borrower, which might be higher or lower than
R depending on the attribute a.
If the borrower does not default, which occurs with probability 1 − ρj , he pays the cost of
the loan; if he defaults, which occurs with probability ρj , he incurs the utility cost of default
δ̃j .16 The expected utility of a type-z̃ borrower in market j who takes a loan with interest
14
A lender should be interpreted as a loan contract rather than a lending (or credit card) company. We do
not model lending companies explicitly.
15
The random allocation of offers across borrowers in an environment with finite numbers of borrowers and
lenders leads to urn-ball matching, which, as the numbers of borrowers and lenders grow large, is approximated
by a Poisson distribution. See Butters (1977) for an early application of urn-ball matching to a similar setting.
16
We assume defaulting occurs independently of any loan features, that is, the interest rate R or the
attribute draw a.
13

δ̃
rate R and attribute a is bj (z̃ − (1 − ρj )(1 + R + a) − ρj δj ), where δj ≡ bjj . The borrower’s
utility from not taking a loan is zero.
We define a borrower’s preference for a loan net of expected default cost and principal
z̃−ρ δ
repayment by z = 1−ρj j j − 1 and note that it is distributed according to Mj (z) = M̃j (1 −

ρj )(z + 1) + ρj δj with support Z = {z1 , ...zNj }. We can therefore rewrite the utility of a
type-z borrower from taking a loan with cost R + a as
bj (1 − ρj ) (z − R − a) .
Anticipating equilibrium behavior, a type-z borrower chooses the loan offer with the lowest
cost among the offers that he examines, conditional on the cost being less than z. A loan offer
with a higher cost generates negative utility, and thus the borrower will never accept it. The
ex-ante value of a type-z borrower in market j equals the expected value of his best loan offer
Vz,j (e) (which depends on effort e) net of the cost of effort, qj (e, Lj ):
Vz,j (e) − qj (e, Lj ). (3)
We denote the optimal (utility-maximizing) effort choice of a type-z borrower in market j by

ej (z).
The expected profits per dollar lent for a type-k̃ lender in market j equal the difference
between the expected revenues, given by the gross interest rate (1 + R) times the market-
specific repayment probability (1 − ρj ), and the costs, given by the lender’s gross cost of funds
(1 + k̃): (1 + R)(1 − ρj ) − (1 + k̃). We define the lender’s expected marginal cost inclusive
of non-repayment risk as k = k̃ + ρj , which means that per-dollar expected profits equal
R(1 − ρj ) − k. We note that for potential lenders k is distributed according to the smooth
distribution Γj (k) = Γ̃j (k − ρj ) with support [k j , kj ], where k j = k̃j + ρj and k j = k̃ j + ρj ; for
entrants, k is distributed according to Gj (k) = G̃j (k − ρj ).
The expected profits of a type-k lender from market j who offers interest rate R, πk,j (R),
are given by the probability of making a loan, denoted by Pj (R), times the loan’s expected
profits:

πk,j (R) = bj R(1 − ρj ) − k Pj (R). (4)
Notice that the idiosyncratic attribute a affects the lender’s payoff only through the probability
of making a loan.
We denote the optimal (profit-maximizing) interest rate choice of a type-k lender in
market j by Rj (k), which, combined with lenders’ entry decisions, determines the interest
rate distribution in market j, FRj (·).
14

We are now ready to define the equilibrium.
Definition 1 An equilibrium consists of borrowers’ effort ej (·) and lenders’ entry and interest
rate choices {Lj , Gj (·), Rj (·)} such that in every market j, borrowers maximize their ex-ante
value (3), lenders maximize their expected profits (4), the expected profits of all entrants exceed
the entry cost χj , and the expected profits of non-entrants would be strictly below χj if they
entered.
To proceed, we first determining borrowers’ and lenders’ optimal choices separately and
then prove the existence of equilibrium. Finally, we characterize the constrained efficient
outcome. Because there is no interaction across markets, we henceforth drop the j subscript
to ease notation. The reader should keep in mind, however, that all equilibrium outcomes are
market specific.
4.1 Borrowers’ Choices

We characterize borrowers’ optimal effort e(·) of examining offers, taking as given the measure
L of lenders in the market and the interest rate offer distribution FR (·). We should point out
that the type distribution of lenders G(·) and interest rate choices R(·) affect borrowers’
choices only through FR (·).
We begin by expressing Vz (e) in a convenient way. Denote the value of a z-borrower from
examining n offers by vz,n , where vz,0 = 0. The expected value for a type-z borrower who
exerts effort e is
∞
X e−eL (eL)n
Vz (e) = vz,n . (5)
n=0
n!
Notice that effort e affects the arrival rate of offers but does not enter vz,n ; therefore, it is
immediate from equation (5) that Vz (e) is continuous and differentiable in e. As a result, the
optimal effort choice e(z) solves
∂q(e, L)
Vz′ (e) = . (6)
∂e
To determine vz,n for n ≥ 1, recall that the borrower chooses the loan offer with the lowest
cost c, if c ≤ z. Let Fc (·) denote the distribution of c. Because the loan cost c is the sum of
two independent random variables (R and a), it is distributed according to
Z R
Fc (c) = Fa (c − R)dFR (R).
R
15

The distribution of the lowest cost out of n ≥ 1 draws from Fc (·) is
F̄c,n (c) = 1 − (1 − Fc (c))n .
Therefore, the value to a z-borrower of examining n ≥ 1 offers is

Z z
vz,n = b(1 − ρ) (z − c) dF̄c,n (c). (7)
−∞
The following proposition characterizes borrowers’ optimal effort e(·) of examining offers,
the resulting distribution of accepted rates and the fraction of borrowers who get a loan,
conditional on lenders’ actions.
Proposition 2 Given FR (·) and L:
1. The optimal effort of a type-z borrower, e(z), is unique and strictly increasing in z and
solves
∞
X e−eL (eL)n ∂q(e, L)
(vz,n+1 − vz,n ) L = , (8)
n=0
n! ∂e
where vz,0 = 0 and equation (7) defines vz,n for n ≥ 1.
2. The distribution of accepted offers equals

RR
−e(z)L R Fa (z−x)dFR (x)
P
1−
z∈Z sz e
HR (R) = R . (9)
P −e(z)L RR Fa (z−x)dFR (x)
1 − z∈Z sz e
3. The fraction of borrowers who get a loan is

X RR
Fa (z−x)dFR (x)
Q = 1− sz e−e(z)L R . (10)
z∈Z
4.2 Lenders’ Choices

We first characterize the optimal interest rate R(k) of a type-k lender, then aggregate the
actions of lenders who enter the market to obtain the interest rate offer distribution FR (·),
and finally characterize lenders’ entry decisions L and G(·) given borrowers’ effort e(·). To ease
notation, we denote the effective arrival rate of offers to a type-z borrower by α(z) ≡ e(z) ∗ L.
A borrower accepts a loan offer with interest rate R if he examines this offer, if this offer
yields the lowest cost from every offer that he examines (taking into account their attributes
16

a), and if this offer yields net positive utility to the borrower. The next lemma characterizes
the probability P (R) that a random borrower accepts a loan offer with interest rate R.
Lemma 3 Given FR (·), L, and e(·), the probability P (R) that borrowers accept a loan offer
with interest rate R is continuous and differentiable in R and equals
z−R RR
X Z
Fa (R+a−x)dFR (x)
P (R) = sz e(z) e−α(z) R dFa (a). (11)
z∈Z −∞
Furthermore, P ′ (R) < 0.
We proceed to characterize the optimal interest rate schedule R(·), the distribution of
interest rate offers FR (·), the distribution of accepted offers HR (·), and the fraction of
borrowers who get a loan.
Proposition 4 Given L, G(·), and e(·):
1. The profit-maximizing interest rate R(k) of a type-k lender is continuous and strictly
increasing in k.
2. R(·) solves the following functional equation:
z−R(k) Rk
X Z
Fa (R(k)+a−R(x))dG(x)
sz e(z) e−α(z) k dFa (a)
z∈Z −∞
z−R(k) Rk k
X
k
Z Z
−α(z) Fa (R(k)+a−R(x))dG(x)
= R(k) − sz e(z) e k α(z) Fa′ (R(k)
1−ρ z∈Z −∞ k
! !
Rk
Fa (z−R(x))dG(x)
+a − R(x)) dG(x) dFa (a) + e−α(z) k Fa′ (z − R(k)) . (12)
3. The interest rate distribution equals FR (x) = G (R−1 (x)) .
The following proposition completes the characterization of lenders’ entry decisions.
Proposition 5 Given borrowers’ effort e(·), lenders’ entry satisfies the following:
1. A cutoff cost k̂ exists such that a lender enters if and only if k ≤ k̂.
2. The measure of lenders in the market equals L = ΛΓ(k̂) and the cost distribution of
Γ(k)
entrants equals G(k) = Γ(k̂)
for k ≤ k̂ and G(k) = 1 for k > k̂.
17

3. The cutoff cost k̂ solves
Z z−R(k̂) R k̂
Fa (R(k̂)+a−R(x)) d Γ(x)
X
−e(z)ΛΓ(k̂)
b R(k̂)(1 − ρ) − k̂ sz e(z) e k Γ(k̂) dFa (a) = χ.
z∈Z −∞
(13)
4.3 Existence of the Equilibrium

We can prove existence, based on the preceding results.
Proposition 6 An equilibrium exists.
4.4 Constrained Efficiency

We now analyze the case of a social planner whose goal is to maximize aggregate welfare
subject to frictions. The planner chooses lenders’ entry decisions, as well as borrowers’
examination effort and trading-decision rules (the interest rate simply redistributes surplus
between borrowers and lenders; thus, it does not matter for the planner’s problem).
We denote by e∗ (z) the planner’s optimal solution for the examination of a type-z borrower.
The entry decision rule is, trivially, a cutoff rule, and we denote the planner’s optimal cutoff
cost by k̂ ∗ . The surplus of a loan from a type-k lender with attribute a to a type-z borrower
k
, where z = z̃−ρδ

is b(1 − ρ) z − a − 1−ρ 1−ρ
− 1 and k = k̃ + ρ, as before. The planner’s optimal
trading-decision rule is that the borrower trades with the lowest-cost lender as long as the
surplus is positive.
The overall surplus given borrowers’ effort e(·) and lenders’ entry cutoff k̂ is
X
W e(·), k̂ = sz Wz e(z), k̂ − q (e(z), L) − χL, (14)
z∈Z
where Wz (e(z), k̂) is the expected surplus that a type-z borrower obtains from the offers that

he receives; q e(z), L is a z-borrower’s examination effort cost; L = ΛΓ(k̂) is the measure
of lenders who enter the market; and χL are aggregate lenders’ entry costs. Notice that
no interaction occurs among borrowers regarding their examination efforts; thus, Wz only
depends on the effort of the type-z borrower and does not depend on the full effort schedule.
k
The cost of a loan for the planner is w ≡ 1−ρ + a. The planner’s loan cost is distributed
18

according to
k̂
k
Z
Fw (w) = dG(k)
Fa w −
k 1−ρ
Z k̂
1 k
= Fa w − dΓ(k).
Γ(k̂) k 1−ρ
The distribution of the lowest w among n offers is

!n
k̂
1 k
Z
F̄w,n (w) = 1 − 1− Fa w − dΓ(k) .
Γ(k̂) k 1−ρ
The social value to a type-z borrower of receiving n offers is:

Z z
Wz,n (k̂) = b(1 − ρ) (z − w)dF̄w,n(w), (15)
−∞
where Wz,0 (k̂) = 0. Notice these terms only depend on k̂ and do not depend on borrowers’
effort e.
The surplus that a type-z borrower who examines offers with effort e generates from the
offers that he receives when lenders’ entry cutoff is k̂ equals
∞
X e−eL (eL)n
Wz (e, k̂) = Wz,n (k̂). (16)
n=0
n!
We now characterize the planner’s optimal solution.

Proposition 7 The constrained-efficient allocation is as follows:
1. The optimal effort for a type-z borrower e∗ (z) satisfies
∞
e−α(e (z),L ) (α(e∗ (z), L∗ ))n
∗ ∗ ∂α(e∗ (z), L∗ )
X ∂q (e∗ (z), L∗ )
Wz,n+1 (k̂ ∗ ) − Wz,n (k̂ ∗ ) = ,
n=0
n! ∂e ∂e
(17)
where Wz,n (k̂ ∗ ) is defined by equation (15), and L∗ = ΛΓ(k̂ ∗ ) is the optimal measure of
lenders in the market. A unique solution e∗ (z) exists for each z.
2. The optimal entry-cost cutoff k̂ ∗ for lenders satisfies

 
X ∂Wz e∗ (z), k̂ ∗ ∂q(e (z), L ) ′ ∗ 
∗ ∗
sz  − ΛΓ (k̂ ) = χΛΓ′ (k̂ ∗ ), (18)
z∈Z ∂ k̂ ∂L
19

where
e−α(e (z),L ) (α(e∗ (z), L∗ ))n

∞ ∗ ∗ Z z
∂Wz (e∗ (z), k̂ ∗ ) X ∂ F̄w,n (w)
= b (z − w)d
∂ k̂ n=0
n! −∞ ∂ k̂

∂α(e (z), L ) ′ ∗
∗ ∗
+ ΛΓ (k̂ ) Wz,n+1 (k̂ ∗ ) − Wz,n (k̂ ∗ ) , (19)
∂L
and
!n−1 !
k̂
∂ F̄w,n (w) 1 k Γ′ (k̂) k̂
Z
= n 1− Fa w − dΓ(k) Fa w− Γ(k̂)
∂ k̂ Γ(k̂) k 1−ρ Γ(k̂)2 1−ρ
Z k̂ !
k
− Fa w − dΓ(k) .
k 1−ρ
The decentralized equilibrium of the economy features two potential sources of inefficiencies
relative to the planner’s allocation. First, for a given measure of lenders, some meetings in
k
which trade is efficient (i.e., z > 1−ρ + a) feature no trade, because the interest rate of lenders
is excessive (i.e., R > z − a) due to lenders’ market power. Second, the measure of lenders is
not optimal (i.e., L 6= L∗ ).
5 Quantitative Analysis
The model does not admit an analytic solution for all endogenous outcomes. Hence, we
choose the parameters that best match moments of the data with the corresponding moments
computed from the model’s numerical solution. We then study the quantitative implications
of the model evaluated at the calibrated parameters.
5.1 Parametric Assumptions

The calibration requires that we make parametric assumptions for each of the four separate
markets, that is, sub-prime, near-prime, prime, and super-prime borrowers.
We borrow some parametric assumptions about the distributions of borrowers’ and lenders’
heterogeneity from papers that structurally estimate search models of the labor market and our
prior work on the retail market for illicit drugs (Galenianos and Gavazza, 2017). Specifically,
given the similarity in modeling frameworks and empirical targets between this paper and
those predecessors, we choose a discretized lognormal distribution with parameters µzj and
σzj and Nj = 20 support points for the distribution Mj (z) of buyers’ preferences z in market j.
Moreover, we assume that the distribution G̃j (k̃) of sellers’ costs k̃ is common across markets
20

and follow a right-truncated Pareto distribution with shape ξ, scale equal to the risk-free rate—
we use the interest rate of the one-year Treasury bill at January 16, 2007, which equals 5.06
percent—and an upper-truncation point k̂. The assumption of a common cost distribution
across markets means that we implicitly assume that the mass of potential lenders Λj varies
across markets.
We normalize the loan size to bj = 1. We further assume the following: 1) the effort cost
of examining offers qj (e, Lj ) equals β0j eβ1 ; 2) the charge-off rate in market j equals ρj ; and
3) attribute a is unobserved in our data, and it follows a normal distribution with mean zero
and standard deviation σaj , and it is uncorrelated with R. Appendix D reports the results of
the calibration in the case in which attribute a is correlated with the interest rate R.
Finally, we assume that the reported accepted rates R̂j and the “true” accepted rate R are
related as R̂j = Rj η, where η is a random variable, identically and independently distributed
across observations, drawn from a lognormal distribution with parameters (µη , ση ), common
across markets, and with mean to equal 1. Hence, reported rates are unbiased and the
parameters (µη , ση ) satisfy µη = −0.5ση2 . The literature that structurally estimates search
models of the labor market frequently assumes that wages are measured with error. In our
application, surveyed borrowers may report the interest rates that they pay on their credit
card debt incorrectly. Moreover, the random variable η may also accounts for some additional
factor that our model does not consider, such as adjustment of the interest rate after the offer
is accepted, as in Nelson (2020). Table 2 shows that the distributions of accepted rates display
a large dispersion, and these random η allow the model to more precisely match this feature
of the data quantitatively.
5.2 Calibration
n o
We choose the vector ψ = Lj , µzj , σzj , ξ, k̂, ρj , σaj , β0j , β1 , ση that minimizes the distance
j∈J
between the target moments m reported in Table 2 and the corresponding moments of the
model. We calibrate two versions of the model: in the first one, we impose ση = 0; in the
second one, ση can take any positive value.
Specifically, for any value of the vector ψ, we solve the model of Section 4 to find its
equilibrium: the distribution FRj (k) of offered interest rates and borrowers’ effective arrival
rate αj (z) in each market j that are consistent with each other. Once we solve for these
policy functions of borrowers and lenders in each market j, we compute the equilibrium
distributions of interest rates of received offers and of accepted offers. In practice, we simulate
these distributions and compute the moments m (ψ) corresponding to those reported in Table
2 on received offers and on accepted offers, as well as the aggregate fraction of credit card
borrowers in each market j. Panel A and Panel C in Table 2 report the distribution of accepted
21

interest rates and the charge-off rate, respectively, for each group j, whereas Panel B reports
moments of the distribution of the number and of the offered rates aggregated for the entire
market. Hence, we use weights ωj corresponding to the population share of each group j (see
footnote 7) to aggregate the number of received offers and the distribution of their interest
rates.
We choose the parameter vector ψ that minimizes the criterion function
(m (ψ) − m)′ Ω (m (ψ) − m) ,
where m (ψ) is the vector of stacked moments simulated from the model evaluated at ψ and
m is the vector of corresponding sample moments. Ω is a symmetric, positive-definite matrix;
in practice, we use the identity matrix.
5.3 Data-Generating Process

Matching the moments reported in Table 2 requires that we account for the fact that the data-
generating process may be unusual, because we combine two separate datasets, collected for
different purposes. Specifically, the dataset on received offers reports all offers that borrowers
in group j receive, whose arrival rate is Lj , and not exclusively the offers that borrowers
examine in equilibrium, which may be lower than the offers received because borrowers’
endogenous examination effort e may be less than full effort e = 1. We derive in Appendix
B the average number of offers and the distribution of the difference between the highest and
lowest offers that borrowers receive under the assumption that the arrival rates of these offers
equal Lj .
However, lenders send these offers anticipating that borrowers receive and examine them
according to their equilibrium ej (z) ∗ Lj . Hence, the moments on the empirical distribution
of accepted offers reflect borrowers’ endogenous examination effort ej (z) ∗ Lj .
5.4 Sources of Identification

The identification of the model is similar to that of other structural search models. Specifically,
although the model is highly nonlinear, so that (almost) all parameters affect all outcomes,
the identification of some parameters relies more heavily on certain moments in the data.
The moments on the number of offers that borrowers receive identify the average offer rate
P
j ωj Lj (where ωj are the known shares of borrowers in each group), and thus contribute to
the identification of group-specific offer rates Lj .17 Similarly, the aggregate charge-off rate is
17
We should point out that the main implications of the model do not particularly rely on the specific
values of Lj , but rather on borrowers’ effective arrival rates α(z) = e(z)Lj , which are identified from the
distributions of costs, of offered rates, and of accepted rates, as we explain shortly. Hence, different values of
22

informative about the group-specific default rates ρj . Moreover, we identify the parameter ξ
of the distribution G(k) of sellers’ heterogeneity from the average funding cost reported by
Standard and Poor’s.
Furthermore, we observe the distribution of the difference between the highest and lowest
offered interest rates R that borrowers receive. We show in Appendix B that this distribution
depends in a precise way on the offer distribution FRj (·), which allows us to recover FRj (·).
With this knowledge, we still have to identify three sets of parameters that determine 1)
the distribution of borrowers’ preferences; 2) borrowers’ examination effort; and 3) the extent
of product differentiation (variation in attributes). Proposition 4 shows that these three sets of
parameters shape three mappings between observable outcomes: A) the mapping between the
distribution G(k) of costs k and the distribution FRj (R) of offered rates R(k); B) the mapping
between the offer distribution FRj (R) and the distribution of accepted rates HRj (R); and C)
the mapping between the offer distribution FRj (R) and the fraction of borrowers who get a
loan Qj . Hence, these three outcomes jointly identify the remaining three sets of parameters.
Intuitively, given lenders’ costs G(k), the dispersion of offers (i.e., mapping A) increases
in the dispersion of borrowers’ preferences, and decreases in the standard deviation σa of
the product attribute a, because lenders (most notably, low-cost lenders) charge similar rates
anticipating that consumers’ choice depends relatively less on interest rates when a displays
larger values. Furthermore, given the offer distribution FRj , the dispersion of accepted interest
rates (i.e., mapping B) increases in examination costs, because borrowers examine fewer offers
when costs are high, and in the standard deviation σa of the product differentiation, because
larger values of a imply that interest rates affect consumers’ choice relatively less than smaller
values. Similarly, given the offer distribution FRj , it is apparent from equation (10) that the
fraction of borrowers (i.e., mapping C) increases as examination effort α(z) increases. The
discussion of the calibrated parameters of Section 5.5 and the comparative statics of Section
5.7 will further clarify how examination costs and product attribute a differentially affect
market outcomes.
Finally, lenders’ free-entry condition (equation (13)) implies that we can recover lenders’
fixed costs χj from the variable profits of the highest-cost lender in each market.
5.5 Calibrated Parameters and Model Fit

Table 3 reports the calibrated parameters of the model. Panel A refers to the version without
measurement error (ση = 0) and Panel B to the version with measurement error (ση > 0).
Overall, the parameters are almost identical across versions, and thus we now discuss on those
in Panel A of Table 3 only. As we recount above and Table 4 shows in detail, the measurement
Lj would imply different equilibrium values of s(z) (and, thus, different costs β0j ), keeping α(z) unchanged.
23

Table 3: Calibrated Parameters
Panel A: No Measurement Error Panel B: Measurement Error

µ z1 3.613 σz1 0.131 µ z1 3.575 σz1 0.121
µ z2 3.518 σz2 0.093 µ z2 3.528 σz2 0.111
µ z3 3.461 σz3 0.131 µ z3 3.447 σz3 0.125
µ z4 3.251 σz4 0.337 µ z4 3.223 σz4 0.192
ξ 3.774 k̂ 10.182 ξ 4.183 k̂ 9.656
L1 1.509 L2 3.779 L1 1.550 L2 3.943
L3 3.219 L4 2.999 L3 3.214 L4 2.983
ρ1 0.033 ρ2 0.020 ρ1 0.040 ρ2 0.030
ρ3 0.015 ρ4 0.011 ρ3 0.020 ρ4 0.010
σa1 0.105 σa2 0.124 σa1 0.081 σa2 0.121
σa3 0.138 σa4 0.153 σa3 0.156 σa4 0.128
β01 8.637 β02 34.714 β01 8.630 β02 42.264
β03 27.001 β04 28.079 β03 29.002 β04 32.478
β1 1.627 ση 0.000 β1 1.741 ση 0.272
Notes: This table reports the calibrated parameters. Panel A refers to the version without
measurement error (ση = 0), and Panel B to the version with measurement error (ση > 0)
error η allows the model to capture the dispersion of accepted offers more precisely.
The parameters µzj and σzj of the distributions of z in group j mean that borrowers’
willingness to pay for credit is, on average, large and displays large heterogeneity within group,
as well as across groups. Specifically, borrowers’ average willingness to pay decreases as their
creditworthiness
r increases. The standard deviation of the willingness to pay, which equals
2
2
2µ +σ σ
e zj zj e zj − 1 , is non-monotonic in creditworthiness, with super-prime borrowers
displaying a standard deviation almost nine times larger than that of near-prime borrowers.
The parameters ξ and k̂ of the distribution of costs k̃ imply that the average costs of all
entrants (not weighted by market shares) equal 636 basis points (the average funding cost
used in the calibration weighs lenders by their market shares, and it equals 613 basis points
at the calibrated parameters). Thus, average costs display a small spread of approximately
130 basis points over the risk-free rate. Moreover, the heterogeneity of lenders’ costs is small,
that is, the standard deviation of costs equals 106 basis points. Thus, the model generates a
large dispersion of offered rates even with a small dispersion of costs.
24

The values of Lj indicate that lenders send, on average, approximately 2.7 credit card offers,
with considerable heterogeneity across groups—sub-prime borrowers receive approximately
half the offers that near-prime, prime, and super-prime borrowers receive. The number of offers
is non-monotonic in the creditworthiness of borrowers, thereby matching the patterns that
Han, Keys, and Li (2018) report. However, the parameters β0j and β1 imply that borrowers
examine only a small fraction of these offers: the cost of effort to examine an average number
of offers per period equal to α = 1 corresponds to approximately 450 basis points, and it
increases by almost 1,000 basis points to examine an average number of offers per period
equal to α = 2.
The value of σaj implies that the standard deviation of the product attribute a is small,
relative to the overall heterogeneity in borrowers’ preferences. Moreover, the value of σaj
increases with borrowers’ creditworthiness, thereby indicating that non-price card attributes
matter relatively more as borrowers’ risk scores increase.
To understand why the calibration results in a sizable role for examination costs and a small
one for product differentiation, we turn to the parts of the model that most contribute to their
identification, namely, the mappings from lenders’ costs to lenders’ offers and from lenders’
offers to borrowers’ outcomes. Key features of the data are that 1) the level of offered interest
rates is high relative to lenders’ costs; 2) the distribution of offered rates is very dispersed; 3)
the accepted rate distribution is similar to the offered rate distribution; 4) borrowers receive
several credit card offers; and 5) the share of households with credit card debt is moderate.
The central question is why borrowers do not systematically end up with low interest rates
given that they receive many offers from a highly dispersed offer distribution.
Examination costs and product differentiation are two features that could rationalize the
fact that individuals do not borrow at low interest rates. If examination costs are high,
borrowers examine few of the offers that they receive—often one only—which results in high
and dispersed accepted rates. If examination costs are low, borrowers examine several offers,
which tends to reduce the level and the dispersion of accepted rates. High accepted rates
could still occur with low examination costs if borrowers choose an offer because the value of
its attribute a is large, which occurs more often when product differentiation is an important
feature of the market, that is, when σa is high. However, the combination of low examination
cost and high product differentiation leads to the following: 1) On the supply side, lenders
set similar interest rates, because they matter less for borrowers’ choices when the variance of
the attribute a is larger; 2) Holding the distribution of offered rates constant, on the demand
side, many borrowers take out loans, because borrowing is attractive due to large values of a.
However, the data show that the dispersion of offered rates is very high, and that the aggregate
fraction of individuals with credit card debt is moderate (approximately 45 percent).
In summary, the model with a large variance of the product-differentiation parameter a
25

cannot simultaneously match the moderate fraction of borrowers, as well as the high dispersion
of offered and of accepted rates that we observe in the data, whereas the model with high
examination costs β0j matches these features quite precisely. The comparative statics of
Section 5.7 further illustrate these issues.
Finally, the calibrated ση equals 0.272, which means that the standard deviation of the
measurement error on the accepted rates equals 0.279. This value is small relative to the
calibrated standard deviations of accepted rates R in the version without measurement error,
which equal 3.50 in the sub-prime market, 3.45 in the near-prime market, 3.08 in the prime
market, and 2.33 in the super-prime market, respectively.
Table 4 presents a comparison between the empirical moments and the moments calculated
from the model at the calibrated parameters reported in Panels A and B of Table 3,
respectively. The model without measurement error matches the data well, though, as
anticipated, it underpredicts the dispersion of accepted rates—i.e., it overpredicts the lower
percentiles and it underpredicts the higher percentiles. It matches reasonably well the
percentiles of the distribution of the difference between the highest and lowest offered interest
rates, and almost perfectly the aggregate statistics on the fraction of credit card borrowers
in each group, thereby reproducing the mild non-monotonicity of the fraction of borrowers as
their creditworthiness increases observed in the data. The model with a small measurement
error on accepted offers matches the data almost perfectly. Perhaps the most notable difference
between the model and the data is the fact that the model underpredicts the aggregate charge-
off rate.
5.6 Model Implications

We study the implications of the model evaluated at the parameters reported in Panel A of
Table 3. Because these parameters are very similar to those of Panel B, the implications of
the model evaluated at the latter parameters are very similar as well.
Figure 1 displays lenders’ and borrowers’ equilibrium policies in each market j, shown in
each row. The left column of Figure 1 displays lenders’ optimal offered rate R(k) (solid line,
left axis) to each group of borrowers as a function of their cost k, as well as the density of
lenders’ cost k (dotted line, right axis) for values of the cost k from the risk-free rate up to the
cutoff value k̂ that the free-entry condition (13) determines. Lenders’ offered rates are strictly
R(k)(1−ρj )−k
increasing in their costs k, as Proposition 4 states. Markups, computed as k
, on
average, exceed 100 percent. They are non-monotonic in borrowers’ creditworthiness: they
equal 117, 154, 149, and 126 percent in the subprime, near-prime, prime, and super-prime
markets, respectively.
The right panel of Figure 1 displays borrowers’ effective arrival rate of offers α(z) (solid
26

Table 4: Model Fit
Data Model Model

ση = 0 ση > 0
10th Percentile Accepted Rate, Sub-prime Borrowers 14.39 18.31 14.64
10th Percentile Accepted Rate, Near-prime Borrowers 13.20 17.01 13.68
10th Percentile Accepted Rate, Prime Borrowers 11.56 15.61 12.43
10th Percentile Accepted Rate, Super-prime Borrowers 10.79 14.17 11.31
Fraction Receiving 2+ Offers (%) 75.00 74.20 74.57
Median Number of Offers Received, Conditional on 2+ Offers 3.00 3.00 3.00
Average Number of Offers Received, Conditional on 2+ Offers 4.00 3.47 3.49
10th Percentile Distribution of Differences in Offered Rates 0.00 1.54 1.19
Fraction with Credit Card Debt, Sub-Prime Borrowers 54.56 55.87 54.52
Fraction with Credit Card Debt, Near-Prime Borrowers 55.33 56.19 55.29
Fraction with Credit Card Debt, Prime Borrowers 54.00 54.93 54.16
Fraction with Credit Card Debt, Super-prime Borrowers 36.02 36.58 36.17
Charge-Off Rate 4.01 1.90 2.28
Average Funding Cost 7.02 6.07 6.00
Criterion Function 139.15 19.02
Notes: This table reports the values of the empirical moments and of the moments calculated at the
calibrated parameters reported in Table 3.
27

Interest Rate R(k) 35 0.8 2 0.15
Arrival Rate α(z)

Pdf of k
Pdf of z
22.5 0.4 1 0.075
10 0 0 0
8.35 10.9 13.5 0 25 50
Cost k Willingness to Pay z
35 0.8 2 0.15
Interest Rate R(k)
Arrival Rate α(z)

Pdf of k
Pdf of z
22.5 0.4 1 0.075
10 0 0 0
7.07 9.6 12.2 0 25 50
35 0.8 2 0.15
Interest Rate R(k)
Arrival Rate α(z)

Pdf of k
Pdf of z
22.5 0.4 1 0.075
10 0 0 0
6.59 9.1 11.7 0 25 50
35 0.8 2 0.15
Interest Rate R(k)
Arrival Rate α(z)

Pdf of k
22.5 0.4 1 0.075 Pdf of z
10 0 0 0
6.18 8.7 11.3 0 25 50
Figure 1: The left panels display lenders’ optimal interest rate R(k) (solid line, left axis) as a function
of their cost k, as well as the density of lenders’ cost k (dotted line, right axis). The right panels
display borrowers’ optimal arrival rate α(z) (solid line, left axis) as a function of their willingness to
pay z, as well as the density of borrowers’ willingness to pay z (dotted line, left axis). The first row
refers to the sub-prime market, the second row to the near-prime market, the third row to the prime
market, and the fourth row to the super-prime market.
28

line, left axis), which is the outcome of borrowers’ optimal examination effort e, as a function
of their willingness to pay z, as well as the density of borrowers’ willingness to pay z (dotted
line, left axis). Because the lowest-valuation borrowers have a willingness to pay that is
below almost all offered interest rates and the product attribute a has a small variance, these
borrowers do not exert any effort to examine offers. More generally, the examination effort
is low—on average, borrowers examine approximately 0.7 offers—and only borrowers whose
willingness to pay z is in the highest 15 percent of the distribution choose α(z) larger than 1.
Figure 2 displays the probabilities Pj (R) that borrowers accept a credit card offer with
an interest rate R. These probabilities are obviously decreasing in R, but perhaps the most
striking features of Figure 2 are that 1) for any R, the probability that subprime borrowers
accept such an offer is at least twice as high as the probability that borrowers in any other
risk group accept it, consistent with the evidence that Agarwal, Chomsisengphet, Mahoney,
and Stroebel (2018) report; and 2) because of borrowers’ low examination effort, they are
quite flat, which means that borrowers’ demand is quite inelastic—the average elasticity
equals approximately −1.70, which is remarkably similar to the elasticity that Nelson (2020)
estimates. These external comparisons seem to suggest that our calibration yields reasonable
parameters. Overall, the average acceptance probability Pj (R) equals 0.17, which, scaled up
PJ
by the mass of lenders j=1 ωj Lj ≈ 2.7, yields the aggregate fraction of individuals with
credit card debt of 45.7 percent.
Figure 3 plots the distributions FRj (R) of offered rates and the distributions HRj (R) of
accepted rates. Of course, the distributions of offered rates first-order stochastically dominate
the distribution of accepted rates. However, the differences between the two distributions
are small. Two reasons account for these small differences: 1) borrowers’ low examination
effort implies that the rate α(z) at which they consider offers is low; and 2) borrowers do not
always accept the offer with the lowest interest rate, because of the differentiation attribute a.
However, this second factor is quantitatively smaller than the first one, because the standard
deviation σa is small and because, for a to have sizable effect, borrowers would need to consider
more than one offer, which happens very infrequently due to their high costs of examining
them. Thus, the mean of the distribution of accepted rates would be almost identical if
borrowers were to always choose the offer with the lowest interest rates.18
The top rows of Table 5 report summary statistics of market outcomes—prices and
quantities—as well as consumer surplus, lenders’ profits, and aggregate welfare in each market.
The calibrated model implies that consumer surplus, lenders’ profits, and aggregate welfare,
although they are all lowest in the super-prime market. The bottom two rows of Table 5
report the optimal number of offers and welfare in the constrained-efficient economy that we
18
Of course, this is not a full equilibrium argument, as the endogenous distribution of offered rates FR (·)
depends on the product attribute a.
29

Prob. of Acceptance P (R) 0.6 0.6
Prob. of Acceptance P (R)

0.3 0.3
0 0
10 22.5 35 10 22.5 35
Interest Rate R Interest Rate R
0.6 0.6

0.3 0.3
0 0
10 22.5 35 10 22.5 35
Figure 2: Probability P (R) that borrowers accept an offer with interest R, for sub-prime borrowers
(top-left panel), near-prime borrowers (top-right panel), prime borrowers (bottom-left panel), and
super-prime borrowers (bottom-right panel).
characterized in Section 4.4. As our theoretical analysis points out, the constrained-efficient
allocation differs from the market allocation. Most notably, it requires a larger number of
offers Lj in each market, on average, by 7 percent; the resulting welfare gains would be large,
ranging from 21 percent in the sub-prime market to 30 percent in the super-prime market.
5.7 Comparative Statics

We further illustrate the working of our model through two comparative statics that vary
the two parameters that are the main focus of our framework, namely the parameter β0j
that affects the effort cost of examining offers, and the standard deviation σaj of the product
attribute a.
We present the results of these comparative statics for near-prime borrowers—that is,
the group for which the model without measurement error matches the data most precisely,
according to Table 4—but the outcomes for the other groups are similar.
30

1 1
CDF
CDF
0.5 0.5
0 0
10 22.5 35 10 22.5 35
1 1
CDF
CDF
0.5 0.5
0 0
10 22.5 35 10 22.5 35
Figure 3: The solid line displays the cumulative distribution function HR (R) of accepted interest
rates and the dotted line displays the cumulative distribution function FR (R) of offered rates.
Cost of Examinining Offers. Figure 4 compares outcomes of the model at the calibrated
parameters (solid line) with those of the model when we decrease the parameters β0j of the
cost of effort by 30 percent (dotted line) while holding all other parameters at their calibrated
values.
The top-left panel shows that the interest rate function R(k) is lower than that in the
baseline case, as all lenders uniformly decrease their interest rates. The decrease is larger for
low-cost lenders than for high-cost lenders, because borrowers accept high-cost lenders’ offers
almost exclusively when borrowers consider one of these high offers only, and thus high-cost
lenders do not need to lower their rates as much as low-cost lenders. The top-right panel
explains why lenders’ offered rates are lower: because the cost of effort is lower, borrowers
increase their search effort.
The bottom-left panel shows that the probability P (R) that borrowers accept an offer with
a given interest rate R is higher than that of the baseline case for low values of R and lower
for high values of R. The reason is that borrowers consider a larger number of offers, and thus
31

Table 5: Market Outcomes and Welfare
Sub- Near- Prime Super-

Average Number of Offers per Borrower 1.51 3.78 3.22 3.00
Average Accepted Rate 22.42 21.20 19.35 16.44
Standard Deviation of Accepted Rates 3.49 3.41 3.07 2.31
Fraction of Borrowers 55.70 55.48 54.27 38.80
Consumer Surplus 5.10 4.18 4.31 3.76
Lender Profits 1.48 1.53 1.49 1.08
Welfare 6.58 5.71 5.80 4.84
Efficient Average Number of Offers per Borrower 1.62 4.07 3.47 3.23
Efficient Welfare 7.99 7.19 7.22 6.29
Notes: This table reports market outcomes and welfare in each market.
their probability of accepting any offer increases, but they are relatively less likely to accept
high-interest-rate offers. Demand becomes more elastic relative to that of the baseline case.
Moreover, because lenders decrease their rates and borrowers accept offers with lower rates
with a higher probability, the fraction of individuals with credit card debt increases relative
to its value in the baseline case—from 55.5 percent to 68.8 percent.
The bottom-right panel of Figure 4 displays the distribution of offered rates (thick lines)
and of accepted rates (thin lines). Both distributions obtained in the model with a lower
β0j (dotted lines) are first-order stochastically dominated by the corresponding distributions
obtained in the model at the calibrated β0j (solid lines). The reason is that low-cost lenders
decrease their offered rates, because borrowers compare more offers if their effort to examine
them is less costly. The average offered and accepted rates equal 20.45 and 19.03, respectively,
and the standard deviation of offered and accepted rates equal 4.21 and 3.93, respectively,
when the cost-of-effort parameter β0j is 30 percent lower than its calibrated value. As the
bottom plots shows, the lower cost of effort affects lower percentiles relatively more than
higher percentiles.
Figure 4 also helps us understand why the calibrated model calls for relatively large effort
costs: if they were smaller, the level of offered and of accepted interest rates would be lower,
and the fraction of borrowers would be higher than those observed in the data.
Product Differentiation. Figure 5 compares outcomes of the model at the calibrated

parameters (solid line) with those of the model with lower search costs (dotted line) and those
of the model when we further increase the standard deviation σaj of the product attribute
a (dashed line), while holding all other parameters at their calibrated values. Because σaj
is calibrated to be small, we increase it by a factor of 30, which makes the value of the
interquartile range of a similar to that of R observed in the data.
The top panels show interesting outcomes. Most notably, the top-left panel shows that
32

Interest Rate R(k) 35 2
Arrival Rate α(z)

22.5 1
10 0
7.1 9.6 12.2 0 25 50
0.4 1
CDFs FR (R) and HR (R)

0.2 0.5
0 0
14.1 21.7 29.3 10 22.5 35
Figure 4: These panels display model outcomes at the calibrated parameters (solid line) and in the
case when β0j′ = 0.7β
0j (dotted line). The top-left panel displays lenders’ optimal interest rate R(k)
as a a function of their cost k; the top-right panel displays borrowers’ effective arrival rate α(z) as
a function of their willingness to pay z; the bottom-left panel displays the probability P (R) that
borrowers accept an offer with interest rate R; and the bottom-right panel displays the distribution
FR (R) of offered rates (thick lines) and the distribution HR (R) of accepted rates (thin lines).
the interest rate function R(k) flattens when product differentiation is more important for
borrowers. The reason is that a larger σaj means that the interest rates affect consumers’
choice across lenders relatively less, and thus all lenders charge similar rates.
The comparison between the dashed and the dotted lines in the top-right panel shows that a
higher σaj has a small effect on borrowers’ search effort. This small change in effort is the result
of opposite effects. Specifically, holding the distribution of offered rates fixed, the increase in
the product-differentiation parameter induces borrowers to search more aggressively, because
they are more likely to receive offers with product features a that they value more. However,
the dispersion of offered interest rates decreases, which decreases borrowers’ incentives to
search. As a result of these offsetting effects, borrowers’ effort to examine offers changes
minimally.
33

Interest Rate R(k) 35 2
Arrival Rate α(z)

22.5 1
10 0
7.1 9.6 12.2 0 25 50
0.4 1

0.2 0.5
0 0
14.1 21.6 29.2 10 22.5 35
Figure 5: These panels display model outcomes at the calibrated parameters (solid line), in the case
when β0j′ = 0.7β (dotted line), and in the case when β ′ = 0.7β and σ ′ = 30σ
0j 0j 0j aj aj (dashed line) for
near-prime borrowers. The top-left panel displays lenders’ optimal interest rate R(k) as a a function
of their cost k; the top-right panel displays borrowers’ effective arrival rate α(z) as a function of
their willingness to pay z; the bottom-left panel displays the probability P (R) that borrowers accept
an offer with interest rate R; and the bottom-right panel displays the distribution FR (R) of offered
rates (thick lines) and the distribution HR (R) of accepted rates (thin lines).
The bottom-left panel displays the probability P (R) that borrowers accept an offer with
interest rate R. Because lenders offer similar interest rates when σaj is higher, the acceptance
probability of an individual offer with a given R increases relative to the case with identical
costs of effort but a lower σaj . Holding the distribution of offered rates fixed, this increase,
cumulated over the range of R, would lead to a non-trivial increase in the fraction of individuals
who borrow on credit cards, as we recount in Section 5.5.
The bottom-right panel displays the distribution of offered rates (thin lines) and of
accepted rates (thick lines). Both distributions obtained in the model with a higher product
differentiation and lower search costs (dashed lines) intersect the corresponding distributions
obtained in the model at the calibrated values (solid lines), as well as those obtained with
34

lower search costs only (dotted lines). This crossing is intuitive, because offered rates—and
thus accepted rates as well—are less dispersed if the product attribute a matters more for
consumers’ choices. The average offer rate and the average accepted rate decrease from 22.12
and 21.20, respectively, in the baseline case at the calibrated parameters to 19.78 and 19.59,
respectively, in the case with lower β0j and higher σaj . However, the most striking effects
are on the standard deviation of offered and accepted rates, which decrease from 3.54 and
3.41, respectively, in the baseline case to 1.32 and 1.23, respectively, in the case with greater
product differentiation and lower costs of effort.
Figure 5 also helps us understand why the calibrated model calls for high costs of effort
and small values of σaj : if costs were low and σaj were larger, the dispersion of interest rates
would be significantly lower than those observed in the data.
6 Policy Experiments
In this section, we use our model to study two policy experiments, motivated by recent
regulatory interventions: 1) a cap on the interest rate—that is, a maximum rate Rmax ; 2)
higher compliance costs for lenders, captured by higher fixed costs χj . The goal of both
experiments is to study how borrowers’ examination effort and lenders’ offered rates respond,
thereby affecting market outcomes and welfare.
6.1 Cap on Interest Rates

As we recount in the Introduction, several countries recently introduced price controls in
markets for some consumer financial products, and are currently considering intervening in
other markets as well. The goal of this section is to study the effects of a interest rate cap on
the equilibrium of our model.
The theoretical literature points out that these caps may have unintended consequences, for
two main reasons. First, caps reduce profit margins and thus may reduce the supply of credit,
most notably to riskier borrowers who have higher default rates. Second, Fershtman and
Fishman (1994) and Armstrong, Vickers, and Zhou (2009) show that, in markets with search
frictions, price caps may have the unintended consequences of increasing the equilibrium prices
that consumers pay. Specifically, they identify two opposing effects: 1) The direct effect of
regulation is to reduce prices for uninformed consumers who, before the regulation, were paying
high prices; and 2) the indirect effect is to reduce price dispersion, which reduces consumers’
incentives to acquire information about prices, thereby increasing suppliers’ market power
and, thus, prices. Armstrong, Vickers, and Zhou (2009) further show that if consumers are
heterogenous in their costs of acquiring information, the introduction of a price cap has an
35

ambiguous effect on the equilibrium price paid by consumers, thereby leading to the possibility
that equilibrium prices may increase. Hence, the relative magnitude of these contrasting effects
is an empirical/quantitative question. Our calibrated model allows us to determine which of
these opposing effect dominates, and thus whether price caps are beneficial to consumers.
To understand these issues, we set a common cap at Rmax = 25 percent. This cap does
not bind in the super-prime market, but it does in all other markets: Table 2 shows that
it corresponds approximately to the 65th, 75th, and 95th percentiles of the distributions of
accepted interest rates in the subprime, near-prime, and prime market, respectively.
We study this counterfactual case in general equilibrium; that is, we require that lenders’
free-entry condition (13) holds. Thus, some lenders may exit the market, in which case we
decrease the aggregate arrival rate of offers to a new value L′j proportionally. Formally, the
new arrival rate equals L′j = Λj G(k̂ ′ ), where k̂ ′ is the marginal cost of the marginal lender—
i.e., the lender that satisfies the free-entry condition (13)—in the counterfactual case (the
marginal cost of the marginal lender in the baseline case equals k̂).
Figure 6 compares outcomes of the model at the calibrated parameters (solid line) with
those of the model with Rmax = 25 percent, while holding all other parameters at their
calibrated values, for the near-prime market.19 The top-left panel shows interesting outcomes.
First, the highest-cost lenders exit the market, even though the cap is above their marginal
cost. Specifically, frictions are such that even if these lenders were to decrease their interest
rates substantially, their market share would not increase enough to allow them to cover their
fixed costs; hence, they exit. Second, all surviving lenders charge lower interest rates, because
the function R(k) lies strictly below that of the baseline case. In particular, the lender with
marginal cost k̂ ′ finds it worthwhile to drop its rate to satisfy the constraint, rather than exit.
Similarly, all other lenders with lower marginal costs charge slightly below their higher-cost
competitors.
The top-right panel shows how borrowers’ effective arrival rate of offers adjusts, displaying
the indirect and direct effects that Fershtman and Fishman (1994) and Armstrong, Vickers,
and Zhou (2009) emphasize. Specifically, because some lenders exit the market, on average,
borrowers receive seven-percent-fewer offers than in the baseline case. Nevertheless, low-
valuation borrowers slightly increase their effort to more than offset the lower arrival rate
of offers, and thus the average effective number of offers α(z) that they examine is higher
than in the baseline. The reason is that the cap reduces the level of interest rates relative
to the baseline case, thereby increasing the expected payoff from a credit card loan for these
19
In particular, we keep the standard deviations σaj of the product attribute a constant. We should point
out that: 1) reducing these standard deviations does not affect our main counterfactual results, as these
standard deviations are very small relative to the standard deviation of Rj ; 2) In the counterfactuals of
Appendix D, which considers the case in which R and a are correlated, the standard deviations of the product
attribute endogenously adjust when R is capped.
36

Interest Rate R(k) 35 1.5
Arrival Rate α(z)

22.5 0.75
10 0
7.1 9.6 12.2 0 25 50
0.4 1

0.2 0.5
0 0
15.3 22.2 29.2 10 22.5 35
Figure 6: These panels display outcomes in near-prime market at the calibrated parameters (solid
line) and in the case when interest rates are capped at 25 percent (dotted line). The top-left panel
displays lenders’ optimal interest rate R(k) as a a function of their cost k; the top-right panel displays
borrowers’ optimal arrival rate α(z) as a function of their willingness to pay z; the bottom-left panel
displays the probability P (R) that borrowers accept an offer with interest rate R; and the bottom-
right panel displays the distribution FR (R) of offered rates (thick lines) and the distribution HR (R)
of accepted rates (thin lines).
lower-valuation borrowers. However, high-valuation borrowers respond differently than low-

valuation borrowers, in that the average number of offers α(z) that they consider is lower
than in the baseline: these borrowers already had positive gains from trade in the baseline
case, but the cap reduces the dispersion of interest rates across lenders and thus reduces the
benefits of examining multiple offers.
The bottom-left panel displays that the probability P (R) that borrowers accept an offer
with a given interest rate R decreases relative to that of the baseline case. The reason is that
the cap reduces lenders’ rates, and thus borrowers are less likely to accept an offer with a
given R and are more likely to accept offers with lower interest rates. The average acceptance
probability across lenders increases and demand is more elastic relative to the baseline case.
37

Table 6: Market Outcomes and Welfare with a Price Cap

Lender Profits 0.47 0.62 0.92 1.00
Welfare 0.97 1.03 1.01 1.00
Notes: This table reports market outcomes and welfare in each market, as ratios of those of the
baseline case.
Moreover, the fraction of individuals with credit card debt increases minimally to 55.97 percent
from 55.47 percent in the baseline case. The reason is that the higher examination effort of
borrowers with a relatively lower lower z leads them to consider and to accept more offers
than in the baseline case. This increase more than offsets the decrease due to borrowers with a
relatively high z, who examine fewer offers and thus are less likely to accept an offer relative to
the baseline. Thus, marginal borrowers (i.e., those with a low valuation z) display a stronger
response to the cap than infra-marginal borrowers (i.e., those with a high valuation z).
The bottom right panel of Figure 6 displays the distribution of offered rates (thick lines)
and of accepted rates (thin lines). Both distributions in the case of an interest rate ceiling
(dotted lines) are first-order stochastically dominated by the corresponding distributions of
the baseline case with no ceiling (solid lines). The average offered and accepted rates equal
19.86 and 19.16, respectively, and the standard deviations of offered and accepted rates equal
2.87 and 2.77, respectively. These values are lower than those of the baseline, suggesting that
the price cap increases the surplus of those who borrow.
Table 6 reports summary statistics of market outcomes, as well as consumer surplus,
lenders’ profits, and welfare for each group of borrowers when interest rates are capped,
as ratios of those of the baseline case. The cap induces a large redistribution of surplus
from lenders to borrowers, but small aggregate welfare effects. Specifically, consumer surplus
increases in all markets affected by the cap, with larger increases in markets in which lender
pricing is more constrained: the increase in consumer surplus equals 11 percent in the subprime
market, 18 percent in the near-prime market, and five percent in the prime market (it is
zero in the super-prime market because the cap is not binding); weighting markets by the
share of borrowers in each of them, the aggregate increase in consumer surplus equals 6.2
percent. Correspondingly, aggregate lender profits decline by 21 percent—i.e., they decline
by 53 percent in the subprime market, by 38 percent in the near-prime market, and by eight
percent in the prime market. As a result, aggregate welfare is almost unchanged—i.e., it
38

declines by less than one percent on aggregate, because it decreases in the sub-prime market
by three percent, whereas it increases by two percent in the near-prime and by one percent
in the prime markets. Appendix D shows that these welfare results are quite similar in the
case in which the attribute a is correlated with the interest rate R, since nevertheless the data
seem to reject that the unobserved attribute, whether correlated with R or not, has a large
variance.
The results reported in Table 6 are broadly consistent with the empirical findings of
Agarwal, Chomsisengphet, Mahoney, and Stroebel (2015), who report an increase in consumer
surplus and a decrease in lender profits after the 2009 Credit Card Act banned overdraft fees
on credit cards. More generally, the aggregate welfare reported Table 6 assigns equal weights
to consumer surplus and to lender profits. Of course, any larger weight assigned to consumer
surplus relative to that assigned to lender profits increases the assessments of the benefits of
the interest rate cap.
6.2 Higher Compliance Costs

A second set of regulations that have been introduced since the Financial Crisis has broadly
increased lenders’ compliance costs. Through the lenses of our model, higher compliance costs
can be interpreted as an increase in lenders’ fixed costs χj .
The increase in the fixed costs shares with our previous counterfactual regarding the
introduction of an interest rate cap the feature that highest-costs lenders will exit the market;
thus, this counterfactual with larger fixed costs allows us to understand how much the results
displayed in Figure 6 obtain because of the exit of these highest-cost lenders. Moreover,
Janssen and Moraga-González (2004) show that a decrease in the number of active sellers
could increase examination effort because fewer sellers may decrease price dispersion, possibly
leading to higher average prices.20 Thus, our model is well suited to understand these effects.
To facilitate the comparison with our price-cap experiment of Figure 6, we increase the
fixed cost χ so that the marginal lender has marginal cost equal to k̂ ′ —i.e., the marginal cost
of the lender that satisfies the free entry (13) condition in the case of the price cap R̄ = 25.
In practice, the new fixed cost χ′ is 13.8-percent larger than that of the baseline case in the
near-prime market (it is 20-percent and 2.2-percent larger in the sub-prime and prime market,
respectively, whereas it does not change in the super-prime market because the cap did not
bind in that market). We further decrease the aggregate arrival rate of offers to a new value
L′j correspondingly; that is, the new arrival rate equals L′j = Λj G(k̂ ′ ).
Figure 7 compares outcomes of the model at the calibrated parameters (solid line) with
20
Similarly, Armstrong and Chen (2009) show that a decrease in the number of sellers could increase welfare
in a search model with inattentive consumers.
39

Interest Rate R(k) 35 1.5
Arrival Rate α(z)

22.5 0.75
10 0
7.1 9.6 12.2 0 25 50
0.4 1

0.2 0.5
0 0
16.5 22.8 29.2 10 22.5 35
Figure 7: These panels display outcomes in near-prime market at the calibrated parameters (solid
line) and in the case when the fixed cost χ′ = 1.138χ (dotted line). The top-left panel displays lenders’
optimal interest rate R(k) as a a function of their cost k; the top-right panel displays borrowers’
effective arrival rate α(z) as a function of their willingness to pay z; the bottom-left panel displays
the probability P (R) that borrowers accept an offer with interest rate R; and the bottom-right panel
displays the distribution FR (R) of offered rates (thick lines) and the distribution HR (R) of accepted
rates (thin lines).
those of the model with a higher fixed cost χ′ for the near-prime market, displaying interesting
patterns. Notably, the exit of high-cost lenders reduces interest rate dispersion (top-left panel),
but it does not reduce the level of interest rates, as surviving lenders increased their rates due
to lower competition. Hence, borrowers consider fewer offers than in the baseline case (top-
right panel) for two reasons: 1) they receive fewer offers—i.e., L′j = 3.50 when compliance
costs are higher versus Lj = 3.79 in the baseline case;21 and 2) they choose not to exert much
effort because price dispersion is lower, and thus the benefits of considering multiple offers are
lower. The bottom-left panel shows that the probability P (R) that borrowers accept an offer
21 ′
Lj in this counterfactual case is identical by construction to that of the counterfactual case of caps on
interest rates.
40

Table 7: Market Outcomes and Welfare with Higher Compliance Costs

Lender Profits 0.39 0.62 0.95 1.00
Welfare 0.75 0.79 0.97 1.00
Notes: This table reports market outcomes and welfare in each market, as ratios of those of the
baseline case.
with a given interest rate R increases relative to the baseline case, because high and low offers
are no longer available. However, the average acceptance probability across lenders decreases
relative to the baseline case—i.e., 0.145 versus 0.147. Similarly, the fraction of borrowers
declines to 0.50 from 0.55 in the baseline.
The bottom-right panel of Figure 7 shows that the distributions of offered rates (thick
lines) and of accepted rates (thin lines) in a market with a higher fixed cost χ′ (dotted line)
intersect the corresponding distributions obtained in the baseline case (solid lines), as lenders
no longer offer the lowest and the highest rates. The average offered and accepted rates
are higher than those of the baseline (22.47 and 21.76 versus 22.12 and 21.20, respectively),
whereas the standard deviations of offered and accepted rates are lower (3.20 and 3.11 versus
3.54 and 3.41, respectively).
Table 7 reports summary statistics of market outcomes, as well as consumer surplus,
lenders’ profits, and aggregate welfare for each group of borrowers when fixed costs are higher,
as ratios of those of the baseline case. Higher fixed costs reduce lender profits, as the price cap
did, but they also decrease consumer surplus, with large negative welfare effects. Specifically,
consumer surplus decreases in all markets in which the cap is binding: the decrease in consumer
surplus equals 14 percent in the subprime market, 15 percent in the near-prime market, and
two percent in the prime market (it is zero in the super-prime market because the fixed cost
is the same as in the baseline case). The aggregate decrease in consumer surplus equals six
percent once we weight markets by their share of borrowers. Similarly, aggregate lender profits
decline by 23 percent—they decline by 61 percent in the subprime market, by 38 percent in
the near-prime market, and by five percent in the prime market. As a result, aggregate welfare
declines by ten percent on aggregate—it declines by 25 percent in the sub-prime market, by
21 percent in the near-prime market and by three percent in the prime market. Appendix
D shows that these results carry through in the case in which the unobserved attribute a is
correlated with the interest rate R.
41

7 Conclusions
This paper develops a framework that captures the observed large number of credit card
offers that individuals receive and the high level and large dispersion of the interest rates that
individuals pay on their credit cards. We focus on two main reasons: endogenous (low) effort
of examining offers and product differentiation. We calibrate the model using data on the
U.S. credit card market, fitting them well. Our analysis implies that low effort of examining
offers mostly accounts for the observed patterns on the data, whereas product differentiation
plays a smaller role. We further use the calibrated model to perform policy experiments.
Most notably, we find that interest rate caps generate quite large gains in consumer surplus,
because they decrease lenders’ market power.
We should point out that these results obtain in a model with some limitations, and thus
future research could enhance it in several ways. As we recount in Section 3, our cross-sectional
data impose some limitations on what our model can identify in the data, and richer data on
borrowers and lenders would allow us to further enrich our current framework. Specifically,
extreme multidimensional heterogeneity is difficult to identify with our data. Many structural
search models share this limitation due to similar data constraints, and one contribution of this
paper is to adapt and to enrich these models to incorporate two key features—i.e., consumer
limited examination effort and product differentiation—that may rationalize the large number
of credit card offers and the large dispersion of interest rates that we observe in the data.
For these main reasons, we view this paper as a first step in quantifying the role of effort
for examining and evaluating offers in search markets. The quantitative analysis clarifies the
data requirements to calibrate/estimate such a model and how the parameters are identified,
and the calibration delivers a sense of the magnitudes involved, allowing us to assess which
forces dominate. Nonetheless, we hope that the future availability of richer data will allow us
to incorporate additional features of retail financial markets.
42

References
Agarwal, S., S. Chomsisengphet, C. Liu, and N. S. Souleles (2015): “Do Consumers
Choose the Right Credit Contracts?,” The Review of Corporate Finance Studies, 4(2), 239–
257.
Agarwal, S., S. Chomsisengphet, N. Mahoney, and J. Stroebel (2015): “Regulating

Consumer Financial Products: Evidence from Credit Cards,” The Quarterly Journal of
Economics, 130(1), 111–164.
(2018): “Do Banks Pass Through Credit Expansions to Consumers Who Want to
Borrow?,” The Quarterly Journal of Economics, 133(1), 129–190.
Agarwal, S., J. C. Driscoll, X. Gabaix, and D. Laibson (2008): “Learning in the

Credit Card Market,” Working Paper 13822, National Bureau of Economic Research.
Allen, J., R. Clark, and J.-F. Houde (2019): “Search Frictions and Market Power in
Negotiated Price Markets,” Journal of Political Economy, 127(4), 1550–1598.
Armstrong, M., and Y. Chen (2009): “Inattentive Consumers and Product Quality,”
Journal of the European Economic Association, 7(2-3), 411–422.
Armstrong, M., J. Vickers, and J. Zhou (2009): “Consumer Protection and the
Incentive to Become Informed,” Journal of the European Economic Association, 7(2-3),
399–410.
Ausubel, L. M. (1991): “The Failure of Competition in the Credit Card Market,” The
American Economic Review, 81(1), 50–81.
Burdett, K., and K. L. Judd (1983): “Equilibrium Price Dispersion,” Econometrica,

51(4), 955–969.
Butters, G. R. (1977): “Equilibrium Distributions of Sales and Advertising Prices,” Review

of Economic Studies, 44(3), 465–491.
Calem, P. S., and L. J. Mester (1995): “Consumer Behavior and the Stickiness of Credit-
Card Interest Rates,” The American Economic Review, 85(5), 1327–1336.
Campbell, G., A. Haughwout, D. Lee, J. Scally, and W. van der Klauuw (2016):
“Just Released: Recent Developments in Consumer Credit Card Borrowing,” Liberty Street
Economics (blog), Federal Reserve Bank of New York.
43

Campbell, J. Y., H. E. Jackson, B. Madrian, and P. Tufano (2011a): “The
Regulation of Consumer Financial Products: An Introductory Essay with Four Case
Studies,” in Moving forward: the future of consumer credit and mortgage finance, ed. by
N. P. Retsinas, and E. S. Belsky, chap. 7, pp. 206–244. Brookings Institution Press.
Campbell, J. Y., H. E. Jackson, B. C. Madrian, and P. Tufano (2011b): “Consumer

Financial Protection,” Journal of Economic Perspectives, 25(1), 91–114.
Consumer Financial Protection Bureau (2012): “Analysis of Differences between

Consumer- and Creditor-Purchased Credit Scores,” Discussion paper.
(2015): “The Consumer Credit Card Market,” Discussion paper.
Crawford, G. S., N. Pavanini, and F. Schivardi (2018): “Asymmetric Information and

Imperfect Competition in Lending Markets,” American Economic Review, 108(7), 1659–
1701.
Cuesta, J. I., and A. Sepúlveda (2019): “Price Regulation in Credit Markets: A Trade-off
between Consumer Protection and Credit Access,” Mimeo, Stanford University.
Ellison, G., and S. F. Ellison (2009): “Search, Obfuscation, and Price Elasticities on
the Internet,” Econometrica, 77(2), 427–452.
Fershtman, C., and A. Fishman (1994): “The ’Perverse’ Effects of Wage and Price
Controls in Search Markets,” European Economic Review, 38(5), 1099–1112.
Galenianos, M., and A. Gavazza (2017): “A Structural Model of the Retail Market for
Illicit Drugs,” American Economic Review, 107(3), 858–896.
Galenianos, M., and J. Nosal (2016): “Segmented Markets and Unsecured Credit,”
Mimeo, Royal Holloway.
Gavazza, A. (2016): “An Empirical Equilibrium Model of a Decentralized Asset Market,”

Econometrica, 84(5), 1755–1798.
Grodzicki, D. (2015): “Competition and Customer Acquisition in the US Credit Card

Market,” Mimeo, Pennsylvania State University.
Han, S., B. J. Keys, and G. Li (2018): “Unsecured Credit Supply, Credit Cycles, and
Regulation,” The Review of Financial Studies, 31(3), 1184–1217.
Hong, H., and M. Shum (2006): “Using Price Distributions to Estimate Search Costs,”
The RAND Journal of Economics, 37(2), 257–275.
44

Hortaçsu, A., and C. Syverson (2004): “Search Costs, Product Differentiation, and
Welfare Effects of Entry: A Case Study of S&P 500 Index Funds,” The Quarterly Journal
of Economics, 119(4), 403–456.
Janssen, M. C., and J. L. Moraga-González (2004): “Strategic Pricing, Consumer

Search and the Number of Firms,” The Review of Economic Studies, 71(4), 1089–1118.
Knittel, C. R., and V. Stango (2003): “Price Ceilings as Focal Points for Tacit Collusion:
Evidence from Credit Cards,” American Economic Review, 93(5), 1703–1729.
Nelson, S. (2020): “Private Information and Price Regulation in the US Credit Card
Market,” Mimeo, University of Chicago.
Ru, H., and A. Schoar (2016): “Do Credit Card Companies Screen for Behavioral Biases?,”
Working Paper 22360, National Bureau of Economic Research.
Salz, T. (2017): “Intermediation and Competition in Search Markets: An Empirical Case

Study,” Mimeo, MIT.
Sirri, E. R., and P. Tufano (1998): “Costly Search and Mutual Fund Flows,” The Journal
of Finance, 53(5), 1589–1622.
Stango, V. (2002): “Pricing with Consumer Switching Costs: Evidence from the Credit
Card Market,” The Journal of Industrial Economics, 50(4), 475–492.
Stango, V., and J. Zinman (2016): “Borrowing High versus Borrowing Higher: Price
Dispersion and Shopping Behavior in the US Credit Card Market,” Review of Financial
Studies, 29(4), 979–1006.
Wildenbeest, M. R. (2011): “An Empirical Model of Search with Vertically Differentiated

Products,” The RAND Journal of Economics, 42(4), 729–757.
Woodward, S. E., and R. E. Hall (2012): “Diagnosing Consumer Confusion and

Sub-Optimal Shopping Effort: Theory and Mortgage-Market Evidence,” The American
Economic Review, 102(7), 3249–3276.
45

APPENDIX
A Additional Empirical Results

In this Appendix, we report additional empirical results that complement those of Section
3. Specifically, we perform some robustness checks that construct the empirical distribution
of interest rates by weighting observations using the revolving balance of each credit card
account. This weighting strengthens the evidence that a pervasive feature of the credit card
market is that borrowers pay very different interest rates on their credit card debt.
Table A1 presents these results. Its construction is similar that of Table 1, with some
differences that we now explain. Column (1) replicates column (1) of Table 1, thus reporting
selected percentiles of the raw data over the entire sample period January 2006-December
2008. Column (2) weighs each observation used in column (1) by its revolving balance in
the corresponding month; hence, the distribution does not include interest rates of accounts
whose balances are paid in full. For all borrower groups, the percentiles of the distributions
reported in column (1) and in column (2) are very similar.
Column (3) replicates column (3) of Table 1, thus restricting the data to January 2007
and excluding introductory “teaser” rates (i.e., low initial rates that reset to higher rates
after an initial offer period). Column (4) weighs each observation used in column (3) by the
average revolving balance calculated over all available months in the panel period January
2006-December 2008. Again, for all borrower groups, the distributions reported in column (3)
and in column (4) display very similar levels and overall dispersions of interest rates.
Column (5) replicates the regressions and the corresponding percentiles of the interest rates
distributions of column (6) of Table 1, thus restricting the sample to cards with a revolving
balance in January 2007. The regressions reported in column (6) of Table A1 weigh each
observation used in the regressions of column (5) by the average revolving balance calculated
over all available months in the panel period January 2006-December 2008. The coefficient
estimates are similar between column (5) and column (6), most notably that of the FICO
score. Moreover, based on these weighted regressions, we construct the residual interest rates
using equation (2). The percentiles reported in the bottom part of column (6) further weigh
these residual interest rates by their average revolving balance over the sample period—i.e.,
both the coefficient estimates and the distribution of residuals weigh each observation by
the amount of the average revolving balance. For all borrower groups, the percentiles of the
distributions reported in column (5) and in column (6) are strikingly similar.
46

Table A1: Interest Rates Weighted by Revolving Balance, by Borrower Group
Subprime Borrowers (1) (2) (3) (4) (5) (6)

FICO Score -0.018 -0.022
(0.006) (0.008)
(0.674) (0.535)
(0.125) (0.116)
(0.140) (0.108)
R2 0.013 0.037
Observations 27,024 27,024 877 877 766 766
10th Percentile 11.90 9.90 14.24 12.99 14.39 13.22
25th Percentile 16.15 15.90 17.24 16.24 17.58 16.43
50th Percentile 20.65 21.40 21.74 22.90 21.93 22.05
75th Percentile 27.49 28.24 27.99 29.23 27.80 27.75
90th Percentile 29.99 29.99 30.24 30.24 30.16 30.27
Near-prime Borrowers
FICO Score -0.052 -0.076
(0.013) (0.014)
(0.565) (0.504)
(0.078) (0.061)
(0.100) (0.072)
R2 0.043 0.090
Observations 27,059 27,059 900 900 661 661
10th Percentile 10.49 9.90 12.99 12.25 13.20 13.73
25th Percentile 14.90 14.24 15.94 15.81 16.55 16.99
50th Percentile 18.24 18.24 19.24 19.24 20.20 20.96
75th Percentile 23.15 24.24 23.30 25.40 25.72 25.67
90th Percentile 28.99 29.74 29.24 29.99 29.16 29.81
Prime Borrowers
FICO Score -0.052 -0.054
(0.015) (0.015)
Reward Card -0.240 -0.614
(0.520) (0.503)
(0.049) (0.045)
(0.059) (0.047)
R2 0.029 0.033
Observations 31,115 31,115 953 953 604 604
10th Percentile 9.90 9.90 11.99 11.24 11.55 11.63
25th Percentile 12.99 12.99 14.31 14.24 14.81 14.73
50th Percentile 16.74 16.99 18.24 18.24 17.90 18.00
75th Percentile 19.99 20.34 20.34 21.24 21.90 21.84
90th Percentile 25.99 28.99 28.15 28.99 28.65 28.88
Super-prime Borrowers
FICO Score -0.024 -0.039
(0.010) (0.011)
(0.471) (0.498)
Credit Limit 0.028 0.030
(0.031) (0.029)
Credit Balance -0.040 0.008
(0.051) (0.040)
R2 0.012 0.028
Observations 56,880 56,880 1,645 1,645 546 546
10th Percentile 9.90 7.99 11.24 10.09 10.79 10.53
25th Percentile 12.99 11.74 14.15 13.49 13.82 13.07
50th Percentile 15.98 15.24 16.99 17.15 16.84 16.63
75th Percentile 18.24 18.74 18.24 19.99 19.54 19.76
90th Percentile 20.24 24.24 20.31 24.24 23.98 24.67
Notes: This table reports coefficient estimates of equation (1) and the corresponding percentiles of the distribution of centered
interest rates as in equation (2). Columns (2), (4), and (6) report percentiles of the distribution weighted by revolving balance.
47

B Auxiliary Results: Distribution of Offers
We pointed out in Section 5.2 that survey respondents may report all offers that they receive,
not only those that they would consider if they were not surveyed. We now derive the
distribution of the number of offers and the distribution of the difference between the offers
with the smallest and the largest interest rates under the assumption that respondents report
all offers that they receive.
The expected number of offers for a borrower who receives n ≥ 2 offers is

L 1 − e−L
E[n|n ≥ 2] = .
1 − e−L − Le−L
Denote the probability distribution of the difference between the highest and the lowest
interest rate of a borrower who receives n ≥ 2 offers by D(x). Denote the probability
distribution of the difference between the highest and the lowest interest rate of a borrower
who receives exactly n offers by Dn (x) and note that
∞
1 X e−L Ln
D(x) = Dn (x).
1 − e−L − Le−L n=2 n!
Consider a borrower who receives n offers. Denote the lowest offer by RL and note that its
distribution follows F̄n (RL ) = 1 − (1 − F (RL ))n . Each of the other n − 1 offers are distributed
iid according to F̂ (R|RL ) = F (R)−F (RL )
1−F (RL )
, for R ≥ RL . The highest among these n − 1 offers is
distributed according to F̂ (RH |RL )n−1 . As a result
R n−1
F (RL + x) − F (RL )
Z
Dn (x) = dF̄n (RL )
R 1 − F (RL )
Z R
= n (F (RL + x) − F (RL ))n−1 F ′ (RL )dRL .
R
Combining the above, we obtain

∞
1 e−L Ln R
X Z
D(x) = n (F (RL + x) − F (RL ))n−1 F ′ (RL )dRL
1 − e − Le n=2 n!
−L −L
R
Z RX ∞
1 e−L Ln
= (F (RL + x) − F (RL ))n−1 F ′ (RL )dRL
1 − e − Le
−L −L
R n=2 (n − 1)!
R
Le−L
Z
L(FR (RL +x)−FR (RL ))
′
= e − 1 FR (RL )dRL .
1 − e−L − Le−L R
48

C Proofs
Proof of Proposition 2. We first show that the cost distribution for a low n first-order
stochastically dominates that for a high n (thereby proving that vz,n is increasing in n) and
that the derivative of the cost distribution for a high n first-order stochastically dominates
that for a low n (thereby proving strictly decreasing differences)
dF̄c,n (c) n
= − 1 − Fc (c) log 1 − Fc (c) > 0,
dn
2
d F̄c,n (c) n 2
= − 1 − Fc (c) log 1 − Fc (c) < 0.
dn2
Therefore, vz,n+1 > vz,n and vz,n+2 − vz,n+1 < vz,n+1 − vz,n for all n.
Differentiating equation (5) with respect to e (and noting that vz,0 = 0)
∞
X e−eL (eL)n e−eL (eL)n−1
Vz′ (e) = − vz,n + vz,n L
n=1
n! (n − 1)!
∞ ∞
X e−eL (eL)n X e−eL (eL)n
= − vz,n + vz,n+1 L
n=0
n! n=0
n!
∞
X e−eL (eL)n
= vz,n+1 − vz,n L > 0.
n=0
n!
As a result, the borrower’s expected value of offers is strictly increasing in their examination
effort, and equation (8) characterizes the optimal choice of effort.
Furthermore, the expected value of loan offers is strictly concave in examination effort:
∞
e−eL (eL)n e−eL (eL)n−1
X
Vz′′ (e) = − + (vz,n+1 − vz,n ) L2
n=1
n! (n − 1)!
∞
X e−eL (eL)n
= (vz,n+2 − vz,n+1 − (vz,n+1 − vz,n )) L2 < 0
n=0
n!
Therefore, equation (6) has a unique solution e(z), which yields the optimal examination effort
for a type-z borrower.
Finally, notice that
Z z
∂vz,n
= b (1 − ρ) dF̄c,n (c) = b (1 − ρ) (1 − (1 − Fc (z))n ) > 0,
∂z −∞
∞
∂Vz (e) X e (eL)n
−eL
⇒ = b (1 − ρ) (1 − (1 − Fc (z))n ) > 0.
∂z n=1
n!
49

Thus, higher-marginal-utility borrowers exert more examination effort, because they gain more
from an increase in the effective arrival rate of offers.
We now calculate the distribution of accepted offers. Denote the probability that a type-z
borrower gets a loan by Qz and the probability that he gets a loan with interest rate less than
R by Qz (R). Note that a type-z borrower gets a loan if he receives at least one offer with cost
below z. Therefore
Qz = 1 − e−e(z)LFc (z)
RR
Fa (z−x)dFR (x)
= 1 − e−e(z)L R ,
RR
−e(z)L Fa (z−x)dFR (x)
Qz (R) = 1 − e R .
Denote the probability that a borrower gets a loan by Q and the probability that he gets
a loan with interest rate less than R by Q(R):
X
Q = Qz
z∈Z
X RR
Fa (z−x)dFR (x)
= 1− e−e(z)L R ,
z∈Z
X RR
Fa (z−x)dFR (x)
Q(R) = 1 − e−e(z)L R .
z∈Z
The distribution of accepted interest rates HR (R) gives the proportion of borrowers who
get a loan with interest rate less than R among the borrowers who get a loan:
Q(R)
HR (R) =
Q
RR
1 − z∈Z e−e(z)L R Fa (z−x)dFR (x)
P
= R .
P −e(z)L RR Fa (z−x)dFR (x)
1 − z∈Z e
The density of the accepted-rate distribution is
1 X −e(z)L RRR Fa (z−x)dFR (x)

HR′ (R) = e e(z)LFa (z − R)FR′ (R).
Q z∈Z
This completes the proof of proposition 2.
Proof of Lemma 3. Denote the probability that a type-z borrower accepts a loan offer with
total cost c by Pc (c, z). If c ≤ z, the borrower accepts the offer if it is the lowest-cost offer
n
received, which occurs with probability 1 − Fc (c) when the borrower examines n additional
50

offers. If c > z, the borrower does not accept that offer. Therefore
∞
X e−α(z) α(z)n n
Pc (c, z) = 1 − Fc (c)
n=0
n!
−α(z)Fc (c)
= e
RR
Fa (c−x)dFR (x)
= e−α(z) R , if c ≤ z, (C1)
Pc (c, z) = 0, if c > z. (C2)
Denote by PR (R, z) the probability that a type-z borrower accepts a loan offer with interest
rate R. A borrower with valuation z accepts this offer if its cost (including the idiosyncratic
attribute) is less than z and if all other offers that he examines have higher costs. Integrating
over the potential values of the idiosyncratic utility draw yields
Z ∞
PR (R, z) = Pc (R + a, z)dFa (a)
−∞
RR
Z z−R
= e−α(z) R Fa (R+a−x)dFR (x) dFa (a). (C3)
−∞
A borrower of type z accepts a loan offer with interest rate R if he examines the offer
(probability e(z)) and the offer is better than any other offer that he examines (probability
PR (R, z)). Therefore, the probability that a randomly drawn borrower accepts a loan with
interest rate R equals
X
P (R) = sz e(z)PR (R, z)
z∈Z
z−R RR
X Z
Fa (R+a−x)dFR (x)
= sz e(z) e−α(z) R dFa (a),
z∈Z −∞
which yields equation (11).

Because Fa (·) is smooth, P (R) is continuous and differentiable in R. Differentiating P (R)
with respect to R yields
!
z−R RR R
X Z Z
Fa (R+a−x)dFR (x)
P ′ (R) = − sz e(z) e−α(z) R α(z) Fa′ (R + a − x)dFR (x) dFa (a)
z∈Z −∞ R
RR

Fa (z−x)dFR (x)
+ e−α(z) R Fa′ (z − R) < 0.
Hence, the probability that borrowers accept a loan is strictly decreasing in the interest rate
R. This completes the proof of lemma 3.
51

Proof of Proposition 4. The optimal interest rate for a type-k lender solves
πk′ (R) = b(1 − ρ)P (R) + b (R(1 − ρ) − k) P ′ (R) = 0.

k k k k
Note that πk (R) < 0 for R < 1−ρ
, πk 1−ρ
= 0, πk′ 1−ρ
= b (1 − ρ) P 1−ρ
> 0 and
k
limR→∞ πk (R) = 0. Therefore, some R̃ > exists such that
1−ρ
= 0, and thus the πk′ (R̃)
optimal choice R(k) exists. In the case of multiple roots, the lender chooses the solution
that yields higher profits. Finally, because πk (R) is continuously differentiable in k, R(k) is
continuous and differentiable in k.
The cross-partial derivative of profits with respect to lender type and interest rate is
positive:
∂πk (R)
= −bP (R),
∂k
∂ 2 πk (R)
= −bP ′ (R) > 0,
∂k∂R
which implies R′ (k) > 0.
Because the optimal interest rate is strictly increasing in the lender’s cost k, we have

FR R(k) = G(k) for k ∈ [k, k]. Hence,

FR (x) = G R−1 (x) .
Using this feature, we can rewrite equation (11) as follows:
z−R(k) Rk
X Z
−α(z) Fa R(k)+a−R(x) dG(x)
P R(k) = sz e(z) e k dFa (a). (C4)
z∈Z −∞
Equation (C4) defines the probability that borrowers accept the loan of the cost-k lender when
all lenders make their equilibrium choice. This probability does not directly depend on the
interest rate distribution, because it incorporates the result that the offered interest rate is
strictly decreasing in a lender cost k.
The profits of a type-k lender who follows the strategy of a type-k̃ lender are
z−R(k̃) Rk
X Z
−α(z) Fa R(k̃)+a−R(x) dG(x)
πk R(k̃) = b R(k̃)(1 − ρ) − k sz e(z) e k dFa (a).
z∈Z −∞
52

Differentiating profits with respect to k̃, we obtain
Rk
Z z−R(k̃)
∂πk (R(k̃))
e−α(z) k Fa (R(k̃)+a−R(x))dG(x) dFa (a)
X
′
= bR (k̃)(1 − ρ) sz e(z)
∂ k̃ z∈Z −∞
Rk
X Z z−R(k̃)
−b R(k̃)(1 − ρ) − k sz e(z) e−α(z) k Fa (R(k̃)+a−R(x))dG(x)
z∈Z −∞
!
Z k
α(z) Fa′ R(k̃) + a − R(x) R′ (k̃)dG(x) dFa (a)
k
!
Rk
Fa (z−R(x))dG(x)
+R′ (k̃)e−α(z) k Fa′ z − R(k̃) .
This derivative equals zero when k̃ = k. Therefore

z−R(k) Rk
X Z
−α(z) Fa R(k)+a−R(x) dG(x)
(1 − ρ) sz e(z) e k dFa (a)
z∈Z −∞
z−R(k) Rk k
Z Z
X −α(z) Fa R(k)+a−R(x) dG(x)
= R(k)(1 − ρ) − k sz e(z) e k α(z) Fa′ R(k) + a
z∈Z −∞ k
Rk

−R(x) dG(x) dFa (a) + e−α(z) k Fa z−R(x) dG(x)

Fa′ z − R(k) ,
which yields equation (12) that defines the interest rate schedule R(k). This completes the
proof of proposition 4.
Proof of Proposition 5. A lender’s expected profits are strictly decreasing in his cost k,
because a lender can always mimic the action of a higher-cost lender and make strictly higher
profits.
Denote the highest-cost lender that enters the market by k̂, where k̂ ≤ k, and note that the
measure of lenders that enter the market is L = ΛΓ(k̂). Denote the profits of the highest-cost
lender by π k̂ :

πk̂ (R(k̂)) = b R(k̂)(1 − ρ) − k̂ P (R(k̂)),
where
Z z−R(k̂) R k̂ Γ(x)
−e(z)ΛΓ(k̂) Fa R(k̂)+a−R(x) d
X
P (R(k̂)) = sz e(z) e k Γ(k̂) dFa (a).
z∈Z −∞
This equation makes explicit the dependence of L and G(·) on k̂.
53

The profits of the highest-cost lender are decreasing in his type:
dπ k̂ ∂π k̂ ′ ∂P (R(k̂))
= R (k̂) − bP R(k̂) + b R(k̂) 1 − ρ) − k̂ ΛΓ′ (k̂),
dk̂ ∂R ∂L
which is negative because the first term equals zero by the envelope theorem, the second
term reflects the cost increase, and the third term reflects that an increase in k̂ increases the
measure of lenders in the market, which reduces the probability that borrowers accept a loan
offer. Therefore, given e(·), a unique k̂ exists that characterizes lenders’ cutoff cost k̂.
The cutoff k̂ is determined by equating the profits of the highest-cost lender with the entry
cost χ, as equation (13) shows.
Proof of Proposition 6. The equilibrium is defined as a fixed point in borrower effort.

Borrower effort is clearly bounded above for all types. Denote the bound by e > 0 and note
that e(z) ∈ [0, e]N which is a non-empty, compact and convex subset of RN . Starting from
some initial e(z) ∈ [0, e]N , Propositions 4 and 5 characterize the lenders’ entry and interest
rate decisions {L, G(·), FR (·)}; given lenders’ actions, Proposition 2 characterizes borrowers’
effort e(·) of examining offers. All of these mappings are continuous. Therefore a fixed point
(and an equilibrium) exists.
Proof of Proposition 7. Differentiating equation (14) with respect to e and equating to

zero for every z, we obtain

∂Wz e∗ (z), k̂ ∗ ∂q e∗ (z), L∗
= .
∂e ∂e
We use equation (16) to rearrange the above equation, obtaining equation (17). The solution
is unique for reasons similar to the decentralized case.
Differentiating equation (14) with respect to k̂ and equating to zero, we obtain equation
(18). Notice that
∞ h −e∗ (z)L∗ ∗ n
∂Wz (e∗ (z), k̂ ∗ ) X e e (z)L∗ ′
= Wz,n (k̂ ∗ )
∂ k̂ n=0
n!
e−e∗ (z)L∗ e∗ (z)L∗ n−1 e∗ (z)ΛΓ′ (k̂ ∗ )
+Wz,n (k̂ ∗ )
(n − 1)!
n
e−e (z)L e∗ (z)ΛΓ′ (k̂ ∗ ) e∗ (z)L∗ i
∗ ∗
−
n!
∞ −e∗ (z)L∗ ∗ n
X e e (z)L∗ ′ ∗ ∗ ′ ∗

∗ ∗

= Wz,n (k̂ ) + e (z)ΛΓ (k̂ ) Wz,n+1 (k̂ ) − Wz,n (k̂ ) .
n=0
n!
54

Furthermore
z ∂ F̄
w,n (w)
Z
′
Wz,n (k̂ ∗ ) = b(1 − ρ) (z − w)d .
−∞ ∂ k̂
Combining the last two equations yields equation (19).
D Correlation between R and a

In this Appendix, we extend the baseline model to consider the case in which the attribute
of a credit card is, in equilibrium, positively correlated with its interest rate R, which might,
in principle, account for some of the interest rate dispersion that we observe in the data. We
derive the equilibrium conditions for this extension of the model. We then calibrate it to
investigate how the extended model with correlation between attribute a and the interest rate
fits the data. Finally, we study the welfare effects of price caps and of higher compliance cost
in this extended model at its calibrated parameters.
D.1 Assumptions and Equilibrium Conditions

The matching process between borrowers and lenders is the same as in the baseline model.
The cost of a loan now consists of three components: the interest rate R, the idiosyncratic
component of the attribute a (distributed according to the zero-mean exogenous distribution
Fa (·)) and a deterministic component τ (k), which depends on the lender’s type and acts as
a mean-shifter the overall attribute realization. The total cost of a loan is c = R − τ (k) + a.
We assume that τ (k) is a smooth function with 0 < τ ′ (k) < 1.
The additional feature of this extension, τ (k), captures the possibility that a lender with
a higher funding cost might offer some additional desirable features that we do not observe in
our data, leading to higher acceptance rate than in the baseline model.
Let R(k) denote the optimal strategy of a type-k lender. The cost of a loan, then, depends
on the draw of two independent random variables: the lender type k from distribution G(·),
which determines R(k) and τ (k), and the draw of the attribute a from distribution Fa (·).
Hence, c is distributed according to
Z k
Fc (c) = Fa c − R(x) + τ (x) dG(x). (D1)
k
Using the amended definition for the cost distribution, the results regarding borrowers’
choice are essentially identical to the baseline model. They are summarized in the following
55

proposition (we omit all proofs, because all derivations are identical to those of the baseline
case).
Proposition 8 Given G(·), R(k) and L, optimal effort e(z) is characterized by the following
equations:
∞
X e−eL (eL)n ∂q(e, L)
(vz,n+1 − vz,n ) L = ,
n=0
n! ∂e
Z z
vz,n = b (z − c) dF̄c,n (c),
−∞
F̄c,n (c) = 1 − (1 − Fc (c))n .
Turning to the side of the lenders, the probability that a loan of cost c is accepted by a
type-z borrower is defined similarly to the baseline case:
Rk
Fa (c−R(x)+τ (x))dG(x)
Pc (c, z) = e−α(z) k if c ≤ z,
Pc (c, z) = 0 if c > z.
A loan from a type-k lender with interest rate R is accepted by a type-z borrower with
probability:
z−R+τ (k) Rk
Z
Fa (R−τ (k)+a−R(x)+τ (x))dG(x)
Pk,R (R, z) = e−α(z) k dFa (a).
−∞
A loan from a type-k lender with interest rate R is accepted with probability:
z−R+τ (k) Rk
X Z
Fa (R−τ (k)+a−R(x)+τ (x))dG(x)
Pk (R) = sz e(z) e−α(z) k dFa (a),
z∈Z −∞
where Pk′ (R) < 0 for the same reasons as in the baseline model.
Notice that, in contrast to the baseline model, a lender’s probability of giving a loan
depends on his type k directly (i.e. over and above his interest rate choice R) because k
determines the value of the mean-shifter. More precisely, the probability of a loan increases
56

in k:
R
Z z−R+τ (k)
∂Pk (R) X −α(z) kk Fa (R−τ (k)+a−R(x)+τ (x))dG(x)

= sz e(z) e α(z) ∗
∂k z∈Z −∞
Z k
′ ′
Fa (R − τ (k) + a − R(x) + τ (x))τ (k)dG(x) dFa (a) +
k
!
Rk
Fa (z−R(x)+τ (x))dG(x)
τ ′ (k)e−α(z) k Fa′ (z − R + τ (k)) > 0,
which is intuitive since higher-k lenders have higher mean-shifters. Furthermore:
∂Pk (R)
= −τ ′ (k)Pk′ (R).
∂k
The expected profits of a lender of type k who offers interest rate R are:
πk (R) = bPk (R) (R(1 − ρ) − k) ,
The optimal choice R(k) is characterized, as before, by:

πk′ (R) = bPk′ (R) R(1 − ρ) − k + bPk (R)(1 − ρ) = 0.
Note that equilibrium profits are declining in k:
πk (R(k)) ∂Pk (R(k))

= b R(k)(1 − ρ) − k − bPk (R(k))
∂k ∂k

= − τ ′ (k)bPk′ (R) R(k)(1 − ρ) − k + bPk (R(k)) < 0.
Hence, lender entry follows a cutoff rule, as in the baseline model.

The cross-partial derivative of profits with respect to R and k is:
∂πk′ (R) ∂Pk′ (R) ∂Pk (R)

= R(1 − ρ) − k) + (1 − ρ) − Pk′ (R).
∂k ∂k ∂k
57

The second and third terms are positive. It is not possible to sign the first term:
Rk
Z z−R+τ (k)
∂Pk′ (R) X
′
= τ (k) sz e(z) e−α(z) k Fa (R−τ (k)+a−R(x)+τ (x))dG(x) − α(z) ∗
∂k z∈Z −∞
Z k
2
Fa′ (R − τ (k) + a − R(x) + τ (x))dG(x) +
k
Z k
α(z) Fa′′ (R − τ (k) + a − R(x) + τ (x))dG(x) dFa (a) −
k
!
Rk
e−α(z) k Fa′′ (z − R + τ (k)) .
We will, from now on, assume that the cross-partial is positive and numerically confirm that
this assumption holds for our parameter values.
Under our assumption, higher-cost lenders choose higher interest rates: R′ (k) > 0. We
characterize the optimal interest rate choice and entry by the lenders in the next proposition,
following the same steps as in the baseline model.
Proposition 9 Lenders’ choices are characterized as follows:
1. Given borrowers’ effort e(·) the optimal interest rate choice by lenders R(·) solves
z−R(k)+τ (k) R k̂
X Z
−α(z) Fa (R(k)−τ (k)+a−R(x)+τ (x))dG(x)
sz e(z) e k dFa (a)
z∈Z −∞
R k̂
X Z z−R(k)+τ (k)
k
= R(k) − sz e(z) e−α(z) k Fa (R(k)−τ (k)+a−R(x)+τ (x))dG(x) ∗
1 − ρ z∈Z −∞
Z k̂
′
α(z) Fa (R(k) − τ (k) + a − R(x) + τ (x))dG(x) dFa (a) + (D2)
k
!
R k̂
e−α(z) k Fa′ (z − R(k) + τ (k)) .
2. The marginal lender type who enters the market k̂ is defined by

Z z−R(k̂)+τ (k̂) R k̂ Γ(x)
Fa (R(k̂)−τ (k̂)+a−R(k̂)+τ (k̂))d
X
−e(z)ΛΓ(k̂)
b R(k̂)(1 − ρ) − k̂ sz e(z) e k Γ(k̂) dFa (a) = χ.
z∈Z −∞
58

Table D1: Calibrated Parameters, Correlation between R and a
Panel B: Large Variance Panel B: Medium Variance Panel B: Small Variance

µz1 3.700 σz 1 0.123 µz1 3.616 σz 1 0.121 µz1 3.607 σz 1 0.116
µz2 3.630 σz 2 0.088 µz2 3.524 σz 2 0.088 µz2 3.519 σz 2 0.089
µz3 3.344 σz 3 0.126 µz3 3.461 σz 3 0.113 µz3 3.458 σz 3 0.122
µz4 3.147 σz 4 0.344 µz4 3.205 σz 4 0.342 µz4 3.208 σz 4 0.353
ξ 3.836 k̂ 9.393 ξ 3.761 k̂ 9.970 ξ 3.805 k̂ 9.903
L1 1.533 L2 3.902 L1 1.525 L2 3.765 L1 1.518 L2 3.806
L3 3.182 L4 3.029 L3 3.211 L4 3.035 L3 3.253 L4 3.067
ρ1 0.062 ρ2 0.055 ρ1 0.042 ρ2 0.024 ρ1 0.035 ρ2 0.021
ρ3 0.055 ρ4 0.050 ρ3 0.018 ρ4 0.014 ρ3 0.017 ρ4 0.011
σa 1 0.110 σa 2 0.126 σa 1 0.112 σa 2 0.130 σa 1 0.112 σa 2 0.130
σa 3 0.136 σa 4 0.147 σa 3 0.131 σa 4 0.135 σa 3 0.132 σa 4 0.155
β01 9.121 β02 38.176 β01 8.844 β02 35.759 β01 8.844 β02 35.124
β03 29.915 β04 30.282 β03 27.627 β04 27.261 β03 27.548 β04 27.481
β1 1.647 γ 0.800 β1 1.638 γ 0.400 β1 1.619 γ 0.200
D.2 Calibration
We calibrate the model by making the same functional form-assumptions that we made in the
baseline case of no correlation. In addition, we specify the function τ (k) to equal γ(k − E(k)),
R k̂
where E(k) = kmin kdG(k) is the average cost.
We perform three calibrations for three separate values of γ: 1) γ = 0.8, which implies
that the variance of the term τ (k) is large; 2) γ = 0.4, which implies that the variance of τ (k)
is intermediate; and 3) γ = 0.2, which implies that the variance of τ (k) is small. Of course,
the baseline calibration of Section 5 corresponds to the case with γ = 0.
Table D1 reports the parameters of these three cases and Table D2 reports how each
case fits the data. These tables show the following: 1) The case with γ = 0.8 fits the data
considerably worse than all other cases, including the baseline case with γ = 0. 2) The best
fit of the data obtains with γ = 0.4, which corresponds to a moderate variance of τ (k). 3)
The values of the other parameters—most notably, those of the cost-of-effort parameters β0j —
obtained in the best-fit case with γ = 0.4 are very similar to those obtained in the baseline
case with γ = 0, thereby leading to similar implications to those of the baseline case.
Tables D3 and D4 report market outcomes and welfare for the counterfactual analyses in
which we cap interest rates at Rmax = 25 percent and in which we increase higher compliance
costs, respectively, using the parameters in Panel B of Table D1 with γ = 0.4. These
counterfactual analyses correspond to those of Section 6, with the only differences being that
59

Table D2: Model Fit, Correlation between R and a
Data Model Model Model

γ = 0.8 γ = 0.4 γ = 0.2
10th Percentile Accepted Rate, Sub-prime Borrowers 14.39 20.24 18.52 18.41
10th Percentile Accepted Rate, Near-prime Borrowers 13.20 19.57 17.08 17.03
10th Percentile Accepted Rate, Prime Borrowers 11.56 13.75 15.60 15.53
10th Percentile Accepted Rate, Super-prime Borrowers 10.79 13.39 13.67 13.66
Fraction Receiving 2+ Offers (%) 75.00 74.65 74.54 74.86
Median Number of Offers Received, Conditional on 2+ Offers 3.00 3.00 3.00 3.00
Average Number of Offers Received, Conditional on 2+ Offers 4.00 3.49 3.48 3.50
10th Percentile Distribution of Differences in Offered Rates 0.00 0.72 1.54 1.55
Fraction with Credit Card Debt, Sub-Prime Borrowers 54.56 56.41 55.38 55.13
Fraction with Credit Card Debt, Near-Prime Borrowers 55.33 56.68 55.16 55.77
Fraction with Credit Card Debt, Prime Borrowers 54.00 51.05 54.26 54.69
Fraction with Credit Card Debt, Super-prime Borrowers 36.02 34.22 35.62 35.86
Charge-Off Rate 4.01 5.52 2.38 2.02
Average Funding Cost 7.02 6.03 6.07 6.05
Criterion Function 436.44 127.82 132.56
60

Table D3: Market Outcomes and Welfare with a Price Cap, Correlation between R and a

Lender Profits 0.41 0.42 0.74 1.00
Welfare 0.85 1.02 1.01 1.00
Notes: This table reports market outcomes and welfare in each market when interest rates are capped
at Rmax = 25 percent and R is correlated with the attribute a.
Table D4: Market Outcomes and Welfare with Higher Compliance Costs, Correlation between
R and a

Lender Profits 0.25 0.45 0.67 1.00
Welfare 0.59 0.69 0.86 1.00
Notes: This table reports market outcomes and welfare in each market when compliance costs are
higher and R is correlated with the attribute a.
they use the parameters reported in Panel B of Table D1 rather than those reported in Panel
A of Table 3.
Tables D3 and D4 confirm the robustness of our results of Section 6 that the price cap
has positive effects on consumer surplus and negative effects on lenders’ profits, whereas
higher compliance costs have negative (and large) effects on consumers and on lenders. More
specifically, Table D3 shows that, on aggregate, a price cap increases consumer surplus by 1.7
percent, decreases lender profits by 27.4 percent, resulting in a 3.3-percent welfare decrease.
Table D4 shows that higher compliance costs decrease consumer surplus by 14.3 percent,
decrease aggregate profits by 31.8 percent, resulting in a 17.3-percent welfare decrease.
61

SSRN Id3405897 PDF

Uploaded by

Copyright:

Available Formats

SSRN Id3405897 PDF

Uploaded by

Document Information

Original Title

Copyright

Available Formats

Share this document

Share or Embed Document

Sharing Options

Did you find this document useful?

Is this content inappropriate?

Copyright:

Available Formats

SSRN Id3405897 PDF

Uploaded by

Copyright:

Available Formats

Regulatory Interventions in Consumer Financial

Markets: The Case of Credit Cards*

We build a framework to understand the effects of regulatory interventions in credit

Electronic copy available at: https://ssrn.com/abstract=3405897

Electronic copy available at: https://ssrn.com/abstract=3405897

Electronic copy available at: https://ssrn.com/abstract=3405897

Electronic copy available at: https://ssrn.com/abstract=3405897

Electronic copy available at: https://ssrn.com/abstract=3405897

3.1 Data Sources

Electronic copy available at: https://ssrn.com/abstract=3405897

3.2 Data Description

Rijt = γX Xit + γZ Zijt + ǫijt , (1)

Electronic copy available at: https://ssrn.com/abstract=3405897

Electronic copy available at: https://ssrn.com/abstract=3405897

Electronic copy available at: https://ssrn.com/abstract=3405897

Subprime Borrowers (1) (2) (3) (4) (5) (6)

Electronic copy available at: https://ssrn.com/abstract=3405897

3.3 Implications for Modeling

Electronic copy available at: https://ssrn.com/abstract=3405897

Panel A: Accepted Offers

Electronic copy available at: https://ssrn.com/abstract=3405897

Electronic copy available at: https://ssrn.com/abstract=3405897

Electronic copy available at: https://ssrn.com/abstract=3405897

Vz,j (e) − qj (e, Lj ). (3)

We denote the optimal (utility-maximizing) effort choice of a type-z borrower in market j by

Electronic copy available at: https://ssrn.com/abstract=3405897

4.1 Borrowers’ Choices

Electronic copy available at: https://ssrn.com/abstract=3405897

F̄c,n (c) = 1 − (1 − Fc (c))n .

Therefore, the value to a z-borrower of examining n ≥ 1 offers is

Proposition 2 Given FR (·) and L:

where vz,0 = 0 and equation (7) defines vz,n for n ≥ 1.

2. The distribution of accepted offers equals

3. The fraction of borrowers who get a loan is

4.2 Lenders’ Choices

Electronic copy available at: https://ssrn.com/abstract=3405897

Furthermore, P ′ (R) < 0.

Proposition 4 Given L, G(·), and e(·):

2. R(·) solves the following functional equation:

3. The interest rate distribution equals FR (x) = G (R−1 (x)) .

The following proposition completes the characterization of lenders’ entry decisions.

Electronic copy available at: https://ssrn.com/abstract=3405897

4.3 Existence of the Equilibrium

Proposition 6 An equilibrium exists.

4.4 Constrained Efficiency

Electronic copy available at: https://ssrn.com/abstract=3405897

The distribution of the lowest w among n offers is

The social value to a type-z borrower of receiving n offers is:

We now characterize the planner’s optimal solution.

2. The optimal entry-cost cutoff k̂ ∗ for lenders satisfies

Electronic copy available at: https://ssrn.com/abstract=3405897

e−α(e (z),L ) (α(e∗ (z), L∗ ))n

5.1 Parametric Assumptions

Electronic copy available at: https://ssrn.com/abstract=3405897

Electronic copy available at: https://ssrn.com/abstract=3405897

(m (ψ) − m)′ Ω (m (ψ) − m) ,

5.3 Data-Generating Process

5.4 Sources of Identification

Electronic copy available at: https://ssrn.com/abstract=3405897