2021 09 Griffin Fintech Lenders PPP Fraud Conference

Summary

Did FinTech Lenders Facilitate PPP Fraud? is an academic paper by John M. Griffin, Samuel Kruger and Prateek Mahajan of the McCombs School of Business, University of Texas at Austin, dated August 13, 2021. The paper analyzes loan-level Paycheck Protection Program data on 10,697,219 loans from 4,902 lenders, using four primary and five secondary indicators of potential misreporting, which the authors state are not proof of misreporting on their own. It reports that FinTech loans are more than 3.5 times as likely to be initiated by someone with a criminal background and that the authors find more than 1.8 million questionable loans representing $76 billion. The paper also compares traditional and FinTech lenders across program rounds. It closes with appendix tables on repayment and DOJ enforcement actions for flagged loans.

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

Did FinTech Lenders Facilitate PPP Fraud?∗
John M. Griffin†

Samuel Kruger‡

Prateek Mahajan§

August 13, 2021
** DO NOT DISTRIBUTE WITHOUT PERMISSION**
Abstract

In the distribution of the Paycheck Protection Program’s (PPP) $780 billion in funds,
FinTech lenders began minimally but ramped up their market share to over 70% of
originated loans by April 2021. We examine metrics related to potential misreporting
including non-registered businesses, multiple businesses at residential addresses, abnormally high implied compensation per employee, and large inconsistencies in jobs
reported with another government program. We assess these four metrics with five additional measures and extensive supporting analysis. FinTech loans exhibit sharp and
discontinuous increases in misreporting at maximum loan thresholds and at round loan
amounts. FinTech loans are more than 3.5 times as likely to be initiated by someone
with a criminal background, strongly cluster in industries-county pairs to a degree that
is infeasible based on U.S. Census data on establishment counts, and frequently exhibit
suspiciously similar loan features within lender-county pairs. Certain FinTech lenders
seem to specialize in dubious loans with more than 45% of their loans experiencing at
least one misreporting indicator. Few of these loans seem to have been detected by
authorities or repaid. FinTech lenders with the highest misreporting in the first two
rounds of the program in 2020 increase both their market share and their misreporting
substantially in the third round in 2021. While FinTech lenders likely expand PPP
access, this may come at the cost of facilitating fraudulent credit.

JEL classification: G21, G23, G28, H12
keywords: FinTech, Paycheck Protection Program (PPP), Misreporting, Fraud.
∗ We are thankful for comments from seminar participants at Hong Kong Baptist University. We thank Integra Research
Group for research support. 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 investigation of financial fraud.
† McCombs School of Business, University of Texas at Austin. Email: John.Griffin@utexas.edu.
‡ McCombs School of Business, University of Texas at Austin. Email: Sam.Kruger@mccombs.utexas.edu.
§ McCombs School of Business, University of Texas at Austin. Email: Prateek.Mahajan@mccombs.utexas.edu.


The melding of financial technology and banking, also known as FinTech lending, has
emerged at a rapid pace in the aftermath of the financial crisis. Buchak et al. (2018) find
that increased regulatory burdens on traditional banks is the predominate driver in the rise
of FinTech lending. A large aspect of the scrutiny and regulation of traditional banking was
their role in the financial crisis, including facilitating wide-scale mortgage fraud as partially
evidenced by over $137 billion in government fines and settlements (Griffin et al. 2019a). FinTech lenders offer a new banking model that replaces traditional lending relationships with
online advertisements, app interfaces, and loan screening algorithms. Are FinTech lenders
able to harness the power of technology to reduce loan maleficence?
The Paycheck Protection Program (PPP), a historic COVID-19 relief program for businesses, rapidly distributed over $780 billion in funds through 10.7 million loans in three short
rounds spread between April 2020 and May 2021. Although FinTech lenders began with a
slow start with less than 10% of loans in round 1, they ramped up their market share to
over 70% of loans by April 2021, highlighting their growing importance. FinTech lending
was hailed for broadening access to PPP loans and for facilitating quick and efficient lending at a time when many small businesses were in dire need due to the COVID pandemic.
However, the rapid expansion of FinTech lending may have come at the expense of underwriting standards. Whereas traditional banks have established borrower relationships and
extensive Bank Secrecy Act (BSA) and Know-Your-Customer (KYC) compliance programs,
many FinTech lenders had few established relationships and may have been light on 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 mortgage lenders not only process
government agency loans faster than traditional banks but 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 access may be able to detect
and prevent PPP applications from fictitious businesses and individuals. Do traditional or
FinTech loans exhibit more features consistent with potential PPP misreporting? And how
does this potential misreporting vary across individual traditional and FinTech lenders?
To investigate these questions, we perform a big data analysis of loan features on over
ten million PPP loans with six disparate databases. We introduce four primary and five
secondary indicators of whether a loan is potentially misstated. Each indicator creates
an inference that a loan is suspicious but is not proof of misreporting on its own. The
four primary measures are non-registered businesses, multiple loans at a residential address,
1


abnormally high implied compensation relative to industry and CBSA norms, and large inconsistencies between the jobs reported by a borrower on its PPP application and another
contemporaneous government program application. The five secondary measures are discontinuities around the maximum PPP compensation level of $100,000, rounded loan amounts,
overrepresentation of PPP loans relative to U.S. Census data on the number of business
entities in a particular industry and county, clustering of loans with similar features within
county-lender pairs, and criminal records for PPP borrowers.
We assess each of the four primary indicators with multiple discontinuity and comparative analyses. First, we find discontinuously elevated levels of potential misreporting, based
on each of the primary indicators, at the $100,000 compensation threshold. These discontinuities are much stronger for FinTech lenders. Second, loans often cluster at rounded
monthly compensation values, and these spikes coincide with higher levels of each of the
four potential misreporting indicators. FinTech lenders have both a higher percentage of
loans at rounded values and larger spikes in misreporting at these values. Third, PPP lending at the industry-county level frequently exceeds the number of businesses listed for that
industry and county in U.S. Census data. Overall, these excess loans represent 19.2% of
first draw business loans. For FinTech lenders, 35.6% of loans exceed industry-county establishment counts, and 28.6% of loans exceed industry-county establishment counts by a
factor of more than two.1 Measures of misreporting monotonically increase as the ratio of
PPP loans-to-businesses documented by the U.S. Census increases, particularly for FinTech
lenders. Fourth, based on the idea that networks in a region may use recurring loan features,
we construct a concentration ratio to measure clustering in loan amounts, number of jobs
(excluding one job), 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 associated with higher levels of potential misreporting. Finally, we collect
criminal background data for a sample of 150,000 individuals. FinTech borrowers are more
than 3.5 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.
Overall, we find more than 1.8 million questionable loans representing $76 billion in
capital. FinTech loans are more than 2.7 times as likely to have at least one primary
indicator of misreporting and 4.7 times as likely to have a primary indicator that is confirmed
1 The corresponding figures for traditional lenders are 13.1% and 7.4%, respectively. 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. For the 28.6%
and 7.4% figures, the interval limits are changed to 2 instead of 1.

2


by an additional primary or secondary indicator. Some regions and lenders stand out with
particularly high rates of suspicious PPP lending, with potential misreporting rates in excess
of 45% for two large FinTech lenders. Moreover, potential misreporting increases over time
with particularly high rates in round 3 even after the Office of the Inspector General for the
Small Business Administration (SBA) flagged PPP fraud as a concern. Network graphs of the
lending space indicate that many FinTech portals switch lenders and utilize multiple lenders
even for the same individual. Several of the FinTech lenders with the highest suspicious loan
rates are new lenders that did not start making PPP loans until round 3, and there is no
evidence that lenders attempted to decrease misreporting over time. Instead, second-draw
loans to borrowers with suspicious first-draw loans by 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 largest three FinTechs, Cross
River, Capital Plus, and Harvest, exhibited high and increasing rates of both misreporting
and lending volume while receiving over $900 million in processing fees each. Finally, FinTech
lenders often doubled, tripled, or even quadrupled their lending in zip codes with high levels
potential misreporting in rounds 1 and 2 while also increasing their misreporting rates.
Our work is related to four literatures. First, there is a rapidly emerging literature on
FinTech lending that highlights its growing importance and positive economic effects through
filling gaps left by traditional banks in both residential (Buchak et al. 2018) and business
lending (Gopal and Schnabl 2020). 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 (2021) examine FinTech lending in the PPP and
finds that FinTech lenders increased access to the PPP by lending more in zip codes with
fewer traditional banks, lower incomes, and higher minority percentages.2 With respect to
FinTech lending before the PPP, Gopal and Schnabl (2020) find that FinTech lenders have
positive economic effects by filling in gaps in lending to small businesses left by traditional
banks following the financial crisis. While most of the FinTech literature finds benefits to
FinTech lending such as increased competition, broader financial access, faster lending speed,
and lower defaults, our paper analyzes a potential cost of FinTech expansion, leaving overall
welfare analysis to future research.
Second, regarding the efficacy of the PPP, Chetty et al. (2020) find that the PPP increased
2 Relatedly, Howell et al. (2020) find that FinTechs were more likely to provide PPP loans to black-owned businesses.

In
contrast, Bartlett et al. (2021) find that FinTech algorithms charge higher interest rates to minorities in residential mortgage
lending.

3


employment at participating firms by only 2% at a cost of $377,000 per job saved, and Autor
et al. (2020) find only slightly higher employment benefits of 2 to 4.5%. Granja et al. (2020)
find small employment effects due to the PPP and a low correlation between regional COVID
variation and PPP funding allocation. Additionally, there is evidence of differential access to
the PPP based on knowledge of the program, distance to the closest bank branch, banking
relationships, and personal banking connections (Amiram and Rabetti 2020; Bartik et al.
2020; Neilson et al. 2020; Duchin et al. 2021; Glancy 2021; Li and Strahan 2021). Our
evidence adds an additional concern regarding the program’s efficacy and fairness.
Third, although we are the first academic paper to examine wide-scale potential PPP
loan misreporting, there have been a number of interesting press and investigative reports regarding suspicious PPP loans (Miami Herald 2020; The Wall Street Journal 2020; Bloomberg
Businessweek 2020; Project on Government Oversight 2020; ProPublica 2021), some of which
feature FinTech lenders. Law enforcement is pursuing investigations into some of these loans
as indicated by U.S. Department of Justice (DOJ) press releases containing information on
162 criminal complaints regarding 355 loans. Concerns about PPP fraud have been flagged
by the Office of the Inspector General for the SBA, and a congressional subcommittee recently launched an investigation into the role of FinTech lenders in PPP fraud.3 Beggs
and Harvison (2021) find that among the 1,090 registered investment advisors who took
PPP loans, those with a history of financial misconduct received unusually large PPP loan
allocations.
Finally, our work relates to a more general literature on loan misreporting and fraud.
Widescale mortgage fraud and misreporting in securitized mortgages prior to the financial
crisis included second-lien and owner-occupancy status misreporting (Piskorski et al. 2015;
Griffin and Maturana 2016), misreported income (Jiang et al. 2014; Mian and Sufi 2017),
misreported assets (Garmaise 2015), and inflated appraisals (Ben-David 2011; Kruger and
Maturana 2020). This fraud involved both smaller, less-known mortgage originators and
large bank underwriters who knowingly passed along these misrepresentations in mortgagebacked securities. FinTech lending emerged and grew against this backdrop and related
regulation increases for traditional banks FinTech (Buchak et al. 2018). Our findings indicate
that replacing traditional lending with FinTech lending may actually amplify misreporting
problems, at least with respect to the PPP.
Our findings also have important practical implications regarding extent and nature of
PPP misreporting, the expanding role of FinTech lending, waste in the PPP, the proliferation
3 The SBA OIG report can be found at https://www.sba.gov/sites/default/files/2021-01/SBA%20OIG%20Report-21-07.
pdf and a press release regarding the investigation by a congressional subcommittee is available at https://coronavirus.house.
gov/news/press-releases/select-subcommittee-launches-investigation-role-fintech-industry-ppp-fraud.

4


of fictitious lending, and the insufficient deterrence of current policies and enforcement. The
potential policy and practical implications of these finding and the role of forensic finance in
examining related issues are discussed further in the conclusion.

1

Data and Summary Statistics

1.1

Data Sources

The basis for our sample is loan-level PPP data released on May 3, 2021 by the Small
Business Administration (SBA). This dataset covers all PPP loans issued from the start
of the program on April 3, 2020 through April 30, 2021 (which is all of rounds 1 and 2
and the majority of round 3) that had not been repaid as of May 3, 2021. At the loanlevel, the data includes business name, address, business type (e.g., corporation, LLC, selfemployed, etc.), NAICS code (industry), loan amount, number of employees, date approved,
loan draw (i.e., initial, first-draw loan or repeat, second draw loan), and lender for 10,697,219
loans originated by 4,902 different lenders and with a total value of $782 billion. We follow
Erel and Liebersohn (2021) and classify lenders as either traditional or FinTech based on
automated name matching with bank identifiers from the Federal Financial Institutions
Examination Council (FFIEC) with hand matching for remaining lenders.4 We also use
Erel and Liebersohn’s (2021) methodology to classify FinTech lenders as online banks or
non-bank lenders.
Concurrently with the PPP, the SBA provided businesses and individuals with the ability
to receive an Economic Injury Disaster Loan (EIDL), with forgivable advances of up to $10k.
EIDL Advance loan-level data was released on December 1, 2020 and covers all EIDL Advance
issued in 2020.5 To check for inconsistencies between the information borrowers provided on
their PPP and the EIDL applications, we match businesses in the PPP and EIDL loan-level
datasets based on zip code and business name.
We also match PPP borrowers with business registry data from OpenCorporates, a nonprofit that maintains a database of companies around the world. OpenCorporates collects
its data directly from state governments and covers 76 million businesses across all US states
except Illinois. The data includes incorporation dates, dissolution dates (if applicable), and,
implicitly, whether the business has ever registered. We match OpenCorporates data to the
PPP loan-level data based on business name and address.
4 We use Erel and Liebersohn’s (2021) classifications for lenders that were active in rounds 1 and 2 (the sample period for
Erel and Liebersohn (2021)) and use the same methodology for classifying round-3 lenders that were not active enough to be
classified in the earlier rounds.
5 The SBA has not released updated EIDL Advance data for 2021.

5


To examine previous criminal and financial activity, we collect criminal background data
from LexisNexis 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
Finally, we use several U.S. governmental data sources for address and demographic information. We standardize addresses and distinguish residential and commercial addresses
from one another based on the Address Validation Application Programming Interface from
the United States Postal Service. For data on number of establishments and average compensation, we use the 2019 County Business Patterns (CBP) data from the US Census
Bureau, aggregated by region (either core-based statistical area (CBSA) or county) and
North American Industry Classification (NAICS) code. The CBP data includes the number
of establishments, number of employees, and total wages for a given industry in a county or
CBSA. Matching between the loan-level data and the CBP data is based the business’s zip
code and the first four digits of its NAICS code.

1.2

Summary Statistics

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 (non-bank FinTech lenders) and cream color (online banks). Six of the
ten top lenders by number of loans are FinTech, with Cross River, Capital Plus, and Harvest
in the top five alongside Bank of America and JP Morgan Chase.7 Due to their larger average
loan size (Erel and Liebersohn 2021), dollar lending volume tends to be higher for 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 1.4% of loans
in the first week of round 1 to 7.6% in the last week of round 1. Round 2 continued the
PPP after a short break of ten 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 49.1% of loans, over the last two weeks, for an overall market share of 5% in round
1 and 20% in round 2. Round 3 of the PPP, which includes both first-draw loans for new
borrowers and second-draw loans for borrowers that already obtained loans in round 1 or 2,
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 is collected only from rounds 1 and 2 loans because
round 3 data was released after the criminal records data was collected.
7 Comparing this figure to Panel A of Figure IA.1 shows how the top lenders differ between the entire sample and solely
rounds 1 and 2. In particular, the growth of Capital Plus and Harvest in round 3 is apparent.

6


started in January 2021 with a low FinTech market share of less than 20% for the first three
weeks as traditional lenders were once again the fastest to originate PPP loans. However,
FinTech share grew rapidly during round 3, reaching over 70% of loans by the end of April
2021 for an overall round 3 market share of 41%.8
Table I reports summary statistics on the 3.1 million FinTech and 7.6 million traditional
bank loans in our sample. FinTech loans have an average loan amount of $27 thousand
compared to $91 thousand for traditional banks loans. Despite these large differences in
means, the median loan sizes of $18.7 and $20.8 thousand are similar. The average FinTech
loan reports supporting 2.8 jobs compared to 10.5 for traditional banks. After normalizing
loan size relative to reported jobs, FinTech loans have higher average ($61 thousand) and
median implied compensation ($62.8 thousand) than traditional bank loans ($47 thousand
average and $38.7k median). The business type distribution for FinTech loans is 22.8%
businesses organized as corporations, S-corporations, or limited liability companies (LLC)
compared to 65.8% for traditional banks lenders. FinTech loans were also less likely to be
repeat loans, with 25.9% of round 3 FinTech loans going to borrowers with previous PPP
loans, compared to 60.1% for round 3 traditional bank loans.

2

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 is not definitive proof of misreporting on its own. In subsequent sections we
validate the measures and explore how they relate to one another and other misreporting
indicators.

2.1

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, 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:
1. Is the business found in the business registry for its home state or in another state
8 Panel B of Figure IA.1 shows the number of loans originated each week of the PPP by type of lender.
9 See loan application at https://www.sba.gov/sites/default/files/2021-03/BorrowerApplication2483ARPrevisions%20%
28final%203-18-21%29-508.pdf.

7


while listing an address in its home state? (“Missing Business”)
2. Was the business dissolved and inactive before being approved for a PPP loan? (“Dissolved Business”)10
3. Is the earliest incorporation or initial filling date for the business after February 15,
2020? (“Late Incorporation/Filing”)
These three subflags are combined to form an overall business registry flag. Panel A of
Figure 2 plots the proportion of corporate and LLC borrowers with missing, dissolved, or late
business registrations. The flag is plotted as a percent of corporation, S-corporation, and LLC
loans because other PPP business entities, such as sole proprietorships, partnerships, and
independent contractors, do not require business registrations. We also exclude businesses in
Illinois because Illinois is missing from the business registry data.11 Missing registrations are
the most common type of business registry flag, representing 3.85% of corporate and LLC
loans. Another 0.72% of corporate and LLC loans are to dissolved entities, and 0.20% have
late registrations for a total business registry flag percentage of 4.77%. Nine of the ten lenders
with the highest rates of business registry flags are FinTech lenders. These lenders have 8 to
25% of their corporate and LLC loans flagged for one of the three business registry issues,
with the vast majority of the flagged loans simply not appearing in the business registry
data. While it is possible that there could be errors in the data, and some businesses may
have names that are difficult to match, detailed manual searches indicate no public records
information for most missing business registrations even after searching for alternative name
variations. Matching issues could explain why all of the lenders have at least some missing
registrations with business registry flag rates of one to five percent common across many
lenders. However, there is not an obvious explanation for why certain lenders (which also
have elevated levels on other indicators) should have disproportionately high matching issues,
particularly since most of the FinTech lenders have broad geographic dispersion in their PPP
lending.

2.2

Multiple Loan Flag

While it is possible that a business owner may have multiple businesses registered to the
same address, the presence of multiple loans at a residential address during the same draw
is also a potential sign of fictitious operations. Using the business address disclosed in the
10 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.
11 Illinois has very restrictive terms and conditions regarding the use of their business registry data and even has legislation
in place to make it a criminal offense to collect or republish the data (see https://www.ilga.gov/legislation/ilcs/ilcs4.
asp?ActID=1876&ChapterID=53&SeqStart=58900000&SeqEnd=59600000).

8


PPP loan-level data, we identify individual residential addresses associated with multiple
loans during the same draw. To do so, we first standardize addresses and identify addresses
that are known business or central addresses (e.g., office and apartment buildings) using
the Address Validation Application Programming Interface from the United States Postal
Service. Then, we find residential (i.e., non-business, non-central) standardized addresses
with multiple loans within the same draw.
As an example, Panel A of Exhibit 1 shows 14 loans given to a single address, all with
colorful business names, almost all in the same industry, most with the same loan amount,
and all backing ten jobs. The address associated with all 14 loans is a modest singlefamily home in suburban Chicago (estimated to have a value of $170k per Zillow). The
borrower associated with the first loan is an LLC that was registered in 2018, but the 13
subsequent loans during July and August of 2020 are to LLCs that were registered only
shortly before the loans were approved, well after the February 15 eligibility cutoff. Detailed
internet searches did not produce information for any of the other 13 business names and no
indication of employees other than the owner. Panel B of Exhibit 1 shows another multipleloan example, this one involving 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 $100,000.12 This income
is at the top of the spectrum for the indicated industries, which have average compensation
$25-46k in the Chicago CBSA according to the US Census CBP. The industries themselves
are also somewhat suspicious in that two are equipment manufacturing and one is auto repair
despite no evidence of these businesses in photos of the property.13 Further, the borrower
in the nail salon industry does not appear to have an Illinois nail technician license. One of
the equipment manufacturing borrowers also switched to the nail salon industry in a second
draw loan during round 3 of the PPP despite also not having a nail technician license. Loan
level inspections of the data reveal numerous other suspicious loans flowing to addresses that
do not seem to be the locations of identifiable businesses despite applications claiming to
employ multiple workers. The multiple loan flag functions as a way to systematically analyze
these loans.
Panel B of Figure 2 shows the percentage of PPP loans that involve multiple loans
to the same residential address by lender. Nine of the ten top lenders with the highest
multiple loan flag rates are FinTech lenders. For these lenders, 12.5 to 17.5% of their loans
involve multiple loans to the same residential address, and most of their flagged loans are to
12 All four of these individuals also received second draw loans for the same amount.
13 SBA guidance asks the borrower for their business address.

9


individual borrowers identified as independent contractors, self-employed, or sole proprietors.
This contrasts with traditional banks, which have fewer flagged loans (5.1% on average).
Further, flagged loans for traditional banks are mainly to formally registered corporations,
S-corporations, and LLCs, consistent with the incentives for owners of multiple legitimate
businesses to formally register their businesses for tax and limited liability purposes. For
example, the largest PPP lender, Bank of America, has 3.46% of its loans flagged for receiving
multiple PPP loans at the same residential address, and most of these loans are to registered
business entities. Some of these differences could be due to differences in loan composition
across lenders composition which we examine in subsequent regression analysis. Interestingly,
three FinTech lenders, Capital One, Square, and Intuit have lower than median levels of
multiple loans at the same address.

2.3

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.14 PPP loan applications
report how many employees the business has, again based on 2019 averages or the same period
over which average monthly payroll costs were determined in most cases. Using loan size and
number of reported employees, we are able to impute implied average annual compensation.
We first focus on the overall level of compensation, and in subsequent sections, we return to
more granular patterns associated with compensation discontinuities and rounding.
Implied compensation at the borrower level is strongly related to average compensation
in the borrower’s industry (NAICS 4-digit) and CBSA (e.g., see Panel B of Figure IA.2).
However, some borrowers have abnormally high implied compensation relative to their industry and geographic area. Panel A of Figure 3 plots the kernel density of implied average
compensation for the borrower normalized by mean compensation across all firms in the borrower’s industry-CBSA pair based on US Census Bureau CBP data separately for FinTech
and traditional loans. The top plot, which is restricted to corporate and LLC borrowers,
shows similar distributions for FinTech and traditional borrowers. The bottom plot, which
includes all loans, reveals larger differences between FinTech and traditional borrowers. For
traditional borrowers, normalized compensation is 1.14 on average, with a median of 0.92
and a standard deviation of 1.97. FinTech borrowers have a higher average and median normalized compensation of 1.83 and 1.39, respectively, with a standard deviation of 1.59. This
difference is largely due to the right tail of the distribution being much larger for FinTech
borrowers. Specifically, 21.5% of FinTech borrowers have normalized compensation above 3,
14 See the Internet Appendix for details on how the loan size was to be calculated and exclusions.

10


compared to 4.1% of traditional borrowers.15
It is instructive to examine how normalized compensation relates to our first two suspicious loan flags. On the right axis of Panel A of Figure 3, we plot the percentage of
loans with the business registry and multiple loan flags across the distribution of normalized
compensation separately for FinTech and traditional lenders.16 Both flags increase significantly as normalized compensation increases for loans made by FinTech lenders. Whereas
6.8% of corporate and LLC FinTech loans with normalized compensation below one have the
business registry flag, 22.8% of loans with normalized compensation above three have the
flag. Similarly, the multiple loan flag increases from 9.2% for FinTech loans with normalized
compensation below one to 16.4% when normalized compensation is above three. Importantly, while FinTech loans exhibit a stronger relation between normalized compensation
and the other loan flags, this pattern is not limited to FinTechs. Traditional bank loans
also have more business registry and multiple loan flags when normalized compensation is
higher suggesting that they also have loan misreporting issues, though at a much lower scale.
Overall, the results show that while some variation in normalized compensation across firms
is to be expected, high implied compensation is strongly related to other suspicious loan
characteristics, particularly for FinTech loans.
For our main measure of high implied compensation, we conservatively only flag loans
where the implied compensation per job reported is more than three times the industryCBSA average compensation (“high implied compensation”). Because compensation is censored at $100,000 for most borrowers, this flag is only possible industry-CBSA pairs with
average annual compensation below $33,333.33.17 Within this set of industry-CBSA pairs,
47.8% of FinTech and 10.2% of traditional lenders have normalized compensation above
three. Panel B of Figure 3 plots how the percentage of loans with the high compensation
flag varies across lenders. Similar to the previous flags, nine of the ten lenders with the highest abnormal compensation percentages are FinTech. For all of these lenders, more than 40%
loans have the high implied compensation flag. By contrast, the 43 of the 68 largest lenders
have less than 10% of their loans flagged, and 41 of these lenders are traditional banks.
Although most of the FinTechs cluster with high rates of abnormally high compensation,
15 Most of this is due to round 3 FinTech loans, which had mean normalized compensation of 2.04 compared to mean
normalized compensation of 1.18 for FinTech loans in rounds 1 and 2. This elevated compensation for FinTech in round 3 is
evident in Panel A of Figure IA.2. Whereas implied compensation is similar for FinTech and traditional borrowers in rounds 1
and 2, FinTech implied compensation is much higher round 3 and appears to be almost completely disconnected from average
industry-CBSA compensation.
16 The business registry flag is calculated as a percentage of corporation, S-corporation, and LLC loans because the business
registry flag can only be determined for these business types whereas the multiple loan flag can apply to any loan.
17 Some loans are also outside of a CBSA or in an industry-CBSA pair that is too small to be included in the US Census
CBP data. In total, 3,297,068 loans are in industry-CBSA pairs with average annual compensation below $33,333.33. The rates
plotted in Panel B of Figure 3 represent percentages of loans within this subsample.

11


Capital One and Square again have low levels of flagged loans.

2.4

EIDL Advance Jobs > PPP Jobs Flag

Concurrently with the PPP, the SBA provided businesses and individuals with the ability
to receive a forgivable Economic Injury Disaster Loan (EIDL) Advance of up to $10,000.18
For all EIDL Advances issued in 2020, the advance amount was calculated as $1,000 per
employee (up to the $10,000 maximum).19 Thus, there was an incentive for borrowers to
inflate the number of jobs reported on their EIDL application. We check for inconsistencies
between implied jobs based on EIDL Advance amounts and reported jobs from PPP loan
applications. For borrowers who take out the maximum EIDL Advance of $10,000, we can
infer that the borrower claimed at least 10 employees of their EIDL Advance application.
We focus on cases where EIDL jobs exceed PPP jobs because the job inflation incentive is
provided by the EIDL Advance program (PPP loans are based on total payroll as opposed to
number of jobs). While the EIDL > PPP jobs flag is primarily an indicator of misreporting
on the EIDL application, applicants who misreport in one area are likely willing to misreport
in other areas too.
Panel A of Figure 4 plots the distribution of differences between EIDL and PPP jobs.
Three patterns stand out. First, consistent with the incentive to inflate EIDL jobs as opposed
to PPP jobs, EIDL exceeds PPP by three or more jobs 8.2% of the time, whereas PPP exceeds
EIDL by three or more jobs only 3.6% of the time. Second, the most common discrepancy
between the programs is a difference of nine jobs, which implies that the borrower claimed 10
jobs and took out the maximum EIDL Advance of $10,000 despite only reporting one PPP
job. Third, EIDL job inflation is much more pronounced in FinTech loans than in traditional
loans. In particular, EIDL exceeds PPP by nine jobs 12.6% of the time for FinTech loans
compared to 0.5% for traditional loans.
Panel B of Figure 4 shows the prevalence of this EIDL job inflation by lender. To
be conservative, we only include cases where the EIDL implied number of jobs is at least
three more than the PPP reported number of jobs. Nine of the ten lenders with most
frequent discrepancies between EIDL and PPP jobs are FinTech. In particular, Capital
Plus, Prestamos CDFI, Harvest, Benworth, Itria, Fountainhead, MBE, Cross River, and
Kabbage (all FinTechs) have job reporting inconsistencies ranging from 10% to 64%. For
all of these lenders except Kabbage and MBE, most of the inconsistencies are a full nine
18 While the EIDL Advance program was billed as a forgivable advance with the potential for a larger non-forgivable loan,
over 65.8% of EIDL Advances involved no additional EIDL loan. EIDL advances were immediately forgiven by the SBA.
19 The EIDL Advance rules changed for 2021 to: A) provide the entire $10,000 regardless of employee count, and B) to target
the advances to low-income communities and those with a demonstrated decrease in revenue. The SBA has not yet reported
data on 2021 EIDL Advances.

12


jobs. Five FinTech lenders (Intuit, WebBank, Square, Capital One, and Live Oak) have
levels of inconsistencies that are similar to traditional banks, which are typically less than
five percent of their loans and a nine-job difference only in rare cases.

2.5

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 types of indicators to occur randomly across loans and
lenders. Therefore, multiple flags for the same loans create a heighted misreporting inference,
and high lender flag rates across multiple indicators may be due to policies and practices
that facilitate more misreporting.
In Table II, we examine how the four flags relate to one another by reporting odds
ratios between each pair of flags. The odds ratios are calculated based on loans for which
data to calculate both flags 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), with
z-statistics calculated based on standard errors double-clustered by zip code and lender in
parentheses. Panel A reports odds ratios for the full sample, all of which are above 1.4
and highly significant. In particular, the odds ratio between the high implied compensation
and EIDL > PPP jobs flags is 14.270 and has a z-statistic of 15.20. Panel B reports odds
ratios separately for FinTech and traditional loans with FinTech loans in the lower triangle
and traditional loans in the upper triangle. The odds ratios are all positive and highly
significant with consistently higher ratios for FinTech loans. To check that these relations
are independent of one another and not explained by loan characteristics, Table IA.I regresses
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. Panel A estimates regressions
without lender fixed effects, and Panel B adds lender fixed effects. With only one exception
(the effect of EIDL > PPP Jobs on the business registry flag and vice versa), the coefficients
between the flags are all positive, economically large relative to the mean flag rates, and
highly statistically significant.
We also find that flag rates are significantly correlated with one another, and the same
lenders frequently have high flag rates across all four flags (as shown in Figure IA.3). In
particular, FinTech lenders Capital Plus, Prestamos CDFI, MBE, and Harvest have flagged
rates in the top 10 for all four flags and Itria, Fountainhead, and Benworth for three of
the flags. In contrast, no traditional lender is consistently in the top 10 for more than two
flags. This pattern is exactly what we would expect if some lenders have looser underwriting
standards and is difficult to explain with random mistakes or errors in the data.

13


2.6

FinTech Differences?

Table III summarizes the percentage of each of the four flags by FinTech and traditional
lenders. The table also summarizes the percent of all loans with at least one flag and with two
or more flags. For each individual measure, the denominator is the loans that could have such
flag (i.e., only corporate and LLC loans for the business registry flag, only loans in industryCBSA pairs with average compensation below $33,333.33 for high implied compensation,
and only loans with a matched EIDL Advance loan for the comparison to EIDL). For the
overall flag measures, the denominator is all loans in the sample, which understates the
incidence of suspicious loans since most of the flags are only applicable to a minority of the
loans. Differences between FinTech and traditional flag percentages are reported in column
(3). For all four individual measures, FinTech lenders have flag rates that are two to four
times as high as traditional lenders, with particularly large differences for the high implied
compensation and EIDL > PPP jobs flags. Overall, 31.4% of FinTech loans have at least
one of the flags, compared to 11.6% for traditional loans. Additionally, 5.03% of FinTech
loans have at least two flags, compared to only 0.61% 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 traditional loans,
column (4) reports adjusted differences that control for geography, business type, and industry based on regressions controlling for loan size and number of jobs with zip code, business
type and industry × CBSA fixed effects.20 After accounting for these effects, the adjusted
difference between FinTech and traditional flag rates is 2.8 ppt for the business registry flag
(which is 65% of the mean rate for traditional loans), 2.9 ppt (57%) for the multiple loan flag,
9.4 ppt (92%) for the high implied compensation flag, and 5.9 ppt (123%) for the EIDL >
PPP jobs flag. These results indicate that even though loan composition explains part of the
difference between FinTech and traditional loans, flag rates remain much higher for FinTech
loans even after controlling for all observable characteristics. It is possible that there are
non-linear characteristics between FinTechs and traditional banks that might explain these
differences. To further control for potentially non-linear loan characteristic effects, we match
FinTech loans with traditional loans based on loan size, industry, county, and business type
in column (5) FinTech with similar results.21 It remains possible that other omitted variables
or unobserved loan characteristics could explain more of the difference between FinTech and
traditional loans, but these effects would have to be large to explain the results. Tests in the
20 Corresponding regressions results with and without the control variables and fixed effects are reported in Table IA.II.
21 Details on the matching process are provided in the Internet Appendix.

14


next section, including grouping around discontinuities and clustering, help to address this
concern.

3

Suspicious Loans or Mistakes?

All of the misreporting indicators in the previous section have potentially innocent explanations. In this section, we develop and analyze five additional measures as external
verification to assess the plausibility of alternative explanations. The additional measures
involve discontinuities, rounded compensation levels, abnormal numbers of loans in a particular industries and geographic areas, clustering of loan features, and criminal records. We
also explore relations between the indicators and differences between FinTech and traditional
lenders that are difficult to rationalize with alternative explanations.

3.1

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.22 This $100,000 cutoff is a hard maximum for
self-employment compensation. For other employees, payroll expenses also include employer
insurance and retirement contributions and unemployment taxes, which can push included
payroll expenses above $100,000 per employee. Someone filling out a fraudulent PPP application might want to maximize their loan amount by submitting payroll expenses at or
close to the $100,000 per employee limit without the additional expenses that are eligible
with proper payroll details.
Figure 5 plots the distribution of implied compensation per employee and shows how it
relates to the misreporting indicators from the previous section. The implied compensation
distributions (up to $130,000) for FinTech and traditional loans are plotted as orange and
gray bars, respectively. Panel A shows that FinTech loans stand out as having more clustering
in business (corporation, S-corporation, and LLC) loans with high implied compensation
right at and slightly under $100,000, and traditional banks have more loans with implied
compensation in the range of $30,000 to 75,000. Business registry flag rates 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 below the
$100,000 compensation bin. As compensation increase from $30 to $100k, the prevalence
of business registry violations for FinTech loans increases from 1.5% to 7%. For traditional
lenders the increase is much smaller and is concentrated in compensation levels that are at
or just below $100,000. There is also a jump in FinTech business registry violations for loans
22 See the Internet Appendix for details on the SBA guidance on how to calculate loan size.

15


with implied compensation per employee of $96,000 to $100,000 and a similar but smaller
jump for traditional loans with implied compensation of $98,000 to $100,000. For loans
with implied compensation above $100,000, there is a sharp drop-off in the business registry
flag, which indicates that businesses that followed the detailed SBA guidelines for including
non-wage payroll expenses for employees with wages above $100,000 are much less likely
to have registry violations. Finally, FinTech and traditional loans both have more business
registry violations as implied compensation decreases below $20,000, which suggests that
misreporting is also elevated for borrowers with abnormally low reported salaries.
Panels B, C, and D repeat the same analysis for the multiple loan, high implied compensation, and EIDL > PPP Jobs flags. The multiple loans plot includes all loans, the high
implied compensation plot includes all loans for which we can determine the industry-CBSA
average compensation, and the EIDL > PPP Jobs plot includes all loans with matched EIDL
Advances. Panel B shows that FinTech loans have a much higher incidence of the multiple loans flag at higher implied compensation levels. The high implied compensation flag in
Panel C naturally increases with implied compensation, but there is a much steeper slope for
FinTech loans than for traditional loans, and FinTech loans have a sharp discontinuity above
$100,000. This pattern for FinTech loans suggests that high compensation levels for FinTech loans are often disconnected from industry-CBSA averages whereas high compensation
traditional loans tend to be in industry-CBSA pairs where one would expect compensation
to be high. The drop in the high implied compensation flag above $100,000 indicates that
borrowers with detailed insurance and tax expenses in excess of $100,000 are also much more
likely to be in industry-CBSA pairs with high compensations more generally. For loans with
EIDL > PPP Jobs in Panel D, there is a sharp increase in job reporting inconsistencies for
FinTech loans with implied compensation from $60,000 to $100,000, with a sharp decrease
above $100,000. In Table IA.III, we formally test for discontinuities at $100,000 of compensation after controlling for loan characteristics including number of employees, loan size,
business type, zip code, and industry × CBSA and find large and highly economically significant discontinuities for FinTech loans, and to a much lesser extent for traditional loans, for
all four measures. Overall, the increasing patterns of flags with compensation and the discontinuities around $100,000 are consistent with suspicious loans maximizing loan amounts.
Innocuous explanations for why there should be more loans with business registry issues,
multiple loans at the same address, abnormally high compensation, and discrepancies in
number of jobs between government programs right before and at, but not above, $100,000
are not readily apparent.

16


3.2

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 documentation.23 It is unlikely that
actual monthly payroll would be a round number, especially after including unemployment
insurance and employer insurance and retirement contributions. Rounded loan amounts
suggest that the numbers are potentially fictitious as opposed to being based on actual documented data. In particular, PPP loan amounts in increments of $1,250, $2,500 or $5,000
imply that average monthly compensation is at a rounded $500, $1,000 or $2,000 increment, a potential red flag for fictitious data. If the flags we have previously identified reflect
misreporting issues, then one might expect both a clustering of loans at round numbers 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 at round numbers.
In Figure 6, we first examine the distribution of the last four digits of loan amounts,
excluding EIDL Refinancing, for FinTech and traditional loans. Loan amounts within 50
cents of a $1,250 increment (which corresponds to $500 of implied monthly compensation)
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 an implied compensation within ±
$1,000 of $100,000 are excluded to make sure these results are distinct from the maximum
compensation result shown in Figure 5. Panel A plots corporate and LLC loans (which are
the relevant sample for the business registry flag). Both FinTech and traditional loans exhibit
rounding at $1,250 increments, particularly at increments of $2,500 ($0, $2,500, $5,000, and
$7,500 in the figure, corresponding to $1,000 increments of implied monthly compensation).
FinTech lenders have moderately more rounding with 10.49% of loans rounded to $1,250
increments compared to 7.59% for traditional lenders.
The right axis of Panel A examines the business registry flag. Business registry flag
prevalence at the $1,250 loan increments is plotted as solid dots and at other loan amounts
(shown in $250 wide bins) is plotted as hollow dots. If rounded loans are more likely to be
misreported, one would expect an elevated level of loans with improper or missing business
23 See the Internet Appendix for details on how the loan size was to be calculated and exclusions.

17


registries at round number thresholds. For FinTech loans, this is exactly what we find.
The business registry flag is 1.23 ppt consistently higher at the rounded increments and
the difference is highly significant, which is easy to see by comparison to the dotted lines
plotting a 95% confidence interval estimated with a third-degree polynomial estimated based
on the non-rounded loans. For traditional lenders, there is only small and weak evidence
of elevated business registry flags in some of the rounded bins. Thus, rounding appears to
capture suspicious loans for FinTech lenders but less so for traditional lenders, which is also
consistent with our findings in Figures 2, 3, and 4.
In Panels B, C, and D we consider the other three primary misreporting indicators.
Panels B plots all loans, Panel C plots loans with an industry-CBSA pairs with average
compensation of less than $33,333.33, and Panel D plots loans with matched EIDL Advances.
Panels B, C, and D show that rounded loans by FinTech lenders also have elevated levels of
the multiple loan, high implied compensation, and EIDL > PPP Jobs flags. For traditional
lenders, the multiple loan flag is slightly elevated levels at every other $1,250 increments (but
not $2,500 increments), and there is no evidence of any relation between rounding and the
other flags. 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.

3.3

Loan Overrepresentation

If there is an organized effort to obtain funds for non-existent businesses, networks of
illegitimate borrowers may fill out multiple applications in a similar manner and could cluster
on characteristics such as industry and geography. Exhibit 2 shows examples of 4,304 $20,000
first draw loans made by Cross River to businesses in the “Insurance Agencies and Brokerage
Industry” in Illinois, mainly in the Chicago area, almost all of which have one employee.
These are followed by examples from 938 $20,000 first draw loans by Cross River to business
engaged in “All Other Miscellaneous Crop Farming,” most of which have exactly one or eight
employees.24 Most of these loans are in urban areas of Chicago, frequently in apartment
dwellings, where it is difficult to see how crop farming is performed. There are also another
3,056 $20,000 first draw loans by Cross River in Illinois to borrowers in other industries
(including 700 to business in “All Other Personal Services,” 347 to “General Freight Trucking,
Local,” 337 to “Other Performing Arts Companies”, and 229 to “New Single-Family Housing
Construction (except For-Sale Builders)”), also primarily in the Chicago area. In addition
to having the same loan amount and similar industries, these $20,000 loans were almost
24 There are an additional 1,574 loans for amounts besides $20,000 by Cross River in Illinois to business in “Insurance Agencies
and Brokerage Industry” and 643 to the “All Other Miscellaneous Crop Farming” industry.

18


non-existent until the very end of round 2. Specifically, 40% were originated in late July
and early August of 2020 during the final two weeks of round 2, and the other 60% were
originated in round 3. Overall, 48.9% of Cross River’s Illinois loans between July 21, 2020 and
August 8, 2020 and 19.8% of Cross River’s Illinois round 3 loans are for $20,000, compared
to 1.14% of Cross River’s Illinois loans before July 21, 2020 and 2.14% of Cross River’s
loans in other states. This pattern is particularly suspicious given that the US Census CBP
reports 2,207 “Insurance Agencies and Brokerage Industry” establishments in Cook County,
Illinois, which is about half the number of first draw loans made in this industries by Cross
River alone (4,388 loans, of which 3,328 are for exactly $20,000).25 Further, excluding Cross
River’s loans, there are 1,657 first draw loans to Cook County businesses in this industry,
which is already 75% of the establishment count provided by the CBP. To systematically
look for similar patterns throughout the PPP data, we compare PPP numbers to overall
establishment counts in the 2019 US Census CBP database. Because the CBP data does
not include self-employed and independent contractors as establishments, we exclude these
loans to these business types from our analysis.
Panel A of Figure 7 plots histograms of FinTech (red bars) and traditional (gray bars)
lender loans by ratio of first-draw PPP loans to census establishments in the loan’s county
and industry. For traditional lenders, 72.1% of loans cluster below one, which means they
tend to be in industry-county pairs where the total number of PPP loans is less than the
number of establishments recorded in that industry-county by the CBP. For FinTech lenders
the distribution is shifted to the right, and 58.9% of their loans are in industry-county pairs
where there are more businesses receiving PPP loans than there are establishments reported
by the CBP. The difference between FinTech and traditional lenders is even more striking for
extreme disconnects between loan and establishment counts. For example, 6.20% of FinTech
loans are in industry-county pairs with ratios of PPP first draw loans to CBP establishments
higher than 10 compared to 0.72% of traditional loans. It is possible that some excess PPP
loans may be due to the missing establishments in the CBP data, industry 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 7 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 dramatically as the
loan-to-establishment ratio increases, particularly for FinTech lenders. Whereas 21.46% of
25 These loan counts are specific to Cook County and exclude self-employed and independent contractors because they are not
included in the Census CBP data. Loan counts for “All Other Miscellaneous Crop Farming” also appear to be high, but the CBP
data does not have a comparable establishment count for this industry because it does not include agricultural establishments.

19


FinTech and 9.46% of traditional loans with a loan-to-establishment ratio at or below one
are flagged, the flag rate is 52.63% for FinTech and 17.90% for traditional loans with loanto-establishment ratios above two.
Panel B of Figure 7 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 than listed in the CBP, the number of suspicious loans flagged
increases dramatically 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 high loan-to-establishment ratio being
particularly suspicious.

3.4

Loan Clustering

In addition to exhibiting geographic and industry clustering, many of the Cross River
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. There will clearly be some loan similarities due to lender specialization and by chance, but it is instructive to quantify how many 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).26 Then, we rescale each of the concentration ratios to have
a median of 1,000 and an interquartile range (IQR) of 300.27 The rescaling is done to ensure that the three concentration ratios have similar impacts on our overall concentration
measure. Finally, we average the three concentration ratios for each lender-county pair.
The bars in Panel A of Figure 8 plots the distribution of scaled concentration ratios
separately for FinTech and traditional loans. High concentration ratios are much more
common for FinTech loans, which have an average scaled concentration ratio of 1,367 with
85.15% of loans in lender-counties with a scaled concentration ratio above 1,000, compared to
an average of 966 for traditional loans with 21.35% of loans in lender-counties with a scaled
concentration ratio above 1,000. The dots in Panel A of Figure 8 plot how the incidence of
the four primary suspicious loan flags changes with concentration ratio. When the scaled
26

PnFor 2example, let i = 1, 2, ..., n represent the n industries in a given lender-county pair, then Concentrationindustry =
s 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
i=1 i
that this concentration ratio is the same as a Herfindahl-Hirschman Index (HHI), which is commonly used to measure market
concentration.
27 Rescaled Concentration

Concentrationindustry −Median[Concentrationindustry ]
industry = 75th Percentile[Concentration
]−25th Percentile[Concentration
industry

20

industry ]

∗ 300 + 1000.


concentration ratio is below 1,000, approximately 16% and 10% of FinTech and traditional
loans, respectively, have at least one flag. However, when scaled concentration ratio is above
1,300, this grows to 48% for FinTech loans and 20% for traditional loans. Panel B of Figure 8
shows that similar patterns hold for each of the four suspicious loan flags individually. The
overall pattern is similar to the previous measures: FinTech lenders have much higher loan
concentration ratios, and high concentration ratios are highly predictive of the suspicious
loan flags, particularly for FinTech loans. This pattern is exactly what one would expect
if the indicators are picking up misreported FinTech loans and is difficult to explain with
innocent mistakes or errors in the data.

3.5

Criminal Records

Recidivism statistics show that individuals with past criminal histories are much more
likely to commit crimes in the future (Alper et al. 2018). The PPP originally prohibited
loans to businesses more than 20% owned by individuals currently subject to criminal charges,
incarceration, probation, or parole or who had been convicted of a felony within the past five
year. 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. 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. Nonetheless, a criminal record
is still a fraud risk factor, especially when combined with other risk flags. To assess the
prevalence of criminal records among PPP borrowers, we collect the 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.
Panel A of Figure 9 plots the percentage of borrowers with felony criminal records in 2000–
2020 within the sample of 150,000 individual borrowers for whom we collected background
information.28 Ninety-five percent confidence intervals based on standard errors clustered
by zip code and lender are plotted on top of the bars. The first takeaway from the figure
is that criminal records are much more common among FinTech borrowers than traditional
borrowers. Whereas 4.92% of non-bank FinTech borrowers and 4.55% of online bank FinTech
borrowers have criminal records, only 1.36% of traditional borrowers have criminal records.
There is also a strong relation between criminal records and the other suspicious loan flags for
FinTech lenders, but not for traditional lenders. For example, FinTech borrowers with one
of the primary suspicious loan flags have an elevated criminal record rate of 7.44%, whereas
28 Panel A of Figure IA.4 replicates this figure using felonies from 2015-2020. While the percentage of borrowers with felonies
is lower across the board, the relative results remain.

21


traditional borrowers with one of the primary flags have a criminal record rate of 1.79% which
is similar to the overall average for traditional borrowers.29 In the Internet Appendix, we
confirm that these relations are robust and statistically significant by regressing an indicator
for having a criminal record on the other primary and secondary risk flags for loans originated
by FinTech lenders.30
Panel B of Figure 9 examines how criminal records vary across lenders with a clear positive
relation between the percentage of a lender’s sampled borrowers with criminal records and
the percentage of its overall loans with one of the primary suspicious loan flags. In particular,
the four lenders with the highest criminal record percentages (MBE, Cross River, Fundbox,
and Kabbage, all of which are FinTech) also have the highest primary flag rates.31

3.6

Relation Between Primary and Secondary Flags

We have already seen that the primary flags are strongly predictive of one another, and
the evidence in Figures 5–9 show strong relations between the primary and secondary flags.
In Table IV, we more formally assess these relations with regression analysis controlling for
loan size and number of jobs with zip code, business type, industry × CBSA, and lender
fixed effects. The dependent variable in the regressions is an indicator variable for the loan
having at least one of the primary flags. Standard errors are double clustered by zip code
and lender. The secondary flags are a) whether the implied compensation of the loan is
within $1,000 of the maximum allowed amount of $100,000, b) whether the loan amount
is rounded to an interval of $1,250 (which represents $500 in monthly compensation) but
does not have implied compensation within $1,000 of $100,000, c) whether the loan is in a
industry-county pair where there are more first draw loans than establishments per the US
Census CBP, d) whether the loan is in a lender-county pair where the concentration ratio
of industry, loan amount, and jobs reported is above the 75th percentile, and e) whether
the borrower has a felony charge on their criminal record from 2000-2020. The secondary
flags are all interacted with an indicator variable for FinTech loans, so the direct coefficients
represent effects for traditional loans. Four of these five effects are positive and significant
29 The primary loan flags for this analysis are multiple loans, high implied compensation, and EIDL > PPP Jobs because the
business registry flag is only relevant to corporate and LLC loans. Because the criminal background analysis is for loans made
to individuals, loans in this sample cannot have the business registry flag.
30 Results are reported in Table IA.IV. The regressions control for loan size and number of jobs with business type, industry
× CBSA, and lender fixed effects. Standard errors are double-clustered by zip code and lender. In all cases except monthly
rounding, the coefficients are positive, statistically significant, and economically large for FinTech loans (Panel A) with almost
no relation between the misreporting indicators and criminal records for traditional loans (Panel B).
31 Panel B of Figure IA.4 replicates this figure using felonies post-2005, post-2010, and post-2015. While the percentage
of borrowers with felonies decreases as the time period is decreased, the relative results remain. Additionally, Panel C of
Figure IA.4 replicates this figure using bankruptcy fillings post-2015 and finds similar results. Lastly, Panel D of Figure IA.4
shows that there is an uptick in the percentage of borrowers with felonies as the implied compensation of the loan approaches
$100k.

22


with magnitudes ranging from 4.6% to 20.0% of mean misreporting rate.
Even more importantly, all of the interactions between the secondary flags and the indicator for FinTech loans are large, positive, and significant.32 As a result, all of the secondary flags strongly predict the primary flags for FinTech loans, with relations that are
much stronger than for traditional loans. For rounded compensation and compensation near
$100,000, the effect for FinTech is over five times the effect for traditional loans. Further,
for high loan concentration and felony, the effects for FinTech are over 1.75 times as large
as the non-FinTech effects. Lastly, for industry overrepresentation, there are strong FinTech
effects despite essentially zero relations for traditional loans. In Figure IA.5 we examine
relations between the primary and secondary flags at the lender level with similar results.
Apart from monthly rounding, lenders with high levels of each secondary flags tend to be
the same lenders that have high levels of the primary flags. Overall, the strong relations
between the primary and secondary flags, particularly for FinTech lenders make it highly
unlikely that the differences in flag rates across lenders are driven by honest mistakes or
errors in the data.

3.7

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. Panel A of Figure 10 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, 1,848,329 loans representing 17.3% of the PPP and totaling $76.3B are
suspicious.33
FinTech lenders are responsible for a disproportionate share of suspicious loans. Combined, non-bank FinTech and online bank FinTech originated 963,868 suspicious FinTech
loans totaling $21.3B. This means FinTech lenders originated 52.1% of flagged loans despite,
substantially outpacing their overall FinTech 28.7% market share of loans.34 As a share of
loans originated by each lender type, 11.6% of traditional loans have at least one of the
primary suspicious loan flags compared to 32.0% for non-bank FinTech and 30.6% for online
bank FinTech.
While some of the loans flagged as suspicious by the primary measure may be honest
32 The prevalence of the secondary flags is also higher for FinTech loans (see Table IA.V).
33 In addition, the EIDL > PPP flag also provides an indication of misreporting in the EIDL and EIDL Advance program; in
particular, 186,149 EIDL Advances (9.6% of those matched to a PPP loan), totaling $1.6B, have potential misreporting.
34 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 overall market share, which is 28.7% of PPP loans and 10.9% of PPP dollar
lending volume.

23


mistakes or errors in the data, these four measures also surely miss many fraudulent loans.
This is particularly true for the business registry and EIDL > PPP Jobs flags, which only
apply to subsets of loans (corporate/LLC loans and loans with matched EIDL Advances,
respectively). Despite having much higher rates for these flags within the relevant subsets of
loans, overall, these flags are relatively uncommon for FinTech lenders because most of their
loans are to individuals and sole proprietorships without EIDL Advances. As a more lenient
measure of suspicious lending that is less sensitive to these restrictions, Panel A of Figure IA.6
plots suspicious loan rates including all loans with any primary or secondary flag. By this
measure, 5,223,126 loans totaling $283B are suspicious. FinTech is again overrepresented
with 2,089,656 suspicious FinTech loans (40% of flagged loans) totaling $50.3B.
As a more conservative estimate, we consider loans that have at least one primary flag
plus an additional primary or secondary flag. While this measure almost certainly misses
a lot of misreporting, it has the benefit of dropping any honest mistakes or errors in the
data that are isolated to a single measure. Under this conservative measure, 1,200,580 loans
totaling $37.9B are suspicious. Of these loans, 786,859 ($17.0B) are FinTech. This is an
even larger FinTech share than for the primary measure because 81.6% of FinTech loans
flagged by at least one primary flag are further confirmed by an additional flag while the
corresponding figure is only 46.8% for traditional loans. The higher confirmation rate for
FinTech loans is consistent with FinTech loans being far more likely to be fraudulent as
opposed to simply reflecting honest explanations or errors in the data.
The last three bars of Panel A plot suspicious lending rates by round with the clear pattern
that suspicious lending increased over time. In round 1, 8.7% are suspicious, compared to
12.0% in round 2 and 23.2% in round 3. The conservative measure with an additional
confirmatory flag follows the same pattern.
In Panel B of Figure 10, we plot suspicious loan rates by lender. The total height of the
bars plot the percent of loans with at least one primary flag, and the solid part of the bars
plot the percent of loans with a primary flag that is confirmed with an additional primary or
secondary flag. Average rates for the two measures are plotted as solid and dashed horizontal
lines, respectively. Disparities across lenders are striking. Using the at least one primary
flag measure, 13 out of 20 FinTech lenders have above average suspicious loan rates, and
the 10 lenders with the most suspicious loans (nine of which are FinTech) all have at least
28.7% of their loans implicated compared to the overall average of 17.3%. In the extreme,
Capital Plus and Prestamos CDFI have primary flag rates of 51.7% and 46.5%, respectively.
Even with the more conservative measure, requiring an additional primary or secondary flag,
these two lenders have flag rates of 50.9% and 45.0%. Capital Plus is particularly striking
24


because it is the second largest FinTech lender and fourth largest lender overall with 380,377
loans. Cross River (largest FinTech lender and second largest overall lender with 471,609
loans) and Harvest (third largest FinTech lender and fifth largest overall lender with 377,620
loans) are also well above the average flag rate with primary flag rates of 28.7% and 38.2%,
respectively. While most of the FinTech lenders cluster among the lenders with the most
suspicious loans, there are a few exceptions. In particular, Square, Intuit, and Capital One
have misreporting rates that are well under the average misreporting rates across all lenders.

3.8

Geography of Suspicious Lending

In addition to varying across lenders, suspicious lending also varies across geographies.
Panel A of Figure 11 plots the percent of loans with at least one primary flag in each county
across the U.S with considerable variation.35 Areas with a particularly high percentage
of flagged loans cluster near New Orleans, Atlanta, and surrounding areas in Mississippi,
Georgia. Chicago and parts of South Carolina also exhibit elevated levels. Many counties in
these areas have suspicious lending rates in excess of 35% whereas large parts of the country
have suspicious loan rates under 15%. The geographic pattern is somewhat regional with
elevated fraud rates in the Southeast, but there are elevated counties scattered across the
country. There are also big differences across large cities. For example, Cook County, IL
has a suspicious loan rate of 35.5% compared to suspicious loan rates of 8.9% in New York
County and 9.4% on Los Angeles County.
In Panel B of Figure 11, we examine the relation between FinTech market share and
suspicious loan rates across counties and zip codes. Each dot represents a zip code. The xaxis plots the percent of loans flagged in at the county level, and the y-axis plots the percent
of loans flagged at the zip code level. There is significant variation across zip codes within
counties with flagged loan rates varying from 20% to 50% in many counties. Additionally,
FinTech market share (represented by the color of the dots) is strongly related to the percent
of flagged loans not only across counties, but also across zip codes within counties—zip codes
with the highest flagged loan rates consistently have the highest FinTech market share.36
Is geographic variation in suspicious lending related to poverty, crime, or culture? Or
does suspicious lending cluster in ways beyond that exceed what can be explained by cultural
and criminal factors? Table V further analyzes the geography of suspicious PPP lending by
considering relations with demographic and cultural measures that are associated with other
forms of financial misconduct (Grullon et al. 2010; Parsons et al. 2018; Griffin et al. 2019b).
35 Panel A Figure IA.7 shows geographic variation in FinTech market share.
36 Within a county, a 10 ppt rise in FinTech market share in a zip code is associated a 4.27 ppt rise in suspicious lending (see

Table IA.VI).

25


The dependent variable is an indicator for whether a loan is flagged by at least one primary
flag and the explanatory variables are county-level cultural and demographic measures.37 In
column (1), public corruption convictions and religious affiliation, have a positive relation
with the probability of a loan being flagged as suspicious and usage of a marital infidelity
website (Ashley Madison) has a negative relation. The strongest relation is for the public
corruption measure. A one standard deviation increase in per capita public corruption
convictions is associated with a 1.0 ppt increase in the suspicious loan rate, which is 5.9%
of the mean. The other two cultural variables are statistically significant but economically
much less important. In column (2), we add county-level demographic control variables for
population density, median income, percentage of non-white population, percentage of adults
who are college educated, and pre-pandemic unemployment. In general, suspicious lending
rates decrease with population density, median income and increase with percentage of nonwhite population and pre-pandemic unemployment. Coefficients on all but percentage of
non-white population are economically small.
In column (3), we add county-level FinTech market share. A one standard deviation
increase (13.6 ppt) in FinTech market share in the county is associated with a 2.4 ppt
increase in the suspicious loan rate, which is 12.6% of the mean misreporting rate.38 This
is a much stronger relation than any of the other county variables, and coefficients for the
cultural and demographic variables generally decrease or become statistically insignificant
once FinTech market share is added to the regression. It is important to note that this
regression specification is not conducive to a causal interpretation but indicates a strong
association.
Why does suspicious lending vary so much across geographies? Strong clustering in
certain counties and zip codes suggests that suspicious borrowing is driven by more than
just the idiosyncratic decisions of individual borrowers. One possibility is that referral fee
programs, agent fees, kickback schemes, or local networks may arise in certain areas to
systematically attract and facilitate suspicious lending.39 Because we do not observe the
37 The regressions are at the loan level to control for jobs reported, loan size, business type and industry code × state fixed
effects. Standard errors are double clustered by zip code and lender. The independent variables are standardized to have mean
of 0 and standard deviation of 1 at the county-level. Thus, the coefficients can be interpreted as the change in the probability
of a loan being flagged for a one standard deviation change in the variable. Table IA.VII shows equivalent regressions at the
county-level.
38 An equivalent regression at the county-level shows that a one standard deviation increase in FinTech market share is
associated with a 0.586 standard deviation increase (3.9 ppt) in suspicious loan rate (see Table IA.VII).
39 For example, Amur Equipment offered a referral fee program and explicitly stated that the referral program required “zero-touch and follow up on your end.” (https://twitter.com/Amur_EF/status/1361801770452672515).
The
PPP allowed lenders to pay agents 1% on loans up to $350k, 0.5% on loans between $350k and $2M, and 0.25% on
loans above $2M (https://home.treasury.gov/system/files/136/PPP%20Lender%20Information%20Fact%20Sheet.pdf).
Further, some people filed PPP loans in return for upfront and backend fees or kickbacks, which was against SBA
rules (e.g., see https://www.justice.gov/usao-ndil/pr/suburban-chicago-tax-preparer-charged-covid-relief-fraud,
https://www.justice.gov/criminal-fraud/file/1380216/download,
and
https://www.justice.gov/usao-sdfl/pr/

26


identity of agents directing or assisting the PPP borrowers, this possibility is difficult to
directly test. That said, the geographic clustering of suspicious loans shown in Figure 11
is what we would expect if agents in certain areas are encouraging suspicious borrowing
and steering suspicious borrowers to FinTech lenders. If agents are utilizing more than one
FinTech lender to clear fraudulent loans, one might expect counties with many potentially
misreported loans by one lender to have elevated levels of suspicious loans and FinTech
more generally. Consistent with this hypothesis, the lower triangle of Figure IA.8 indicates
that that when one FinTech lender has a high flagged loan rate in a county, other FinTech
lenders also tend to have elevated flagged loan rates. The upper triangular shows that
FinTech lenders with the highest overall flagged loan rates have a high correlation in their
market shares at the county-level, whereas market-share correlations between most other
lenders are slightly negative.

4

Why Does Suspicious Lending Concentrate in FinTech?

FinTech lenders on average have much higher suspicious lending rates than traditional
lenders, and Tables III, IA.II, and IA.V show that their elevated suspicious lending is not
explained by observable facets of loan composition. What could be driving the elevated flag
rates for FinTech lenders?

4.1

FinTech Lender Background and Incentives

4.1.1

FinTech Lender Background

Differences across FinTech lenders give a first clue to this puzzle. While most FinTech
lenders have high suspicious loan rates, Square and Intuit have among the lowest suspicious
loan rates of all lenders. Online lending does not appear to be the problem in and of itself.
One thing that sets Square and Intuit apart is that they have established relationships with
customers based on a broad suite of payment, accounting, payroll, and other financial support
services.
By contrast, the largest FinTech PPP lender, Cross River, is a small community bank in
New Jersey that acts as conduit for partner FinTechs. Similarly, Capital Plus Financial, the
second largest FinTech PPP lender, is a small mortgage lender in Texas that traditionally focused on supporting Hispanic home ownership but now appears to be almost entirely focused
on PPP lending. The number three FinTech lender, Harvest Small Business Finance, is also
southern-district-florida-takes-sweeping-action-against-cares-act-fraud).

27


a small lender with limited history. Now that the PPP has ended, Harvest’s only current
product appears to be SBA 7(a) commercial real estate loans. Benworth, Fountainhead, and
Itria, the other FinTech lenders in the top 10 by number of PPP loans originated, follow a
similar pattern with high suspicious flag rates and limited business outside of PPP lending.
Table IA.VIII systematically examines the relation between suspicious loan flag rates and
past SBA lending with a similar finding. 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 lending (in particular new FinTechs) all have higher rates of flagged loans.
The six largest FinTech lenders all primarily originated loans that were sourced from other
FinTech platforms. Cross River adopted this business model early in round 1 by partnering
with other FinTechs such as Intuit and Kabbage to originate PPP loans (New York Times
2020). The other large FinTech lenders almost exclusively originated loans sourced by two
marketing FinTechs that did not do any PPP lending until round 3, Womply and BlueAcorn
(New York Times 2021). 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 Capital Plus and Prestamos CDFI. Both firms relied heavily
on online advertising promoting easy access to PPP money, and anecdotal discussions on
forums such as Reddit suggest that both employed identity verification that was not rigorous.
4.1.2

FinTech Fluidity

To examine relationships between these lenders, Panel A, Figure 12 plots the network
of relationships between lenders based on originating loans in the same draw to the same
non-commercial address as identified by the multiple loan flag. The edges between lenders
represent the number of addresses to which both lenders originated a loan within the same
draw. Node size is based on the number of loans at addresses flagged for having multiple
loans. The thicker edges between FinTech lenders shows that FinTech borrowers receiving
multiple loans often received funds from more than one FinTech lender even within the same
draw. Specifically, 66.0% of FinTech borrowers with multiple loans to the same address split
their loans across multiple lenders. Shared FinTech lending to the same address is largely
explained by FinTech portals sourcing loans for multiple lenders. For example, the plot
shows a strong relationship between Prestamos CDFI and Capital Plus, the two lenders that
partnered with BlueAcorn. This is likely from borrowers applying for multiple loans through
BlueAcorn, some of which were originated by Prestamos CDFI while others were directed to
Capital Plus. Similarly, there are strong relationships between Harvest, Benworth, Capital
28


Plus, and Fountainhead, all of which are Womply partners. By contrast, there are fewer
relationships between traditional lenders, and 70% percent of traditional borrowers who took
out multiple loans received all their loans from the same lender. Some exceptions include
Customers Bank, Amur Equipment, and Bank of America. The first two at least have known
FinTech affiliations even though they do not meet the formal FinTech criteria.40 Overall,
205 thousand FinTech loans are at the same address as a loan originated by a different
FinTech lender with the same draw and 45 thousand by a traditional lender. For traditional
banks, there were only 81 thousand loans at addresses where the borrower received another
loan from a different traditional bank within the same draw and 42 thousand loans from a
FinTech lender.
As another way to examine relationships between lenders, we track borrowers switching
lenders between their first and second PPP loan draws. If a borrower already submitted
information in the first draw through a particular lender, obtaining the second draw only
required refreshing the application with some additional information.41 This provided a
strong incentive for borrowers to use the same lender. For borrowers with first draw loans
flagged for potential misreporting that subsequently took out second draw loans, Panel B
of Figure 12 shows the movement of these loans across rounds to different lenders.42 The
thickness of the edges between lenders is proportional to the number of flagged loans that
changed lenders between the first and second draws. The switches between the first and
second draw are clockwise. Node size is based on the number of first draw loans (with a
matching second draw) and second draw loans by each lender. The large movements and
connections between FinTech lenders may reflect online lending portals switching lenders.
For example, many FinTech originators, such as Intria and Intuit, cleared loans though Cross
River in the first two rounds but became SBA authorized lenders by round 3, and Kabbage
originated many loans for Customers Bank in the first two rounds. Overall, the graphs
highlight the fluid nature of the FinTech space where online portals originating loans can
easily originate their loans through different lenders, and suspicious borrowers can utilize
several platforms or switch platforms. The lack of relationship banking within the FinTech
space may be advantageous to expand access to capital (Erel and Liebersohn 2021), but it
also appears to be expedient for dubious lending.
40 Customers Bank directly worked with multiple FinTech lenders,
in particular Kabbage and Cross River
(https://newsroom.kabbage.com/wp-content/uploads/2020/07/Kabbage-Paycheck-Protection-Program-PPP-Report.pdf).
41 For example, question 55 of the January 29, 2021 PPP FAQ (https://www.sba.gov/sites/default/files/2021-01/PaycheckProtection-Program-Frequently-Asked-Questions.pdf) says “information a lender obtained from a borrower in connection with
a First Draw PPP Loan can be relied upon by that lender for a Second Draw PPP Loan application.”
42 Figure IA.9 replicates this figure using all second draw loans.

29


4.1.3

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 increased to the lesser of 50% or $2,500 for loans below
$50,000.43 Based on this fee schedule, we estimated that PPP lending generated $36.2B of
lender processing fees, $7.2B of which went to FinTech lenders (see Table IA.IX). The top
3 FinTech lenders alone likely generated $2.9B in processing fees, including $1.0B to Cross
River, $926M to Capital Plus, and $925M to Harvest, all of which appear to have little
business other than PPP lending. On a percentage basis, these figures imply that the average
processing fee for FinTech PPP loans was 12.8% of the loan balance, largely driven by the
high processing fees for small loans in round 3.44 We lack data on cost structure associated
with PPP lending and do not observe how lender fees are shared with partner organizations
that FinTech lenders used to source the loans such as Womply and BlueAcorn. Nonetheless,
costs were presumably lower for FinTech lenders than for traditional lenders given their
limited interactions with borrowers, and their number of employees, where observable, are
small.45 The large scale of PPP lending from small and relatively unknown PPP lenders also
suggests that costs were probably limited.46
PPP lenders were required to follow SBA lending guidelines but did not bear any credit
risk themselves. While the lenders were required to collect documentation from loan applicants and to follow Bank Secrecy Act requirements, they were explicitly allowed to rely on
borrower certifications and representations and do not face liability for borrower misstatements.47 Up-front processing fees on a per-loan basis combined with no credit risk potentially
created an incentive for lax underwriting standards, particularly for specialized PPP lenders
43 See https://home.treasury.gov/system/files/136/Updated-Guidance-PPP-Lender-Processing-Fee-Payment-1502-Reporting-Process.

pdf.
44 The average processing fee for round 1 and 2 was 4.9%. This dramatically increased to 20.2% in round 3.
45 For example, Capital Plus (the second largest FinTech lender and fourth largest lender overall) received a PPP loan of
$376,800, reportedly to cover payroll for its 28 employees. The loan was approved in April 2020, potentially before their business
opportunities as a PPP lender, most of which occurred in round 3, were apparent. Similarly, Benworth Capital Partners (fourth
largest FinTech and eighth 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 on April 27, 2020, and Amur Equipment was approved for
a PPP loan of $2,817,846 on May 2, 2020 but then repaid/canceled its loan 12 days later.
46 Accounting guidance from the FDIC (https://www.fdic.gov/coronavirus/smallbusiness/faq-sb.pdf) indicates that
PPP loan processing fees should be recognized over the life of the loan or upon its sale. As a result, it is likely too early
to see most of the impact of PPP processing fees on bank Call Reports. Nonetheless, Cross River’s net profits have already
more than quadrupled to $156M for March 31, 2020 to 2021 compared to $34M for March 31, 2019 to 2020 per their Call Reports
filled with the FFIEC. Similarly, a financial disclosure by Capital Plus’s parent company shows that it received $464.1M in
PPP fees during the second quarter of 2021 (http://crossroads.mediaroom.com/2021-06-14-Crossroads-Systems-Reports-FiscalSecond-Quarter-2021-Financial-Results?pagetemplate=widgetpopup).
47 See and https://home.treasury.gov/system/files/136/PPP--IFRN%20FINAL.pdf and https://home.treasury.gov/
system/files/136/Paycheck-Protection-Program-Frequently-Asked-Questions.pdf.

30


with little reputation risk to other business interests.

4.2

Did FinTech Lenders Improve Standards Over Time?

While a full exploration of the incentives and culpability of PPP lenders is beyond the
scope of this paper, we can gain insight into the economic structure of PPP lending by
studying patterns of suspicious lending over time. We consider two potential scenarios under
which suspicious lending could arise:
• 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.
• Scenario B: The lender is aware of the existence of or potential for fraud within its PPP
loans but ignores this risk because they are earning fees from the loans and there is
little if any downside to fraud for the lender. This may be particularly true for lenders
with little reputation or other business to protect.
Under scenario A, we would expect that when lenders are new to PPP lending, they may
facilitate questionable loans, but over time as they experience more loans with improbable
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 B, in which the lender
is willing to turn a blind eye or encourage fictitious loans, the amount of suspicious lending
could grow through time as lenders develop a reputation for attracting more suspicious
borrowers.
Scenario A predicts:
1. Loan misreporting will decrease over time as lenders become more aware and develop
systems to screen out suspicious loans.
2. Suspicious borrowers will be less likely to receive a repeat loan from the same lender
compared to other borrowers.
3. Regions with high misreporting in rounds 1 and 2 will face extra scrutiny from lenders,
which will decrease round-3 misreporting.
Scenario B predicts:
1. Loan misreporting will grow over time as borrowers learn about the potential for fraud
and lenders do little to stop it.
2. Borrowers with suspicious first draw loans in rounds 1 and 2 will be able to obtain
second draw loans in round 3 from the same lenders.
31


3. Regions with high misreporting in rounds 1 and 2 will have the same or more misreporting in round 3 because lenders are maximizing loan volume with little regard to
potential fraud.
Did lenders improve their loan screening 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 13
plots more granular suspicious loan rates on a weekly basis separately for non-bank FinTech,
online bank FinTech, and traditional lenders. For the FinTech 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, with around 40% of loans
flagged as suspicious in the closing weeks of round 3. Suspicious lending by traditional
lenders also grew over time, but at a much lower rate. FinTech and traditional lenders both
started the PPP with suspicious loan rates of around 10%, but by the end of the program
the FinTech suspicious loan rate was close to 40%, more consistent with scenario B.48
We also examine lending growth and suspicious loan rates in the different rounds by
lender. Panel A of Figure IA.11 shows that almost all lenders had higher suspicious lending
rates in round 3 than in rounds 1 and 2.49 Additionally, many of the FinTech lenders
with the highest suspicious loan rates in round 3 also had the most growth. In particular,
Fountainhead, Harvest, and Itria originated almost all their loans in round 3, and more
than 30% of their loans are flagged. Overall, other than Kabbage which was acquired by
American Express before round 3,50 In Table VI, we regress indicators for the four primary
flags in round 3, individually and combined, on lenders’ rounds 1 and 2 misreporting rates
for the same flags. As in previous regressions, we control for loan size and jobs with zip
code, business type, industry × CBSA fixed effects. For FinTech lenders, we find highly
economically and statistically significant relations across the board with weaker relations
for traditional lenders. In other words, the suspicious lending behavior of FinTech lenders
is highly persistent over time. Overall, there is little evidence that suggest most FinTech
lenders are taking steps to screen out dubious loans.
Regarding prediction 2, if lenders are taking steps to screen out questionable loans, then
48 Panel B of Figure IA.6 shows similar trends for each primary flag individually.
49 Panel B of Figure IA.11 shows that that the growth of misreported loans is also present, and even stronger for some lenders,

when considering only first draw loans.
50 Kabbage had a high suspicious loan rate for rounds 1 and 2 but a lower suspicious loan rate in round 3 that is also
accompanied by lower lending volume. Two other large round 3 FinTech lenders, Capital Plus and Benworth, also have round
3 suspicious lending rates in excess of 35% but are not included in the plot because they did not have enough earlier lending to
calculate a suspicious loan rate for rounds 1 and 2.

32


their borrowers with questionable first draw loans in rounds 1 and 2 may get rejected when
they apply for a second draw loan in round 3. To examine this, we estimate regressions to
determine whether a first draw borrower is more or less likely to receive a second draw loan
from the same lender if its first draw loan is flagged by one of the primary misreporting
indicators. Table VII shows that traditional loans which 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 1.83 ppt (with t-stat of -9.64) and FinTechs have a statistically
insignificant increase of 0.77 (-1.83 + 2.61) ppt (with t-stat of 0.81). 51 This provides
some indication that traditional banks were less likely to continue lending to borrowers with
previous suspicious borrowing, but FinTechs do not seem to be screening or implementing
procedures which make it less likely for questionable borrowers to continue receiving funds
in the form of a second draw. Columns (3) and (4) of Table VII condition on the borrower
receiving a second draw (either from the same or different lender) with similar results.52
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 Figure 13 plots the percentage of
loans flagged in rounds 1 and 2 in each zip code on the x-axis, and the percentage of flagged
loans in round 3 in the same zip code on the y-axis. The left subpanel uses all loans, middle
uses FinTech loans, and right uses traditional loans. Each dot represents a zip code, and the
size of the dots corresponds to the number of loans in the zip code. Purple to blue colors
indicate that a zip code had fewer loans in round 3 than in rounds 1 and 2, light blue to
light yellow colors indicate that a zip code had about the same number of loans in round
3 compared to rounds 1 and 2, and orange to red colors correspond to an increase in the
number of loans in the zip code. Darker colors correspond to higher degrees of decline or
growth.
The figure displays three interesting findings. First, most zip codes (90.9%) are above
the 45-degree line in the left subpanel, indicating that misreporting rates increased in round
3 almost everywhere. In many zip codes (34.3%), the percentage of loans that are flagged
as suspicious in round 3 is more than twice as high as in rounds 1 and 2. Second, the zip
codes with the highest suspicious loan rates experienced the most growth in lending. Many
zip codes with the highest level of flagged loans in round 3 have more than three times
the number of loans in round 3 compared to rounds 1 and 2, suggesting that significant
51 These results are based on Same Lender being set to 0 if the borrower did not get a second draw at all. Columns (3) and
i
(4) of Table VII condition on the borrower receiving a second draw loan and show similar results.
52 In Figure IA.10, 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
towards flagged and nonflagged loans rather than systematic changes in the lender’s behavior. For most traditional lenders,
borrowers with a flagged first-draw loan are less likely to receive a second draw loan from the same lender, but for several
FinTech lenders, suspicious first-draw borrowers are slightly more likely to receive a second draw.

33


portions of zip code level loan growth in round 3 may be due to suspicious lending practices.
Third, the middle and right subpanels differentiate between FinTech and traditional lenders
and show that lending growth and increased misreporting rates are almost entirely from
FinTech lenders. Traditional lenders had only small increases in suspicious loan indicators,
and their lending generally a decreased. In contrast, FinTech lenders increased the number
of loans they originated and increased their suspicious lending rates in almost all zip codes.
Additionally, FinTech growth was highest in zip codes with the highest misreporting rates.53
We also test these results at the zip code-lender level in Table IA.X 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 an 18.9 ppt increase in lending for a FinTech lender and
an insignificant increase of 1.5 ppt for a traditional lender. There is also strong persistence
of suspicious lending across rounds within zip code-lender pairs.

4.3

Repayments and Enforcement Actions

The economics of crime depend crucially on a crime’s expected penalty and probability
of detection (Becker 1968). The US Department of Justice is pursuing criminal complaints
alleging PPP fraud, and some borrowers have voluntarily repaid their loans without applying
for loan forgiveness. However, the magnitude of these enforcement actions is tiny. Compared
to the 1.8 million loans we identify as suspicious, the DOJ has publicized 162 criminal
complaints regarding only 355 loans, and SBA data indicates that only 16,930 round 1 and
2 loans were repaid between December 1, 2020 and June 30, 2021.54 In Figure IA.12, we
descriptively summarize enforcement actions and repayments that we are able to match to
the PPP loan data with the result that both are elevated for FinTech lenders and for loans
flagged as suspicious by our measures.55 While it is possible that more enforcement actions
will be forthcoming, there appears to be no penalty for most suspicious lending thus far.

5

Conclusion

We examine four primary and five secondary metrics related to potentially misreported
loans. FinTech loans are highly suspicious at a rate of almost five times that for traditional
lenders. Nine of the ten lenders with the highest rates of suspicious loans are FinTech
lenders and the remaining traditional bank (Amur Equipment) is in many ways similar to
a FinTech lender. We estimate the total amount of potential misreporting at 1.8 million
53 Results are similar at the county and state level. See Panels C and D of Figure IA.11. Panel B of Figure IA.7 also plots
lending growth by county.
54 Of the DOJ enforcement action loans with enough data to be matched to the PPP loan level data, 153 loans were originated
by FinTech lenders and 126 were originated by traditional 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.
55 Regressions in Table IA.XI also show evidence of elevated repayment and enforcement action rates for flagged loans.

34


loans with a balance of $76.3 billion based on the four primary metrics, and $37.9 billion
(1.2 million loans) under a more conservative estimate requiring an additional indicator.
The total amount of misreporting is likely larger than either estimate because many of our
indicators are only available for a subset of loans, and 5.2 million loans with a balance of
$283 billion are flagged at least one of our nine indicators. In the early stages of the PPP,
about 10% of FinTech loans were potentially misreported, but the percentage of suspicious
FinTech loans increased to more than 40% by the end of round 3. These findings highlight
the large costs of low oversight and lack of sufficient negative ramifications to borrowers and
lenders for poor lending practices in the PPP.
Our findings have important practical policy implications. First, with the focus on rapid
distribution of funds, the PPP did not include robust verification requirements. Traditional
banks may have been more apt to follow standard practices anyway. The lack of rigorous
verification seems to have led to substantial costs to taxpayers. Second, FinTech lending,
though quite successful at adapting to new environments and quickly disbursing funds, 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 evidence, along with convincing evidence that the PPP saved relatively few jobs at an
extremely high cost (Autor et al. 2020; Chetty et al. 2020; Granja et al. 2020), provides
growing evidence that the PPP seems to have been a poor allocation of capital. Fourth,
incentives in the PPP appear misaligned in that FinTech lenders made billions of dollars
dispersing loans with widespread indicators of misreporting. While there are limitations to
what our data and analysis can discern, the sheer scope of the tens and hundreds of thousands
of suspicious loans originated by many FinTech lenders suggests that many lenders either
encouraged such loans, turned a blind eye to them, or had extremely lax oversight procedures.
Finally, the increasing scope of the misreporting through time indicates that current
penalty and enforcement systems are not effective. If the system is not changed, the most
likely outcome is more of the same. This paper is also an example of how forensic finance
research can more fully investigate the rent-seeking dimension of finance (Zingales 2015).
Government agencies can assist 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.

35


References
Alper, Mariel, Matthew R. Durose, and Joshua Markman, 2018, 2018 update on prisoner
recidivism: A 9-year follow-up period (2005-2014), Special Report by Bureau of Justice
Statistics: Office of Justice Programs.
Amiram, Dan, and Daniel Rabetti, 2020, The relevance of relationship lending in times of
crisis, Working Paper.
Autor, David, David Choz, Leland D. Cranez, Mita Goldarx, Byron Lutzz, Joshua Montesz,
William B. Petermanz, David Ratnerz, Daniel Villarz, and Ahu Yildirmazx, 2020, An
evaluation of the Paycheck Protection Program using administrative payroll microdata,
Working Paper.
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
expectations, Proceedings of the National Academy of Sciences 117, 17656–17666.
Bartlett, Robert, Adair Morse, Richard Stanton, and Nancy Wallace, 2021, Consumerlending discrimination in the FinTech era, Journal of Financial Economics.
Becker, Gary S., 1968, Crime and punishment: An economic approach, Journal of Political
Economy 76, 169–217.
Beggs, William, and Thuong Harvison, 2021, Fraud and abuse in the Paycheck Protection
Program? Evidence from investment advisory firms, Working Paper.
Ben-David, Itzhak, 2011, Financial constraints and inflated home prices during the real
estate boom, American Economic Journal: Applied Economics 3, 55–87.
Bloomberg Businessweek, 2020, PPP scammers made fintech companies their lenders of
choice, Bloomberg Businessweek.
Buchak, Greg, Gregor Matvos, Tomasz Piskorski, and Amit Seru, 2018, Fintech, regulatory
arbitrage, and the rise of shadow banks, Journal of Financial Economics 130, 453–483.
Chetty, Raj, John N. Friedman, Michael Hendren, Nathaniel amd Stepner, and The Opportunity Insights Team, 2020, The economic impacts of COVID-19: Evidence from a new
public database built using private sector data, Working Paper.
Duchin, Ran, Xiumin Martin, Roni Michaely, and Ivy Wang, 2021, Concierge treatment
from banks: Evidence from the Paycheck Protection Program, Working Paper.
Erel, Isil, and Jack Liebersohn, 2021, Does fintech substitute for banks? Evidence from the
Paycheck Protection Program, Working Paper.
Fuster, Andreas, Matthew Plosser, Philipp Schnabl, and James Vickery, 2019, The role of
technology in mortgage lending, The Review of Financial Studies 32, 1854–1899.
Garmaise, Mark J., 2015, Borrower misreporting and loan performance, The Journal of
Finance 70, 449–484.

36


Glancy, David, 2021, Bank relationships and the geography of PPP lending, Working Paper.
Gopal, Manasa, and Philipp Schnabl, 2020, The rise of finance companies and fintech lenders
in small business lending, Working Paper.
Granja, João, Christos Makridis, Constantine Yannelis, and Eric Zwick, 2020, Did the Paycheck Protection Program hit the target?, Working Paper.
Griffin, John M., Samuel Kruger, and Gonzalo Maturana, 2019a, Do labor markets discipline? Evidence from RMBS bankers, Journal of Financial Economics 133, 726–750.
Griffin, John M., Samuel Kruger, and Gonzalo Maturana, 2019b, Personal infidelity and
professional conduct in 4 settings, Proceedings of the National Academy of Sciences 116,
16268–16273.
Griffin, John M., and Gonzalo Maturana, 2016, Who facilitated misreporting in securitized
loans?, The Review of Financial Studies 20, 384–419.
Grullon, Gustavo, George Kanatas, and James Weston, 2010, Religion and corporate
(mis)behavior, Working Paper.
Howell, Sabrina T., Theresa Kuchler, and Johannes Stroebe, 2020, Which lenders had the
highest minority share among their Payment Protection Program (PPP) loans?, Research
Note.
Jagtiani, Julapa, and Catharine Lemieux, 2019, The roles of alternative data and machine
learning in fintech lending: Evidence from the LendingClub consumer platform, Financial
Management 48, 1009–1029.
Jiang, Wei, Ashlyn A. Nelson, and Edward Vytlacil, 2014, Liar’s loan? effects of origination
channel and information falsification on mortgage delinquency, The Review of Economics
and Statistics 96, 1–18.
Keys, Benjamin J., Tanmoy Mukherjee, Amit Seru, and Vikrant Vig, 2010, Did securitization
lead to lax screening? Evidence from subprime loans, The Quarterly Journal of Economics
125, 307–362.
Kruger, Samuel, and Gonzalo Maturana, 2020, Collateral misreporting in the residential
mortgage-backed security market, Management Science 67, 2657–3320.
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, Forthcoming.
Miami Herald, 2020, Copycat websites, newly minted companies, double dippers cashed in
on PPP loans, Miami Herald.
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.

37


Neilson, Christopher, John Eric Humphries, and Gabriel Ulyssea, 2020, Information frictions
and access to the Paycheck Protection Program, Journal of Public Economics 190.
New York Times, 2020, The tiny bank that got pandemic aid to 100,000 small businesses,
New York Times.
New York Times, 2021, How two start-ups reaped billions in fees on small business relief
loans, New York Times.
Parsons, Christopher A., Johan Sulaeman, and Sheridan Titman, 2018, The geography of
financial misconduct, The Jorrnal of Finance 73, 2087–2137.
Piskorski, Tomsaz, Amit Seru, and James Witkin, 2015, Asset quality misrepresentation by
financial intermediaries: Evidence from the RMBS market, The Journal of Finance 70,
2635–2678.
Project on Government Oversight, 2020, Lamborghinis, strip clubs, bogus companies, and
lies: The first 56 Paycheck Protection Program fraud cases, Project on Government Oversight.
ProPublica, 2021, Hundreds of PPP loans went to fake farms in absurd places, ProPublica.
Purnanandam, Amiyatosh, 2011, Originate-to-distribute model and the subprime mortgage
crisis, Review of Financial Studies 24, 1181–1915.
The Wall Street Journal, 2020, Evidence of PPP fraud mounts, officials say, The Wall Street
Journal.
Zingales, Luigi, 2015, Presidential address: Does finance benefit society?, The Journal of
Finance 70, 1327–63.

38


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 Round 1,
2, and 3 of the PPP on the left axis and the total number of loans originated each week on the the
right axis. In both panels, red represents non-bank fintech lenders, cream represents online bank
fintech lenders, and grey 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.
Panel A. Number of Loans and Dollar Value of Loans, by Lender (Top 75)

Panel B. Fintech Market Share, by Week

39


Figure 2. Business Registry and Multiple Loans Flags
This figure shows the prevalence of the business registry and multiple loans flagged loans by lender.
Panel A shows the percentage of loans flagged for being incorporated after February 15, 2020 (“Late
Incorporation/Filling”), being dissolved and inactive before approved for a PPP loan (“Dissolved
Business”), or not being found in the business registry for its home state or in any other state while
listing an address in its home state (“Missing Business”). Panel B shows the percentage of loans
flagged as being located at a non-business, non-central (e.g., not an apartment or office building)
address that received more than one loan within the given loan’s draw (i.e., the first or second
draw). For Panel A, only loans to businesses organized as a corporation, subchapter S corporation,
or LLC and not based in Illinois or a territory are considered; lenders originating at least 10,000
loans fitting these criteria are shown. For Panel B, all loans are considered and lenders originating
at least 15,000 loans are shown. In both panels, red represents non-bank fintech lenders, cream
represents online bank fintech lenders, and grey represents traditional lenders.
Panel A. Business Registry Flagged Loans, by Lender

Panel B. Multiple Loans Flagged Loans, by Lender

40


Figure 3. High Implied Compensation Flag
This figure shows the relation between a loan’s implied compensation per employee and the average
compensation in the loan’s industry (represented by NAICS [North American Industry Classification System] code) and region (represented by CBSA [core-based statistical area]). We define
normalized compensation by dividing the implied compensation of the loan by the average compensation in the loan’s industry-CBSA. Panel A shows the relation between normalized compensation
and the business registry (top subpanel) and multiple loans flags (bottom subpanel). Panel B shows
the percentage of loans with normalized compensation above 3 (i.e., implied compensation is more
than three times the NAICS/CBSA average). For Panel A, the left axis (distribution) shows the
kernel density of loans and the right axis (dots) shows the percentage of flagged loans in each bin,
where each bin is 0.2 units wide. The solid lines are third-degree polynomial fits for the percentage
flagged and the dashed lines are the 95% confidence intervals. For the business registry subpanel,
only loans to businesses organized as a corporation, subchapter S corporation, or LLC and not
based in Illinois are considered. For Panel B, only loans where the average compensation in the
loan’s industry-CBSA is less than $33,333.33 are considered; lenders with at least 5,000 loans fitting
this criterion are shown.
Panel A. Business Registry and Multiple Loans Flags, by Normalized Compensation

Panel B. High Implied Compensation, by Lender

41


Figure 4. EIDL > PPP Jobs Flag
This figure shows the difference between the number of employees implied by a business’s EIDL
Advance amount (“EIDL Implied Jobs”) and the number of jobs reported by the business on its
PPP application (“PPP Reported Jobs”). Panel A shows the lower bound (in the absolute value
sense) of the difference between the EIDL implied jobs and PPP reported jobs by lender type.
Panel B shows the percentage of loans by each lender where the EIDL implied jobs is at least
three more than the PPP reported jobs. In both panels, only loans with a matched EIDL Advance
are consider; lenders with at least 5,000 loans fitting this criterion are shown in Panel B. In both
panels, red represents non-bank fintech lenders, cream represents online bank fintech lenders, and
grey represents traditional lenders.
Panel A. Difference Between EIDL Implied Jobs and PPP Reported Jobs

Panel B. EIDL > PPP Jobs, by Lender

42


Figure 5. Discontinuities at $100,000
This figure shows the prevalence of various flags by the implied compensation per employee. Panel
A shows the prevalence of the business registry flag, Panel B the multiple loans flag, Panel C the
high implied compensation flag, and Panel D the EIDL > PPP jobs flag. For all panels, loans are
binned into $2,000 wide bins (i.e., ($0k, $2k], ... , ($98k, $100k], ... ,($128, $130k]), the left axis
shows the percentage of loans that in each bin (bars) and the right axis shows the percentage of the
loans in the bin that have the given flag (dots). Loans are filtered to corporation, S-corporation,
and LLC loans for Panel A, loans for which we can determine CBSA/NAICS average compensation
for Panel C, and loans with a matched EIDL Advance for Panel D. The solid lines are third-degree
polynomial fits (weighted based on number of loans in the each bin), which are separately fitted
for loans below $98,000 and loans above $100,000, and the dashed lines are the 95% confidence
intervals. Red represent fintech loans and grey represent traditional loans.
Panel A. Business Registry, by Implied Compensation

Panel B. Multiple Loans

Panel C. High Implied Compensation

Panel D. EIDL > PPP Jobs

43


Figure 6. Rounded Compensation
This figure shows the prevalence of each flags by whether the total monthly implied compensation
of a loan is rounded to an interval of $500 (i.e., loan amount is within ± 50 cents of an interval
of $1,250). Panel A shows the business registry flag, Panel B the multiple loans flag, Panel C the
high implied compensation flag, and Panel D the EIDL > PPP jobs flag. The top subpanel is for
fintech loans and the bottom for traditional loans. For all panels, the last four digits of the loan
amount is considered (i.e., $123,456.78 → $3,456.78). 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
percentage of loans that are flagged within each $1 bins for monthly rounded (solid dots) and $250
wide bins for non-rounded (hollow dots). Loans are filtered to corporation, S-corporation, and LLC
loans for Panel A, loans with CBSA/NAICS average compensation less than $33,333.33 for Panel
C, and loans with a matched EIDL Advance for Panel D. 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 from all Panels. The solid lines are third-degree polynomial fits for the
percentage flagged in the non-rounded bins and the dashed lines are the 95% confidence intervals.
Panel A. Business Registry Flag, by Lender Type and Rounding

Panel B. Multiple Loans

Panel C. High Implied Compensation

Panel D. EIDL > PPP Jobs

44


Figure 7. Overrepresentation of Industries in Counties
This figure shows overrepresenation of loans to industry-county pairs. We define normalized number
of loans as the number of first draw loans divided by the number of establishments (per the 2019 US
Census County Business Patterns dataset) in an industry (represented by NAICS [North American
Industry Classification System] code) and county. Panel A shows the relationship between normalized number of loans and our four main flags combined together as at least one flag and Panel B
shows the relationship for each flag separately. Since the CBP does not include self-employed and
independent contractors as establishments, we exclude 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.
Panel A. Percentage Flagged, by Normalized Number of Loans in Industry-County Pair

Panel B. Individual Flags, by Normalized Number of Loans in Industry-County Pair

45


Figure 8. Clustering Within Lenders and Counties
This figure shows clustering of loans within lenders-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, rescaled 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,
P
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 ]
∗ 300 + 1000. Panel A shows the rela75th Percentile[Concentrationindustry ]−25th Percentile[Concentrationindustry ]
tionship between the average rescaled concentration ratio and our four main flags combined together
as at least one flag and Panel B shows the relationship 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% traditional loans are outside the average rescaled concentration ratio range 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 each 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.
Panel A. Percentage Flagged, by Average Rescaled Concentration Ratio in Lender-County Pair

Panel B. Individual Flags, by Average Rescaled Concentration Ratio in Lender-County Pair

46


Figure 9. Criminal Records
This figure shows criminal records for a sample of 150,000 Round 1 and 2 loans to self-employed
individuals, independent contractors, and sole-proprietorships. Panel A shows the percentage of
loans where the borrower has a felony from 2000 or after on their record by lender type and whether
the loan has various features. The error bars denote 95% confidence intervals (based on standard
errors clustered by zip code and lender) for each percentage. Panel B shows the relationship
between the percentage of loans in this sample that are flagged by at least one primary flags and
the percentage of borrowers that have a felony from 2000 or after on their record by lender. Lenders
with at least 0.2% of the sample (300 loans) are shown. The dashed line is a linear fit and correlation
is shown in the bottom left corner. In both panels, red represents non-bank fintech lenders, cream
represents online bank fintech lenders, and grey represents traditional lenders.
Panel A. Percentage Felony, by Lender Type and Presence of Misreporting Indicators

Panel B. Percentage Flagged vs. Felonies

47


Figure 10. Lender Level
This figure shows the variation in percentage of loans flagged. Panel A shows the percentage and
dollar amounts of flagged loans by lender type and round. Panel B shows the percentage of flagged
loans by lender for the top 75 lenders (by number of loans). In both panels, the plain solid filled in
section represents the percentage of loans flagged by one primary flag and an additional flag (either
another primary or a secondary) and the entire bar (plain and stripped combined) represents loans
flagged by at least one primary flag. In Panel A, the set of numbers to the left of each bar represent
the number of loans and dollar value of loans flagged one primary flag and an additional flag and
the set on top of each bar by at least one primary. The markers within each bar represent the
percentage of loans flagged by each of the primary flags. In Panel B, the two horizontal lines
represent the overall sample percentage of loans flagged by each measure (dashed for loans flagged
by one primary flag and an additional flag and solid by at least one primary). In both panels,
red represents non-bank fintech lenders, cream represents online bank fintech lenders, and grey
represents traditional lenders.
Panel A. Percentage of Loans Flagged, by Lender Type and Rounds

Panel B. Percentage of Loans Flagged, by Lender

48


Figure 11. Geography of Flagged Loans
This figure shows the geographic variation in flagged loans. Panel A shows the percentage of flagged
loans in each county and Panel B shows within county variation. In Panel A, counties are colored
based on the color scheme shown in the bar to the right of the map and counties with fewer than
100 loans are colored grey. Panel B shows the percentage of flagged loans in each zip code on the
vertical axis and the percentage of flagged loans in the corresponding county on the horizontal axis.
Dots are colored by the percentage of fintech loans in each zip code and sized based on the number
of loans in the zip code. Zip codes with at least 100 loans are shown. The dashed line is a linear
fit and the correlation is shown in at the bottom left corner.
Panel A. Percentage of Flagged Loans, by County

Panel B. Within County Variation

49


Figure 12. Lender Network
This figure shows connections between lenders. Panel A shows connections between lenders that
were used by borrowers at the same address within the same draw. Panel B shows connections
between lenders used by the flagged first draw borrowers across draws. In Panel A, node size is
proportional to the number of loans with the multiple loans at the same address flag originated by
each lender, edges are not directed, and edge width is proportional to the number of addresses that
used both lenders. In Panel B, node size is proportional to the number of first draw loans (that
also got a second draw loan from the same or different lender) and second draws originated by the
lender, edges are directed, edge width is proportional to the number of flagged first draw borrowers
moving clockwise from the first draw lender to the second draw lender. In both panels, red nodes
are fintech lenders and grey nodes are traditional lenders. Further, pure red edges are between
two fintech lenders, pure grey edges are between two traditional lenders, and darker red edges are
between a fintech and traditional lender. Top 100 lenders (by the ssme measure that node size is
based on) are shown and the remainder are combined into the “Other" nodes (one for other fintech
lenders and one for other traditional lenders).
Panel A. Multiple Lenders at Same Address

Panel B. Changes Between Draws


Figure 13. Persistence and Growth Across Rounds
This figure shows the persistence and growth of flagged loans across the PPP 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 Panels
B, the percentage of loans flagged in rounds 1 and 2 are shown on the horizontal axis and round 3
on the vertical axis. For Panel A, the vertical dotted lines split each subpanel into the three PPP
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 (either another primary or a
secondary). For Panel B, the left subpanel uses all loans, middle uses fintech loans, and right uses
traditional loans. Zip codes with at least 100 loans and, for the fintech and traditional subpanels,
25 loans by the given lender type are shown. The black line is a 45-degree line and the correlation
is presented in the bottom of each panel. The circle size corresponds to the number of loans in the
zip code by the given lender type and color corresponds the growth/decline in lending in the zip
code.
Panel A. By Lender Type

Panel B. By Zip Code

51


Exhibit 1. Examples of Suspicious Loans
This exhibit shows some examples of suspicious loans.
Panel A. 14 Loans to The Same Address, 13 Incorporated Late
Business Name
(Redacted)

Date
Incorporated

Date
Approved

Lender

Loan Amount

Jobs
Reported

FDML
JTBCL
GGITL
EIEL
FML
DYJNL
CTWIL
CAYL
BLNL
ATYL
STWL
KYHUL
LTTBTL
DREL

4/2/2018
8/5/2020
7/30/2020
7/22/2020
7/27/2020
7/22/2020
7/23/2020
7/21/2020
7/19/2020
7/19/2020
7/16/2020
7/15/2020
7/15/2020
7/7/2020

5/13/2020
8/5/2020
8/1/2020
7/30/2020
7/30/2020
7/26/2020
7/23/2020
7/22/2020
7/21/2020
7/21/2020
7/17/2020
7/16/2020
7/15/2020
7/8/2020

Celtic
Indep. Artists, Writers, Performers
Kabbage
Misc. Schools & Instruction
Kabbage
Misc. Schools & Instruction
Kabbage
Misc. Schools & Instruction
Kabbage
Misc. Schools & Instruction
Kabbage
Misc. Schools & Instruction
Kabbage
Misc. Schools & Instruction
Kabbage
Misc. Schools & Instruction
Kabbage
Misc. Schools & Instruction
Kabbage
Misc. Schools & Instruction
Kabbage
Misc. Schools & Instruction
Kabbage
Misc. Schools & Instruction
Kabbage
Misc. Schools & Instruction
Kabbage
Musical Groups & Artists

$62,083
$53,229
$53,229
$53,229
$53,229
$53,229
$53,229
$53,229
$53,229
$53,229
$53,229
$91,770
$53,229
$53,125

10
10
10
10
10
10
10
10
10
10
10
10
10
10

Individual Name
(Redacted)

Age

Date
Approved

Lender

Industry

Loan Amount

Jobs
Reported

A. P.
O. P.
G. P.
T. P.

20
21
46
49

7/15/2020
7/10/2020
7/15/2020
7/10/2020

Kabbage
Kabbage
Kabbage
Kabbage

Nail Salons
Lawn/Garden Equipment Manuf.
Other Automotive Repair
Lawn/Garden Equipment Manuf.

$20,833
$20,833
$20,833
$20,833

1
1
1
1

Industry

Panel B. Four Loans to Same Household

Exhibit 2. Cross River Case Study
This exhibit shows examples of $20,000 first draw loans by Cross River Bank in Illinois.
Individual Name
(Redacted)

Date
Approved

Lender

Industry

Loan Amount

Jobs
Reported

J. C.
R. J.

7/29/2020
7/29/2020

Cross River
Cross River

Insurance Agencies and Brokerages
Insurance Agencies and Brokerages

$20,000
$20,000

1
1

4,304 $20k Loans by Cross River in Illinois (Mostly in Chicago Area) to Individuals/Businesses
(98% with 1 Employee) in "Insurance Agencies and Brokerages" Industry.
C.C.
M. A.

7/29/2020
7/29/2020

Cross River
Cross River

All Other Miscellaneous Crop Farming
All Other Miscellaneous Crop Farming

$20,000
$20,000

8
8

938 $20k Loans by Cross River in Illinois (Mostly in Chicago Area) to Individuals/Businesses
(56% with 8 Employee and 22& with 1 Employee) in "All Other Miscellaneous Crop Farming" Industry.
C.M.
K. K.

8/7/2020
7/30/2020

Cross River
Cross River

Other General Government Support
All Other Miscellaneous Manufacturing

$20,000
$20,000

50
9

3,056 $20k Loans by Cross River in Illinois (Mostly in Chicago Area) to Individuals/Businesses
in Various Industries (Including 700 to “All Other Personal Services,” 347 to “General Freight Trucking, Local,”
337 to “Other Performing Arts Companies”, and 229 to “New Single-Family Housing Construction.”)

52


Table I. Summary Statistics
This table presents summary statistics for our sample. The sample includes all PPP loans approved
from the start of the program (March 2020) through most of Round 3 (April 2021) that have not
been repaid as of May 3, 2021. Fintech lenders are determined following Erel and Liebersohn
(2021). 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 from the US Census
CBP and Normalized Comp. is the ratio of Implied Comp. and CBSA/NAICS Avg. Comp.. 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, Frac.
Matched EIDL Advance is the percentage of loans with a matching EIDL Advance, and Frac. $10k
EIDL Adv. (Given Matched) is the percentage of loans with a matching EIDL Advances that are
for $10,000 (the maximum possible). Frac. Fintech (Either Type), Frac. Non-bank Fintech, and
Frac. Online Bank Fintech are the percentages of loans that are originated by the given type of
lender.
Fintech
Mean
Num. Loans [Pct. Loans]
Loan Amount
Jobs Reported
Implied Comp.
CBSA/NAICS Avg. Comp
Normalized Comp.
Num. Loans (Within Draw) at Address
Frac. Corp, S Corp, LLC
Frac. Second Draw (Given Round 3)
Frac. Matched EIDL Advance
Frac. $10k EIDL Adv. (Given Matched)

Num. Loans [Pct. Loans]
Loan Amount
Jobs Reported
Implied Comp.
CBSA/NAICS Avg. Comp
Normalized Comp.
Num. Loans (Within Draw) at Address
Frac. Fintech (Either Type)
Frac. Non-bank Fintech
Frac. Online Bank Fintech
Frac. Corp, S Corp, LLC
Frac. Second Draw
Frac. Matched EIDL Advance
Frac. $10k EIDL Adv. (Given Matched)

SD

Traditional
Median

3,071,586 [28.7%]
27,708 112,175
18,750
2.842
11.151
1.000
61,598
40,664
62,840
47,288
38,702
37,453
1.826
1.593
1.338
1.314
0.816
1.000
0.228
0.259
0.202
0.228

Mean

SD

Median

7,625,633 [71.3%]
91,427 308,121 20,833
10.483
29.507
3.000
47,008
67,228
38,775
49,727
37,343
42,800
1.142
1.979
0.917
1.192
0.789
1.000
0.658
0.601
0.270
0.257

Round 1

Round 2

Round 3

1,619,446 [15.1%]
198,875
20.470
48,024
49,342
1.178
1.275
0.048
0.025
0.023
0.829
0.292
0.385

3,523,411 [32.9%]
57,852
7.912
43,841
51,829
1.055
1.208
0.204
0.075
0.129
0.656
0.301
0.178

5,554,362 [51.9%]
46,160
4.976
56,788
47,012
1.609
1.247
0.409
0.265
0.145
0.372
0.461
0.206
0.262

53


Table II. Odds Ratios
In this table, we present the odds ratios between each of our four main indicators. Panel A shows
the odds ratios for fintech and traditional loans combined. Panel B shows the odds ratios for fintech
loans only in the lower triangular and traditional loans only in the upper triangular. Note that
odds ratios are symmetric, which is why only values for the lower triangular are provided. Robust
standard errors are double clustered by zip code and lender.
Panel A. Fintech and Traditional Loans Combined
Business
Registry
Business Registry

Multiple
Loans

High Implied
Comp.

EIDL >
PPP Jobs

-

Multiple Loans

1.607***
(6.52)

-

High Implied Comp.

2.572***
(8.75)

2.830***
(14.47)

-

EIDL > PPP Jobs

1.403***
(6.07)

3.027***
(10.80)

14.270***
(15.20)

-

Panel B. Fintech Loans (Lower Triangular) and Traditional Loans (Upper Triangular)
Business
Registry

Multiple
Loans

High Implied
Comp.

EIDL >
PPP Jobs

-

1.262***
(10.64)

1.836***
(11.28)

1.177***
(6.09)

Multiple Loans

2.349***
(12.42)

-

1.550***
(4.58)

1.405***
(7.53)

High Implied Comp.

3.548***
(11.48)

1.886***
(4.81)

-

5.636***
(20.11)

EIDL > PPP Jobs

1.964***
(5.45)

2.606***
(11.75)

17.240***
(27.14)

-

Business Registry

z-statistics in parentheses
* p<0.10, ** p<0.05, *** p<0.010

54


Table III. Prevalence of Flags by Lender Type
This table presents the percentage of loans flagged by the four main flags, at least one of 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, and
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 jobs
and loan size controls, and column (5) shows the differences between matched pairs of fintech and
traditional loans. The N values show is 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 differences and adjusted differences are reported in
Panel A and Panel B, respectively, of Table IA.II.
(1)

(2)

(3)
Unadjusted
Difference

(4)
Adjusted
Difference

(5)
Matched
Difference

Fintech

Traditional

Business
Registry

0.0837
N = 671,298

0.0427
N = 4,777,168

0.0409***
(3.12)

0.0276***
(3.47)

0.0204***
(3.20)

Multiple
Loans

0.119
N = 3,071,586

0.0513
N = 7,625,633

0.0681***
(7.16)

0.0290***
(3.85)

0.0479***
(2.68)

High Implied
Comp.

0.478
N = 1,215,857

0.102
N = 2,081,163

0.376***
(6.27)

0.0943***
(5.74)

0.0776***
(2.99)

EIDL > PPP
Jobs

0.194
N = 621,785

0.0478
N = 2,057,611

0.146***
(3.80)

0.0590***
(3.76)

0.0686***
(3.98)

At Least One
Flag

0.314
N = 3,071,586

0.116
N = 7,625,633

0.198***
(6.19)

0.0733***
(6.31)

0.0682***
(5.21)

At Least Two
Flags

0.0503
N = 3,071,586

0.00605
N = 7,625,633

0.0442***
(5.42)

0.0164***
(5.90)

0.0259***
(3.89)

t-statistics in parentheses
* p<0.10, ** p<0.05, *** p<0.010

55


Table IV. Secondary Flags
In this table, we examine the relationship 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 five secondary flags as independent variables. Each specification
also include an interaction between the secondary flag and an indicator for whether the loan was
originated by a fintech lender. $100k Implied Comp. is a dummy variable equal 1 if the implied
compensation per job is within ± $1,000 of $100,000. Monthly Rounding is a dummy variable equal
1 if the loan amount is within ± 50 cents of an interval of $1,250. Overrep. in County/NAICS is a
dummy variable equal 1 if the number of first draw loans to businesses not listed as self-employed
and independent contractors in a loan’s county/NAICS pair exceeds the number of establishments
in the county/NAICS according to the US Census CBP. High Concentration is a dummy variable
equal 1 if the average rescaled concentration ratio in the loan’s county/lender pair is above the
75th percentile. Felony Post-2000 is a dummy variable equal 1 if the borrower has a felony on
their criminal record from 2000 or after. 1(Fintech) is a dummy variable equal 1 if the loan was
originated by a fintech lender. 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 bottom of each column. Robust
standard errors are double clustered by zip code and lender.
Dep. Variable: At Least One Flag
(1)
$100k Implied Comp.

(2)

(3)

(4)

0.0258**
(2.29)

Monthly Rounding

0.00467***
(5.60)

Overrep. in County/NAICS

-0.00225
(-0.72)

High Concentration

0.0207***
(5.96)

Felony Post-2000

× 1(Fintech)

(5)

0.0387***
(3.12)
0.118***
(11.19)

0.0187***
(2.74)

0.0509***
(6.86)

0.0166***
(2.75)

0.0292*
(1.73)

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

Yes
Yes
Yes
Yes
Yes
Yes

Yes
Yes
Yes
Yes
Yes
Yes

Yes
Yes
Yes
Yes
Yes
Yes

Yes
Yes
Yes
Yes
Yes
Yes

Yes
Yes
No
Yes
Yes
Yes

Observations
Num. Lenders
R2
Mean of Dep. Variable

10,043,853
4,860
0.274
0.176

5,077,092
4,775
0.102
0.080

5,788,553
4,808
0.314
0.180

6,938,580
4,213
0.300
0.192

123,745
2,582
0.281
0.194

t-statistics in parentheses
* p<0.10, ** p<0.05, *** p<0.010

56


Table V. County Cultural Features
In this table, we examine the relationship between our four main flags, which we combine to form
At Least One Flag, and cultural/regional features. We estimate OLS regressions with At Least One
Flag as the dependent variable and the cultural/regional features as independent variables. All
independent variables are rescaled at the county level to have a mean of 0 and standard deviation
of 1. Political Corruption is the number of public corruption convictions per million residents in
2004-13 (as reported by the DOJ). Religious Adherence is the percent of the county’s population
with a religious affiliation as of 2010 (as reported by the Association of Religious Data Archives).
Ashley Madison Usage is the paid Ashley Madison usage rate in the county (as reported by Griffin
et al. (2019b)). Population Density is the population per square mile as of 2019, Median Income
is the median household income as of 2019, Pct. Non-White is the percentage of the population
that is not white as of 2019, College Educated is the percentage adults with a bachelor’s degree or
higher as of 2015-19, and 2019 Unemployment is the unemployment rate as of 2019 (all from the
Economic Research Service of the U.S. Department of Agriculture). Pct. Fintech is the percentage
of PPP loans in the county originated by a fintech lender. Fixed effects are as indicated at bottom
of each column. Robust standard errors are double clustered by zip code and lender.
Dep. Variable: At Least One Flag
(1)

(2)

(3)

Public Corruption

0.0103***
(5.53)

0.00760***
(4.86)

0.00328***
(2.71)

Religious Affiliation

0.00180*
(1.93)

-0.000870
(-0.82)

0.0000323
(0.03)

Ashley Madison Usage

-0.00505***
(-7.36)

0.00393**
(2.30)

-0.00224
(-1.56)

Population Density

-0.00160***
(-6.26)

-0.00143***
(-4.93)

Median Income

-0.00219**
(-2.03)

0.000213
(0.22)

Pct. Non-White

0.0188***
(8.20)

-0.00162
(-1.17)

College Educated

-0.00750***
(-4.97)

-0.00325***
(-3.18)

2019 Unemployment

0.00203
(1.40)

0.00292**
(2.10)

Pct. Fintech
ln(Jobs Reported)
ln(Loan Amount)
Business Type FE
NAICS × State FE
Observations
Num. Lenders
R2
Mean of Dep. Variable
t-statistics in parentheses
* p<0.10, ** p<0.05, *** p<0.010

0.0237***
(7.31)
Yes
Yes
Yes
Yes

Yes
Yes
Yes
Yes

Yes
Yes
Yes
Yes

10,589,326
4,886
0.194
0.174

10,589,114
4,886
0.195
0.174

10,589,114
4,886
0.196
0.174


Table VI. 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 individually (specifications (1) through (4)) and at least one of them (specification (5)) as the dependent
variables 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 loam was originated by a fintech or
traditional lender are include in all specification. Specifications (1) through (4), loans are filtered
to the sets for which we can determine the flag (same as in Figures 2-4). Further, to ensure we have
accurate measures of past behavior, we require that each lender have at least 100 loans in Round
1 and 2 (combined) for which we can determine the given flag. Fixed effects are as indicated at
bottom of each column. Robust standard errors are double clustered by zip code and lender.
(1)
Business
Registry

(2)
Multiple
Loans

(3)
High Implied
Comp.

(4)
EIDL >
PPP Jobs

(5)
At Least
One Flag

× 1(Fintech)

0.794***
(3.59)

0.776***
(5.95)

0.369***
(3.23)

0.750***
(3.08)

0.619***
(5.69)

× 1(Traditional)

0.209
(1.07)

0.421***
(3.30)

-0.277*
(-1.89)

-0.111
(-0.93)

0.164*
(1.52)

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

Yes
Yes
Yes
Yes
Yes

Yes
Yes
Yes
Yes
Yes

Yes
Yes
Yes
Yes
Yes

Yes
Yes
Yes
Yes
Yes

Yes
Yes
Yes
Yes
Yes

Observations
Num. Lenders
R2
Mean of Dep. Variable

1,753,188
2,481
0.124
0.065

4,562,655
3,142
0.097
0.082

1,512,743
1,580
0.642
0.326

938,143
1,440
0.317
0.119

4,562,655
3,142
0.307
0.219

Dep. Variable:
Past Pct. This Flag

t-statistics in parentheses
* p<0.10, ** p<0.05, *** p<0.010

58


Table VII. Second Draw
In this table, we examine whether lenders were more/less likely to provide a second draw loan to
a borrower who’s 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 borrower did
not receive a second draw loan, the dependent variable is set to 0, and in specifications (3) and
(4), only borrowers that received both a first and second draw loans are included in the sample. In
the even specification, an interaction with whether the first draw loan was originated by a fintech
lender is included. Fixed effects are as indicated at bottom of each column. Robust standard errors
are double clustered by zip code and lender.
Dep. Variable: 1(First and Second Draw by Same Lender)

First Draw Flagged

(1)
(2)
Unconditional of Receiving
Second Draw

(3)
(4)
Conditional on Receiving
Second Draw

-0.0123***
(-3.68)

-0.00715
(-1.40)

× 1(Fintech)

-0.0183***
(-9.64)
0.0261**
(2.55)

-0.0144***
(-5.86)
0.0294
(1.44)

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

Yes
Yes
Yes
Yes
Yes
Yes

Yes
Yes
Yes
Yes
Yes
Yes

Yes
Yes
Yes
Yes
Yes
Yes

Yes
Yes
Yes
Yes
Yes
Yes

Observations
Num. Lenders
R2
Mean of Dep. Var.

4,853,956
4,742
0.122
0.275

4,853,956
4,742
0.122
0.275

1,552,196
4,523
0.427
0.846

1,552,196
4,523
0.427
0.846

t-statistics in parentheses
* p<0.10, ** p<0.05, *** p<0.010

59


Internet Appendix
Loan Size Calculation
Average monthly payroll expenses are to be based on the full 2019 calendar year for most loans.
Exceptions include seasonal businesses which may use average monthly payroll for any twelve-week
period between February 15, 2019 and February 14, 2020, and new businesses may use average
monthly payroll over the period from January 1, 2020 to February 29, 2020. Additionally, second
draw loans by businesses in the hospitality industry (NAICS starting with 72) were allowed to
receive 3.5 times their average monthly payroll, which we account for while computing implied
compensation. See SBA guidance entitled “How to Calculate First Draw PPP Loan Amounts,”
which was updated over time. A list of all versions can be found at https://www.sba.gov/
document/support-how-calculate-first-draw-ppp-loan-amounts.
In determining the high implied compensation flag and for the $100k discontinuity and rounding
analyses, we subtract any portion of the loan amount used to refinance an EIDL. Second draw loans
to the hospitality industry represent about 2% of loans and we exclude these loans for the rounding
analysis.

Matching Analysis
The matching analysis reported in Table III is based on a combination of propensity score
matching and exact matching. First, we estimate a propensity score for whether a loan is originated
by a FinTech lender using a logistic regression and the following variables: 4-digit NAICS code
(industry), CBSA, business type, loan amount, jobs reported, lending draw, lending round, week
of loan approval, and whether the borrower received an EIDL Advance. Second, for each loan
originated by a FinTech lender, we identify loans made by traditional lenders such that the loans
were made to borrowers in the same industry, CBSA, business type, and year (either 2020 or 2021)
and either both received EIDL Advances or did not. Exactly matching on these characteristics
ensures that we can determine each of our flags for both loans. Finally, among the loans that match
exactly on these features, we match the loans that have smallest absolute difference in propensity
scores. In total, 2,791,287 of 3,071,586 FinTech loans are matched. Note that a traditional loan
may be matched to more than one FinTech loan.

Repayment and Enforcement Action Data
To examine round 1 and 2 loans that have been repaid, we use PPP loan level data released
by the SBA on December 2, 2020 and data from USASpending.gov as of May 15, 2021. The
USASpending data provides monthly updates on the PPP loans, which allows us to observe which
loans have been repaid by the borrower, but does not have all the loan level details. The earlier
version of the PPP loan level data provides the same details as the main PPP data and covers all
loans that had not been repaid as of December 2020.
To examine enforcement actions by the government, we collect information from Department of
Justice crime complaints against PPP borrowers that purportedly committed fraud based on https:
//www.justice.gov/criminal-fraud/cares-act-fraud and https://www.arnoldporter.com/
en/general/cares-act-fraud-tracker. In total, we collect data on 162 complaints involving 355

1


loans. Of these 355 loans, 279 include enough information to be matched to the December 2020
version of the PPP loan level data (most of the unmatched loans were repaid before December 2020
and thus are not in the loan level data).

2


Figure IA.1. Fintech Market Share
This figure shows the role of fintech lenders during the PPP (expanding on Figure 1). Panel A
replicates Figure 1, Panel A based on loans from only rounds 1 and 2. Panel B shows the number
of loans originated by lender type during each week of the PPP. In both panels, red represents nonbank fintech lenders, cream represents online bank fintech lenders, and grey 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.
Panel A. Rounds 1 and 2 Lenders (Top 75)

Panel B. Lender Composition, by Week

3


Figure IA.2. Implied Compensation
This figure shows the distribution of implied compensation. Panel A compares the implied compensation of loans to the average compensation in its industry (NAICS) and CBSA. Panel B shows
the percentage of loans by annualized implied compensation and demonstrates the presence of
rounding. In Panel A, loans are split by round (rounds 1 and 2 in left subpanel and round 3 in
right subpanel) and lender type, sorted based on their CBSA/NAICS average compensation, and
binned into bins of 50,000 (10,000 for fintech in rounds 1 and 2) loans. The median annualized
compensation and CBSA/NAICS average compensation of each bin is shown. The solid line is a
45-degree line. In Panel B, annualized implied compensation is binned it into $1,000 wide bins (e.g.,
$36,500 to $37,500) and loans with annualized implied compensation within ± $1 of an interval
of $1,000 (orange dots) are split from those that are not (blue dots). The dashed lines are every
$12,000 ($1,000 in monthly compensation) and the dotted lines are every $6,000 ($500 in monthly
compensation). Note that 1.6% to 11% on the vertical axis is excluded since there are no data
points in this region.
Panel A. Implied Compensation vs. CBSA/NAICS Average Compensation

Panel B. Rounding of Implied Compensation

4


Figure IA.3. Relation Between Flags
This figure shows the relationship between the primary 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 flag by one of the flags on each axis.
Loans are filtered to the sets for which we can determine each flag (same as in Figures 2-4) for each axis separately (i.e., we do not require
that both flags be able to 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. Red triangles represent non-bank fintech lenders,
cream squares represent online bank fintech lenders, and grey circles represent traditional lenders.

5


Figure IA.4. Criminal Records
This figure shows additional features (expanding on Figure 9) of the criminal records for a sample
of Round 1 and 2 loans. Panel A replicates Figure 9, Panel A using felonies post-2015, Panel B
replicates Figure 9, Panel B using various time cutoffs, Panel C replicates Figure 9, Panel B using
bankruptcies (2015 or after), and Panel D shows the percentage of felonies (2000 or after) by lender
type and across implied compensation. In Panels B and C, lenders with at least 0.2% of the loans
in the sample (300 loans) are shown. The dashed lines are linear fits and correlations are in the
bottom corner. In Panel D, loans are binned into $4,000 wide bins, solid lines are third-degree
polynomial fits, and the dashed lines are 95% confidence intervals.
Panel A. Percentage Felony, by Lender Type and Various Features (2015 or After)

Panel B. Percentage Flagged vs. Felony, Different Time Cutoffs

6


Panel C. Percentage Flagged vs. Bankruptcies

Panel D. Percentage with Felony, by Implied Compensation

7


Figure IA.5. Secondary Flags
This figure shows the relationship between the secondary flags and the percentage flagged by
lender. For maxed loans, loans with null/zero jobs reported are excluded; for monthly rounding,
loans with null/zero or one job reported are exclude; for overrepresentation, loans to self-employed
and independent contractors, second draw loans, and loans in a county/NAICS not in the CBP
are excluded; for high HHI, second draw loans and loans in a county/lender with fewer than 25
first draw loans are excluded. No loans are excluded for percentage flagged. Lenders with at least
5,000 loans are shown. The dashed line is a linear fit and the correlation is in the bottom corner of
each panel. Red triangles represent non-bank fintech lenders, cream squares represent online bank
fintech lenders, and grey circles represent traditional lenders.

8


Figure IA.6. Lender Level
This figure shows additional variation in percentage of loans flagged. Panel A replicates Figure 10,
Panel A and adds loans flagged by at least one primary or secondary flag as the dashed, non-shaded
portions of the bar. In Panel B, each subpanel shows a lender type and each series is the percentage
of loans flagged by the given flag across time. The vertical dotted lines split each subpanel into the
three PPP lending rounds. The loans used to calculate each series are filtered to the sets for which
we can determine each flag (same as in Figures 2-4). Note that mid-August through December
2020 is not shown in Panel B since no PPP loans were originated during this period. In both
panels, red represents non-bank fintech lenders, cream represents online bank fintech lenders, and
grey represents traditional lenders.
Panel A. Percentage of Loans Flagged with Broader Flags, by Lender

Panel B. Percentage of Loans Flagged Over Time, by Lender Type

9


Figure IA.7. Geography
This figure shows additional geographic variation (extending Figure 11). Panel A shows the fintech
market share in each county. Panel B shows the growth in lending between rounds 1-2 and round
3 in each county. In both panels, counties are colored based on the color scheme shown in the bar
to the right of the maps and counties with fewer than 100 loans are colored grey.
Panel A. Fintech Market Share, by County

Panel B. Lending Growth, by County

10


Figure IA.8. Pairwise Lender Correlations
This figure shows pairwise lender correlations using data at the county-lender level. The lower
triangular shows correlations between the percentage of loans flagged by at least one primary flag
and the upper triangular shows correlations between the lenders’ market shares across counties.
Lenders are order such that those with the highest percentage of flagged loans (across the entire
sample) are at the top on the vertical axis and the left on the horizontal axis. Labels are colored red
for non-bank fintechs, orange for online bank fintechs, and black for traditional lenders. Coloring
of each square in the matrix is based pairwise correlation and the coloring scheme is shown at
the bottom with darker red representing higher positive correlation and darker blue representing
higher negative correlation. For each pairwise correlation, counties are filtered to the set that have
25 loans by both lenders. The top 75 lenders (by number of loans) are shown.

11


Figure IA.9. Lender Network
This figure replicates Panel B of Figure 12 using all loans where the borrower received a first and
second draw loan. Node size is proportional to the number of first draw loans (which also received
either a second draw from the same or different lender) and second draw originated by the lender.
Edges are directed and have a width proportional to the number of loans moving clockwise from
the first draw lender to the second draw lender. Red nodes are fintech lenders and grey nodes are
traditional lenders. Pure red edges are between two fintech lenders, pure grey edges are between
two traditional lenders, and darker red edges are between a fintech and traditional lender. Top 100
first draw lenders are shown and the remainder are combined into the “Other" nodes (one for other
fintech lenders and one for other traditional lenders).

12


Figure IA.10. Second Draw Loans
This figure examines whether lenders were more/less likely to provide a second draw loan to a
borrower who’s first draw loan is flagged by at least one of our primary flags. Similar to Table VII,
we estimate a OLS regression 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 interacted with an indicator for each lender as the
independent variable. The regression includes controls for loan size and jobs and zip code, business
type, NAICS × CBSA, and lender fixed effects. For Panel A, if a borrower did not receive a second
draw loan, the dependent variable is set to 0, and in Panel B, only borrowers that received both a
first and second draw loans are included in the sample. In both panels, the hollow dots show the
point estimates from the regression and the error bars show 95% confidence intervals corrected for
multiple comparisons using a Bonferroni correction. Lenders with at least 10,000 loans in rounds
1-2 are considered. Red error bars and labels represents non-bank fintech lenders, cream represents
online bank fintech lenders, and grey represents traditional lenders.
Panel A. Unconditional of Receiving Second Draw

Panel B. Conditional on Receiving Second Draw

13


Figure IA.11. Persistence and Growth Across Rounds
This figure shows the persistence and growth of flagged loans across the PPP lending rounds. Panel
A shows this by lender using all loans and Panel B by lender using first draw loans only. Panel C
and D shows this by county and state (by zip code is shown as Panel B of Figure 13). For all panels,
the percentage of loans flagged in rounds 1 and 2 are shown on the horizontal axis and round 3 on
the vertical axis. For Panel A, lenders with at least 1,000 loans in rounds 1 and 2 combined and
in round 3 are shown. For Panel B, lenders with at least 1,000 loans in rounds 1 and 2 combined
and 250 first draw loans in round 3 are shown. In Panel C, counties with at least 100 loans in
round 1 and 2 combined and in round 3 are shown. In all panels, the circle size corresponds to
the total number of loans, the black line is a 45-degree line, and the correlation is presented in the
bottom of each panel. For Panels A and B, the percentage of the circle that is shaded represents
the proportion of loans that each lender provided/in each state in round 3 relative to in round 1
and 2. For Panel C and D, the circles are colored based lending growth in the county/state with
the color scheme shown in the bar to the right of each panel.
Panel A. By Lender

Panel B. By Lender, First Draw Loans Only

14


Panel C. By County

Panel D. By State

15


Figure IA.12. Repayment
This figure shows the percentage of first draw loans that have been repaid between December 1,
2020 and June 30, 2021 (left axis, blue bars) and that are part of a DOJ enforcement action (right
axis, purple bars). The first draw sample is filtered based on the criteria at the bottom of each set
of bars. In total, this figure is based on 16,930 loans that were repaid between December 1, 2020
and June 30, 2021 and 279 loans that are part of DOJ enforcement actions.

16


Table IA.I. Cross Verification of Flags
In this table, we examine the relationships between our four main flags. We estimate OLS regressions with dummies for each of our main flags as dependent variables and dummies for the other
three flags as independent variables. Panel A shows the relationships without lender fixed effects
and Panel B shows the relationship within lenders by adding lender fixed effects. In both panels,
specification (1) uses business registry as the dependent variable, (2) uses multiple loans, (3) uses
high implied compensation, and (4) uses EIDL > PPP jobs. Loans are filtered to the sets for which
we can determine the flag being used as the dependent variable (same as in Figures 2-4). Fixed
effects are as indicated at bottom of each column. Robust standard errors are double clustered by
zip code and lender.
Panel A. Without Lender Fixed Effect

Dep. Variable:

(1)
Business
Registry

Business Registry

(2)
Multiple
Loans

(3)
High Implied
Comp.

(4)
EIDL > PPP
Jobs

0.00872***
(4.81)

0.0211***
(9.17)

0.00281*
(1.89)

0.0293***
(6.54)

0.0307***
(11.32)

Multiple Loans

0.0138***
(5.47)

High Implied Comp.

0.0363***
(6.87)

0.0243***
(8.65)

EIDL > PPP Jobs

-0.00545***
(-4.34)

0.0217***
(11.65)

0.0701***
(18.96)

ln(Jobs Reported)
ln(Loan Amount)
Zip Code FE
Business Type FE
NAICS × CBSA FE
1(EIDL Adv. Matched)
Lender FE

Yes
Yes
Yes
Yes
Yes
Yes
No

Yes
Yes
Yes
Yes
Yes
Yes
No

Yes
Yes
Yes
Yes
Yes
Yes
No

Yes
Yes
Yes
Yes
Yes
No
No

Observations
Num. Lenders
R2
Mean of Dep. Variable

5,192,997
4,731
0.086
0.048

10,043,880
4,885
0.088
0.069

3,291,095
4,779
0.597
0.241

2,545,320
4,736
0.261
0.083

t-statistics in parentheses
* p<0.10, ** p<0.05, *** p<0.010

17

0.215***
(16.30)


Panel B. With Lender Fixed Effect

Dep. Variable:

(1)
Business
Registry

Business Registry

(2)
Multiple
Loans

(3)
High Implied
Comp.

(4)
EIDL > PPP
Jobs

0.00762***
(5.52)

0.0117***
(3.86)

-0.000433
(-0.26)

0.0192***
(5.46)

0.0192***
(6.78)

Multiple Loans

0.00960***
(6.94)

High Implied Comp.

0.0281***
(8.67)

0.0166***
(7.31)

EIDL > PPP Jobs

-0.00559***
(-4.30)

0.0182***
(10.05)

0.0589***
(14.13)

ln(Jobs Reported)
ln(Loan Amount)
Zip Code FE
Business Type FE
NAICS × CBSA FE
1(EIDL Adv. Matched)
Lender FE

Yes
Yes
Yes
Yes
Yes
Yes
Yes

Yes
Yes
Yes
Yes
Yes
Yes
Yes

Yes
Yes
Yes
Yes
Yes
Yes
Yes

Yes
Yes
Yes
Yes
Yes
No
Yes

Observations
Num. Lenders
R2
Mean of Dep. Variable

5,192,915
4,654
0.099
0.048

10,043,853
4,860
0.095
0.069

3,291,009
4,695
0.615
0.241

2,545,208
4,631
0.290
0.083

t-statistics in parentheses
* p<0.10, ** p<0.05, *** p<0.010

18

0.175***
(17.52)


Table IA.II. Prevalence of Flags by Lender Types
Panel A shows the full results for the unadjusted and adjusted differences provided in Table III.
Fixed effects are as indicated at bottom of each column. Robust standard errors are double clustered
by zip code and lender.
Panel A. Unadjusted Percentages
(1)
Business
Registry

(2)
Multiple
Loans

(3)
High
Comp.

(4)
EIDL >
PPP Jobs

(5)
At Least
One

(6)
At Least
Two

Fintech

0.0409***
(3.12)

0.0681***
(7.16)

0.376***
(6.27)

0.146***
(3.80)

0.198***
(6.19)

0.0442***
(5.42)

Observations
Num. Lenders
R2
Mean of Dep. Var.

5,448,466
4,771
0.004
0.048

10,697,211
4,902
0.014
0.071

3,297,020
4,779
0.180
0.241

2,679,396
4,837
0.051
0.082

10,697,211
4,902
0.056
0.173

10,697,211
4,902
0.022
0.019

Dep. Variable:

Panel B. Adjusted Percentages
(1)
Business
Registry

(2)
Multiple
Loans

(3)
High
Comp.

(4)
EIDL >
PPP Jobs

(5)
At Least
One

(6)
At Least
Two

0.0276***
(3.47)

0.0290***
(3.85)

0.0943***
(5.74)

0.0590***
(3.76)

0.0733***
(6.31)

0.0164***
(5.90)

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

Yes
Yes
Yes
Yes
Yes

Yes
Yes
Yes
Yes
Yes

Yes
Yes
Yes
Yes
Yes

Yes
Yes
Yes
Yes
Yes

Yes
Yes
Yes
Yes
Yes

Yes
Yes
Yes
Yes
Yes

Observations
Num. Lenders
R2
Mean of Dep. Var.

5,192,997
4,731
0.083
0.048

10,043,880
4,885
0.089
0.069

3,291,095
4,779
0.601
0.241

2,545,320
4,736
0.250
0.083

Dep. Variable:
Fintech

t-statistics in parentheses
* p<0.10, ** p<0.05, *** p<0.010

19

10,043,880 10,043,880
4,885
4,885
0.260
0.122
0.173
0.019


Table IA.III. Discontinuity at $100k
In this table, we examine the relationship between our four main flags and implied compensation.
We estimate OLS regressions with dummies for each of our main flags as dependent variables
and dummies for $5k wide bins (i.e., ($0k, $5k], ..., ($95k, $100k], ..., ($125k, $130k]) of implied
compensation as independent variables. The dummy variable for the ($0k, $5k] bin is used as
the baseline. Panel A shows the results for fintech loans and Panel B for traditional loans. In
both panels, specification (1) uses business registry as the dependent variable, (2) uses multiple
loans, (3) uses high implied compensation, and (4) uses EIDL > PPP jobs. Loans are filtered to
corporation, S-corporation, and LLC loans for specification (1), all loans for (2), loans for which
we can determine CBSA/NAICS average compensation for (3), and loans with a matched EIDL
Advance for (4). Fixed effects and control variables are as indicated at bottom of each column.
Robust standard errors are double clustered by zip code and lender.
Panel A. Fintech Loans
(1)
Business
Registry

Dep. Variable:
($0k, $5k]
($5k, $10k]
($10k, $15k]
($15k, $20k]
($20k, $25k]
($25k, $30k]
($30k, $35k]
($35k, $40k]
($40k, $45k]
($45k, $50k]
($50k, $55k]
($55k, $60k]
($60k, $65k]
($65k, $70k]
($70k, $75k]
($75k, $80k]
($80k, $85k]
($85k, $90k]
($90k, $95k]
($95k, $100k]
($100k, $105k]
($105k, $110k]
($110k, $115k]
($115k, $120k]
($120k, $125k]
($125k, $130k]

(2)
Multiple
Loans

(3)
High Implied
Comp.

(4)
EIDL > PPP
Jobs

—————————————— Used as Baseline ——————————————
-0.00786** (-2.30)
0.0112*** (3.25)
0.0189** (2.56)
-0.0106** (-2.56)
-0.0154*** (-2.88)
0.0156*** (3.24)
0.0252** (2.20)
-0.0125** (-2.44)
-0.0133** (-2.34)
0.0199*** (3.10)
0.0384*** (2.72)
-0.00352 (-0.55)
-0.0162** (-2.25)
0.0228*** (3.19)
0.0651*** (3.71)
0.000367 (0.05)
-0.0108 (-1.38)
0.0237*** (3.03)
0.0829*** (4.10)
0.00670 (0.82)
-0.0122 (-1.53)
0.0267*** (3.16)
0.0953*** (4.45)
0.0164** (2.20)
-0.0111 (-1.27)
0.0297*** (3.42)
0.104*** (4.50)
0.0298*** (3.54)
-0.00806 (-0.92)
0.0314*** (3.56)
0.126*** (4.95)
0.0286*** (3.24)
-0.00562 (-0.59)
0.0384*** (4.06)
0.123*** (4.75)
0.0544*** (4.93)
-0.000251 (-0.02)
0.0363*** (3.78)
0.149*** (5.33)
0.0476*** (4.71)
-0.00344 (-0.34)
0.0387*** (3.97)
0.171*** (5.86)
0.0587*** (5.35)
0.00184 (0.18)
0.0441*** (4.14)
0.200*** (6.47)
0.0737*** (5.59)
0.00314 (0.28)
0.0463*** (4.20)
0.250*** (7.02)
0.0900*** (6.46)
0.00107 (0.11)
0.0470*** (4.45)
0.310*** (7.92)
0.0987*** (5.67)
0.0116 (1.21)
0.0535*** (5.04)
0.385*** (8.32)
0.127*** (6.28)
0.0139 (1.42)
0.0519*** (4.93)
0.422*** (8.39)
0.146*** (6.77)
0.0112 (1.08)
0.0580*** (5.75)
0.471*** (9.71)
0.177*** (8.01)
0.0178 (1.57)
0.0615*** (6.05)
0.516*** (10.74)
0.246*** (9.93)
0.0433*** (3.07)
0.0638*** (6.00)
0.540*** (11.20)
0.307*** (16.80)
-0.00553 (-0.37)
0.0639*** (2.88)
0.455*** (10.94)
0.115*** (6.20)
-0.00488 (-0.39)
0.0605*** (3.49)
0.445*** (10.57)
0.121*** (6.07)
-0.00838 (-0.64)
0.0562*** (4.85)
0.443*** (9.52)
0.0940*** (5.98)
-0.0134 (-1.00)
0.0526*** (3.50)
0.466*** (10.19)
0.103*** (6.22)
-0.00159 (-0.10)
0.0500*** (3.39)
0.489*** (9.84)
0.0809*** (5.13)
-0.00347 (-0.20)
0.0403*** (3.39)
0.463*** (10.06)
0.0672*** (3.18)

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

Yes
Yes
Yes
Yes
Yes

Yes
Yes
Yes
Yes
Yes

Yes
Yes
Yes
Yes
Yes

Yes
Yes
Yes
Yes
Yes

Observations
Num. Lenders
R2
Mean of Dep. Var.

628,213
77
0.208
0.0831

2,983,832
77
0.093
0.121

2,692,483
77
0.681
0.214

590,845
77
0.443
0.197

t-statistics in parentheses
* p<0.10, ** p<0.05, *** p<0.010

20


Panel B. Traditional Loans
(1)
Business
Registry

Dep. Variable:
($0k, $5k]
($5k, $10k]
($10k, $15k]
($15k, $20k]
($20k, $25k]
($25k, $30k]
($30k, $35k]
($35k, $40k]
($40k, $45k]
($45k, $50k]
($50k, $55k]
($55k, $60k]
($60k, $65k]
($65k, $70k]
($70k, $75k]
($75k, $80k]
($80k, $85k]
($85k, $90k]
($90k, $95k]
($95k, $100k]
($100k, $105k]
($105k, $110k]
($110k, $115k]
($115k, $120k]
($120k, $125k]
($125k, $130k]

(2)
Multiple
Loans

(3)
High Implied
Comp.

(4)
EIDL > PPP
Jobs

—————————————— Used as Baseline ——————————————
-0.00758*** (-4.28)
-0.00459** (-2.51)
-0.00645 (-1.56)
-0.0110** (-2.39)
-0.0119*** (-4.88)
-0.00832*** (-3.43)
-0.00349 (-0.63)
-0.0120** (-2.22)
-0.0137*** (-4.82)
-0.0103*** (-3.63)
0.00511 (0.85)
-0.00718 (-1.39)
-0.0152*** (-4.76)
-0.0129*** (-4.07)
0.0162** (2.50)
-0.00161 (-0.29)
-0.0150*** (-4.12)
-0.0144*** (-4.06)
0.0277*** (3.90)
0.00399 (0.71)
-0.0143*** (-3.68)
-0.0149*** (-4.11)
0.0401*** (5.21)
0.00947 (1.53)
-0.0153*** (-3.61)
-0.0165*** (-4.01)
0.0513*** (6.18)
0.0127** (2.11)
-0.0145*** (-3.27)
-0.0170*** (-4.04)
0.0643*** (7.20)
0.0178*** (2.72)
-0.0137*** (-3.02)
-0.0159*** (-3.62)
0.0771*** (8.11)
0.0196*** (2.93)
-0.0132*** (-2.68)
-0.0178*** (-3.72)
0.0915*** (9.06)
0.0241*** (3.34)
-0.0121** (-2.47)
-0.0175*** (-3.68)
0.108*** (10.13)
0.0314*** (3.98)
-0.0129** (-2.50)
-0.0161*** (-3.34)
0.123*** (10.87)
0.0374*** (4.02)
-0.0128** (-2.40)
-0.0160*** (-3.15)
0.144*** (11.99)
0.0452*** (4.90)
-0.0110** (-2.05)
-0.0160*** (-2.95)
0.160*** (12.81)
0.0393*** (4.92)
-0.0104* (-1.87)
-0.0153*** (-2.73)
0.170*** (13.01)
0.0333*** (4.19)
-0.0109* (-1.94)
-0.0150*** (-2.71)
0.184*** (12.93)
0.0383*** (4.63)
-0.00975* (-1.73)
-0.0133** (-2.25)
0.203*** (12.85)
0.0456*** (5.46)
-0.00757 (-1.33)
-0.0110* (-1.80)
0.218*** (11.78)
0.0531*** (4.89)
0.0148** (2.42)
-0.0149** (-2.39)
0.248*** (10.78)
0.0694*** (3.97)
-0.0128* (-1.83)
-0.0155*** (-2.59)
0.210*** (12.20)
0.0555*** (5.03)
-0.0151** (-2.43)
-0.0142** (-2.20)
0.216*** (14.42)
0.0545*** (6.08)
-0.0130** (-2.01)
-0.0172*** (-2.61)
0.234*** (14.91)
0.0627*** (6.96)
-0.0150** (-2.29)
-0.0150** (-2.32)
0.255*** (16.93)
0.0674*** (6.37)
-0.0151** (-2.25)
-0.0151** (-2.35)
0.271*** (16.44)
0.0629*** (6.45)
-0.0161** (-2.30)
-0.0162** (-2.57)
0.292*** (18.48)
0.0694*** (6.91)

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

Yes
Yes
Yes
Yes
Yes

Yes
Yes
Yes
Yes
Yes

Yes
Yes
Yes
Yes
Yes

Yes
Yes
Yes
Yes
Yes

Observations
Num. Lenders
R2
Mean of Dep. Var.

4,466,915
4,652
0.073
0.0426

6,922,683
4,805
0.088
0.0465

5,735,350
4,774
0.295
0.0360

1,907,883
4,659
0.105
0.0464

t-statistics in parentheses
* p<0.10, ** p<0.05, *** p<0.010

21


Table IA.IV. Criminal Records (Fintech)
In this table, we examine the relationship between each of our flags and the borrower’s criminal
records for loans by fintech lenders. We estimate OLS regressions with a dummy for whether the
borrower has a felony from 2000 or after on their record as the dependent variable and dummies
for whether the loan is flagged by each of our flags individually as the independent variable. Panel
A shows the relationship for fintech loans and Panel B for traditional loans. Loans are filtered
to the sets for which we can determine the flag (same as in Figures 2-9). Note that the business
registry flag is not included since we can only determine criminal records for loans to individuals
while the business registry flag can only be determined for corporations and LLCs. Fixed effects
are as indicated at bottom of each column. Robust standard errors are double clustered by zip
code and lender.
Panel A. Fintech Loans
Dep. Variable: Felony Post-2000
(1)
Multiple Loans at
Address

(2)

(3)

(4)

(5)

(6)

(7)

0.00669**
(2.14)

High Implied
Comp.

0.0198***
(3.18)

EIDL Jobs >
PPP Jobs

0.0675***
(9.66)

$100k Implied.
Comp.

0.0132***
(5.11)

Monthly
Rounding

0.00735
(1.46)

Overrep. in
County/NAICS

0.0105***
(3.14)

High
Concentration

0.00588*
(1.70)

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

Yes
Yes
Yes
Yes
Yes

Yes
Yes
Yes
Yes
Yes

Yes
Yes
Yes
Yes
Yes

Yes
No
Yes
Yes
Yes

Yes
Yes
Yes
Yes
Yes

Yes
Yes
Yes
Yes
Yes

Yes
Yes
Yes
Yes
Yes

Observations
Num. Lenders
R2
Mean of Dep. Var.

54,372
70
0.127
0.047

17,479
45
0.112
0.056

10,626
47
0.185
0.047

54,372
70
0.126
0.047

54,372
70
0.126
0.047

23,737
68
0.150
0.047

52,486
63
0.123
0.048

t-statistics in parentheses
* p<0.10, ** p<0.05, *** p<0.010

22


Panel B. Traditional Loans
Dep. Variable: Felony Post-2000
(1)
Multiple Loans at
Address

(2)

(3)

(4)

(5)

(6)

(7)

0.00269
(1.54)

High Implied
Comp.

0.00712
(1.21)

EIDL Jobs >
PPP Jobs

0.0352*
(1.74)

$100k Implied
Comp.

-0.00167
(-1.32)

Monthly
Rounding

0.00205
(0.73)

Overrep. in
County/NAICS

0.00393
(1.25)

High
Concentration

0.00366
(0.92)

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

Yes
Yes
Yes
Yes
Yes

Yes
Yes
Yes
Yes
Yes

Yes
Yes
Yes
Yes
Yes

Yes
No
Yes
Yes
Yes

Yes
Yes
Yes
Yes
Yes

Yes
Yes
Yes
Yes
Yes

Yes
Yes
Yes
Yes
Yes

Observations
Num. Lenders
R2
Mean of Dep. Var.

64,297
2,470
0.222
0.014

11,345
1,054
0.307
0.017

7,095
603
0.358
0.016

64,297
2,470
0.222
0.014

64,297
2,470
0.222
0.014

32,924
1,649
0.252
0.014

57,872
2,187
0.216
0.013

t-statistics in parentheses
* p<0.10, ** p<0.05, *** p<0.010

23


Table IA.V. Prevalence of Secondary Flags by Lender Types
In this table, we examine the propensity of the secondary flags by lender type. We estimate OLS
regressions with each of the secondary flags as the dependent variable and a dummy for whether
the lender is a fintech as the independent variable. Loans are filtered to the sets for which we can
determine each flag (same as in Figures 5-9). Fixed effects and additional controls are as indicated
at bottom of each column. Robust standard errors are double clustered by zip code and lender.
Panel A. Unadjusted Differences
(1)
$100k
Comp.

(2)
Monthly
Rounding

(3)
Overrep. in
County/NAICS

(4)
High
Concentration

Fintech

0.172***
(4.62)

0.0297**
(2.02)

0.294***
(9.22)

0.400***
(3.58)

0.0334***
(8.64)

Observations
Num. Lenders
R2
Mean of Dep. Var.

10,697,211
4,902
0.054
0.129

5,365,570
4,883
0.001
0.080

5,981,065
4,887
0.070
0.415

7,351,176
4,397
0.210
0.205

150,000
3,657
0.010
0.027

(5)

Dep. Variable:

(5)
Felony

Panel B. Adjusted Differences
(1)
$100k
Comp.

(2)
Monthly
Rounding

(3)
Overrep. in
County/NAICS

(4)
High
Concentration

0.0568***
(3.65)

0.0197
(1.39)

0.123***
(8.74)

0.293***
(3.83)

0.0179***
(8.50)

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

Yes
Yes
Yes
Yes
Yes

Yes
Yes
Yes
Yes
Yes

Yes
Yes
Yes
Yes
No

Yes
Yes
Yes
Yes
Yes

Yes
Yes
No
Yes
Yes

Observations
Num. Lenders
R2
Mean of Dep. Var.

10,043,880
4,885
0.297
0.127

5,077,162
4,839
0.041
0.080

5,975,404
4,886
0.208
0.415

6,939,796
4,232
0.525
0.192

124,476
3,194
0.128
0.027

Dep. Variable:
Fintech

t-statistics in parentheses
* p<0.10, ** p<0.05, *** p<0.010

24

Felony


Table IA.VI. Within State/County Variation
This table examines the degree to which geographic variation in flagged loans can be explained
by fintech market share. We estimate OLS regressions with the percentages of flagged loans in
each zip code as the dependent variable and the fintech markey share in each zip code as the
independent variable. Specification (1) examines the relationship across all zip codes, (2) examines
the relationship within states, and (3) examines the relationship with counties. Zip codes with at
least 100 loans are considered. Fixed effects are indicated at the bottom of each column. Robust
standard errors are clustered at the county level.
Dep. Variable: Percentage Flagged in Zip Code

Fintech Market Share
State FE
County FE
Observations
Num. Counties
R2
Mean of Dep. Var.

(1)

(2)

(3)

0.272***
(13.82)

0.351***
(23.95)

0.427***
(19.71)

No
No

Yes
No

No
Yes

15,457
2,880
0.350
0.161

15,457
2,880
0.600
0.161

14,599
2,022
0.771
0.163

t-statistics in parentheses
* p<0.10, ** p<0.05, *** p<0.010

25


Table IA.VII. County Cultural Features
This table replicates Table V at the county level rather than the loan-level. The variables are as
defined in Table V. All variables (independent and dependent) are rescaled to have a mean of 0
and standard deviation of 1. Counties with at least 100 loans are considered. Fixed effects are
indicated at the bottom of each column. Robust standard errors are clustered at the state level.
Dep. Variable: At Least One Flag
(1)

(2)

(3)

Public Corruption

0.106
(1.48)

0.0767
(1.21)

0.00313
(0.10)

Religious Affiliation

0.0206
(0.72)

0.00994
(0.40)

0.0464**
(2.41)

Ashley Madison Usage

0.0703***
(2.88)

0.0422
(1.35)

-0.0665**
(-2.65)

Population Density

-0.0179*
(-1.85)

-0.0487***
(-4.40)

Median Income

-0.0275
(-0.83)

-0.0431
(-1.87)

Pct. Non-White

0.389***
(6.52)

0.101***
(2.80)

College Educated

0.00571
(0.20)

0.00430
(0.18)

2019 Unemployment

-0.169***
(-2.95)

-0.185***
(-3.99)

Pct. Fintech

0.586***
(7.00)

State FE

Yes

Yes

Yes

Observations
R2

3,002
0.381

3,001
0.469

3,001
0.577

t-statistics in parentheses
* p<0.10, ** p<0.05, *** p<0.010

26


Table IA.VIII. Previous SBA Lending
In this table, we examine lender level relationship between suspicious lending in the PPP and
previous SBA lending. ln(Num. 7(a) Loans Pre-2020 + 1) is the natural log of the number of SBA
7(a) loans originated by the lender pre-2020. Num. Year Since First 7(a) Loan is the number of
years between when the lender originated its first SBA 7(a) loan and 2020 (we have data going
back to 1990, so this variable can take a maximum value of 30). New Lender is a dummy that
takes one a value of 1 if the lender had not originated any SBA 7(a) loans pre-2020. 1(Fintech)
and 1(Traditional) are indicator functions for whether the lenders is a fintech or traditional lender,
respectively. Lenders with at least 1,000 PPP loans are considered. Robust standard errors are
used.
Dep. Variable: Percentage Flagged
(1)
ln(Num. 7(a) Loans Pre-2020 + 1)

(2)

(3)

(4)

-0.00730***
(-8.32)

Num. Year Since First 7(a) Loan

-0.00125***
(-4.93)

New Lender

0.0514***
(4.41)

× 1(Fintech)

0.0733**
(2.03)

× 1(Traditional)

0.0373***
(3.64)

1(Fintech)

0.0307**
(2.43)

Observations
R2
Mean of Dep. Var.

1,111
0.099
0.125

1,111
0.050
0.125

t-statistics in parentheses
* p<0.10, ** p<0.05, *** p<0.010

27

1,111
0.061
0.125

1,111
0.095
0.125


Table IA.IX. Fees
In this table, we show the dollars value of loans originated, number of loans originated, and the
estimated fees received by the top 75 PPP lenders (by number of loans). The list is sorted in
descending order by estimated fees. The first forty lender are on this page and the remaining 35
are on the next page.
Lender

Lender Type

Dollars Lent

Number of Loans

Estimated Fees

JPMChase
BoA
Cross River
Capital Plus
Harvest
Wells Fargo
Customers Bank
Benworth
PNC
Branch
Fountainhead
Itria
U.S. Bank
TD Bank
KeyBank
Zions
Prestamos CDFI
M&T
Readycap
Huntington
Citizens Bank
Fifth Third
Regions Bank
Celtic
First Horizon
WebBank
Citibank
Kabbage
BMO Harris
City National
First-Citizens
Frost
Pinnacle
Bank of the West
Northeast
Comerica Bank
Synovus
BBVA
Square
First National PA

Traditional
Traditional
Fintech
Fintech
Fintech
Traditional
Traditional
Fintech
Traditional
Traditional
Fintech
Fintech
Traditional
Traditional
Traditional
Traditional
Fintech
Traditional
Fintech
Traditional
Traditional
Traditional
Traditional
Fintech
Traditional
Fintech
Traditional
Fintech
Traditional
Traditional
Traditional
Traditional
Traditional
Traditional
Traditional
Traditional
Traditional
Traditional
Fintech
Traditional

$41,443,991,367
$34,237,075,966
$12,771,048,507
$6,600,599,313
$7,604,665,609
$13,767,834,276
$7,052,505,022
$3,766,809,946
$17,347,783,902
$16,632,176,414
$3,461,750,875
$5,476,551,657
$10,774,383,938
$12,181,766,237
$11,064,228,582
$9,838,138,445
$2,388,759,974
$9,586,986,927
$4,834,860,801
$8,572,031,801
$7,165,123,714
$7,351,389,855
$6,406,068,642
$4,624,587,686
$5,775,813,972
$3,155,253,163
$4,750,119,052
$3,327,154,152
$6,098,989,540
$5,937,998,585
$4,421,177,219
$4,679,275,623
$4,252,036,436
$4,373,311,173
$3,398,482,352
$4,930,860,923
$3,809,538,913
$4,047,201,081
$645,832,846
$3,658,484,414

435,357
489,325
471,609
380,377
377,620
277,949
279,880
275,920
118,639
118,281
233,226
193,336
173,024
131,879
68,580
75,854
145,583
58,211
101,710
58,298
85,246
65,094
77,868
167,098
50,462
117,110
47,572
180,291
35,902
25,476
35,364
32,186
37,628
30,779
34,462
20,897
27,526
30,545
68,747
30,227

$1,678,309,528
$1,465,504,191
$1,016,828,757
$926,869,448
$925,992,546
$682,814,302
$633,299,177
$620,067,286
$599,553,379
$587,181,993
$544,969,507
$529,482,556
$500,996,114
$489,897,697
$373,012,393
$354,088,405
$350,692,221
$331,490,976
$299,981,250
$298,254,178
$292,990,290
$267,637,340
$265,140,975
$235,465,688
$216,970,355
$208,997,502
$189,980,270
$188,787,526
$185,406,250
$185,259,885
$176,372,145
$170,165,268
$169,374,487
$163,264,968
$160,045,643
$141,288,472
$140,926,651
$139,796,313
$137,441,297
$136,728,828

28


Lender

Lender Type

Dollars Lent

Number of Loans

Estimated Fees

People’s United
South State
Hancock Whitney
Umpqua
Valley National
MUFG
Newtek
TCF National
Fulton
Banco Popular
United Community
Glacier Bank
FirstBank
MBE
Prosperity
Atlantic Union
First Bank
BancorpSouth
Arvest
United Bank
Blue Ridge
Webster
Capital One
Old National
Santander
Peoples Bank
First Insterstate
Amur Equipment
First State
A10Capital
Legacy
American Lending
DreamSpring
Funding Circle
Intuit

Traditional
Traditional
Traditional
Traditional
Traditional
Traditional
Fintech
Traditional
Traditional
Traditional
Traditional
Traditional
Traditional
Fintech
Traditional
Traditional
Traditional
Traditional
Traditional
Traditional
Traditional
Traditional
Fintech
Traditional
Traditional
Traditional
Traditional
Traditional
Traditional
Traditional
Fintech
Fintech
Fintech
Fintech
Fintech

$3,684,138,310
$3,231,854,545
$3,323,651,191
$2,896,383,321
$3,235,550,835
$3,110,280,065
$2,251,798,067
$2,672,060,267
$2,732,197,822
$1,715,514,179
$2,218,846,800
$2,005,785,629
$1,979,278,631
$977,705,708
$2,006,818,115
$2,205,251,421
$1,855,736,212
$1,755,438,034
$1,650,380,829
$2,033,793,736
$1,160,073,893
$1,981,355,729
$1,755,412,442
$2,103,975,416
$1,766,843,960
$1,621,567,864
$1,635,315,725
$638,236,364
$987,800,593
$627,843,832
$884,489,617
$536,325,746
$270,698,683
$583,939,117
$639,524,177

30,575
27,892
21,416
26,302
19,736
21,215
33,191
24,329
16,698
43,916
22,125
24,401
26,087
40,531
18,692
16,858
19,820
24,031
27,387
17,572
24,490
18,371
24,534
15,844
19,897
18,908
18,733
26,517
18,992
18,099
15,938
19,875
24,707
17,072
18,561

$134,384,063
$131,439,149
$116,900,675
$111,438,561
$110,968,604
$110,220,137
$109,857,021
$101,875,992
$94,982,595
$91,221,028
$90,448,450
$87,708,029
$86,396,463
$81,168,106
$80,413,268
$80,411,666
$80,129,789
$79,769,179
$79,327,337
$77,850,182
$77,795,525
$76,955,064
$76,891,585
$76,149,828
$73,154,956
$71,227,793
$71,199,493
$69,401,078
$54,941,435
$53,636,427
$53,078,274
$52,714,858
$47,508,876
$44,210,843
$30,670,444

29


Table IA.X. Persistence and Growth by Lender-Region Pairs
In this table, we examine persistence and growth of suspicious lending by lender-region pairs.
Panel A looks at lender-zip code pairs and Panel B looks at lender-county pairs. We estimate OLS
regressions with the percentage of flagged loans during rounds 1-2 within each lender-region pair
interacted with whether the lender is a fintech or traditional lender as the independent variables.
In specification (1), the dependent variable is whether the lender increased its lending within the
region (as a percentage of its overall lending) between rounds 1-2 and round 3; in specification
(2), the dependent variable is the percentage change in the percentage of the lender’s loans that
are in the region between rounds 1-2 and round 3; in specification (3), the dependent variable is
the percentage of flagged loans during round 3 within each lender-region pair. In both panels.
lender-region pars with at least 25 loans in rounds 1-2 (combined) are considered. Fixed effects are
as indicated at bottom of each column. Robust standard errors are double clustered by region (zip
code in Panel A and county in Panel B) and lender.
Panel A. By Lender-Zip Code
(1)
1(Lending
Growth)

(2)
Lending
Pct. Change

(3)
Pct. Flagged
in Round 3

× 1(Fintech)

1.124***
(4.41)

1.890***
(5.65)

0.549***
(7.83)

× 1(Traditional)

0.113
(1.34)

0.148
(1.33)

0.155***
(5.89)

Zip Code FE
Lender FE

Yes
Yes

Yes
Yes

Yes
Yes

Observations
Num. Lenders
R2
Mean of Dep. Var.

37,091
1,632
0.409
0.352

37,091
1,632
0.571
-0.068

37,091
1,632
0.342
0.158

Dep. Variable:
Pct. Flagged in Rounds 1-2

t-statistics in parentheses
* p<0.10, ** p<0.05, *** p<0.010

30


Panel B. By Lender-County
(1)
1(Lending
Growth)

(2)
Lending
Pct. Change

(3)
Pct. Flagged
in Round 3

× 1(Fintech)

1.291***
(4.89)

2.104***
(3.50)

0.490***
(4.51)

× 1(Traditional)

0.369***
(4.10)

0.804***
(4.88)

0.206***
(5.83)

County FE
Lender FE

Yes
Yes

YesYes
Yes

Yes

Observations
Num. Lenders
R2
Mean of Dep. Var.

20,031
2,347
0.371
0.430

20,031
2,347
0.541
0.102

20,031
2,347
0.356
0.181

Dep. Variable:
Pct. Flagged in Rounds 1-2

t-statistics in parentheses
* p<0.10, ** p<0.05, *** p<0.010

31


Table IA.XI. Repayment and Enforcement
In this table, we examine elevated repayment and enforcement actions for flagged loans. We
estimate OLS regressions with a dummy for whether the loan is flagged by at least one primary
flag as the dependent variable and a dummy for whether the loan is repaid (Panel A) or part of a
DOJ enforcement action (Panel B) as the independent variable. Even columns include lender fixed
effects and odd columns do not. Since loans originated by MBE Capital Partners makes up 42%
of the repaid loans, Panel A includes specifications across all loans and excluding loans by MBE
Capital Partners. Fixed effects and additional controls are as indicated at bottom of each column.
Robust standard errors are double clustered by zip code and lender.
Panel A. Repayment
Dep. Variable: 1(Repaid)
(1)

(2)
All Loans

Flagged

(3)
(4)
Ex. MBE Capital Partners

0.00122**
(2.08)

0.0000469
(0.05)

0.000680***
(3.19)

0.000905***
(4.37)

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

Yes
Yes
Yes
Yes
Yes
No

Yes
Yes
Yes
Yes
Yes
Yes

Yes
Yes
Yes
Yes
Yes
No

Yes
Yes
Yes
Yes
Yes
Yes

Observations
Num. Lenders
R2
Mean of Dep. Variable

4,866,890
4,823
0.031
0.00328

4,866,836
4,769
0.159
0.00328

4,843,296
4,822
0.027
0.00192

4,843,296
4,768
0.049
0.00192

Panel B. DOJ Enforcement
Dep. Variable: 1(DOJ Enforcement Action)
(1)

(2)

0.0000989***
(3.17)

0.0000961***
(3.08)

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

Yes
Yes
Yes
Yes
Yes
No

Yes
Yes
Yes
Yes
Yes
Yes

Observations
Num. Lenders
R2
Mean of Dep. Variable

4,866,890
4,823
0.025
0.0000541

4,866,836
4,769
0.026
0.0000541

Flagged

t-statistics in parentheses
* p<0.10, ** p<0.05, *** p<0.010

32

File and source

File

2021-09-griffin-fintech-lenders-ppp-fraud-conference.pdf

Size

2,460,278 bytes

SHA-256

bf263e7aedb854559593b541270de77ceb418937d61238a17718a5f80693d740

Our copy

2021-09-griffin-fintech-lenders-ppp-fraud-conference.pdf

Original

bpb-us-e2.wpmucdn.com