Did FinTech Lenders Facilitate PPP Fraud? — The Journal of Finance, Vol. LXXVIII, No. 3 (June 2023)

Summary

Did FinTech Lenders Facilitate PPP Fraud?, an article by John M. Griffin, Samuel Kruger and Prateek Mahajan of the McCombs School of Business, The University of Texas at Austin, in The Journal of Finance, Vol. LXXVIII, No. 3, June 2023. The authors examine metrics related to potential misreporting in the $793 billion Paycheck Protection Program, including nonregistered businesses, multiple businesses at residential addresses, abnormally high implied compensation per employee and inconsistencies with jobs reported in another government program. They report that these measures concentrate in certain FinTech lenders and that suspicious lending by FinTechs in 2021 is four times the level at the start of the program. They state that FinTech lenders rose from less than 5% of loans in round 1 to over 80% by May 2021.

Summary drafted by a model from the document's text below and checked by script against that text before publication. It is a navigation aid, not a reading of what the document proves. Where AI is used

Full text

THE JOURNAL OF FINANCE • VOL. LXXVIII, NO. 3 • JUNE 2023




      Did FinTech Lenders Facilitate PPP Fraud?

         JOHN M. GRIFFIN, SAMUEL KRUGER, and PRATEEK MAHAJAN*

                                            ABSTRACT
       In the $793 billion Paycheck Protection Program, we examine metrics related to po-
       tential misreporting including nonregistered businesses, multiple businesses at res-
       idential addresses, abnormally high implied compensation per employee, and large
       inconsistencies with jobs reported in another government program. These measures
       consistently concentrate in certain FinTech lenders and are cross-verified by seven
       additional measures. FinTech market share increased significantly over time, and
       suspicious lending by FinTechs in 2021 is four times the level at the start of the
       program. Suspicious loans are being overwhelmingly forgiven at rates similar to
       other loans.




THE MELDING OF FINANCIAL TECHNOLOGY and banking, also known as Fin-
Tech lending, has emerged at a rapid pace in the aftermath of the financial
crisis. Buchak et al. (2018) find that an increase in regulatory burdens for

   * John M. Griffin, Samuel Kruger, and Prateek Mahajan are all at McCombs School of Business,

The University of Texas at Austin. We thank Stefan Nagel (the editor), an anonymous associate ed-
itor, two anonymous referees, Andres Almazan, Zahi Ben-David, Bo Bian, Bruce Carlin, Jonathan
Cohn, Matt Denes, Mark Egan, Isil Erel, Michael Faulkender, William Fuchs, Arpit Gupta, Umit
Gurun, Ben Hébert, Jessica Jeffers, Mark Johnson, Jack Liebersohn, Zack Liu, Christos Makridis,
Thorsten Martin, Alex Priest, Alberto Rossi, Laura Starks, Boris Vallée, Brian Wolfe, Constantine
Yannelis, Yao Zeng, Harold Zhang, Eric Zwick, BlueAcorn, Capital Plus, Itria Ventures, anony-
mous industry participants, and many others for helpful comments. We are also thankful for com-
ments from seminar and conference participants at the ASU Sonoran Winter Finance Conference,
the Federal Reserve Bank of Atlanta, the Finance Research Association Conference, Georgetown
University, Hong Kong Baptist University, the Lone Star Finance Conference, the Office of the
Comptroller of the Currency, Pontificia Universidad Católica de Chile, the RCFS Winter Confer-
ence, the Society for Financial Studies Cavalcade, the Texas Finance Festival, the U.S. Government
Accountability Office, The University of Texas at Austin, the UT Dallas Fall Finance Conference,
and the Western Finance Association Conference. We thank Integra Research Group for research
support and the University of Texas McCombs Research Excellence Grant. Griffin is an owner
of Integra Research Group, Integra FEC, and Integra REC, which engage in research, financial
consulting, and recovery on a variety of issues related to the investigation of financial fraud. A
disclosure statement is available on The Journal of Finance website.
   Correspondence: John M. Griffin, Department of Finance, The University of Texas at Austin,
2110 Speedway, Austin, TX 78705; e-mail: John.Griffin@utexas.edu.

This is an open access article under the terms of the Creative Commons Attribution-
NonCommercial-NoDerivs License, which permits use and distribution in any medium, provided
the original work is properly cited, the use is non-commercial and no modifications or adaptations
are made.
DOI: 10.1111/jofi.13209
© 2023 The Authors. The Journal of Finance published by Wiley Periodicals LLC on behalf of
American Finance Association.
                                                 1777
1778                       The Journal of Finance®

traditional banks is the predominant driver in the rise of FinTech lending. A
large aspect of the scrutiny and regulation of traditional banking was its per-
ceived role in the financial crisis, which included facilitating widescale mort-
gage fraud, as partially evidenced by over $137 billion in government fines
and settlements (Griffin, Kruger, and Maturana (2019)). FinTech lenders of-
fer a new banking model that replaces traditional lending relationships with
online advertisements, application programming interfaces, and loan screen-
ing algorithms.
   The Paycheck Protection Program (PPP), a historic COVID-19 relief program
for businesses, rapidly distributed over $793 billion of funds in three short
rounds spread between April 2020 and May 2021. Although FinTech lenders
began with a slow start, with less than 5% of loans in round 1, they ramped
up to over 80% of loans by May 2021, highlighting their growing importance.
FinTech lending was recognized for broadening access to PPP loans, partic-
ularly to smaller firms without preexisting lending relationships with tradi-
tional banks, and for facilitating quick and efficient lending when many small
businesses were in need of funds due to the COVID-19 pandemic. However,
the rapid expansion of FinTech lending may have come at the expense of un-
derwriting standards. While traditional banks have established borrower rela-
tionships and extensive Bank Secrecy Act (BSA) compliance programs, many
FinTech lenders had few established relationships and may have been less dili-
gent when establishing formal procedures, with little reputation to protect.
   Alternatively, FinTech lenders have been shown to use financial data with
increased speed and accuracy. Fuster et al. (2019) find that FinTech mort-
gage lenders not only process government agency loans faster than tradi-
tional banks but also have fewer defaults, indicating potentially superior loan
screening. Peer-to-peer FinTech platforms utilize a rich set of alternative data
and machine learning to optimize credit decisions (Jagtiani and Lemieux
(2019)). If used effectively, this enhanced technology and increased data ac-
cess may be able to detect and prevent PPP applications from fictitious busi-
nesses and individuals. Did FinTech lenders prevent or facilitate fraud in the
PPP? And how does potential fraud and misreporting vary across individual
lenders?
   To investigate these questions, we perform a big data analysis of loan
features on the 11.5 million PPP loans with eight disparate data sets. We
introduce four primary and four secondary indicators of whether a loan is
potentially misstated, which we then validate with three independent exter-
nal measures. Each indicator creates an inference that a loan is suspicious
but is not proof of misreporting on its own. The four primary measures are
nonregistered businesses, multiple loans at a residential address, abnormally
high implied compensation relative to industry by core-based statistical area
(CBSA) averages, and large inconsistencies (as large as 10-fold) between the
jobs reported by borrowers on their PPP application and jobs reported to
another contemporaneous government program application with a different
incentive structure. FinTech loans are 3.23 times as likely to have at least
one primary indicator of misreporting compared to traditional loans and 6.52
                    Did FinTech Lenders Facilitate PPP Fraud?                        1779

times as likely to have a primary indicator that is confirmed by an additional
primary, secondary, or external indicator. Suspicious lending rates vary sub-
stantially across lenders, with potential misreporting rates in excess of 25%
for 10 large FinTech lenders. Suspicious lending is most pronounced for un-
incorporated businesses, which highlights the challenge of quickly extending
widespread aid to small entities with limited documentation.1 FinTech lenders
exhibited consistently higher misreporting rates across all borrower types
even after controlling for loan and borrower characteristics.
   We assess each of the four primary indicators with multiple discontinuity
and comparative analyses. First, there is substantial cross-validation across
the four main indicators. For example, loans that report abnormally high com-
pensation relative to the U.S. Census average compensation in the loan’s indus-
try and CBSA also have higher incidences of nonregistered businesses, mul-
tiple loans at the same address, and inconsistencies in jobs reported. These
patterns are substantially elevated in FinTech lenders but are also present
for traditional lenders, indicating that misreporting is not just confined to
FinTech. Second, someone receiving a fictitious loan might wish to maximize
or come close to maximizing their proceeds. We find monotonically increas-
ing levels of indicators when approaching the perceived maximum compen-
sation threshold and a sharp discontinuity at the threshold with lower lev-
els of suspicious lending just above the threshold. These patterns are present
for all four main indicators; discontinuities are present both for traditional
and FinTech lenders but are much more pronounced for FinTechs with an
increase in potential misreporting of over 6.5 times when approaching the
threshold from below. Third, even though PPP loan amounts were supposed to
be based on historical past compensation with detailed supporting documen-
tation, rounded monthly compensation values are common and coincide with
higher levels of each of the four primary misreporting indicators for FinTech
lenders.
   Fourth, PPP lending at the industry-county level frequently exceeds the
number of establishments listed for that industry and county in U.S. Census
data. For FinTech lenders, 39.3% of loans exceed industry-county establish-
ment counts, and 32.6% of loans exceed industry-county establishment counts
by a factor of more than two.2 Measures of misreporting monotonically increase
as the ratio of PPP loans to businesses documented by the U.S. Census in-
creases, particularly for FinTech lenders. Fifth, because networks in a region
may use recurring loan features, we construct a concentration ratio to measure
clustering in loan amounts, number of jobs, and industries within each lender-
county pair. Like the other secondary measures, FinTech lenders have higher
levels of clustering along loan features, and clustering is monotonically asso-
ciated with higher levels of potential misreporting. Sixth, we collect criminal
background data for a sample of 150,000 individuals. FinTech borrowers are

   1 Fraud was also widespread in pandemic unemployment insurance programs, potentially for

similar reasons (Podkul (2021)).
   2 The corresponding figures for traditional lenders are 14.0% and 8.2%, respectively.
1780                        The Journal of Finance®

more than 3.4 times as likely to have a felony record, and borrowers flagged
for potential misreporting based on the primary and other secondary measures
are also more likely to have felony records. Seventh, the flags strongly corre-
late with loans flagged in online crowd-sourced data collected from PPP Detec-
tive. Finally, the PPP coincides with a surge in regulatory suspicious activity
reports (SARs) from banks and other financial institutions related to small
business lending, and these reports exhibit strong geographic correlation with
our suspicious loan measures.
   Overall, we find 1.41 million questionable loans representing $64.2 billion
in capital based on our primary measures. These measures inevitably contain
some false positives, which would lead to overstatements, and some flagged
loans may have been legitimately eligible for smaller loans. However, the
measures also miss many forms of suspicious lending, and sensitivity analysis
indicates that this total is likely substantially understated. Slightly lowering
the threshold on the high implied compensation and considering excess loans
in industry-county pairs beyond the number of establishments reported by the
U.S. Census results in a total suspicious lending estimate of $117.3 billion.
Moreover, this sensitivity analysis is along only two limited dimensions, and
most of our metrics are conservative and only apply to subsets of loans, further
highlighting that these estimates are ranges that, like most fraud estimates,
have considerable uncertainty.
   Proponents of the PPP often point to the urgency of getting money out
quickly as a potential rationale for tolerating a high level of fraud (e.g., McAr-
dle (2022)). However, this urgency mainly applies to the initial rollout of the
program, and potential misreporting increased over time with particularly
high rates in the last month of round 3 (25.0%), even after the Office of the
Inspector General for the Small Business Administration (SBA) flagged PPP
fraud as a concern in October 2020. Several of the FinTech lenders with the
highest suspicious loan rates are new lenders who did not start making PPP
loans on their own until round 3. There is no evidence that lenders decreased
misreporting over time. Instead, second-draw loans to borrowers with suspi-
cious first-draw loans from the same lender are common, and lenders with
high rates of misreporting in rounds 1 and 2 increased both their misreporting
rates and their loan volume in round 3. For example, the four largest FinTech
lenders, Cross River, Prestamos, Harvest Small Business Finance, and Capital
Plus, exhibited high rates of misreporting and large lending volume growth
while generating approximately a billion dollars in processing fees each. Fi-
nally, FinTech lenders often doubled or tripled their potential misreporting
rates in round 3 compared to rounds 1 and 2.
   Overall, our findings indicate that misreporting was common in the PPP,
especially for FinTech lenders. This result comes with two important caveats.
First, not all FinTech lenders have high misreporting rates. In particular,
Square and Intuit, which are quintessential FinTech lenders but also benefit
from established reputations and broad existing customer relationships be-
fore the PPP, have among the lowest rates of potential misreporting. Thus,
online lending in and of itself does not appear to be the problem. Second,
                     Did FinTech Lenders Facilitate PPP Fraud?                             1781

our findings may not translate to other settings because of unique features
of the PPP. In particular, the PPP involved almost no balance sheet risk for
lenders. Although many FinTech lenders were susceptible to misreporting in
the PPP, this vulnerability may not be a characteristic of FinTech lending
under different incentive structures.
   Our work is related to four main literatures. First, there is a rapidly emerg-
ing literature on FinTech lending that highlights its growing importance and
positive economic effects achieved by filling gaps left by traditional banks
in both residential (Buchak et al. (2018)) and business lending (Gopal and
Schnabl (2022)). Fuster et al. (2019) find that FinTech mortgage lenders
process loans faster and increase the odds of borrowers refinancing their
loans at lower rates, all with fewer defaults, indicating that FinTechs are not
simply engaged in lax screening, as was the case for securitized lending in
the run-up to the financial crisis (Keys et al. (2010), Purnanandam (2011)).
Erel and Liebersohn (2022) examine FinTech lending in the PPP and find that
FinTech lenders increased access to the PPP by lending more in ZIP Codes
with fewer traditional banks, lower incomes, and higher minority percentages.
Chernenko and Scharfstein (2022) show that Black- and Hispanic-owned firms
were less likely to receive PPP loans from traditional lenders. Howell et al.
(2022) find that FinTechs were more likely to provide PPP loans to Black-
owned businesses.3 With respect to FinTech lending before the PPP, Gopal and
Schnabl (2022) show that FinTech lenders have positive economic effects by
filling in gaps in lending to small businesses left by traditional banks following
the financial crisis. Although most of the FinTech literature finds benefits to
FinTech lending such as increased competition, broader financial access, less
discrimination, faster lending speed, and lower defaults, our paper analyzes a
potential cost of FinTech expansion and differential practices across FinTechs.
These costs are an important consideration when evaluating FinTech PPP
lending, but we leave overall welfare analysis to future research.
   Second, regarding the efficacy of the PPP, Chetty et al. (2022) show that
the PPP increased employment at participating firms by only 2% at a cost
of $377,000 per job saved, and Autor et al. (2022) find costs of $170,000 to
$257,000 per job retained. Granja et al. (2022) find small employment effects
due to the PPP and a low correlation between regional COVID-19 variation and
PPP funding allocation. In contrast, Faulkender, Jackman, and Miran (2021)
find that the program was much more effective with an estimated 18.6 mil-
lion jobs saved at an average cost of $28,000, and Denes, Lagaras, and Tsout-
soura (2021) show that firms with short-term delays in PPP access experienced

   3 Atkins, Cook, and Seamans (2022) find that FinTechs helped close a gap in loan size be-

tween Black- and white-owned businesses. In contrast, Bartlett et al. (2022) find that FinTech
algorithms charge higher interest rates to minorities in residential mortgage lending but price
discriminate less than traditional banks. Begly et al. (2022) show that the SBA disaster-relief
home loan program denies more loans to minorities and subprime borrowers due to the program’s
risk-insensitive pricing. Ben-David, Johnson, and Stulz (2022) find that FinTech lending to small
businesses through a major lending platform contracted with the onset of COVID-19 in March
2020 due to financial constraints.
1782                             The Journal of Finance®

fewer visits and higher shutdown rates. Additionally, there is evidence of dif-
ferential access to the PPP based on knowledge of the program, distance to the
closest bank branch, banking relationships, and personal banking connections
(Rabetti (2022), Bartik et al. (2020), Neilson, Humphries, and Ulyssea (2020),
Duchin et al. (2022), Glancy (2022), Li and Strahan (2021)). Our evidence adds
an additional concern regarding the PPP’s efficacy and fairness. We are the
first academic paper to examine widescale potential PPP loan misreporting,
but there have been interesting press and investigative reports regarding sus-
picious PPP loans (e.g., Wieder, Ben and Bobrowsky, Meghan (2020), Willis and
DePillis (2021)), some of which feature FinTech lenders.4
   Third, assessment of the PPP also relates to a broader literature on fraud,
waste, and abuse in government programs (e.g., Glaeser and Goldin (2006)),
including waste due to poor program design and administration (Bandiera,
Prat, and Valletti (2009)). Duflo (2017) emphasizes the importance of program
details. In the PPP, the lack of direct tools for validating eligibility and limited
incentives for high-quality underwriting by lenders with no skin in the game
may have significantly increased fraud and abuse. Chetty, Friedman, and Saez
(2013) find that the fraction of people who manipulate self-reported income in
the Earned Income Tax Credit program grows over time. A similar mechanism
of fraud growing over time seems to be at work here. Hanson et al. (2020)
argue that direct relief for small businesses from the Internal Revenue Service
could have been more targeted and efficient than the PPP’s external lender
model.
   Finally, our work relates to forensic economics and loan misreporting.
Zitzewitz (2012) surveys the literature on forensic economics, noting that a
common thread in this literature is quantifying activity about which there
was previously only anecdotal evidence, in large part because agents have an
incentive to keep it hidden. Widescale mortgage fraud and misreporting in
securitized mortgages (Piskorski, Seru, and Witkin (2015), Garmaise (2015),
Griffin and Maturana (2016), Mian and Sufi (2017), Kruger and Maturana
(2021)) prior to the financial crisis involved both smaller, less-known mortgage
originators and large bank underwriters who knowingly passed along these
misrepresentations in mortgage-backed securities. FinTech lending emerged
and grew against this backdrop as related regulation increased for traditional
banks (Buchak et al. (2018)). Our findings indicate that replacing traditional
lending with FinTech lending amplified misreporting problems, at least with
respect to the PPP.
   Our findings also have important practical implications regarding the ex-
tent and nature of PPP misreporting, the expanding role of FinTech lending,
waste in the PPP, the proliferation of fictitious lending, and the insufficient

   4 Concerns about PPP fraud have been flagged by the Office of the Inspector General for the SBA

(see report at https://www.sba.gov/sites/default/files/2021-01/SBA%20OIG%20Report-21-07.pdf).
Beggs and Harvison (2022) find that among the 2,999 registered investment advisors who took
PPP loans, those with a history of financial misconduct received unusually large PPP loan alloca-
tions.
                      Did FinTech Lenders Facilitate PPP Fraud?                              1783

deterrence of current policies and enforcement. The implications of these find-
ings are further discussed in the conclusion.


                           I. Data and Summary Statistics
A. Data Sources
   The basis for our sample is loan-level PPP data released on January 2, 2022
by the SBA. This data set covers all PPP loans issued from the start of the
program on April 3, 2020 through its end on June 30, 2021 that had not been
repaid or canceled as of January 2, 2022. At the loan level, the data include
business name, address, business type (e.g., corporation, LLC, self-employed,
etc.), NAICS code (industry), loan amount, number of employees, date ap-
proved, loan draw (i.e., initial, first-draw loan or repeat, second-draw loan),
and lender for 11,469,801 loans originated by 4,809 different lenders and with
a total value of $793 billion. We follow Erel and Liebersohn (2022) and clas-
sify lenders as either traditional or FinTech (consisting of online banks and
nonbank lenders) using the same methodology.5
   In addition to the PPP data, we use multiple other data sources. Economic
Injury Disaster Loans (EIDL) were provided by the SBA to businesses and indi-
viduals and included forgivable advances of up to $10,000. EIDL Advance loan-
level data were released on December 1, 2020 and cover all EIDL Advances
issued in 2020. Data on state business registrations are from OpenCorporates,
which collects its data directly from state governments and covers 76 million
businesses across all U.S. states except Illinois. The data include incorporation
dates, dissolution dates (if applicable), and, implicitly, whether the business
has ever been registered. Criminal background data are collected from Lex-
isNexis based on the borrower’s name and address for a random sample of
150,000 round 1 and 2 loans made to individuals (12.9% of rounds 1 and 2 PPP
loans made to individuals).6 We also utilize data on 147,662 loans reported on
PPP Detective (https://www.pppdetective.com), a crowd-sourced platform that
allows individuals to search PPP loans and report PPP loans as “potentially
being fraudulent.” The BSA requires banks and other financial institutions to

    5 We use Erel and Liebersohn’s (2022) classifications for lenders who were active in rounds 1

and 2 (the sample period for Erel and Liebersohn (2022)), and we use the same methodology for
classifying round 3 lenders who were not active enough to be classified in the earlier rounds. Clas-
sification of FinTech lenders can be difficult because traditional banks with multiple branch loca-
tions may also originate loans from other lenders or online portals. See the Internet Appendix for
additional details. The Internet Appendix is available in the online version of the article on The
Journal of Finance website. Results are robust to different definitions of FinTech lenders, includ-
ing restricting FinTechs to online banks or nonbanks and dropping large community development
lenders who operated as FinTechs in the PPP (as shown in Table IA.II in the Internet Appendix).
    6 Because the LexisNexis searches require an individual’s name, only loans with an individual

name listed as the borrower (rather than a business name) and where the business type is a self-
employed individual, an independent contractor, or a sole proprietor are included in this criminal
search. The criminal records data are collected only from round 1 and 2 loans because round 3
data were released after the criminal records data were collected.
1784                            The Journal of Finance®

report suspicious activities to the Financial Crimes Enforcement Network
(FinCEN), a division of the U.S. Treasury Department. We utilize summary
data from FinCEN on SARs from 2014 to 2021. Finally, we use several U.S.
governmental data sources for address, demographic, and business infor-
mation. Additional details on all of the data sources are described in the
Internet Appendix.


B. Summary Statistics
   Summary statistics for the 3.7 million FinTech and 7.7 million traditional
bank loans in our sample are reported in Table I. FinTech loans have an av-
erage loan amount of $23,000 compared to $91,000 for traditional bank loans.
Despite these large differences in means, the median loan sizes of $19,000
and $21,000 are similar. The average FinTech loan reports supporting 2.3 jobs
compared to 10.5 for traditional banks. After normalizing loan size relative
to reported jobs, FinTech loans have a higher average ($63,000) and median
($68,000) implied compensation than traditional bank loans ($47,000 average
and $39,000 median). For FinTech loans, 19.5% of borrowers are organized as
corporations, S-corporations, or limited liability companies (LLC), compared
to 65.8% for traditional bank lenders. FinTech loans were also less likely to
be repeat loans, with 27.1% of round 3 FinTech loans going to borrowers with
previous PPP loans, compared to 60.8% for round 3 traditional bank loans.
   Panel A of Figure 1 shows the number of loans originated on the left axis
and the total amount lent on the right axis by each of the top 75 PPP lenders.
FinTech lenders are highlighted in red (nonbank FinTech lenders) and cream
(online banks). Six of the 10 top lenders by number of loans are FinTechs,
with Cross River, Prestamos, and Harvest in the top five alongside Bank of
America and JPMorgan Chase.7 Due to their larger average loan size (Erel and
Liebersohn (2022)), dollar lending volume is higher for most traditional banks.
   Panel B of Figure 1 shows the total FinTech market share during each week
throughout the three rounds of PPP lending. Total FinTech market share grew
from only 1.1% of loans in the first week of round 1 to 5.9% in the last week
of round 1. Round 2 continued the PPP after a short break of 10 days in May
2020 with new funding for borrowers who did not receive a loan in round 1.
By the end of round 2 in August 2020, FinTech market share grew to 50.1% of
loans over the last two weeks, for an overall market share of 4.2% in round 1
and 20.1% in round 2. Round 3 of the PPP started in January 2021 with a low
FinTech market share of 12.3% in the first three weeks, but reached over 86%
of loans for the second half of May 2021, for an overall round 3 market share
of 46.4%.8

   7 Comparing this figure to Panel A of Figure IA.1 in the Internet Appendix shows how the top

lenders differ between the entire sample and solely rounds 1 and 2. In particular, the growth of
Prestamos, Harvest, and Capital Plus in round 3 is apparent.
   8 Panel B of Figure IA.1 in the Internet Appendix shows the number of loans originated each

week of the PPP by type of lender.
                      Did FinTech Lenders Facilitate PPP Fraud?                                1785

                                              Table I
                                   Summary Statistics
This table presents summary statistics for our sample. The sample includes all PPP loans ap-
proved from the start of the program on April 3, 2020 through its end on June 30, 2021 that had
not been repaid or canceled as of January 2, 2022. FinTech lenders are determined following Erel
and Liebersohn (2022). Loan Amount is the initial approved amount minus any portion used to
refinance an EIDL loan. Implied Comp. is determined following the guidelines in place when the
loan was approved and is based on loan amount and jobs reported. CBSA/NAICS Avg. Comp.
is the average compensation in the loan’s industry-CBSA based on the U.S. Census CBP data.
CBSA/NAICS Avg. Receipts is the average receipts (for business types that were able to use gross
income to calculate loan size) to nonemployer businesses in the loan’s industry-CBSA based on the
U.S. Census NES data. Normalized Comp. is the ratio of Implied Comp. and either CBSA/NAICS
Avg. Comp. or, if the business was able to use gross income to calculate its loan amount, the larger
of CBSA/NAICS Avg. Comp. and CBSA/NAICS Avg. Receipts. Loans (Within Draw) at Address is
the number of loans (within the loan’s draw) at the same residential address. Frac. Corp, S Corp,
LLC is the percentage of loans to these business types, Frac. Second Draw is the percentage of
round 3 loans that are the borrower’s second draw from the PPP, and Frac. Matched EIDL Ad-
vance is the percentage of loans with a matching EIDL Advance. Frac. FinTech (Either Type), Frac.
Nonbank FinTech, and Frac. Online Bank FinTech are the percentages of loans that are originated
by the given type of lender.

                                                   FinTech                      Traditional

                                          Mean       SD      Median     Mean       SD         Median

Num. Loans [Pct. Loans]                       3,732,133 [32.54%]            7,737,668 [67.46%]
Loan Amount                               23,200 84,470 18,500         91,211 305,665 20,833
Jobs Reported                              2.322    9.000     1.000    10.493    29.547     3.000
Implied Comp.                             62,984 37,357 67,594         46,633    46,321     38,743
CBSA/NAICS Avg. Comp                      46,921 38,895 36,942         49,798    37,459     42,849
CBSA/NAICS Avg. Receipts                  46,141 31,432 35,732         59,458    37,254     53,291
Normalized Comp.                           1.676    1.377     1.227     1.091     1.221     0.897
Num. Loans (Within Draw) at Address        1.321    0.794     1.000     1.193     0.801     1.000
Frac. Corp, S Corp, LLC                    0.195                        0.658
Frac. Second Draw (Round 3 Loans)          0.271                        0.608
Frac. Matched EIDL Advance                 0.182                        0.269

                                              Round 1             Round 2             Round 3

Num. Loans [Pct. Loans]             1,618,989 [14.1%] 3,517,397 [30.7%] 6,333,415 [55.2%]
Loan Amount                              197,351           57,424            42,766
Jobs Reported                            20.468            7.915             4.561
Implied Comp.                             47,173           43,364            57,945
CBSA/NAICS Avg. Comp                      49,339           51,830            46,935
CBSA/NAICS Avg. Receipts                     –                –              50,003
Normalized Comp.                          1.156            1.040             1.477
Num. Loans (Within Draw) at Address       1.277            1.212             1.261
Frac. FinTech (Either Type)               0.0424           0.201             0.467
Frac. Nonbank FinTech                     0.0265           0.0789            0.410
Frac. Online Bank FinTech                 0.0159           0.122             0.0574
Frac. Corp, S Corp, LLC                   0.829            0.657             0.342
Frac. Second Draw                            –                –              0.451
Frac. Matched EIDL Advance                0.292            0.301             0.195
1786                            The Journal of Finance®




Figure 1. FinTech market share. This figure shows the role that FinTech lenders played in the
PPP. Panel A shows the number of loans (bars) and dollar value of loans (dots) originated by the
top 75 lenders (by number of loans). Panel B shows the percentage of loans originated by FinTech
lenders during each week of rounds 1, 2, and 3 of the PPP on the left axis and the total number
of loans originated each week on the right axis. In both panels, red represents nonbank FinTech
lenders, cream represents online bank FinTech lenders, and gray represents traditional lenders.
Note that mid-August through December 2020 is not shown in Panel B since no PPP loans were
originated during this period. (Color figure can be viewed at wileyonlinelibrary.com)


                     II. Primary Suspicious Loan Measures
  We introduce four primary indicators that a loan is potentially misstated.
In this section, we define and introduce the indicators. Each indicator creates
an inference that a loan is suspicious but may also contain false positives. In
subsequent sections, we validate the measures and explore how they relate to
one another and other misreporting indicators.


A. Business Registry Flag
  Businesses organized as corporations, S-corporations, and LLCs are required
to file an article of incorporation or LLC filing with a state, either as a domestic
company in their home state or as a foreign company in another state. Further,
                      Did FinTech Lenders Facilitate PPP Fraud?                                 1787

the SBA required businesses to be “in operation on February 15, 2020…[and]
not permanently closed.”9 Based on these requirements, we check the following
conditions for all corporation, S-corporation, and LLC borrowers:

    (i) Is there a matching business in the business registry data? (“Missing
        Business”)10
   (ii) Was the business dissolved and inactive before being approved for a
        PPP loan? (“Dissolved Business”)11
  (iii) Is the earliest incorporation or initial filing date for the business after
        February 15, 2020? (“Late Incorporation/Filing”)

These three subflags are combined to form an overall business registry flag.12
Additional details are described in the Internet Appendix.

B. Multiple Loan Flag
   Although a business owner may have multiple businesses registered to the
same address, the presence of multiple loans at an individual residential ad-
dress during the same draw is also a potential sign of fictitious operations. We
first standardize addresses and identify those that are known business or cen-
tral addresses (e.g., office and apartment buildings) using the Address Valida-
tion Application Programming Interface from the United States Postal Service.
We identify residential (i.e., nonbusiness, noncentral) standardized addresses
with multiple loans within the same draw. To be conservative, we only flag
addresses with three or more loans.
   As an example, a modest suburban home north of Chicago with an estimated
home value of $170,000 per Zillow received 14 loans at a single address, all
with colorful business names, almost all in the same industry, most with the
   9 See     loan     application     at    https://www.sba.gov/sites/default/files/2021-03/Borrower-
Application2483ARPrevisions%20%28final%203-18-21%29-508.pdf.
   10 We take a conservative approach to quantifying missing businesses by considering all busi-

nesses registered to do business in the state of the PPP loan as well as all businesses in other states
that list a registration address in the state of the PPP loan. We are also conservative in our name
matching and count any business name with a match ratio of 75% or more as a potential match.
   11 To be flagged, the dissolution date of the business must be before the PPP loan approval date

and, to screen out businesses that may be administratively dissolved (e.g., for not filling some
paperwork), the business status must be listed as inactive. To be conservative, this flag is only
applied when the dissolved business registration matches the PPP borrower with a higher level
of confidence (with at least a 90% match ratio as opposed to the 75% ratio cutoff used for the
existence of a registration). If multiple business registrations match with ratios of at least 90%,
the flag is only applied if they have all been dissolved. The late incorporation flag also employs the
same conservative methodology.
   12 As external validation of this flag for a smaller sample, we also compare PPP borrower names

to data from the Florida Department of Business and Professional Regulation, following Cher-
nenko and Scharfstein (2022) (see Figure IA.2 in the Internet Appendix). Loans flagged as a miss-
ing business based on the overall business registry are over 6.8 times less likely to have a potential
match in the restaurant data as compared to loans that are not flagged. Loans flagged as having
missing business registrations that also have another flag (primary or secondary) are over 13.9
times less likely to have a potential match in the restaurant data.
1788                             The Journal of Finance®

same loan amount, and all backing 10 jobs (as shown in Panel A of Exhibit IA.1
in the Internet Appendix).13 Another multiple-loan example involves loans to
four people in the same household, again in a modest suburban Chicago home,
all of whom received loans for the same amount, $20,833, which corresponds
to the PPP’s maximum annual compensation of $100,000 (as shown in Panel B
of Exhibit IA.1 in the Internet Appendix).14 Random loan-level inspections of
the data reveal numerous other examples of multiple suspicious loans flowing
to addresses that do not seem to be the locations of identifiable businesses.
The multiple loan flag functions as a way to systematically analyze these
loans.


C. High Implied Compensation Flag
  PPP loan size is limited to 2.5 times a business’s average monthly payroll
expenses, including up to $100,000 in annual compensation per employee.
PPP loan applications report how many employees the business has based
on the same time period used to calculate average payroll expenses (2019
in most cases). Using loan size and number of reported employees, we can
impute implied average annual compensation. Implied compensation at the
borrower level is strongly related to average compensation in the borrower’s
industry (North American Industry Classification System (NAICS) four-digit)
and CBSA (e.g., see Figure IA.3 in the Internet Appendix).15
  For our main measure of high implied compensation, we conservatively
only flag loans for which the implied compensation per job reported is more
than three times the industry-CBSA average compensation/receipts (“high
implied compensation”). Because compensation is censored at $100,000 for




   13 The first loan is to an LLC that was registered in 2018, but the 13 subsequent loans during

July and August 2020 are to LLCs that were registered only shortly before the loans were ap-
proved, well after the February 15 eligibility cutoff. Detailed Internet searches did not produce
information for 13 of the business names or any indication of employees other than the owner.
   14 This income is high for the indicated industries, which have average compensation of $25,000

to $46,000 in the Chicago CBSA according to the U.S. Census County Business Patterns (CBP). All
four individuals also received second-draw loans for the same amount. SBA guidelines ask the bor-
rower for their business address. The industries themselves are also suspicious in that one of the
loans is for an auto repair business and two are for equipment manufacturing despite no evidence
of these businesses in photos of the property. The borrower in another loan for a nail salon does
not appear to have an Illinois nail technician license. Another borrower at the address changed
industries from equipment manufacturing to the nail salon industry during round 3 despite also
not having a nail technician license.
   15 Schedule C filers also had the option of using gross income instead of net income for owner

compensation after March 3, 2021. To conservatively account for the option, we compare implied
compensation for sole proprietor, independent contractor, self-employed, and single member LLC
loans after March 3, 2021 to the greater of industry/CBSA average compensation and indus-
try/CBSA average receipts for single-employee firms. This adjustment is also used in calculating
the high implied compensation measure below.
                    Did FinTech Lenders Facilitate PPP Fraud?                            1789

most borrowers, this flag is only possible in industry-CBSA pairs with average
annual compensation/receipts below $33,333.33.16


D. EIDL Advance Jobs > PPP Jobs Flag
   Concurrently with the PPP, the SBA provided businesses and individuals
with the ability to receive a forgivable EIDL Advance of up to $10,000.17 For all
EIDL Advances issued in 2020, the advance amount was calculated as $1,000
per employee (up to the $10,000 maximum).18 Thus, there was an incentive for
borrowers to inflate the number of jobs reported on their EIDL applications.19
We focus on cases in which EIDL jobs exceed PPP jobs because the job inflation
incentive is provided by the EIDL Advance program; there is no direct incen-
tive to inflate the number of jobs on the PPP application since PPP loans are
based on total payroll as opposed to the number of jobs reported. To make it
less likely that differences are driven by reporting or timing differences in the
number of employees reported, we only consider differences of three or more
jobs. Most of the flagged loans are even more egregious than this. As discussed
further below, 14.1% of matched FinTech loans have an EIDL-PPP jobs differ-
ence of a full nine jobs, indicating that the borrower reported one job to the
PPP but maxed out the EIDL Advance with at least 10 reported jobs.


E. FinTech Differences?
  In Panel A of Figure 2, we plot suspicious loan rates by lender for the top 75
lenders. FinTech lenders tend to cluster at the left of the graph and traditional

   16 Some loans are also outside of a CBSA or in an industry-CBSA pair that is too small to be

included in the U.S. Census CBP or Nonemployer Statistics (NES) data. In total, 3,313,359 loans
(28.9% of the sample) are in industry-CBSA pairs with average annual compensation/receipts
below $33,333.33. Because this compositional construction censors loans in higher compensa-
tion industry-CBSAs, we analyze this flag within the subset of industry-CBSAs with compen-
sation/receipts below $33,333.33.
   17 Although the EIDL Advance program was billed as a forgivable advance with the potential

for a larger nonforgivable loan, 65.8% of EIDL Advances involved no additional EIDL loan. EIDL
Advances were immediately forgiven by the SBA.
   18 The EIDL Advance rules changed for 2021 to: (i) provide the entire $10,000 regardless of

employee count, and (ii) target the advances to low-income communities and those with a demon-
strated decrease in revenue.
   19 Note that the misreporting identified by this flag is committed on the EIDL Advance

application. The flag does not directly imply misreporting on the PPP loan; however, indi-
viduals who engaged in financial misconduct in one area are five times as likely to com-
mit subsequent misconduct (Egan, Matvos, and Seru (2019)). For borrowers who take out the
maximum EIDL Advance of $10,000, we can infer that the borrower claimed at least 10 em-
ployees on their EIDL Advance application. Subsequent to the release of the first public ver-
sion of this paper, an October 7, 2021 report by the SBA OIG found that over 700,000 EIDL
recipients applied for and received advances for multiple employees even though they only had
a single employee, resulting in $4.5 billion of improper EIDL Advance payments (see report at
https://www.sba.gov/sites/default/files/2021-10/SBA%20OIG%20Report%2022-01%20.pdf).
1790                              The Journal of Finance®




Figure 2. Misreporting rates. This figure shows the variation in the percentage of loans flagged
across lenders. In both panels, the bars show the percentage of loans originated by each lender that
are flagged by at least one of the primary flags and the horizontal line represents the percentage
of loans originated by any lender that are flagged by at least one primary flag. The top 75 lenders
by number of loans (across the entire sample) are shown. Panel A is based on all loans by each
lender and Panel B is based on loans by each lender to specific business types (as denoted by each
subpanel title). (Color figure can be viewed at wileyonlinelibrary.com)
                      Did FinTech Lenders Facilitate PPP Fraud?                               1791

lenders cluster in the middle and to the right. The 10 lenders with the most
suspicious loans are all FinTechs and have at least a quarter of their loans
implicated, compared to the overall average of 12.3%. In the extreme, Lendis-
try, Capital Plus, Prestamos, Itria, and MBE have primary flag rates of 33.8%,
31.8%, 29.8%, 29.4%, and 29.3%, respectively. Although most of the FinTech
lenders cluster among the lenders with the most suspicious loans, there are a
few exceptions. In particular, Capital One, Intuit, and Square have misreport-
ing rates that are well under the average misreporting rates across all lenders.
   Panel B of Figure 2 plots the suspicious lending rates by lender separately
for different business types. FinTech lenders consistently cluster to the left
in each plot, indicating that the patterns of suspicious lending are not iso-
lated to only a few business types. Suspicious lending to unincorporated enti-
ties is consistently higher, reaching near or over 40% for some lenders’ LLC,
self-employed, and sole proprietorship loans. However, for corporate entities
(C-Corp and S-Corp), the levels are not trivial, as they reach as high as
20% and 15%, respectively. It is important to note that most of the suspi-
cious loan indicators only apply to subsets of loans. To confirm that selec-
tion along these dimensions is not driving the results, we examine lender
fixed effects from regressions with business type fixed effects and indica-
tors for loans with matching EIDL loans and loans in industry-CBSAs with
compensation/receipts below $33,333.33 and find similar dispersion across
lenders.20
   Figure 3 plots lender-level flag rates separately for each indicator. Each plot
shows how two of the individual flags relate to one another at the lender level.
For example, the plot on the top left shows the relation between the business
registry and multiple loan flags, and the plot on the bottom right shows the
high implied compensation and EIDL > PPP jobs flags. All of the flags are
plotted as a percentage of loans that could potentially be flagged by the mea-
sure (corporate and LLC loans for the business registry flag, all loans for the
multiple loan flag, loans in industry-CBSAs with compensation/receipts below
$33,333.33 for the high implied compensation flag, and loans with matching
EIDL Advances for the EIDL > PPP jobs flag). Three related patterns are ap-
parent in the plots. First, FinTech lenders are consistently near the top end
of each individual flag measure. Second, a similar set of FinTech lenders clus-
ters near the top for each of the individual measures. Third, the flag rates

   20 As shown in Panel A of Figure IA.4 in the Internet Appendix, regressions with these con-

trols result in large and significant lender fixed effects (with Bonferroni corrections for multiple
testing and for standard errors double clustered by lender and ZIP Code) with a dispersion of
30.5 percentage points (ppt) across the top 75 lenders, which is only slightly below the uncondi-
tional dispersion of 32.9 ppt shown in Panel A of Figure 2. We also calculate lender fixed effects
in more saturated regressions with all of the control variables and fixed effects used in column (4)
of Table II. Although fixed effects in these regressions remain large and significant, the additional
control variables decrease their dispersion to 14.6 ppt across lenders. To the extent that fraud
clusters heavily in certain ZIP Codes and industry-CBSA pairs as displayed later in the paper,
this fully saturated regression with industry × CBSA and ZIP Code fixed effects may understate
differences across lenders.
1792                               The Journal of Finance®




Figure 3. Relation between primary flags. This figure shows the relation between the pri-
mary flags (business registry, multiple loans, high implied compensation, or EIDL > PPP jobs
flags) at the lender level. Each subpanel is a scatterplot with the percentage of loans flagged by
one of the flags on each axis. Loans are filtered to the sets for which we can determine each flag
for each axis separately (i.e., we do not require that both flags can be determined for a given loan).
Lenders with at least 5,000 loans are shown; for the subpanels with the EIDL > PPP jobs flag, we
additionally require that the lender have at least 1,000 loans with a matched EIDL Advance. The
dashed line is a linear fit and the correlation is shown in the bottom left corner of each subpanel.
(Color figure can be viewed at wileyonlinelibrary.com)




are significantly correlated with one another across lenders. In particular,
FinTech lenders Capital Plus, Lendistry, Prestamous, and Harvest are con-
sistently at the top of each graph, and several other FinTech lenders are not
                      Did FinTech Lenders Facilitate PPP Fraud?                                1793

                                               Table II
                        Prevalence of Flags by Lender Type
This table presents the percentage of loans flagged by the four main flags, at least one of the four
flags, and at least two of the four flags. In Panel A, column (1) shows the percentage of FinTech
loans with the given flag; column (2) shows the percentage of traditional loans with the given
flag; column (3) shows the difference between the FinTech and traditional percentages; column (4)
shows the adjusted differences with ZIP Code, business type, and NAICS × CBSA fixed effects and
controlling for number of jobs reported, loan size, and whether we can match the loan to an EIDL
Advance; and column (5) shows the differences between matched pairs of FinTech and traditional
loans. The N values show the number of loans for which the flag can be determined and robust
standard errors are double clustered by ZIP Code and lender. For the matched differences, robust
standard errors are four-way clustered by the ZIP Code and lender of both matched loans. The
full regression results for the unadjusted and adjusted differences are reported in Panels A and
B, respectively, of Table IA.I in the Internet Appendix. t-statistics are in parentheses. Significance
levels are indicated by ***p < 0.010.

                        (1)              (2)              (3)               (4)              (5)
                                                      Unadjusted        Adjusted         Matched
                     FinTech         Traditional      Difference        Difference       Difference

Business             0.103            0.0429               0.0605∗∗∗    0.0372∗∗∗         0.0296∗∗∗
Registry          N = 697,283      N = 4,838,905          (3.12)       (3.76)            (3.19)
Multiple             0.0330           0.00961              0.0234∗∗∗    0.0111∗∗∗         0.0199∗∗∗
Loans             N = 3,723,133    N = 7,737,668          (9.96)       (6.33)            (5.29)
High Implied          0.474           0.0857               0.388∗∗∗     0.100∗∗∗          0.134∗∗∗
Comp.             N = 1,285,292    N = 2,028,067          (8.14)       (7.97)            (5.35)
EIDL > PPP           0.213            0.0473               0.166∗∗∗     0.0672∗∗∗         0.0956∗∗∗
Jobs              N = 677,896      N = 2,084,437          (4.57)       (4.22)            (4.52)
At Least One          0.230           0.0713               0.159∗∗∗     0.0602∗∗∗         0.0754∗∗∗
Primary Flag      N = 3,723,133    N = 7,737,668          (8.62)       (8.25)            (7.95)
At Least Two         0.0239           0.00329              0.0206∗∗∗    0.00744∗∗∗        0.0137∗∗∗
Primary Flags     N = 3,723,133    N = 7,737,668          (8.00)       (5.76)            (7.92)



far behind. In contrast, no traditional lender is in the top 10 for more than one
flag.21
   Table II summarizes the percentage of loans with each of the four flags sepa-
rately for FinTech and traditional lenders. The table also summarizes the per-
centage of all loans with at least one flag and with two or more flags. For each
individual measure, the denominator is the number of loans that could have
the flag. For the overall flag measures, the denominator is all loans in the sam-
ple, which understates the incidence of suspicious loans since most of the flags
are only applicable to a minority of loans. Differences between FinTech and
traditional flag percentages are reported in column (3). For all four individual

   21 Graphical rankings of lenders for each separate measure similar to Panel A of Fig-

ure 2 are shown in Figure IA.5 in the Internet Appendix. Additional details and supplemen-
tal analysis regarding how the individual flags vary across lenders are further discussed in the
Internet Appendix.
1794                             The Journal of Finance®

measures, FinTech lenders have flag rates that are 2.40 to 5.53 times as high
as traditional lenders, with particularly large differences for the high implied
compensation and EIDL > PPP jobs flags. Overall, 23.0% of FinTech loans
have at least one of the flags, compared to 7.1% for traditional loans. These
differences are all highly significant with standard errors double-clustered
by ZIP Code and lender to conservatively allow for potential geographic and
within-lender correlations.
  To account for potential compositional differences between FinTech and tra-
ditional lenders, column (4) reports adjusted differences based on regressions
that control for loan size and number of jobs and include ZIP Code, business
type, and industry × CBSA fixed effects.22 After accounting for these effects,
the adjusted difference between FinTech and traditional flag rates is 3.7 ppt
for the business registry flag (which is 85% of the rate for traditional loans),
1.1 ppt (116%) for the multiple loan flag, 10.0 ppt (117%) for the high implied
compensation flag, and 6.7 ppt (142%) for the EIDL > PPP jobs flag. These
results indicate that, even though loan composition explains part of the dif-
ference between FinTech and traditional loans, flag rates remain much higher
for FinTech loans even after controlling for all observable characteristics. To
further control for potentially nonlinear loan characteristic effects, we match
FinTech loans with traditional loans based on loan size, industry, county,
and business type in column (5) with similar results.23 Despite the extensive
matching, control variables, and fixed effects, it remains possible that other
omitted variables or unobserved loan characteristics could explain some
of the differences between FinTech and traditional loans, but these effects
would have to be large to explain the results. To control for any unobserved
differences across households, Table IA.III in the Internet Appendix considers
a restricted sample of residential addresses with multiple loans within the
same draw. Consistent with results in Table II, flag rates are elevated for
FinTech loans across all of the potential misreporting measures, with highly
statistically significant differences in all specifications. Tests in the next
section, including grouping around discontinuities and clustering, also help to
address omitted variable concerns.


F. Are the Suspicious Loan Flags Related to One Another?
  If the above indicators of potential misreporting are due to random data
errors or honest mistakes, one might expect different indicators to occur
randomly across loans and lenders. Therefore, multiple flags for the same
loan create a heightened misreporting inference, and high lender flag rates


  22 The differences and standard errors reported in columns (3) and (4) are based on regressions

with and without control variables and fixed effects. Results for these regressions are reported in
more detail in Table IA.I in the Internet Appendix. Table IA.II in the Internet Appendix shows
how these results change with the inclusion of different fixed effects, different standard error
clustering, and different definitions of FinTech lenders.
   23 Details on the matching process are provided in the Internet Appendix.
                     Did FinTech Lenders Facilitate PPP Fraud?                              1795

across multiple indicators may be due to policies and practices that facilitate
more misreporting.
   In addition to their relations to one another, the granularity of the flags es-
tablishes patterns that would be difficult to explain based on errors or honest
mistakes. This is particularly true for differences for the EIDL > PPP jobs
flag. Panel A of Figure 4 shows the frequency of particular differences between
EIDL Advance and PPP jobs. The asymmetry of the plot is consistent with
an incentive to inflate EIDL Advance to increase EIDL Advance payments,
which were $1,000 per employee, up to $10,000. Even more strikingly, the most
common discrepancy between the programs is a difference of nine jobs, which
implies that the borrower claimed 10 or more jobs and took out the maxi-
mum EIDL Advance of $10,000 despite only reporting one PPP job. In par-
ticular, EIDL jobs exceed PPP jobs by nine 14.3% of the time for FinTech loans
compared to 0.5% for traditional loans.
   We next examine the distribution of normalized compensation in more detail
and assess how it relates to our other three suspicious loan flags. To capture
abnormally high compensation, we examine the distribution of implied av-
erage compensation for the borrower normalized by mean compensation or
receipts across all firms in the borrower’s industry and CBSA based on U.S.
Census Bureau CBP data. The distribution is shown separately for FinTech
and traditional loans in the top left plot in Panel B of Figure 4. FinTech bor-
rowers have a much fatter right tail of the abnormal compensation distribution
within CBSA-industry pairs; 16.4% of FinTech borrowers have normalized
compensation above 3, compared to 2.5% of traditional borrowers.24
   The remaining three plots in Panel B of Figure 4 show the percentage
of loans with the business registry, multiple loans, and EIDL > PPP loan
flags, respectively, separately for FinTech and traditional lenders.25 All flags
increase significantly as normalized compensation increases for loans made
by FinTech lenders. While 7.9% of corporate and LLC FinTech loans with
normalized compensation below one have the business registry flag, 27.0%
of loans with normalized compensation above three have the flag. Similarly,
the multiple loan flag increases from 2.3% for FinTech loans with normalized
compensation below one to 4.9% when normalized compensation is above
three. An even larger increase from 8.1% to 63.6% is observed for the EIDL >
PPP flag. Importantly, while FinTech loans exhibit a stronger relation between
normalized compensation and the other loan flags, this pattern is not limited

   24 Most of this is due to round 3 FinTech loans, as is evident in Figure IA.3, Panel A, in the

Internet Appendix. FinTech implied compensation is much higher in round 3 and appears to be
almost completely disconnected from average industry-CBSA compensation and receipts. Panel
B of Figure IA.3 in the Internet Appendix shows that this differential pattern for FinTech and
traditional lenders in round 3 is also evident even when the sample is restricted to Schedule
C borrowers, both before and after Schedule C borrowers were permitted to use gross income,
starting on March 3, 2021.
   25 The plots, as well as subsequent analysis. are calculated based on loans for which data to

calculate each flag are available (e.g., corporate and LLC loans for the business registry flag and
loans with matched EIDL Advances for the EIDL > PPP jobs flag).
1796                              The Journal of Finance®




Figure 4. Additional features of primary flags. This figure shows additional features of the
primary flags. Panel A shows the difference between the number of employees implied by a busi-
ness’s EIDL Advance amount (“EIDL Implied Jobs”) and the number of jobs reported by the busi-
ness on its PPP application (“PPP Reported Jobs”) by lender type. Panel B shows the distribution
of normalized compensation and the relation between it and the other primary flags. We define
normalized compensation as the implied compensation of the loan divided by the average compen-
sation/receipts in the loan’s industry-CBSA. In the top left subpanel of Panel B, the kernel density
of loans is shown, and in the other three subpanels, the percentage of loans flagged by the given
flag in each bin is shown, where each bin is 0.2 units wide. The solid lines are third-degree poly-
nomial fits for the percentage flagged and the dashed lines are 95% confidence intervals. In both
panels, only loans for which the given flag can be determined are considered. (Color figure can be
viewed at wileyonlinelibrary.com)



to FinTechs. Traditional bank loans also have more flags when normalized
compensation is higher, suggesting that at least some traditional bank loans
also have misreporting, though at a much lower scale. Overall, the results
show that while some variation in normalized compensation across firms is to
                      Did FinTech Lenders Facilitate PPP Fraud?                              1797

be expected, high implied compensation is strongly related to other suspicious
loan characteristics, particularly for FinTech loans.
   To assess loan-level relations between the flags, we calculate odds ratios be-
tween each pair of flags. For each odds ratio, z-statistics calculated based on
standard errors double-clustered by ZIP Code and lender are in parentheses.
The odds ratios for the full sample are all above 1.48 and highly significant
(as shown in Panel A of Table IA.IV in the Internet Appendix). In particular,
the odds ratio between the high implied compensation and EIDL > PPP jobs
flags is 14.43 and has a z-statistic of 20.46, indicating that loans flagged by the
high implied compensation flag are over 14 times as likely to be flagged by the
EIDL > PPP jobs flag as well and vice versa. To examine whether these rela-
tions can be explained by observed loan characteristics, we regress each of the
flags jointly on the other flags, controlling for loan size and number of jobs with
ZIP Code, business type, and industry × CBSA fixed effects, with and without
lender fixed effects (as shown in Table IA.V in the Internet Appendix). Except
for the relation between the business registry flag and EIDL > PPP jobs flags,
the coefficients between the flags are all positive, economically large relative
to the mean flag rates, and highly statistically significant.
   As an additional examination of the primary flags, we also compare them
to direct evidence of loan size inflation for nonprofits based on comparing loan
sizes to nonprofit compensation disclosed on IRS Form 990. Loan size infla-
tion by nonprofits is increasing and highly related to the primary flags (see
Figure IA.7 and the Internet Appendix for details on this analysis).


                   III. Secondary Suspicious Loan Measures
  In addition to misreporting, the primary indicators introduced in the pre-
vious section also have potentially innocent explanations. In this section,
we develop and analyze four additional measures as added validation. The
additional measures involve discontinuities, rounded compensation levels,
abnormal numbers of loans in industry-county pairs, and clustering of loan
features within lender-county pairs. We also examine how these measures
relate to the primary measures and the differences between FinTech and
traditional lenders.


A. Discontinuities at $100,000 Compensation
  PPP loan size is calculated as 2.5 times a borrower’s average monthly
payroll, including up to $100,000 in wages per employee, with the ability to
also include nonwage benefits and payroll expenses in excess of $100,000.26
A borrower might want to maximize their loan amount by submitting payroll

   26 See the Internet Appendix for details on the SBA guidance for how to calculate loan size. This

$100,000 cutoff is a hard maximum for self-employment compensation. For other employees, pay-
roll expenses also include employer insurance and retirement contributions and unemployment
taxes, which can push included payroll expenses above $100,000 per employee.
1798                              The Journal of Finance®

expenses at or close to the $100,000 per employee limit without the additional
expenses that are eligible with proper payroll details.
   Panel A of Figure 5 plots the distribution of implied compensation per em-
ployee and shows how it relates to the misreporting indicators from the pre-
vious section. The implied compensation distributions (up to $130,000) for
FinTech and traditional loans are plotted as orange and gray bars, respec-
tively. FinTech loans stand out as having more loans with implied compensa-
tion right at and slightly under $100,000, and traditional banks have more
loans with implied compensation between $10,000 and $75,000. The percent-
age of loans with one of the four primary flags for FinTech and traditional loans
are plotted as orange and gray dots along with third-degree polynomials and
their associated 95% confidence intervals estimated separately above and be-
low the $100,000 compensation bin. As compensation increases from $40,000
to $100,000, the prevalence of the primary flags for FinTech loans increases
from 6.5% to 43.5%. For traditional lenders, the increase is also present but
much smaller. For FinTech loans with implied compensation above $100,000,
there is a sharp drop-off in the flag rate, which indicates that businesses that
followed the detailed SBA guidelines for including nonwage payroll expenses
for employees with wages above $100,000 are less likely to have one of the
primary misreporting flags. Similar patterns exist for each of the flags individ-
ually (see Panel A of Figure IA.8 in the Internet Appendix).27
   Additionally, the SBA used a loan amount cutoff of $150,000 for more
streamlined processing (fewer calculations and less documentation) of loan for-
giveness applications.28 Consistent with applicants or lenders being aware of
the threshold and trying to avoid scrutiny, the percentage of flagged loans is
high for loans up to $150,000 and decreases after the threshold. This is true for
both traditional and FinTech lenders, but much more pronounced for FinTech
lenders (see Figure IA.9 in the Internet Appendix).

B. Rounded Loan Amounts
  The PPP loan application instructs borrowers to enter their average monthly
compensation and to calculate their loan amount as:
 Loan Amount = Average Monthly Payroll × 2.5 + EIDL Refinance Amount.
Applicants are instructed to calculate average monthly payroll based on
historical compensation (in 2019 in most cases) with detailed supporting doc-
umentation.29 It is unlikely that actual monthly payroll would be a round
  27 In Table IA.VI in the Internet Appendix, we formally test for discontinuities at $100,000 of

compensation using a fully saturated regression and find large and highly economically significant
discontinuities for FinTech loans for three of the four measures and for the combined measure. The
discontinuities are much smaller for traditional loans.
   28 The shorter form and reduced requirements for loans of $150,000 or below to receive forgive-

ness are outlined at https://www.sba.gov/sites/default/files/2021-07/PPP%20–%20Forgiveness%
20Application%20and%20Instructions%20–%203508S%20%287.30.2021%29-508.pdf.
   29 See the Internet Appendix for details on how the loan size was to be calculated and exclusions.
                      Did FinTech Lenders Facilitate PPP Fraud?                               1799




Figure 5. Discontinuities at $100,000 and rounding. This figure shows the prevalence of the
primary flags across implied compensation and loan amounts. Panel A shows the relation between
implied compensation per employee and being flagged by at least one of the four main flags. Loans
are binned into $2,000-wide bins (i.e., ($0k, $2k], … , ($98k, $100k], … ,($128, $130k]). The left
axis shows the percentage of loans in each bin (bars) and the right axis shows the percentage of
loans in each bin that are flagged (dots). The solid lines are third-degree polynomial fits (weighted
based on the number of loans in each bin), which are separately fitted for loans below $98,000 and
loans above $100,000, and the dashed lines are 95% confidence intervals. Panel B shows the preva-
lence of loans being flagged by at least one of the four main flags by whether the total monthly
implied compensation of a loan is rounded to an interval of $500 (i.e., the loan amount is within
± 50 cents of an interval of $1,250) and lender type. Specifically, the last four digits of the loan
amount are considered (i.e., $123,456.78 → $3,456.78). The top subpanel shows FinTech loans and
the bottom shows traditional loans. Further, the left axis shows the percentage of loans in each
$1-wide bin (bars for rounded compensation are thickened) and the right axis shows the percent-
age of loans that are flagged within each $1-bin for monthly rounded (solid dots) and $250-wide
bin for nonrounded (hollow dots). Additionally, loans with one job reported, loans with implied
compensation within ± $1,000 of $100,000, and second-draw loans to hospitality businesses are
excluded. The solid lines are third-degree polynomial fits for the percentage flagged in the non-
rounded bins and the dashed lines are 95% confidence intervals. (Color figure can be viewed at
wileyonlinelibrary.com)
1800                             The Journal of Finance®

number, especially after including unemployment insurance, employer insur-
ance, and retirement contributions. Thus, rounded loan amounts may suggest
that the loans lack actual documentation and may be more likely to involve
misreporting, as shown in other contexts (Eid, Maltby, and Talavera (2017),
Nigrini (2018)). If the flags we have previously identified reflect misreport-
ing issues, then one might expect both a clustering of loans at round num-
bers and elevated flags at round numbers. However, if round numbers are
simply a result of a borrower with valid documentation rounding numbers
slightly downward to simplify calculations, then one would expect no elevated
reporting issues.
   In Panel B of Figure 5, we first examine the distribution of the last four dig-
its of loan amounts, excluding EIDL refinancing, for FinTech and traditional
loans.30 Both FinTech and traditional loans exhibit rounding at $1,250 incre-
ments, particularly at increments of $2,500 (corresponding to $500 and $1,000
increments of implied monthly payroll). FinTech lenders have moderately more
rounding, with 9.7% of loans rounded to $1,250 increments compared to 7.4%
for traditional lenders.
   The right axis of Panel B of Figure 5 examines the prevalence of the primary
misreporting flags. The percentage of loans with a primary flag is plotted as
a solid dot at the $1,250 loan increments and as a hollow dot at other loan
amounts (shown in $250-wide bins). If rounded loans are more likely to be mis-
reported, one would expect an elevated flag rate at round number thresholds.
For FinTech loans, this is exactly what we observe. At rounded increments,
the flag rates are consistently higher, by 2.14 ppt on average. This difference is
highly significant, which can be seen by comparison to the dotted lines plotting
a 95% confidence interval estimated with a third-degree polynomial estimated
based on the nonrounded loans. For traditional lenders, there is economically
small and statistically weak evidence of elevated flags in some of the rounded
bins. Results for each flag individually show that rounded loans by FinTech
lenders have elevated levels of all four primary misreporting flags (see Panel
B of Figure IA.8 in the Internet Appendix). Overall, the fact that all of the loan
flags are elevated at round loan amounts for FinTech loans provides additional
validation for the suspicious behavior underlying these loans.


C. Loan Overrepresentation
  If there is an organized effort to obtain funds for nonexistent businesses, net-
works of illegitimate borrowers may fill out multiple applications in a similar
manner and could cluster on characteristics such as industry and geography.
Exhibit IA.2 in the Internet Appendix shows an example that appears to fit this

  30 Loan amounts within 50 cents of a $1,250 increment (which corresponds to $500 of implied

monthly payroll) are plotted as thicker and slightly darker bars, with all other loans binned into
$1-wide bins plotted as the thinner, lighter bars. Loans with total implied compensation within
± $1,000 of an interval of $100,000 are excluded to make sure these results are distinct from the
maximum compensation result shown in Panel A of Figure 5.
                      Did FinTech Lenders Facilitate PPP Fraud?                               1801

pattern: a group of 4,299 $20,000 first-draw loans made by Cross River to busi-
nesses in the “Insurance Agencies and Brokerage Industry” in Illinois, mainly
in the Chicago area, almost all of which have one employee (discussed further
in the Internet Appendix). These are followed by examples from 938 $20,000
first-draw loans by Cross River to businesses engaged in “All Other Miscel-
laneous Crop Farming” primarily located in urban areas of Chicago, most of
which have exactly one or eight employees. In addition to having the same
loan amount and similar industries, these $20,000 loans cluster in the last two
weeks of round 2 and in round 3.
   Panel A of Figure 6 plots histograms of FinTech (red bars) and traditional
(gray bars) lender loans by the ratio of first-draw PPP loans to Census CBP
establishments counts in the loan’s industry-county pair. For FinTech lenders,
39.3% of loans exceed industry-county establishment counts, and this occurs
14.0% of the time for traditional lenders. For loans in the right tail, the differ-
ences are even more extreme, with 32.6% of loans exceeding industry-county
establishment counts by a factor of more than two for FinTech lenders and
8.2% for traditional lenders. Even further in the right tail, 7.4% of FinTech
loans exceed industry-county establishment counts by a factor of more than 10
as compared to 0.8% of traditional loans.31 This analysis excludes loans to self-
employed individuals and individual contractors because these business types
are not included in CBP establishment counts. Abnormal levels of loans rel-
ative to CBP establishment counts are largely driven by sole proprietorships,
highlighting potential issues with lending to nonregistered entities.32 Some
excess PPP loans may be due to missing establishments in the CBP data, in-
dustry misclassifications, or other errors in the data. Nonetheless, the large
excess loan rate for FinTech lenders is difficult to explain, particularly since it
is so much higher than traditional lenders.
   Panel A of Figure 6 also plots, for FinTech and traditional lenders separately,
the percentage of loans flagged by one of the four primary suspicious loan flags
by the ratio between PPP first-draw loans and CBP establishments. The flag
rate increases substantially as the loan-to-establishment ratio increases, par-
ticularly for FinTech lenders. While 13.9% of FinTech and 6.7% of traditional
loans in industry-county pairs with a loan-to-establishment ratio at or below
one are flagged, the flag rate is 41.9% for FinTech and 10.7% for traditional
loans when the loan-to-establishment ratio is above two.
   Panel B of Figure 6 plots separate rates for each of the four suspicious loan
flags, with consistent results for all measures. As one moves to ratios above
one, indicating more PPP loans in an industry-county pair than listed in the

   31 Excess loan percentages are calculated by assigning a weight to each loan based on the

inverse of its industry-county’s loan-to-establishment ratio. Specifically, let r be the loan-to-
establishment ratio in the loan’s industry-county pair, the weight is 0 if r ≤ 1 and 1 − 1/r if r
> 1. The interval limits are changed to 2 (10) instead of 1 for the 32.6% (7.4%) and 8.2% (0.8%) fig-
ures.
   32 The majority of loans in overrepresented industry-county pairs are generated by elevated

levels of sole proprietorships, particularly as one reaches increasing levels of abnormal establish-
ments (as shown in Figure IA.10 in the Internet Appendix).
1802                             The Journal of Finance®




Figure 6. Overrepresentation of industries in counties. This figure shows overrepresen-
tation of loans within industry-county pairs. We define the normalized number of loans as the
number of first-draw loans divided by the number of establishments (per the 2019 U.S. Census
County Business Patterns data set) in an industry (represented by NAICS code) and county pair.
Panel A shows the relation between the normalized number of loans and our four main flags com-
bined together as at least one flag and Panel B shows the relation for each flag separately. Since
the CBP does not include self-employed and independent contractors as establishments, we ex-
clude loans to these business types. Note that 6.20% of FinTech and 0.72% of traditional loans
are in industry-county pairs with ratios of at least 10; these loans are represented in Panel A by
the bars and dots at the far right labeled “≥ 10.” In both panels, loans are binned into 0.25-unit-
wide bins. The solid lines are third-degree polynomial fits for the percentage of flagged loans
and the dashed lines are 95% confidence intervals. (Color figure can be viewed at wileyonlineli-
brary.com)




CBP, the number of suspicious loans flagged increases substantially for all of
the suspicious loan measures. This is true for both FinTech and traditional
lenders, but the increase is generally steeper for FinTech lenders, consistent
with FinTech loans in industry-county pairs with a high loan-to-establishment
ratio being particularly suspicious.
                       Did FinTech Lenders Facilitate PPP Fraud?                                  1803

D. Loan Clustering
   In addition to exhibiting geographic and industry clustering, many of
the examples discussed above also feature identical loan amounts and job
numbers. If networks submitting fictitious loan applications repeat the same
application information across multiple loans, lenders may have many loans in
a geographic region with similar industries, loan amounts, or jobs reported.33
There will clearly be some loan similarities by chance and due to lender
specialization, but it is instructive to quantify how frequently loans cluster.
For each lender-county pair with at least 25 loans, we calculate concentration
ratios for the industry, loan amount (rounded to $100), and reported jobs
(excluding one because it is common across all lenders and counties). The
concentration ratios are based on the sum of squared shares of loans with a
characteristic.34 Then, we rescale each of the concentration ratios to have a
median of 1,000 and an interquartile range of 300 so that the three concentra-
tion ratios have similar impacts on the overall concentration measure. Finally,
we average the three concentration ratios for each lender-county pair.
   The bars in Panel A of Figure 7 plot the distribution of scaled concentra-
tion ratios separately for FinTech and traditional loans. High concentration
ratios are much more common for FinTech loans: 88.6% of FinTech loans in
lender-county pairs with a scaled concentration ratio above 1,000, compared
to 20.3% of loans for traditional banks. The dots in Panel A of Figure 7 plot
how the incidence of loans being flagged by at least one of the four primary
suspicious loan flags changes with the concentration ratio. When the scaled
concentration ratio is below 1,000, 10.2% of FinTech loans and 6.6% of tradi-
tional loans are flagged by at least one flag. When the scaled concentration
ratio is above 1,300, this grows to 35.7% for FinTech loans and is similar
(6.7%) for traditional loans. Panel B of Figure 7 shows that similar patterns
hold for each of the four suspicious loan flags individually. The overall pattern
is similar to the previous secondary measures: FinTech lenders have much
higher loan concentration ratios, and high concentration ratios are highly re-
lated to the suspicious loan flags, particularly for FinTech loans. This pattern
is what one would expect if the indicators are picking up misreported Fin-
Tech loans and is difficult to explain with innocent mistakes or errors in the
data.


                                   IV. External Validation
   The primary and secondary measures developed in the previous two sec-
tions allow us to systematically flag individual loans as suspicious. In this

   33 It is not clear why fictitious loan applications frequently repeat the same information, but

the pattern is clear in the examples, and there is no other obvious explanation for this clustering.
   34 For example, let i = 1, 2, . . . , n represent the n industries in a given lender-county pair, then
                         
Concentrationindustry = ni=1 s2i , where si is the percentage of loans in the lender-county that are
in industry i times 100 (e.g., 6.2 for 6.2%). Note that this concentration ratio is the same as a
Herfindahl-Hirschman Index (HHI), which is commonly used to measure market concentration.
1804                                          The Journal of Finance®




Figure 7. Clustering within lenders and counties. This figure shows clustering of loans
within lender-county pairs. We calculate the concentration ratios of industries, loan amount
(rounded to $100), and jobs reported (excluding 1) for first-draw loans in each lender-
county pair, rescale each concentration ratio to a median of 1,000 and IQR (interquartile
range) of 300, and then take the average of the three rescaled concentration ratios. For
example, let i = 1, 2, . . . , n represent the n industries in a given lender-county pair, then
                         
Concentrationindustry = ni=1 s2i , where si is the percentage of loans in the lender-county pair
that are in industry i times 100 (e.g., 6.2 for 6.2%). Then, Rescaled Concentrationindustry =
           Concentrationindustry −Median[Concentrationindustry ]
75th Percentile[Concentrationindustry ]−25th Percentile[Concentrationindustry ]
                                                                                  × 300 + 1, 000. Panel A shows the rela-
tion between the average rescaled concentration ratio and our four main flags combined together
as at least one flag and Panel B shows the relation for each flag separately. In both panels, only
lender-county pairs with at least 25 loans are considered. Note that 2.4% of FinTech loans and
0.5% of traditional loans are outside the range of the average rescaled concentration ratio shown
in Panel A. In both panels, loans are binned into 50-unit-wide bins; in Panel B, bins with fewer
than 100 loans for which the given flag can be determined are excluded. The solid lines are third-
degree polynomial fits for the percentage of flagged loans and the dashed lines are 95% confidence
intervals. (Color figure can be viewed at wileyonlinelibrary.com)
                      Did FinTech Lenders Facilitate PPP Fraud?                               1805

section, we validate these measures based on comparisons with detailed
subsets of data for three external measures of suspicious lending: criminal
records data, crowd-sourced data on PPP fraud, and official SARs filed by
financial institutions.


A. Criminal Records
   Recidivism statistics show that individuals with past criminal histories are
highly likely to commit crimes in the future (Alper, Durose, and Markman
(2018)). The PPP originally prohibited loans to businesses more than 20%
owned by individuals currently subject to criminal charges, incarceration, pro-
bation, or parole or who had been convicted of a felony within the past five
years. These restrictions were relaxed somewhat in June 2020 to permit loans
to businesses owned by individuals facing misdemeanor charges and those
with convictions, probation, or parole for most felonies more than a year in the
past.35 We do not evaluate the design and efficacy of this aspect of the program
or lenders’ compliance with its specific prohibitions. Instead, our analysis uses
criminal records as a potential additional validation and risk factor. To assess
the prevalence of criminal records among PPP borrowers, we collect criminal
histories for a random sample of 150,000 round 1 and 2 loans to individual
names in the PPP data that can be matched to LexisNexis public records data.
   The left subpanel of Panel A of Figure 8 plots the percentage of borrowers
with felony charges on their criminal records between 2000 and 2020 within
the sample of 150,000 individual borrowers for whom we collected background
information.36 Felony charges are present for 4.7% of FinTech borrowers com-
pared to 1.4% of traditional borrowers. There is also a strong relation between
criminal records and both the primary and secondary indicators. A loan that
is flagged by at least one primary flag is 2.2 times as likely to be to a borrower
with a felony charge for FinTech lenders and 1.4 times for traditional lenders.
The EIDL misreporting indicator seems to capture the highest percentage of
felony charges. We confirm that these relations are robust and statistically
significant by regressing an indicator for having a felony charge on the other
primary and secondary risk flags for loans originated by FinTech lenders.37

   35 The five-year criminal record prohibition was only retained for financial crimes such as fraud

and embezzlement. As a result, many individuals with criminal records were legally eligible for
PPP loans.
   36 Ninety-five percent confidence intervals based on standard errors clustered by ZIP Code and

lender are plotted on top of the bars. Panel A of Figure IA.11 in the Internet Appendix shows ro-
bustness to using convictions rather than charges; charges are used for the main analysis since
convictions are more difficult to identify in the data. Panel B of Figure IA.11 in the Internet Ap-
pendix shows how felony rates vary across loans flagged by each flag individually, and Panel C of
Figure IA.11 in the Internet Appendix replicates Panel B using felony charges from 2015 to 2020.
We find significant results for FinTech based on nearly all of the flags, and while the percentage of
borrowers with felonies is lower across the board in Panel C, the relative results remain.
   37 Results are reported in Table IA.VII in the Internet Appendix. The regressions control for

loan size and number of jobs with business type, industry × CBSA, and lender fixed effects. Stan-
dard errors are double-clustered by ZIP Code and lender. In all cases, the coefficients are positive,
1806                             The Journal of Finance®




Figure 8. External validation. This figure shows relations between the external validation
measures and whether the loan is flagged by at least one primary flag. Panel A shows the relation
between the borrower having a felony charge from between 2000 and 2020 on their record and
their loan being flagged. This is based on a sample of 150,000 round 1 and 2 loans to self-employed
individuals, independent contractors, and sole-proprietors. Panel B shows the relation between
the loan being reported on PPP Detective and the loan being flagged. The left panel of Panel C
shows the number of SARs from banks and other financial institutions to FinCEN per month
from 2014 to 2021, and the right subpanel shows the relation between the percentage flagged
(separately for FinTech and traditional borrowers) and the average number of business loan SARs
in 2020 and 2021 per 100,000 population (truncated at the 95th percentile) across counties. In the
left subpanels of Panels A and B, the error bars denote 95% confidence intervals (standard errors
double-clustered by ZIP Code and lender). In the right subpanels of Panels A and B, lenders with
at least 0.2% of the sample and the top 75 lenders by number of loans, respectively, are shown.
In all right subpanels, the lines are linear fits and correlations are shown in the bottom left. In
the right subpanel of Panel C, the dashed lines are 95% confidence intervals. (Color figure can be
viewed at wileyonlinelibrary.com)
                     Did FinTech Lenders Facilitate PPP Fraud?                              1807

   The right subpanel of Panel A of Figure 8 examines how felony charges
vary across lenders with a clear positive relation between the percentage
of a lender’s sampled borrowers with felony charges and the percentage of
its overall loans with at least one of the primary suspicious loan flags. In
particular, the four lenders with the highest felony charge percentages (MBE,
Cross River, Fundbox, and Kabbage, all of which are FinTech) also have the
highest primary flag rates.38

B. PPP Detective
   We utilize data from the PPP Detective, a crowd-sourced platform that al-
lows users to report loans as “potentially being fraudulent.” In total, 147,662
loans have at least one PPP Detective report as of the end of February 2022,
which represents 1.2% of PPP loans overall. Although it is unclear what in-
formation is used to report loans on PPP Detective and the number of loans
reported on PPP Detective is relatively small compared to what our flags iden-
tify, the measure provides an external, crowd-sourced qualitative measure to
compare our more systematic quantitative approach against.
   Panel B of Figure 8 compares PPP Detective to our primary suspicious lend-
ing flags. The left plot first compares primary flag rates for loans with and
without PPP Detective reports. Loans with no PPP Detective reports have a
flag rate of 12.0%, compared to 33.1% for loans with at least one PPP Detective
Report and 40.7% for loans with two or more PPP Detective reports. The plot
also repeats the same comparison separately for FinTech and traditional loans.
In both cases, flag rates are considerably higher for loans with PPP Detective
reports, with particularly large differences for FinTech loans. This pattern in-
dicates that our flags are strongly correlated with PPP Detective at the loan
level, particularly for FinTech loans. The right plot in Panel A assesses cor-
relations across lenders by plotting lender-level flag rates and PPP Detective
report rates. The two measures have a highly significant correlation coefficient
of 0.9275, and the same cluster of FinTech lenders who have high flag rates
also have high PPP Detective report rates, again providing strong validation
for our suspicious loan flags.39 A potential concern with this validation is that
PPP Detective users could flag loans based on measures that are similar to the
flags in this paper, particularly if they read or saw media coverage of early ver-
sions of the paper. This concern is mitigated by the PPP Detective interface’s
statistically significant, and economically large for FinTech loans with almost no relation between
the misreporting indicators and criminal records for traditional loans.
   38 Panel D of Figure IA.11 in the Internet Appendix replicates this figure using felonies post-

2005, post-2010, and post-2015. Although the percentage of borrowers with felony charges de-
creases as the time period is decreased, the relative results remain. Additionally, Panel E of Fig-
ure IA.11 in the Internet Appendix replicates this figure using bankruptcy filings post-2015 and
finds similar results. The rate of felony charges also increases with implied compensation for Fin-
Tech loans but not for traditional loans, adding to the evidence of inflated compensation within
FinTech loans (Panel F of Figure IA.11 in the Internet Appendix).
   39 Table IA.VIII in the Internet Appendix shows that these patterns hold in a fully saturated

regression framework and for each primary flag.
1808                            The Journal of Finance®

focus on flagging individual loans and by the consistent results we find across
flags, some of which would be difficult to replicate (as shown in Figure IA.VIII
in the Internet Appendix). Nevertheless, it is possible that the PPP Detective
estimates are not an entirely independent verification.

C. Suspicious Activity Reports
   We next examine summaries of SARs to the Financial Crimes Enforcement
Network. Panel C of Figure 8 plots SARs over time and across different geo-
graphical areas. The left side plots the number of SARs per month over time
from 2014 to 2021. The dashed green line plots the number of SARs that are
related to mortgage loans, which has been relatively stable over time with be-
tween 3,000 and 4,000 reports per month. By contrast, SARs associated with
business loans were relatively rare until April 2020 and then jump to over
2,500 per month from August to October 2020 with another large jump to over
5,000 per month from May to August 2021. These jumps follow shortly after
the initial launch of the PPP in April 2020 and its further expansion starting in
January 2021, with small lags likely due to delays in reporting. Interestingly,
most of these SARs are from banks as opposed to nonbanks, potentially sug-
gesting that banks were more rigorous in identifying and reporting suspicious
PPP loan applications.40
   On the right side of Panel C, we examine how business loan SARs correlate
geographically with our flag rates at the county level, which is the lowest level
of aggregation for SARs reported by FinCEN. For FinTech loans, the binscat-
ter shows a relation between the geography of SARs and the percentage of
loans flagged, with flag rates of around 15% in counties with low business SAR
rates compared to flag rates in excess of 30% in counties with the highest SAR
rates. By contrast, there is little, if any, geographic correlation for traditional
loans, suggesting that the SAR reports are likely regarding FinTech loans. This
pattern is consistent with many of our other measures and provides another
strong external validation for the flagged FinTech loans.

D. Relation between Primary and Secondary Flags and External Validation
  We have already seen that the primary flags are strongly predictive of one
another, and the evidence in Figures 5 to 8 shows strong relations between
the primary and secondary flags and external validation. In Table III, we
more formally assess these relations with a regression analysis that controls
for loan size and number of jobs reported and includes ZIP Code, business
type, industry × CBSA, and lender fixed effects. The dependent variable is an
   40 We thank Mark Egan for pointing out this pattern in the SARs data. The FinCEN SAR data

do not distinguish between traditional and online banks, so some of the bank SARs could be from
online bank FinTechs. A conversation with a senior employee in the Financial Intelligence Unit
of a major traditional bank indicated that the unit filed many SAR reports with the FinCEN
regarding customers receiving and depositing apparently fraudulent PPP loans initiated through
other lenders.
                       Did FinTech Lenders Facilitate PPP Fraud?                            1809

                                            Table III
                      Secondary Flags and External Validation
In this table, we examine the relation between our four main flags, which we combine to form At
Least One Flag, and the secondary flags. We estimate OLS regressions with At Least One Flag
as the dependent variable and the six secondary/external flags as independent variables. Each
specification also includes an interaction between the secondary/external flag and an indicator for
whether the loan was originated by a FinTech lender. $100k Implied Comp./Receipts is a dummy
variable equal to 1 if the implied compensation/receipts per job is within ± $1,000 of $100,000.
Monthly Rounding is a dummy variable equal to 1 if the loan amount is within ± 50 cents of an
interval of $1,250. Overrepresentation is a dummy variable equal to 1 if the number of first-draw
loans to businesses not listed as self-employed and independent contractors in a loan’s industry-
county pair exceeds the number of establishments in the industry-county pair according to the
U.S. Census CBP data. High Concentration is a dummy variable equal to 1 if the average rescaled
concentration ratio in the loan’s lender-county pair is above the 75th percentile. Felony Post-2000
is a dummy variable equal to 1 if the borrower has a felony charge on their criminal record from
between 2000 and 2020. PPP Detective is a dummy variable equal to 1 if the loan has at least
one report on PPP Detective. For all specifications, loans are filtered to the sets for which we
can determine the secondary flag. Further, for specification (2), one-job loans and loans where
1($100k Implied Comp.) = 1 are excluded. Fixed effects are as indicated at the bottom of each
column. Robust standard errors are double clustered by ZIP Code and lender. t-statistics are in
parentheses. Significance levels are indicated by *p < 0.10, **p < 0.05, ***p < 0.010.

Dep. Variable: 1(At Least One Primary Flag)

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

1($100k Implied Comp.)     0.0192∗∗
                            (2.45)
1(Monthly Rounding)                    0.00274∗∗∗
                                          (4.27)
1(Overrepresentation)                               −0.00609∗∗
                                                     (−1.98)
1(High Concentration)                                            0.0116∗∗∗
                                                                   (3.68)
1(Felony Post-2000)                                                          0.0165∗
                                                                              (1.82)
1(PPP Detective)                                                                        0.0316∗∗∗
                                                                                          (7.89)

× 1(FinTech)               0.123∗∗∗     0.0103∗∗    0.0592∗∗∗    0.0189∗∗∗ 0.0507∗∗∗    0.0557∗∗∗
                            (16.21)      (2.08)       (9.32)       (3.21)    (4.15)       (6.91)

ln(Jobs Reported)            Yes          Yes           Yes         Yes        Yes         Yes
ln(Loan Amount)              Yes          Yes           Yes         Yes        Yes         Yes
ZIP Code FE                  Yes          Yes           Yes         Yes        No          Yes
Business Type FE             Yes          Yes           Yes         Yes        Yes         Yes
NAICS × CBSA FE              Yes          Yes           Yes         Yes        Yes         Yes
Lender FE                    Yes          Yes           Yes         Yes        Yes         Yes

Observations              10,792,833   5,152,691    6,060,817    7,382,631   123,651   10,794,163
Num. Lenders                 4,769       4,687        4,719        4,146      2,557      4,769
R2                           0.310       0.102        0.332        0.333      0.383      0.294
Mean of Dep. Var.            0.128      0.0663        0.137        0.136      0.105      0.128
1810                             The Journal of Finance®

indicator variable for the loan having at least one of the primary flags. The
secondary flags are all interacted with an indicator variable for FinTech loans,
so the direct coefficients represent effects for traditional loans. Five of these
six effects for traditional loans are positive and significant, with magnitudes
ranging from 4.1% to 24.6% of the mean misreporting rate. Further, all of the
interactions between the secondary flags and the indicator for FinTech loans
are large and positive, and all are significant. As a result, all of the secondary
flags strongly relate to the primary flags for FinTech loans, with relations that
are much stronger than for traditional loans. For compensation near $100,000
and rounded compensation, the effects for FinTech loans are 7.41 and 4.76
times as high as those for traditional loans, respectively. For criminal records,
PPP Detective, and high loan concentration, the effects for FinTech loans are
4.07, 2.76, and 2.63 times as large as the traditional loan effects, respectively.
Finally, for industry overrepresentation, there is a strong effect for FinTechs
despite essentially no relation for traditional loans. We also examine relations
between the primary and secondary flags at the lender level (see Figure IA.13
in the Internet Appendix) and find that, except for monthly rounding, lenders
with high levels of each secondary flag tend to be the same lenders who have
high levels of the primary flags.


                   V. How Many PPP Loans Are Suspicious?
  In this section, we quantify ranges of suspicious loans based on the primary
and secondary flags developed in the previous two sections. There are two com-
peting issues: (i) despite all the validation, the flag indicators likely contain
some false positives, but (ii) the indicators only apply to subsets of loans and
are incomplete. We seek to obtain a sense of both of these issues while also
caveating that quantifying the extent of fraud often involves large estimate
ranges.41 Panel A of Figure 9 plots flag rates for each of the four primary flags
along with overall suspicious lending rates. Our primary measure consists of
loans that have at least one primary flag, plotted as the total height of the bars.
By this measure, 1,410,193 loans representing 12.3% of the PPP and totaling
$64.2 billion are suspicious.42 This valuation of suspicious loans includes the
full value of flagged loans, including borrowers who may have been eligible for
smaller loans.43

   41 For example, despite a large and rigorous academic literature and many government inves-

tigations and settlements, there is still no consensus on the exact quantification of mortgage and
other fraud leading up to the financial crisis (Griffin (2021)).
   42 Unincorporated sole proprietorships, self-employed individuals, and independent contractors

account for 838,830 of these flagged loans, representing $16.7 billion. In addition, the EIDL >
PPP flag also provides an indication of misreporting in the EIDL and EIDL Advance program. In
particular, 199,019 EIDL Advances (10.1% of those matched to a PPP loan), totaling $1.71 billion,
have potential misreporting. Importantly, this does not include other types of misreporting in the
EIDL Advance program.
   43 For loans that are flagged by the high implied compensation flag without another primary,

secondary, or external flag, Figure IA.14 in the Internet Appendix considers alternative compensa-
                      Did FinTech Lenders Facilitate PPP Fraud?                               1811




Figure 9. Overall misreporting flag rates. This figure shows the variation in the percentage
of loans flagged. Panel A shows the percentage and dollar amounts of flagged loans overall, by
lender type, and by round. The plain section of each bar represents the percentage of loans flagged
by one primary flag and an additional flag (either another primary flag, secondary flag, or exter-
nal validation) and the entire bar (plain and stripped sections combined) represents loans flagged
by at least one primary flag. The set of numbers to the left of each bar represent the number of
loans and dollar value of loans flagged by one primary flag and an additional flag and the set on
top of each bar by at least one primary flag. The markers within each bar represent the percentage
of loans flagged by each of the primary flags (unconditional of whether a flag can be determined
for a given loan). Panel B shows the dollar value of suspicious lending using alternative measures.
The leftmost bar is the dollar value flagged by at least one primary flag, the second bar is the
incremental dollar value of suspicious lending when the high compensation flag is changed from
3 to 2.5 times the industry-CBSA average compensation/receipts, the third bar is the incremental
dollar value of suspicious lending based on excess loans when the loans previously flagged are
considered excess first and the dollar value of any additional excess is the average of the flagged
loans in the industry-county, and the fourth bar is sum of the first three bars. (Color figure can be
viewed at wileyonlinelibrary.com)
1812                              The Journal of Finance®

   FinTech lenders are responsible for a disproportionate share of suspicious
loans. FinTech lenders originated 858,820 suspicious loans totaling $19.4 bil-
lion. This means that FinTech lenders originated 60.9% of flagged loans, sub-
stantially outpacing their 32.5% loan market share.44 As a share of loans orig-
inated by each lender type, 7.1% of traditional loans have at least one of the
primary suspicious loan flags compared to 23.0% for Fintech loans.
   As a more conservative estimate, we consider loans that have at least one
primary flag plus an additional primary flag, secondary flag, or external vali-
dation, plotted as the solid part of the bars in Panel A of Figure 9. Although
this measure almost certainly misses considerable misreporting, it has the
benefit of dropping sincere mistakes or errors in the data that are isolated
to a single measure. Under this more conservative measure, 949,999 loans to-
taling $32.4 billion are suspicious. Of these loans, 720,870 ($15.8 billion) are
FinTech. This is an even larger FinTech share than for the primary measure
because 83.9% of FinTech loans with a primary flag are further confirmed by
an additional flag while the corresponding figure is only 41.6% for traditional
loans. The higher confirmation rate for FinTech loans is consistent with flagged
FinTech loans being far more likely to be fraudulent as opposed to simply
reflecting honest explanations or errors in the data. Lender-level differences
with the conservative measure (see Panel A of Figure IA.15 in the Internet
Appendix) are also even more pronounced than with the baseline measure. In
particular, similar lenders cluster at the top of the graph, and the five Fin-
Tech lenders with the highest percentages of loans flagged by at least one flag
have flag rates of 29.4%, 29.8%, 28.1%, 25.0%, and 23.5%, respectively, under
this more conservative measure. Prestamos is particularly striking because
it is the third largest lender overall, with 444,789 loans. Cross River (largest
FinTech lender and second largest overall lender, with 478,866 loans) and Har-
vest (third largest FinTech lender and fifth largest overall lender, with 408,450
loans) are also well above the average flag rate with primary flag rates of 20.0%
and 27.6%, respectively.
   The last three bars of Panel A plot suspicious lending rates by rounds of the
program with the clear pattern that suspicious lending increased over time. In
round 1, 6.3% are suspicious, compared to 7.8% in round 2 and 16.3% in round
3. Differences across rounds are even more pronounced with the conservative
measure, which is 2.6% in round 1, 4.1% in round 2, and 12.1% in round 3.
   Although some of the loans flagged as suspicious by the primary measures
may be sincere mistakes or errors in the data, the four primary measures
also surely miss many fraudulent loans. For example, because the high im-
plied compensation flag only flags loans with implied compensation that is

tion thresholds and calculates dollar valuations using only loan value attributed to compensation
above the flag thresholds. Using this approach decreases the overall suspicious lending estimate
from $64.2 billion to $52.1 billion.
   44 Additionally, FinTech lenders originated 30.2% of the dollar volume of flagged loans, substan-

tially outpacing their 10.9% dollar market share. FinTech represents a larger share of suspicious
loans than suspicious loan dollar volume because FinTech loans tend to be smaller. The same
pattern is reflected in FinTech market share.
                     Did FinTech Lenders Facilitate PPP Fraud?                              1813

over three times the mean compensation or receipts in the loan’s industry
and CBSA and the PPP has a maximum eligible compensation of $100,000,
the indicator is only applicable in industry-counties with mean annual com-
pensation/receipts of $33,333 or less. These industries and counties represent
28.9% of PPP loans, suggesting that there is probably substantial exaggerated
compensation in other loans that we cannot detect. Similarly, the business
registry and EIDL > PPP jobs flags apply to 48.3% and 24.1% of loans, re-
spectively. These limitations are particularly important for estimating FinTech
suspicious lending because FinTech loans were less common for corporate/LLC
borrowers and borrowers with matching EIDL Advances.45 In Panel B of Fig-
ure 9, we evaluate potential undercounting of suspicious lending with sensitiv-
ity analysis to adding other suspicious loans that we might be missing. We first
consider changing the threshold for high implied compensation from 3 times
to 2.5 times the mean compensation in the loan’s industry-CBSA pair. This en-
ables us to potentially detect inflated compensation in another 7.3% of loans
that are in industries and counties with mean compensation between $33,333
and $40,000. This change flags another 265,849 loans, totaling $18.7 billion,
as suspicious.
   Another indication of potential undercounting is the magnitude of excess
loans beyond the number of business entities in an industry-county pair
as per counts from the U.S. Census CBP data. Unlike the primary suspicious
loan flags, this measure does not directly identify individual loans. Its mag-
nitude is also inherently limited because loans can be suspicious even if they
do not exceed Census establishment counts. The excess loan measure requires
us to exclude loans to self-employed individuals and independent contractors
because they are not included in Census establishment counts, and we also
omit second-draw loans to prevent double counting. Due to these restrictions,
the excess loan measure only applies to 54.6% of loans, representing 65.0% of
total dollars lent. Nonetheless, excess lending is independent of our primary
measures and can be used as a separate quantification of the total amount of
suspicious loans even though it is also likely an underestimate. In the third bar
of Panel B of Figure 9, we calculate how many additional loans would have to
be suspicious in each industry-county pair to explain the magnitude of excess
lending. For example, if an industry-county pair has 100 excess loans (e.g., 200
PPP first-draw business loans compared to 100 Census business entities), and
our primary measures flag 75 loans as suspicious in that industry-county, we
infer that at least 25 other loans are also suspicious. To estimate the size of
these excess loans, we use the average loan amount of the flagged loans in the
industry-county pair.46 When subtracting flagged loans, we use the 2.5 times

  45 The business registry flag only applies to corporate/LLC loans and the EIDL > PPP jobs flag

only applies to loans with matched EIDL Advances. The business registry flag is applicable to
18.6% of FinTech loans (compared to 62.5% of traditional loans), and the EIDL > PPP jobs flag is
applicable to 18.2% of FinTech loans (compared to 26.9% of traditional loans). The high implied
compensation flag is more applicable to FinTech loans than traditional loans, but the difference is
moderate (34.4% of FinTech loans are applicable, compared to 26.2% of traditional loans).
  46 In industry-counties where there are no flagged loans, we use the average of all loans.
1814                       The Journal of Finance®

mean compensation version of the high compensation flag so that the excess
loan measure is additive with the additional loans identified with the 2.5 times
version of the high compensation flag. Based on this process, we estimate that
excess loans represent another $34.4 billion of suspicious lending. Combined
with the 2.5 times version of the high compensation flag, this results in a total
suspicious lending estimate of $117.3 billion. It is worth noting that this sen-
sitivity analysis is along only two limited dimensions that do not apply to all
loans and thus likely still misses significant suspicious lending.


       VI. Why Might Suspicious Lending Concentrate in FinTech?
   FinTech lenders on average have much higher suspicious lending rates than
traditional lenders, and Tables II, IA.1, and IA.IX show that their elevated
suspicious lending is not explained by observable facets of loan composition.
Did FinTech lending improve over time? What could be driving the elevated
flag rates for FinTech lenders? And what are the consequences in terms of loan
forgiveness and enforcement?


A. Did FinTech Lenders Improve Standards over Time?
  In the early stages of the COVID pandemic, getting money out quickly was
a primary focus, and it is possible that FinTech lenders making SBA loans for
the first time may have lacked some of the capabilities necessary to review
loan applications for signs of fraud. As the program progressed, the urgency of
processing loans quickly likely diminished, and lenders and the SBA had more
time to develop fraud-detection programs. Additionally, the Office of Inspector
General for the SBA issued a report in October 2020 warning that fraud may
be widespread in the PPP. All else equal, if lenders and the SBA improved
standards over time, we should expect fraud to decrease. On the other hand,
people exploiting the system may have also learned over time.
  To assess these competing forces, we consider two potential scenarios under
which suspicious lending could arise:

  r Scenario A: The lender does not want to facilitate fictitious loans but is not
    performing great due diligence. As it learns over time, the lender cracks
    down on the fraud.
  r Scenario B: The lender is aware of the existence of or potential for fraud
    within its PPP loans but ignores this risk because there is little down-
    side for the lender. This may be particularly true for lenders with little
    reputation or other business to protect.

  Under scenario A, when lenders are new to PPP lending, they may facilitate
questionable loans, but over time as they experience more loans with improba-
ble features, they should originate fewer of these loans. In this case, borrowers
who wish to commit loan fraud would need to rotate among lenders. In scenario
                     Did FinTech Lenders Facilitate PPP Fraud?                       1815

B, the amount of suspicious lending could grow over time as lenders develop a
reputation for rapid and unquestioning approval.
  Scenario A predicts:

    (i) Loan misreporting will decrease over time as lenders become more
        aware and develop systems to screen out suspicious loans.
   (ii) Suspicious first-draw borrowers will be less likely to receive a second-
        draw loan from the same lender than nonsuspicious borrowers.
  (iii) Regions with high misreporting in rounds 1 and 2 will face extra
        scrutiny from lenders, which will decrease misreporting in round 3.

Scenario B predicts:

    (i) Loan misreporting will grow over time as borrowers learn about the
        potential for fraud.
   (ii) Borrowers with suspicious first-draw loan will be able to obtain second-
        draw loans from the same lender.
  (iii) Regions with high misreporting in rounds 1 and 2 will have the same or
        more misreporting in round 3.

   Did lenders improve their loan screening over time? Although we do not
observe denied applications or specific lender practices, we can observe how
loans that were approved and funded changed over time. We have already
seen that the overall rate of suspicious lending grew over time from round
1 to round 3. Panel A of Figure 10 plots more granular suspicious loan rates
on a weekly basis separately for FinTech and traditional lenders. For the Fin-
Tech lenders, loans became more suspicious over time throughout rounds 1 and
2. The rate of suspicious lending dropped at the beginning of round 3, likely
due to pent-up demand for second-draw loans from legitimate borrowers. Most
round 3 FinTech lending occurred later in round 3 (see Figure 1), and as round
3 progressed, the suspicious loan rate rose dramatically to around 30% of loans
flagged as suspicious during April and early May of 2021. PPP funds for most
loans were exhausted on May 4, 2021 (Cowley (2021)). The suspicious loan
rate fell toward the end of May 2021, but this could be due to loan composition
since funding after May 4 was only available for prioritized community finan-
cial institutions and some loans that were already under review prior to May
4. Suspicious lending by traditional lenders also grew over time, but at a much
lower rate. FinTech and traditional lenders both started the PPP with suspi-
cious loan rates of less than 10%, but by the end of the program the FinTech
suspicious loan rate was close to 30%, more consistent with scenario B.47
   We also examine lending growth and changes in suspicious loan rates across
rounds at the lender level. Most FinTech lenders had higher suspicious lend-
ing rates in round 3 than in rounds 1 and 2 (Panel A of Figure IA.16 in the

  47 Panel B of Figure IA.15 in the Internet Appendix shows similar trends for each primary

flag individually.
1816                              The Journal of Finance®




Figure 10. Persistence and growth across rounds. This figure shows the persistence and
growth of flagged loans across lending rounds. Panel A shows this by lender type and Panel B by
ZIP Code. In Panel A, each subpanel shows a lender type and each series is the percentage of loans
flagged by the given measure across time. In Panel B, the percentage of loans flagged in rounds 1
and 2 is shown on the horizontal axis and in round 3 on the vertical axis. For Panel A, the vertical
dotted lines split each subpanel into the three lending rounds. The solid lines are loans flagged by
at least one primary flag and the dashed line is loans flagged by at least one primary flag and an
additional flag (another primary, secondary, or external). For Panel B, the left subpanel uses all
loans, the middle uses FinTech loans, and the right uses traditional loans. ZIP Codes with at least
100 loans in rounds 1 and 2 combined are shown. The black line is a 45◦ line. (Color figure can be
viewed at wileyonlinelibrary.com)




Internet Appendix). Additionally, many of the FinTech lenders with the high-
est suspicious loan rates in rounds 1 and 2 also had the most growth in lend-
ing and the most growth in suspicious loans in round 3. In Table IV, we ask
whether lenders appear to be learning by regressing indicators for the four
primary flags in round 3, individually and combined, on lenders’ rounds 1 and
2 misreporting rates for the same flags. For FinTech lenders, we find large
and statistically significant positive relations across the board with largely
                       Did FinTech Lenders Facilitate PPP Fraud?                                   1817

                                               Table IV
                 Persistence of Lender Behavior across Rounds
In this table, we examine the persistence of lender behavior across rounds. We estimate OLS
regressions with dummies for whether each round 3 loan is flagged by our four main flags individ-
ually (specifications (1) to (4)) and at least one of them (specification (5)) as the dependent variable
and the percentage of the lender’s loans were flagged by the same flag in rounds 1 and 2 as the
independent variable. Interactions with whether the loan was originated by a FinTech or tradi-
tional lender are included in all specifications. For specifications (1) to (4), loans are filtered to the
sets for which we can determine the flag. Further, to ensure that we have accurate measures of
past behavior, we require that each lender have at least 100 loans in rounds 1 and 2 (combined)
for which we can determine the given flag. Fixed effects are as indicated at the bottom of each
column. Robust standard errors are double clustered by ZIP Code and lender. t-statistics are in
parentheses. Significance levels are indicated by **p < 0.05, ***p < 0.010.

                               (1)             (2)              (3)               (4)            (5)
Dep. Variable:              1(Business     1(Multiple     1(High Implied       1(EIDL >       1(At Least
                             Registry)       Loans)           Comp.)           PPP Jobs)      One Flag)

Past Pct. This Flag
  × 1(FinTech)                0.905∗∗∗       0.627∗∗∗           0.361***          0.861∗∗∗       0.456∗∗∗
                             (4.67)         (3.69)             (2.72)            (2.97)         (4.50)
  × 1(Traditional)            0.167          0.211∗∗          −0.453***         −0.173         −0.185
                             (0.73)         (2.20)           (−3.83)           (−1.26)        (−1.55)

ln(Jobs Reported)               Yes            Yes               Yes               Yes            Yes
ln(Loan Amount)                 Yes            Yes               Yes               Yes            Yes
ZIP Code FE                     Yes            Yes               Yes               Yes            Yes
Business Type FE                Yes            Yes               Yes               Yes            Yes
NAICS × CBSA FE                 Yes            Yes               Yes               Yes            Yes

Observations                1,830,109       5,280,796        1,574,978         1,010,713       5,280,796
Num. Lenders                  2,432           3,078            1,547             1,409           3,078
R2                           0.141           0.0512            0.630             0.346           0.366
Mean of Dep. Variable        0.0705          0.0208            0.341             0.135           0.164



insignificant and several negative relations for traditional lenders. For tradi-
tional lenders, there is overall no relation between their flag rates in rounds 1
and 2 and the likelihood of their loans being flagged in round 3. By contrast,
FinTech lenders have persistent and increasing levels of suspicious loans over
time, consistent with scenario B above.
   To assess prediction 2, we estimate regressions to test whether a first-draw
borrower is more or less likely to receive a second-draw loan from the same
lender if their first-draw loan is flagged by at least one of the primary mis-
reporting indicators. Table V shows that borrowers of traditional loans that
are flagged in the first two rounds have a statistically significant decrease in
the probability of receiving a second-draw loan from the same lender of 3.09
ppt (with t-statistics of 11.41) and FinTechs have a smaller and statistically
insignificant decrease of 1.02 ppt (with t-statistics of −0.84).48 This indicates

  48 These results are based on SameLender
                                                 i being set to 0 if the borrower did not get a second
draw at all.
1818                               The Journal of Finance®

                                              Table V
                 Likelihood of Receiving a Second-Draw Loan
In this table, we examine whether lenders were more/less likely to provide a second-draw loan
to a borrower whose first-draw loan is flagged by at least one of our primary flags. We estimate
OLS regressions with a dummy for whether the same lender provided the first- and second-draw
loans as the dependent variable and a dummy for whether the first-draw loan was flagged by at
least one of the primary flags as the independent variable. In specifications (1) and (2), if a bor-
rower did not receive a second-draw loan, the dependent variable is set to 0, and in specifications
(3) and (4), only borrowers who received both a first- and second-draw loans are included in the
sample. In the even specifications, interactions for whether the first-draw loan was originated by
a FinTech or a traditional lender are included. Fixed effects are as indicated at the bottom of each
column. Robust standard errors are double clustered by ZIP Code and lender. t-Statistics are in
parentheses. Significance levels are indicated by *p < 0.10, ***p < 0.010.

Dep. Variable: 1(First and Second Draw by Same Lender)

                                   (1)                (2)                 (3)                (4)
                                 Unconditional of Receiving              Conditional on Receiving
                                       Second Draw                            Second Draw

1(First Draw Flagged)          −0.0261***                              −0.00878*
                                (−6.54)                                 (−1.84)
  × 1(FinTech)                                      −0.0102                                 0.0122
                                                     (−0.84)                                (0.55)
  × 1(Traditional)                                 −0.0309***                             −0.0156***
                                                    (−11.41)                               (−4.86)

ln(Jobs Reported)                  Yes                 Yes                Yes                 Yes
ln(Loan Amount)                    Yes                 Yes                Yes                 Yes
Zip Code FE                        Yes                 Yes                Yes                 Yes
Business Type FE                   Yes                 Yes                Yes                 Yes
NAICS × CBSA FE                    Yes                 Yes                Yes                 Yes
Lender FE                          Yes                 Yes                Yes                 Yes

Observations                    4,847,988           4,847,988          1,579,897           1,579,897
Num. Lenders                      4,647               4,647              4,446               4,446
R2                                0.121               0.121              0.415               0.415
Mean of Dep. Var.                 0.278               0.278              0.836               0.836



that traditional banks were less likely to continue lending to borrowers with
previous suspicious borrowing, but FinTechs are not actively learning in a sim-
ilar manner. Columns (3) and (4) of Table V condition on the borrower receiving
a second draw (either from the same or different lender) with similar results.49
   To assess prediction 3, we examine whether areas with high misreporting
in rounds 1 and 2 had higher or lower misreporting in round 3. Panel B of
Figure 10 plots ZIP Code level binscatters of the percentage of loans flagged
in rounds 1 and 2 on the horizontal axis and the percentage of flagged loans

  49 In Figure IA.17 in the Internet Appendix, we show results separately for individual lenders

with lender fixed effects and lender interactions. The inclusion of the lender fixed effects ensures
that the reported coefficient is due solely to differences in the lender’s behavior toward flagged and
nonflagged loans rather than systematic changes in the lender’s behavior.
                      Did FinTech Lenders Facilitate PPP Fraud?                              1819

in round 3 on the vertical axis. The left subpanel uses all loans, the middle
uses FinTech loans, and the right uses traditional loans. The figure displays
three interesting findings. First, all of the binscatter dots in the left plot are
above the 45◦ line, indicating that misreporting rates increased in round 3
almost everywhere. Second, the ZIP Codes with the highest suspicious loan
rates in rounds 1 and 2 also had the highest suspicious loan rates in round
3. Third, growth in potential misreporting rates was particularly pronounced
for FinTech lenders. Instead of scrutinizing areas with high suspicious lending
in rounds 1 and 2, FinTech lenders facilitated even more suspicious lending
in these areas in round 3.50 We also test these results at the ZIP Code-lender
level with ZIP Code and lender fixed effects and find that a 10-ppt increase in
flagged loans in a ZIP Code-lender pair in rounds 1 and 2 is associated with
a 5.2-ppt increase in round 3 flagged loans for FinTech lenders, with a much
weaker relation for traditional lenders and similar results at the county-lender
level (see Table IA.X in the Internet Appendix).


B. FinTech Lender Background, Fluidity, and Incentives
  The largest FinTech lender, Cross River, is a small community bank in New
Jersey that acts as a conduit for partner FinTechs. The second largest FinTech
PPP lender, Prestamos, is a Community Development Financial Institution
with locations in Arizona, Nevada, and New Mexico. Harvest Small Business
Finance, Capital Plus Financial, Benworth, and Fountainhead, the other Fin-
Tech lenders in the top 10 by number of PPP loans originated, follow a similar
pattern of limited business outside of PPP lending (see Table IA.XI in the In-
ternet Appendix for background information on the largest FinTech lenders).51
We systematically examine this relation more formally and find that lenders
who have fewer SBA loans pre-pandemic, have lent in SBA programs for
fewer years, and for whom the PPP was their first experience with SBA lend-
ing (in particular, new FinTechs) all have higher rates of flagged loans (as
shown in Table IA.XII in the Internet Appendix). FinTech lenders also relied
more heavily on liquidity support from the Federal Reserve than traditional
lenders.52



    50 Results are similar at the county and state levels. See Panels C and D of Figure IA.16 in the

Internet Appendix.
    51 MBE Capital Partners, the fourteenth largest FinTech lender with the fifth highest suspicious

loan rate, is another interesting example. MBE scaled up to fund PPP loans with a combination
of financing from an insurance company affiliated with Magic Johnson and loans from the Federal
Reserve. Its CEO is now being criminally charged for fraud in MBE’s application to become an
approved PPP lender (Omeokwe (2022)).
    52 Financing for FinTech PPP loans was in part provided with a credit facility, the Paycheck

Protection Program Liquidity Facility (PPPLF), in which the Federal Reserve extended credit to
lenders using PPP loans as collateral. While most traditional lenders did not use the PPPLF at
all, it was a major source of funding for some of the largest FinTech lenders (Figure IA.18 in the
Internet Appendix).
1820                            The Journal of Finance®

                                  B.1. FinTech Fluidity
   The six largest FinTech lenders primarily originated loans that were sourced
from FinTech platforms. Cross River adopted this business model early in
round 1 by partnering with other FinTechs such as Intuit and Kabbage to orig-
inate PPP loans (Cowley (2020)). The other large FinTech lenders originated
loans sourced by two start-ups that did not do any PPP lending until round 3,
Womply and BlueAcorn (Cowley and Koeze (2021)).53
   To further understand the network of relationships between lenders we first
consider the use of multiple lenders for loans in the same draw to the same
noncommercial address as identified by the multiple loan flag. For FinTech
borrowers with multiple loans to the same address, 60.2% of loans are split
across multiple lenders, with connections between lenders seemingly explained
by FinTech portals sourcing loans for multiple lenders (as shown in Panel A,
Figure IA.19 in the Internet Appendix). For example, there is a strong rela-
tionship between Prestamos and Capital Plus, the two lenders who partnered
with BlueAcorn. There are also strong relationships between Harvest, Ben-
worth, Capital Plus, and Fountainhead, all of which are Womply partners. By
contrast, 82.7% of traditional borrowers who took out multiple loans received
all their loans from the same lender.
   For a borrower that already received a first-draw loan, obtaining a second-
draw loan from the same lender only required refreshing the application with
some additional information. This provided a strong incentive for borrowers to
use the same lender. Nonetheless, there are large movements between FinTech
lenders that likely reflect online lending portals switching lenders (Panel B,
Figure IA.19 in the Internet Appendix), which again highlights the fluid nature
of the FinTech space. The lack of relationship banking within FinTech may be
advantageous for expanding access to capital (Erel and Liebersohn (2022)), but
it may also increase dubious lending.


                                 B.2. FinTech Revenue
  PPP lending had the potential to be a profitable business for lenders.
Lenders were initially compensated with processing fees of 5% for loans up
to $350,000, 3% for loans between $350,000 and $2,000,000, and 1% for loans
of $2,000,000 or more. For loans made in 2021, fees for small loans were in-
creased to the lesser of 50% or $2,500 for loans below $50,000.54 Based on this
fee schedule, we estimate that PPP lending generated $38.0 billion of lender
processing fees, $8.6 billion of which went to FinTech lenders (see Table IA.XIII

   53 Womply is a marketing technology firm with no lending history before participating in

the PPP. It launched a platform called Fast Lane to facilitate PPP applications that were then
originated by partner lenders including Harvest, Capital Plus, Benworth, and Fountainhead.
BlueAcorn was founded in April 2020 exclusively to source PPP loans in partnership with Cap-
ital Plus and Prestamos.
   54 See fee schedule at https://home.treasury.gov/system/files/136/Updated-Guidance-PPP-

Lender-Processing-Fee-Payment-1502-Reporting-Process.pdf.
                     Did FinTech Lenders Facilitate PPP Fraud?                             1821




Figure 11. Enforcement and forgiveness. This figure shows the forgiveness of and DOJ en-
forcement actions regarding PPP loans. The left subpanel shows the percentage of round 1 loans
that have been completely forgiven as of January 3, 2022. The right subpanel shows the percent-
age of round 1 and 2 loans that are part of DOJ enforcement actions as of March 1, 2022. In both
panels, the overall percentage and the percentage by lender type and whether the loan is flagged
by at least one primary measure are shown. (Color figure can be viewed at wileyonlinelibrary.com)


in the Internet Appendix). The top four FinTech lenders alone likely generated
$4.03 billion in processing fees, including $1.06 billion to Prestamos, $1.03 bil-
lion to Cross River, $1.02 billion to Harvest, and $938 million to Capital Plus.
The average processing fee for FinTech PPP loans was 18.5% of the loan bal-
ance, largely driven by the high processing fees for small loans in round 3.55 We
lack data on cost structure and lender fee sharing with partner organizations
used to source the loans such as Womply and BlueAcorn.56

C. Forgiveness and Enforcement Actions
  PPP loans were designed to be fully forgiven in most cases. The left plot of
Figure 11 shows the percent of round 1 loans that have been fully forgiven as
of January 2, 2022. Approximately 90% of loans have already been forgiven,
even those that are flagged as suspicious. Loans in rounds 2 and 3 are also
on track for high levels of forgiveness (Figure IA.20 in the Internet Appendix).
These high forgiveness levels make it doubtful that most dubious loans will
ever be paid back.
   55 The average FinTech processing fee for rounds 1 and 2 was 4.97%. This dramatically in-

creased to 22.1% in round 3.
   56 Crossroads Systems, the owner of Capital Plus (the fourth largest FinTech lender and sixth

largest lender overall), reported $932.7 million of revenue in 2021, $897.8 million of which was
due to PPP lending fees, compared to $36.6 million of total revenue in 2020, before Capital Plus
starting originating PPP loans on a widespread basis (see press release at https://www.prnew-
swire.com/news-releases/crossroads-systems-reports-fiscal-fourth-quarter-and-fiscal-year-2021-
financial-results-301443835). Earlier in the pandemic, Capital Plus received a PPP loan of
$376,800, reportedly to cover payroll for its 28 employees. Similarly, Benworth Capital Partners
(the fifth largest FinTech and seventh largest lender overall) received a PPP loan of $100,600
for its 13 employees on April 5, 2020; DreamSpring received a PPP loan of $757,753 for its
54 employees.
1822                               The Journal of Finance®

   The economics of crime depend crucially on a crime’s expected penalty and
probability of detection (Becker (1968)). The U.S. Department of Justice (DOJ)
is pursuing criminal complaints alleging PPP fraud, and some borrowers have
voluntarily repaid their loans without applying for loan forgiveness or had
their loan canceled. However, the magnitude of these enforcement actions is
tiny. Compared to the 1.41 million loans we identify as suspicious, as of March
1, 2022, the DOJ has publicized 258 criminal complaints regarding only 502
loans. The right plot in Figure 11 shows the percent of loans with DOJ enforce-
ment actions by lender type and whether or not the loan is flagged as suspi-
cious by at least one of the four primary measures. Consistent with our flags
identifying fraud, enforcement actions are more common for FinTech loans and
loans that have been flagged, but the enforcement rate is well under 0.1% in
all cases.57
   An earlier version of this paper based on the May 3, 2021 SBA data re-
lease was strongly criticized by several PPP lenders for including loans that
were in the SBA data but were eventually canceled. Lenders indicated that
many loans were approved by the SBA but later denied due to their internal
fraud detection process. It is important to note that the current version of the
paper is entirely based on the January 2, 2022 SBA data release and thus
does not include these earlier loans that were approved but later canceled.58
Overall, we detect 234,727 canceled loans between the May 3, 2021 and Jan-
uary 3, 2022 data releases. Of these canceled loans, 30.7% are flagged by at
least one primary suspicious loan measure. This again implies that the true
amount of fraud is likely substantially greater than our indicators can iden-
tify. Overall, enforcement action, repayment, and cancellation rates are all rare
but somewhat elevated for flagged loans (see Table IA.XIV and Figure IA.21
in the Internet Appendix). Although more enforcement actions may be forth-
coming, there appears to be little penalty for most suspicious lending thus
far.


D. Economic Discussion
   What have we learned that may speak to why suspicious lending concen-
trated so heavily in FinTech lenders? There were clearly large financial incen-
tives for loan origination, and many FinTech lenders with small pre-pandemic
operations generated substantial processing fees. There seems to have been

   57 See the Internet Appendix for additional details on enforcement actions, repayments, and

cancelations. Of the DOJ enforcement action loans with enough data to be matched to the PPP
loan-level data, 216 loans were originated by FinTech lenders and 185 were originated by tradi-
tional banks. There are likely other cases that are still sealed, are in early stages of investigation,
or are not included on the DOJ website for other reasons. We focus on loans from rounds 1 and 2
for this analysis to allow more time for repayments and enforcement actions. SBA data indicate
that only 16,930 round 1 and 2 loans were repaid between December 1, 2020 and June 30, 2021.
   58 Interestingly, dropping the canceled loans led to only minuscule reductions in suspicious loan

rates, even for the lenders who raised the issue as a major concern.
                     Did FinTech Lenders Facilitate PPP Fraud?                              1823

little cost for bad lending practices since all credit risk was borne by the gov-
ernment, and financial penalties for lenders with poor lending practices, at
least currently, seem very low. Although most FinTech lenders have high sus-
picious loan rates, Square and Intuit have among the lowest suspicious loan
rates of all lenders, indicating that online lending does not appear to be the
problem in and of itself. Square and Intuit had reputational capital to protect
and established relationships with customers. Since Square, Intuit, and tradi-
tional lenders likely had access to additional information from applicants (e.g.,
customer payments, payroll, and bank account details), the lack of customer
relationships may have also made other FinTechs more attractive targets for
fraudulent applications.
   The fact that new FinTechs and lenders who have fewer SBA loans pre-
pandemic and have lent in SBA programs for fewer years all have higher rates
of suspicious lending points to poor lending practices playing a larger role when
there is a lack of reputational capital at stake. Traditional lenders may have
followed standard, pre-pandemic SBA lending standards, in which the govern-
ment only partially reimbursed lenders for losses. The one-time nature of the
PPP with immediate profits in the present and a low likelihood of future repeat
business also removed a natural market disciplinary force. Interestingly, in the
buildup to the mortgage crisis, considerable fraudulent origination was simi-
larly concentrated in small originators with little prior histories (Piskorski,
Seru, and Witkin (2015), Griffin and Maturana (2016)).59
   The fact that suspicious lending through FinTechs substantially increased
from 2020 to 2021 suggests that poor practices cannot be solely attributed
to the urgency of distributing funds or the lack of existing lending processes
at the onset of the PPP. Although it may not be possible to precisely pin
down which factors led to poor FinTech lending practices, the growing mag-
nitudes of fraud throughout the program highlight that the lack of immediate
ramifications for borrowers or lenders encouraged growth in fraudulent
practices. Policymakers should carefully consider the effects of government
programs that remove traditional market-based disciplinary forces and rely
solely on discipline from reputational concerns and potential future law
enforcement.




   59 There is some evidence that ZIP Codes with mortgage fraud preceding the financial crisis also

have higher levels of PPP fraud, though the relation is weaker with the inclusion of demographic
controls and county fixed effects (as shown in Table IA.XV in the Internet Appendix). Interest-
ingly, in the 2003 to 2007 period, the largest and most reputable banks created some of the worst
performing structured products (mortgage-backed securities and collateralized debt obligations),
indicating that the financial incentives for short-term profits were seemingly large enough that
even these banks were willing to tarnish their reputation (Griffin, Lowery, and Saretto (2014)).
With the PPP, the fact that traditional banks had low levels of suspicious lending better fits the
traditional reputation model (Chemmanur and Fulghieri (1994)).
1824                          The Journal of Finance®

                                  VII. Conclusion
   We examine four primary and four secondary measures related to poten-
tially misreported loans with further validation from three additional external
measures. FinTech loans are highly suspicious at a rate of over six times that
of traditional loans. The top 12 lenders with the highest rates of suspicious
loans are all FinTech lenders. Our analysis of the four primary suspicious loan
indicators estimates the total amount of potential misreporting as 1.41 million
loans with a balance of $64.2 billion, and these flags only apply to subsets of
loans. Supplemental analysis indicates that likely misreporting may be twice
as large, and suspicious lending more than quadrupled from the early to later
stages of the PPP. Round 1 suspicious loans have already been overwhelm-
ingly forgiven at a rate similar to nonsuspicious loans, and extremely few have
been prosecuted.
   Our findings have important policy implications. First, the PPP did not in-
clude robust verification requirements, and traditional banks may have been
more apt to follow standard lending practices than new FinTech PPP lenders.
The lack of rigorous verification for PPP loans seems to have led to substan-
tial costs to taxpayers, especially in 2021 when there was likely also less ur-
gency to the loans. Second, FinTech lending, though found in other papers to
be successful at adapting to new environments and quickly disbursing funds,
seemingly needs to improve due diligence practices. Two established FinTech
lenders persistently have low rates of misreporting, indicating that FinTech
lending need not be substandard. Third, our findings, along with evidence that
the PPP saved relatively few jobs at a high cost (Autor et al. (2022), Chetty
et al. (2022), Granja et al. (2022)), provide growing evidence that the PPP may
not have been an efficient source of capital allocation. Fourth, incentives in the
PPP appear misaligned in that FinTech lenders with widespread indicators
of misreporting made billions of dollars dispersing loans with apparently lax
oversight procedures.
   Finally, the increasing scale of misreporting through time indicates that
current penalty and enforcement systems are not effective. If the system is
not changed for future programs, the most likely outcome is even more of the
same. This paper is also an example of how forensic research can more fully
investigate the rent-seeking dimension of finance (Zingales (2015)). Govern-
ment agencies can assist with this transparency goal by making detailed data
available to the public. We hope to see future research with additional forensic
investigation of the PPP as well as other recent government and private
lending programs.

                 Initial submission: March 11, 2022; Accepted: July 28, 2022
                Editors: Stefan Nagel, Philip Bond, Amit Seru, and Wei Xiong
                     Did FinTech Lenders Facilitate PPP Fraud?                            1825

                                     REFERENCES
Alper, Mariel, Matthew R. Durose, and Joshua Markman, 2018, 2018 update on prisoner recidi-
    vism: A 9-year follow-up period (2005–2014), Special Report by Bureau of Justice Statistics:
    Office of Justice Programs.
Atkins, Rachel M. B., Lisa D. Cook, and Robert Seamans, 2022, Discrimination in lending? Evi-
    dence from the Paycheck Protection Program, Small Business Economics 58, 843–865.
Autor, David, David Cho, Leland Crane, Mita Goldar, Byron Lutz, Joshua Montes, William Pe-
    terman, David Ratner, Daniel Villar, and Ahu Yildirmaz, 2022, The $800 billion Paycheck
    Protection Program: Where did the money go and why did it go there?, Journal of Economic
    Perspectives 36, 55–80.
Bandiera, Oriana, Andrea Prat, and Tommaso Valletti, 2009, Active and passive waste in gov-
    ernment spending: Evidence from a policy experiment, American Economic Review 99, 1278–
    1308.
Bartik, Alexander W., Marianne Bertrand, Zoe Cullen, Edward L. Glaeser, Michael Luca, and
    Christopher Stanton, 2020, The impact of COVID-19 on small business outcomes and expec-
    tations, Proceedings of the National Academy of Sciences 117, 17656–17666.
Bartlett, Robert, Adair Morse, Richard Stanton, and Nancy Wallace, 2022, Consumer-lending dis-
    crimination in the FinTech era, Journal of Financial Economics 143, 30–56.
Becker, Gary S., 1968, Crime and punishment: An economic approach, Journal of Political Econ-
    omy 76, 169–217.
Beggs, William, and Thuong Harvison, 2022, Fraud and abuse in the Paycheck Protection Pro-
    gram? Evidence from investment advisory firms, Journal of Banking and Finance 147,
    106444.
Begly, Taylor A., Umit G. Gurun, Amiyatosh Purnanandam, and Daniel Weagley, 2022, Disaster
    lending: “Fair” prices, but “unfair” access, Working paper, Georgia Institute of Technology.
Ben-David, Itzhak, Mark J. Johnson, and René M. Stulz, 2022, Why did small business FinTech
    lending dry up during March 2020? Working paper, NBER.
Buchak, Greg, Gregor Matvos, Tomasz Piskorski, and Amit Seru, 2018, FinTech, regulatory arbi-
    trage, and the rise of shadow banks, Journal of Financial Economics 130, 453–483.
Chemmanur, Thomas J., and Paolo Fulghieri, 1994, Investment bank reputation, information pro-
    duction, and financial intermediation, Journal of Finance 49, 57–79.
Chernenko, Sergey, and David S. Scharfstein, 2022, Racial disparities in the Paycheck Protection
    Program, Working paper, NBER.
Chetty, Raj, John N. Friedman, Nathaniel Hendren, and Michael Stepner, 2022, The economic
    impacts of COVID-19: Evidence from a new public database built using private sector data,
    Working paper, NBER.
Chetty, Raj, John N. Friedman, and Emmanuel Saez, 2013, Using differences in knowledge across
    neighborhoods to uncover the impacts of the EITC on earnings, American Economic Review
    103, 2683–2721.
Cowley, Stacy, 2020, The tiny bank that got pandemic aid to 100,000 small businesses, New York
    Times, June 23.
Cowley, Stacy, 2021, The Paycheck Protection Program is out of money, New York Times, May 4.
Cowley, Stacy, and Ella Koeze, 2021, How two start-ups reaped billions in fees on small business
    relief loans, New York Times, June 27.
Denes, Matthew, Spyridon Lagaras, and Margarita Tsoutsoura, 2021, First come, first served: The
    timing of government support and its impact on firms, Working paper, SSRN.
Duchin, Ran, Xiumin Martin, Roni Michaely, and Ivy Wang, 2022, Concierge treatment from
    banks: Evidence from the Paycheck Protection Program, Journal of Corporate Finance 72,
    102124.
Duflo, Esther, 2017, The economist as plumber, American Economic Review 107, 1–26.
Egan, Mark, Gregor Matvos, and Amit Seru, 2019, The market for financial adviser misconduct,
    Journal of Political Economy 127, 233–295.
Eid, Nourhan, Josephine Maltby, and Oleksandr Talavera, 2017, Income rounding and loan per-
    formance in the peer-to-peer market, Working Paper, SSRN.
1826                              The Journal of Finance®

Erel, Isil, and Jack Liebersohn, 2022, Does FinTech substitute for banks? Evidence from the Pay-
    check Protection Program, Journal of Financial Economics 146, 90–118.
Faulkender, Michael W., Robert Jackman, and Stephen Miran, 2021, The job preservation effects
    of Paycheck Protection Program loans, Working paper, Office of Economic Policy.
Fuster, Andreas, Matthew Plosser, Philipp Schnabl, and James Vickery, 2019, The role of technol-
    ogy in mortgage lending, Review of Financial Studies 32, 1854–1899.
Garmaise, Mark J., 2015, Borrower misreporting and loan performance, Journal of Finance 70,
    449–484.
Glaeser, Edward L., and Claudia Goldin, 2006, Corruption and Reform: Lessons from America’s
    Economic History (University of Chicago Press, Chicago).
Glancy, David, 2022, Bank relationships and the geography of PPP lending, Working paper, SSRN.
Gopal, Manasa, and Philipp Schnabl, 2022, The rise of finance companies and fintech lenders in
    small business lending, Review of Financial Studies 35, 4859–4901.
Granja, João, Christos Makridis, Constantine Yannelis, and Eric Zwick, 2022, Did the Paycheck
    Protection Program hit the target?, Journal of Financial Economics 145, 725–761.
Griffin, John M., 2021, Ten years of evidence: Was fraud a force in the financial crisis?, Journal of
    Economic Literature 59, 1293–1321.
Griffin, John M., Samuel Kruger, and Gonzalo Maturana, 2019, Do labor markets discipline? Evi-
    dence from RMBS bankers, Journal of Financial Economics 133, 726–750.
Griffin, John M., Richard Lowery, and Alessio Saretto, 2014, Complex securities and underwriter
    reputation: Do reputable underwriters produce better securities?, Review of Financial Studies
    27, 2872–2925.
Griffin, John M., and Gonzalo Maturana, 2016, Who facilitated misreporting in securitized loans?,
    Review of Financial Studies 20, 384–419.
Hanson, Samuel G., Jeremy Stein, Adi Sunderam, and Eric Zwick, 2020, Business credit programs
    in the pandemic era, Brookings Papers on Economic Activity Fall 2020, 3–60.
Howell, Sabrina T., Theresa Kuchler, David Snitkof, Johannes Stroebe, and Jun Wong, 2022, Au-
    tomation in small business lending can reduce racial disparities: Evidence from the Paycheck
    Protection Program, Journal of Finance (forthcoming).
Jagtiani, Julapa, and Catharine Lemieux, 2019, The roles of alternative data and machine learn-
    ing in FinTech lending: Evidence from the LendingClub consumer platform, Financial Man-
    agement 48, 1009–1029.
Keys, Benjamin J., Tanmoy Mukherjee, Amit Seru, and Vikrant Vig, 2010, Did securitization lead
    to lax screening? Evidence from subprime loans, Quarterly Journal of Economics 125, 307–
    362.
Kruger, Samuel, and Gonzalo Maturana, 2021, Collateral misreporting in the residential mortgage
    backed security market, Management Science 67, 2728–2750.
Li, Lei, and Philip E. Strahan, 2021, Who supplies PPP loans (and does it matter)? Banks,
    relationships and the COVID crisis, Journal of Financial and Quantitative Analysis 56,
    2411–2438.
McArdle, Megan, 2022, Billions were stolen from the Paycheck Protection Program. It was a huge
    success anyway, The Washington Post, March 29.
Mian, Atif R., and Amir Sufi, 2017, Fraudulent income overstatement on mortgage applications
    during the credit expansion of 2002 to 2005, Review of Financial Studies 30, 1832–1864.
Neilson, Christopher, John Eric Humphries, and Gabriel Ulyssea, 2020, Information frictions and
    access to the Paycheck Protection Program, Journal of Public Economics 190, 104244.
Nigrini, Mark J., 2018, Round numbers: A fingerprint of fraud, Journal of Accountancy May 2018,
    36–42.
Omeokwe, Amara, 2022, A Magic Johnson affiliate funded a small-business lender. That lender’s
    CEO now faces fraud charges, Wall Street Journal, March 6.
Piskorski, Tomsaz, Amit Seru, and James Witkin, 2015, Asset quality misrepresentation by finan-
    cial intermediaries: Evidence from the RMBS market, Journal of Finance 70, 2635–2678.
                     Did FinTech Lenders Facilitate PPP Fraud?                             1827

Podkul, Cezary, 2021, How unemployment insurance fraud exploded during the pandemic, ProP-
    ublica, July 26.
Purnanandam, Amiyatosh, 2011, Originate-to-distribute model and the subprime mortgage crisis,
    Review of Financial Studies 24, 1181–1915.
Rabetti, Daniel, 2022, Non-information asymmetry benefits of relationship lending, Working pa-
    per, Tel Aviv University.
Wieder, Ben and Meghan Bobrowsky, 2020, Copycat websites, newly minted companies, double
    dippers cashed in on PPP loans, Miami Herald, September 11.
Willis, Derek, and Lydia DePillis, 2021, Hundreds of PPP loans went to fake farms in absurd
    places, ProPublica, May 18.
Zingales, Luigi, 2015, Presidential address: Does finance benefit society? Journal of Finance 70,
    1327–1363.
Zitzewitz, Eric, 2012, Forensic economics, Journal of Economic Literature 50, 731–769.


                               Supporting Information
Additional Supporting Information may be found in the online version of this
article at the publisher’s website:
Appendix S1: Internet Appendix.
Replication Code.
Disclosure Statement.


File and source

File

Griffin-Kruger-Mahajan_did-fintech-lenders-facilitate-ppp-fraud_jf2023.pdf

Size

4,575,963 bytes

SHA-256

ac75b5aff8452b3ad3e98ea9c82519a2f54b5759da821f2fdc4ec36af405b079

Our copy

Griffin-Kruger-Mahajan_did-fintech-lenders-facilitate-ppp-fraud_jf2023.pdf

Original

No public link identified.