The Role of Technology in Mortgage Lending

This paper presents preliminary findings and is being distributed to economists

and other interested readers solely to stimulate discussion and elicit comments.

The views expressed in this paper are those of the authors and do not necessarily

reflect the position of the Federal Reserve Bank of New York or the Federal

Reserve System. Any errors or omissions are the responsibility of the authors.

Federal Reserve Bank of New York

Staff Reports

The Role of Technology in Mortgage Lending

Andreas Fuster

Matthew Plosser

Philipp Schnabl

James Vickery

Staff Report No. 836

February 2018

The Role of Technology in Mortgage Lending

Andreas Fuster, Matthew Plosser, Philipp Schnabl, and James Vickery

Federal Reserve Bank of New York Staff Reports, no. 836

February 2018

JEL classification: D14, D24, G21, G23

Abstract

Technology-based (“FinTech”) lenders increased their market share of U.S. mortgage lending

from 2 percent to 8 percent from 2010 to 2016. Using market-wide, loan-level data on U.S.

mortgage applications and originations, we show that FinTech lenders process mortgage

applications about 20 percent faster than other lenders, even when controlling for detailed loan,

borrower, and geographic observables. Faster processing does not come at the cost of higher

defaults. FinTech lenders adjust supply more elastically than other lenders in response to

exogenous mortgage demand shocks, thereby alleviating capacity constraints associated with

traditional mortgage lending. In areas with more FinTech lending, borrowers refinance more,

especially when it is in their interest to do so. We find no evidence that FinTech lenders target

marginal borrowers. Our results suggest that technological innovation has improved the

efficiency of financial intermediation in the U.S. mortgage market.

Key words: mortgage, technology, prepayments, nonbanks

_________________

Fuster, Plosser, and Vickery: Federal Reserve Bank of New York (emails:

andreas.fuster@ny.frb.org, matthew.plosser@ny.frb.org, james.vickery@ny.frb.org). Schnabl:

NYU Stern School of Business, NBER, and CEPR (email: schnabl@stern.nyu.edu).The authors

thank an anonymous reviewer, Sudheer Chava, Scott Frame, Itay Goldstein, Wei Jiang, Andrew

Karolyi, Chris Mayer, Stephen Zeldes, and seminar and conference participants at Columbia

(RFS FinTech conference), NYU Stern, Kellogg School of Management, University of St.

Gallen, the Federal Reserve Bank of Atlanta’s 2017 Real Estate Conference, the Homer Hoyt

Institute, and the University of Technology, Sydney, for helpful comments. They also thank a

number of anonymous mortgage industry professionals for providing information about

institutional details and industry trends. Katherine di Lucido, Patrick Farrell, Eilidh Geddes,

Drew Johnston, April Meehl, Akhtar Shah, Shivram Viswanathan, and Brandon Zborowski

provided excellent research assistance. The views expressed in this paper are those of the authors

and do not necessarily reflect the position of the Federal Reserve Bank of New York or the

Federal Reserve System.

I Introduction

The U.S. residential mortgage industry is experiencing a wave of technological innovation as

both start-ups and existing lenders seek ways to automate, simplify and speed up each step of

the mortgage origination process. At the forefront of this development are FinTech lenders,

which have a complete end-to-end online mortgage application and approval process that

is supported by centralized underwriting operations, rather than the traditional network of

local brokers or “bricks and mortar” branches. For example, Rocket Mortgage from Quicken

Loans, introduced in 2015, provides a tool to electronically collect documentation about

borrower’s income, assets and credit history, allowing the lender to make approval decisions

based on an online application in as little as eight minutes.

In the aftermath of the 2008 ﬁnancial crisis, FinTech lenders have become an increasingly

important source of mortgage credit to U.S. households. We measure “FinTech lenders”

as lenders that oﬀer an application process that can be completed entirely online. As of

December 2016, all FinTech lenders are stand-alone mortgage originators that primarily

securitize mortgages and operate without deposit ﬁnancing or a branch network. Their

lending has grown annually by 30% from $34bn of total originations in 2010 (2% of market) to

$161bn in 2016 (8% of market). The growth has been particularly pronounced for reﬁnances

and for mortgages insured by the Federal Housing Administration (FHA), a segment of the

market which primarily serves lower income borrowers.

In this paper, we study the eﬀects of FinTech lending on the U.S. mortgage market. Our

main hypothesis is that the FinTech lending model represents a technological innovation that

reduces frictions in mortgage lending, such as lengthy loan processing, capacity constraints,

ineﬃcient reﬁnancing, and limited access to ﬁnance by some borrowers. The alternative

hypothesis is that FinTech lending is not special on these dimensions, and that FinTech

lenders oﬀer services that are similar to traditional lenders in terms of processing times and

scalability. Under this explanation, there are economic forces unrelated to technology that

explain the growth in FinTech lending (e.g., regulatory arbitrage or marketing).

It is important to distinguish between these explanations to evaluate the impact of techno-

logical innovation on the mortgage market. If FinTech lenders do indeed oﬀer a substantially

diﬀerent product from traditional lenders, they may increase consumer surplus or expand

credit supply, at least for individuals who are comfortable obtaining a mortgage online. If,

however, FinTech lending is driven primarily by other economic forces, there might be little

beneﬁt to consumers. FinTech lending may even increase the overall risk of the U.S. mort-

gage market (e.g., due to lax screening). In addition, the results are important for evaluating

the broader impact of recent technological innovation in loan markets. Mortgage lending is

arguably the market in which technology has had the largest economic impact thus far, but

other loan markets may undergo similar transformations in the future.

Our analysis identiﬁes several frictions in U.S. mortgage markets and examines whether

FinTech lending alleviates them. We start by examining the eﬀect of FinTech lending on

loan outcomes. We focus particularly on the time it takes to originate a loan as a measure of

eﬃciency. FinTech lenders may be faster at processing loans than traditional lenders because

online processing is automated and centralized, with less scope for human error. At the same

time, this more automated approach may be less eﬀective at screening borrowers; therefore,

we also examine the riskiness of FinTech loans using data on loan defaults.

We ﬁnd that FinTech lenders process mortgages faster than traditional lenders, measured

by total days from the submission of a mortgage application until the closing. Using loan-

level data on the near-universe of U.S. mortgages from 2010 to 2016, we ﬁnd that FinTech

lenders reduce processing time by about 10 days, or 20% of the average processing time.

In our preferred speciﬁcations, this eﬀect is larger for reﬁnance mortgages (14.6 days) than

purchase mortgages (9.2 days). The result holds when we restrict the sample to non-banks,

indicating that it is not solely due to diﬀerences in regulation. The results are also robust

to including a large set of borrower, loan, and geographic controls; along with other tests

we conduct, this suggests that faster processing is not explained by endogenous matching of

“fast” borrowers with FinTech lenders.

Faster processing times by FinTech lenders do not result in riskier loans. We measure

loan risk using default rates on FHA mortgages, which is the riskiest segment of the market

in recent years. We ﬁnd that default rates on FinTech mortgages are about 25% lower than

Many industry observers believe that technology will soon disrupt a wide range of loan markets includ-

ing small business loans, leveraged loans, personal unsecured lending, and commercial real estate lending

(Goldman Sachs Research, 2015).

those for traditional lenders, even when controlling for detailed loan characteristics. There

is no signiﬁcant diﬀerence in interest rates. These results speak against a “lax screening”

hypothesis, and instead indicate that FinTech lending technologies may help attract and

screen for less risky borrowers.

We also ﬁnd that FinTech lenders respond more elastically to changes in mortgage de-

mand. Existing research documents evidence of signiﬁcant capacity constraints in U.S. mort-

gage lending.

FinTech lenders may be better able to better accommodate demand shocks

because they collect information electronically and centralize and partially automate their

underwriting operations. To empirically identify capacity constraints across lenders, we use

changes in nationwide application volume as a source of exogenous variation in mortgage

demand and trace out the correlation with loan processing times.

Empirically, we ﬁnd that a doubling of the application volume raises the loan processing

time by 13.5 days (or 26%) for traditional lenders, compared to only 7.5 days for FinTech

lenders. The result is robust to including a large number of loan and borrower observables,

restricting the sample to nonbanks, or using an interest rate reﬁnancing incentive or a Bartik-

style instrument to measure demand shocks. The estimated eﬀect is larger for reﬁnances,

where FinTech lenders are particularly active. We also document that FinTech lenders reduce

denial rates relative to other lenders when application volumes rise, suggesting that their

faster processing is not simply due to credit rationing during peak periods.

Given that FinTech lenders particularly focus on mortgage reﬁnances, we next study

their eﬀect on household reﬁnancing behavior. Prior literature has shown that many U.S.

households reﬁnance too little or at the wrong times (e.g., Campbell, 2006; Keys et al., 2016).

FinTech lending may encourage eﬃcient reﬁnancing by oﬀering a faster, less cumbersome

loan process. We examine this possibility by studying the relationship between the FinTech

lender market share and reﬁnancing propensities across U.S. counties.

We ﬁnd that borrowers are more likely to reﬁnance in counties with a larger FinTech

lender presence (controlling for county and time eﬀects). An 8 percentage point increase

Fuster et al. (2017b) show that increases in aggregate application volumes are strongly associated with

increases in processing times and higher interest rate margins, thereby attenuating the pass-through of lower

mortgage rates to borrowers. Sharpe and Sherlund (2016) and Choi et al. (2017) also ﬁnd evidence of

capacity constraints, which they argue alter the mix of loan applications that lenders attract.

in the lagged market share of FinTech lenders (which corresponds to moving from the 10

percentile to the 90

percentile in 2015) raises the likelihood of reﬁnancing by about 10% of

the average. This increase in reﬁnancing appears to be most pronounced among borrowers

estimated to beneﬁt from reﬁnancing. Our ﬁndings suggest that FinTech lending, by reducing

reﬁnancing frictions, increases the pass-through of market interest rates to households.

We also analyze cross-sectional patterns in who borrows from FinTech lenders. We ﬁnd

that FinTech borrowing is higher among more educated populations, and surprisingly among

older borrowers who may be more familiar with the process of obtaining a mortgage. We

ﬁnd little evidence that FinTech lenders disproportionately target marginal borrowers with

low access to ﬁnance. We ﬁnd no consistent correlation between FinTech lending and local

Internet usage or speed; similarly, using the entry of Google Fiber in Kansas City as a natural

experiment, we ﬁnd no evidence that improved Internet access increases FinTech mortgage

take-up. These results mitigate concerns about a digital divide in mortgage lending.

Taken together, our results suggest that recent technological innovations are improving

the eﬃciency of the U.S. mortgage market. We ﬁnd that FinTech lenders process mortgages

more quickly without increasing loan risk, respond more elastically to demand shocks, and

increase the propensity to reﬁnance, especially among borrowers that are likely to beneﬁt

from it. We ﬁnd, however, little evidence that FinTech lending is more eﬀective at allocating

credit to otherwise constrained borrowers.

Our results do not necessarily predict how FinTech lending will evolve in the future.

FinTech lenders are nonbanks who securitize most of their mortgages—their growth could

be aﬀected by regulatory changes or reforms to the housing ﬁnance system. There is also

uncertainty as to how the increased popularity of machine learning techniques, which FinTech

lenders may be using more intensely, will inﬂuence the quantity and distribution of credit.

Related to this issue, although we ﬁnd no evidence FinTech lenders select the highest-quality

borrowers (“cream skim”), which could reduce credit for other borrowers, these results could

change as technology-based lending becomes more widespread. Lastly, FinTech lenders use

a less personalized loan process that relies on hard information, which could reduce credit

See Bartlett et al. (2017) and Fuster et al. (2017a) for recent studies of these issues in the context of the

U.S. mortgage market.

to atypical applications.

Our research contributes to several strands of the literature. Although a large body of

research has studied residential mortgage lending (see Campbell, 2013 and Badarinza et al.,

2016 for surveys), much of the recent work focuses on securitization and the lending boom

prior to the U.S. ﬁnancial crisis.

Our paper instead focuses on how technology aﬀects the

structure of residential mortgage lending after the crisis. Most closely related to this paper,

Buchak et al. (2017) study the recent growth in the share of nonbank mortgage lenders,

including FinTech lenders. While there is some overlap between the descriptive parts of

our analyses, and we use similar approaches to classify FinTech lenders, the two papers are

otherwise strongly complementary. Buchak et al. focus on explaining the growth of non-

bank lending, using reduced-form analysis and a calibrated structural model. Our paper

focuses on how technology impacts frictions in the mortgage origination process, such as

slow processing times, capacity constraints and slow or suboptimal reﬁnancing.

Our ﬁndings also inform research on the role of mortgage markets in the transmission

of monetary policy (e.g., Beraja et al., 2017; Di Maggio et al., 2017). If lenders constrain

the pass-through of interest rates (Agarwal et al., 2017; Drechsler et al., 2017; Fuster et al.,

2017b; Scharfstein and Sunderam, 2016), or borrowers are slow to reﬁnance (Andersen et al.,

2015; Agarwal et al., 2015), changes in interest rates will not be fully reﬂected in mortgage

rates and originations. Our ﬁndings suggest that technology may be easing these frictions,

potentially improving monetary policy pass-through in mortgage markets.

Finally, our paper contributes to a growing literature on the role of technology in ﬁnance

(see Philippon, 2016, for an overview), and a broader literature on how new technology can

lead to productivity growth (see e.g. Syverson, 2011 and Collard-Wexler and De Loecker,

2015). In our case, the “productivity” or “eﬃciency” measures we consider are processing

times, supply elasticity, default and reﬁnancing propensities, and we are the ﬁrst to document

that FinTech lending appears to lead to improvements along these dimensions.

See, for example, Mian and Suﬁ (2009), Keys et al. (2010), Purnanandam (2010), Acharya et al. (2013),

or Jiang et al. (2014). Aside from this paper, research focusing on mortgage lending in the post-crisis

environment includes D’Acunto and Rossi (2017), DeFusco et al. (2017), and Gete and Reher (2017).

We also study loan defaults and mortgage pricing in a similar way to Buchak et al., but focus on the

riskier FHA segment of the market; they primarily study loans insured by Fannie Mae and Freddie Mac.

II Who is a FinTech Lender?

A. Deﬁning FinTech lenders

A central feature of our study is the distinction between FinTech mortgage originators and

other lenders. While many mortgage lenders are adopting new technologies to varying de-

grees, it is clear that some lenders are at the forefront of using technology to fundamentally

streamline and automate the mortgage origination process. The deﬁning features of this busi-

ness model are an end-to-end online mortgage application platform and centralized mortgage

underwriting and processing augmented by automation.

How does the FinTech business model aﬀect the mortgage origination process in prac-

tice?

Online applications mean that a borrower can be approved for a loan without talking

to a loan oﬃcer or visiting a physical location. The online platform is able to directly access

the borrower’s ﬁnancial account statements and tax returns to electronically collect informa-

tion about assets and income. Other supporting documents can be uploaded electronically,

rather than by being sent piecemeal by mail, fax or email.

This automates a labor-intensive

process, speeds up information transfer, and can improve accuracy, for example by elimi-

nating transcription errors (Goodman 2016, Housing Wire 2015). The online platform also

allows borrowers to customize their mortgage based on current lender underwriting standards

and real-time pricing.

Supporting and complementing this online application process, FinTech mortgage lenders

The discussion of institutional details in this section draws upon extensive conversations with mortgage

industry professionals, market economists within the Federal Reserve, and other industry experts. For more

detail on how technology is reshaping the mortgage market, see Oliver Wyman (2016), The Economist

(2016), Goodman (2016), Goldman Sachs Research (2015) and Housing Wire (2015, 2017).

Obtaining a purchase mortgage involves three main steps (see e.g., Freddie Mac, 2016). (1) An initial

application and pre-approval—a pre-approval letter is nonbinding, but is indicative of a borrower’s credit

capacity and is often required to make an oﬀer on a home. (2) Processing and underwriting, which is usually

undertaken after a property has been identiﬁed and sale price agreed upon. This step involves veriﬁcation

of all supporting documentation, often involving many interactions between the processor, loan oﬃcer and

borrower, and can take from 1-2 days to several weeks or more (known as the “turn time”). (3) Closing,

when the funds and property deed are transferred. FinTech lenders partially automate the ﬁrst two steps

and allow them to be completed online. Recently, some lenders have also digitized the third and ﬁnal step

by creating an electronic mortgage note (e.g., see Quicken Loans, 2017a).

FinTech lenders also oﬀer email and phone support. The key distinction to traditional lenders is that

borrowers can process the entire application without using paper forms, email, or phone support. In practice,

the degree of automation is much larger among FinTech lenders relative to other lenders, even if some FinTech

borrowers communicate via email or over the phone with their lender.

have developed “back-end” processes to automatically analyze the information collected dur-

ing the application. For example, borrower information is compared against employment

databases, property records, as well as marriage and divorce records; additionally, algo-

rithms can examine whether recent bank account deposits are consistent with the borrower’s

paystubs. Optical character recognition and pattern recognition software can be used to

process documents uploaded by the borrower and ﬂag missing or inconsistent data. These

systems make the mortgage underwriting process more standardized and repeatable, and

may help identify fraud (Goodman, 2016).

This approach does not eliminate the role of human underwriters, but does make mort-

gage processing less labor-intensive. In contrast with more hub-and-spoke loan origination

operations that put loan oﬃcers and underwriters in branches, FinTech lenders centralize

their processing operations, which allows for labor specialization in the underwriting process.

Lenders have told us anecdotally that this makes it easier to train new workers and to adjust

labor supply in response to demand shocks.

Against these advantages, there may also be important disadvantages of a more auto-

mated approach to mortgage underwriting. For example, poorly designed online platforms

may confuse borrowers or lead to errors, and a lack of personal interaction may impede

the transmission of soft information, resulting in less eﬀective borrower screening or credit

rationing.

Our empirical analysis examines both the beneﬁts and costs of the FinTech

mortgage lending model.

We emphasize that automation and online applications are not entirely new.

For exam-

ple many lenders in recent years have allowed borrowers to initiate a mortgage application

online. However, the online application is often just a ﬁrst step before directing applicants to

speak to a loan oﬃcer who then initiates a more traditional loan application process. Simi-

larly, although online mortgage rate comparison services such as LendingTree and BankRate

have been a feature of the mortgage market for many years, these services simply provide

information and connect borrowers and lenders; they do not automate the mortgage origi-

A substantial academic literature has emphasized the role of soft versus hard information in lending

(e.g., Petersen and Rajan, 2002; Stein, 2002).

More generally, the use of information technology in mortgage lending and servicing is not a recent

phenomenon—see e.g. LaCour-Little (2000) for a discussion of developments in the 1990s.

nation process or put it online.

The emergence of several stand-alone FinTech ﬁrms as major lenders over the last few

years is a strong indicator that fundamental change is underway. These ﬁrms are at the

technological frontier and focus exclusively on the new business model. In contrast, estab-

lished lenders with branch-based mortgage origination processes face signiﬁcant obstacles

in recalibrating their operations away from branches and loan oﬃcers. For this reason, the

vanguard of FinTech lenders is composed of nonbanks, which do not have access to deposit

ﬁnance and therefore do not retain originated loans on balance sheet. Like other nonbanks,

the vast majority of FinTech lenders sell their loans through established channels supported

by government guarantee programs (FHA, VA, Fannie Mae, and Freddie Mac).

That said, a signiﬁcant and growing number of mortgage lenders are at present incorpo-

rating aspects of the “FinTech model,” and the current distinction between FinTech origi-

nators and other ﬁrms, including banks, may be temporary. The current market structure

presents a window of opportunity to study the impact of FinTech on mortgage origination,

and to draw inferences about what is likely to happen to the mortgage industry as a whole

as these technologies diﬀuse more broadly.

B. Classifying FinTech lenders

For our empirical analysis, we classify an originator as a FinTech lender if they enable a

mortgage applicant to obtain a pre-approval online. We believe this classiﬁcation distin-

guishes FinTech lenders from more traditional mortgage originators that may use “online

applications” for marketing purposes but still require interaction with a loan oﬃcer.

Our classiﬁcation should be viewed as a proxy, since an online application platform is only

one dimension of the FinTech “model”. Even so, it is an important component, and is also

easily measurable in a consistent way across a large number of mortgage lenders. In practice,

the set of lenders classiﬁed as FinTech by our approach matches up well with ﬁrms considered

by industry observers and media to be at the frontier of technology-based mortgage lending.

It also matches quite closely with the independent classiﬁcation by Buchak et al. (2017).

Our classiﬁcation and empirical analysis closely follows the methodology in our proposal to the RFS

FinTech initiative submitted on March 15, 2017. Our proposal was submitted before we and Buchak et al.

We implement our classiﬁcation by ﬁrst compiling lists of the top 100 non-bank lenders

for purchase loans and for reﬁnancings over the analysis period.

The resulting list includes

135 lenders. We then manually initiate a mortgage application with each lender and analyze

whether it is possible to obtain a pre-approval online. Most lenders halt the online application

prior to the pre-approval and ask the borrower to directly contact a loan oﬃcer or broker. We

classify the lender as a FinTech lender if we are able to continue with the online process until

we get to the pre-approval decision that is based on a hard credit check of the applicant’s

credit score.

Our ﬁnal classiﬁcation is based on an analysis completed in June 2017. To construct a

panel, we go back in time using a database that archives websites (“Wayback Machine”).

Using the database from 2010 to 2017, we evaluate at which point in time a lender appears

to have adopted their qualifying online lending process. We cannot always conduct a full

evaluation because online application processes often rely on a technological process that

evaluates information in real time. However, we can use the archived website to evaluate

when a lender adopted an application which resembles the qualifying application in 2017.

We use this information to determine the year in which a lender adopted a FinTech lending

model. We corroborate our results using industry reports.

FinTech lenders exhibit several other distinguishing characteristics relative to their com-

petitors. For example, FinTech ﬁrms typically require a Social Security Number and conduct

a hard credit check online, unlike most traditional mortgage originators we classiﬁed. Fin-

Tech lenders also tend to orient their marketing eﬀorts around their website or mobile phone

app. In particular, FinTech lender advertisements emphasize the functionality and ease of

use of their website or app, and direct borrowers to those platforms. Other lenders may in-

clude their website in their marketing material but do not emphasize it to the same degree,

and may primarily use it for “lead generation.”

Figure 1 plots the number of FinTech lenders by year based on our classiﬁcation. The

became aware of each others’ work and pre-dates the ﬁrst public version of their working paper.

We also examined several top depository bank lenders, but did not classify any of them as FinTech

through 2016 (although some began oﬀering online pre-approvals in 2017). As discussed above, entrenched

bank business models may slow their ability to integrate new technology into their existing branch-based

mortgage origination process.

We ﬁnd no instance of a lender that stopped oﬀering online processing during the analysis period.

number increases from two ﬁrms in 2010 to 18 lenders by 2017. In Table 1 we list the top 20

lenders in 2016, along with other FinTech lenders in the data in that year. The three largest

originators identiﬁed as FinTech lenders are Quicken, LoanDepot.com, and Guaranteed Rate.

All of the primary analyses in this paper use this classiﬁcation, although we have veriﬁed

that our main results are robust to the alternative classiﬁcation of Buchak et al. (2017).

Table 2 provides summary statistics of mortgage originations and applications, in to-

tal and by lender type, based on data collected under the Home Mortgage Disclosure Act

(HMDA). HMDA data report characteristics of individual residential mortgage applications

and originations from the vast majority of U.S. banks and non-banks. Data include the

identity of the lender, loan amount, property location, borrower income, race and gender,

though not credit score or loan-to-value ratio (LTV). Based on known local conforming loan

limits, we impute whether each loan has “jumbo” status and thus cannot be securitized by

Fannie Mae, Freddie Mac, or Ginnie Mae. The processing time of loan applications, one of

our main outcome variables of interest, can only be computed from a restricted version of

the dataset available to users within the Federal Reserve System.

We include loans with

application dates between January 2010 and June 2016.

First, we see that in terms of basic

risk characteristics, non-bank lenders originate loans to borrowers with relatively low-income

and high loan to income (LTI) ratio relative to banks. Similarly, FinTech lenders and other

non-bank lenders have a much higher share of FHA and VA loans, but a lower share of jumbo

mortgages, than banks. FinTech lenders originate many more reﬁnance loans (as opposed

to loans used for a home purchase) than banks and other non-bank lenders.

We also see that FinTech lenders have shorter average processing times than both banks

and other non-bank lenders. In the next section, we study whether this result persists once

Our classiﬁcation is similar tho one proposed by Buchak et al. (2017). There are only minor diﬀerences

with respect to the classiﬁcation of a few smaller lenders.

This restricted version of the data records the exact date the lender receives an application, as well as

the date on which the application was resolved (e.g. origination of the loan or denial or withdrawl of the

application). The publicly available HMDA data only contains the year. All other variables are the same.

We end the sample in June because for applications submitted later in the year, processing times may

be biased downward. This is due to the fact that only applications for which an action (origination, denial,

etc.) was taken by the end of 2016 are included in the HMDA data available at the time of writing.

As Buchak et al. (2017) also note, FinTech lenders have a higher fraction of applications where appli-

cant race or gender information is missing. We understand this is because borrowers can complete online

applications without being required to provide this information.

we control for loan characteristics and location-time ﬁxed eﬀects. In Section VII we will

study diﬀerences in borrower and location characteristics between FinTech and non-FinTech

mortgages more systematically, building on Table 2.

III Is FinTech Lending Faster?

Our ﬁrst research question is whether FinTech lenders are able to process mortgage appli-

cations more quickly than other lenders. We measure processing time by the number of

days between application and origination date, as in Fuster et al. (2017b). We estimate the

following OLS regression using loan-level HMDA data:

Processing Time

ijct

= δ

+ βFinTech

+ γControls

ijct

+ 

ijct

(1)

where Processing Time

ijct

is for loan i issued by lender j in census tract c for an application

received in month t, FinTech

is an indicator variable equal to one for FinTech lenders and

zero otherwise, δ

is a vector of census-tract-month ﬁxed eﬀects, and Controls

ijct

includes

loan and borrower controls.

We winsorize the top and bottom 1% of processing times and

cluster standard errors at the lender-month level.

Our regression includes a large number of observable loan and borrower characteristics

to control for factors other than lender eﬃciency that may aﬀect processing time (e.g., local

laws, housing market conditions, the complexity of the loan, borrower, and property, and

the speed of obtaining a property appraisal). We expect that our rich set of controls should

account well for these factors. In particular, census-tract-month ﬁxed eﬀects control in a

highly disaggregated way for common geographic and time variation in processing times.

We conduct the analysis separately for home purchase mortgages and reﬁnances because the

latter do not require the homeowner to move and the application process is simpler.

The control variables are the natural logarithm of borrower income, the natural logarithm of total loan

amount, indicator variables for race and gender, an indicator variable for whether there is a coapplicant,

an indicator variable for whether a pre-approval was requested, indicator variables for the occupancy and

lien status of the loan, indicator variables for property type, indicator variables for whether the loan is

insured by the FHA or the Department of Veterans Aﬀairs (VA) and an indicator variable for loans above

the conforming loan limit (i.e. jumbo loans), and an indicator variable in case applicant income is missing.

A. Processing time results

Panel A of Table 3 presents the results for purchase mortgages. In column (1), we ﬁnd that

FinTech lenders process loans 7.9 days faster than non-FinTech lenders. This eﬀect is large,

corresponding to 15% of average home purchase processing time of 52 days. The result is

slightly larger in magnitude and remains statistically signiﬁcant when we include loan and

borrower controls (col. 2), census tract-month ﬁxed eﬀects (col. 3), and both (col. 4, where

the estimated eﬀect corresponds to 18% of the average processing time). The results are also

robust to dropping deposit-taking banks from the sample (col. 5), which suggests that the

results are not driven by regulatory factors or the diﬀerent funding model of banks.

Panel B of Table 3 ﬁnds even larger eﬀects for reﬁnances. Across speciﬁcations, FinTech

lenders process mortgages 9.3 to 14.6 days faster than other lenders. The eﬀect corresponds

to 17%-29% of the average reﬁnance loan processing time of 51 days. Again, the result is

robust to comparing FinTech lenders only to other nonbanks, which suggests that it is not

driven by regulation or funding. The FinTech advantage for reﬁnance loans might be larger

because reﬁnances oﬀer more scope for automation than home purchase loans. For example,

home mortgage loans always require an appraisal, which is administered locally and is not

(yet) automated. This interpretation is consistent with the fact that FinTech lending growth

has been larger for reﬁnances relative to home purchase loans.

While these regressions capture average eﬀects, it is instructive to study the entire dis-

tribution of processing times across lender types. We do so in Figure A.1 in the Internet

Appendix, where we plot the cumulative distributions of processing times for both purchase

and reﬁnance mortgages, after accounting for census-tract-month ﬁxed eﬀects and loan char-

acteristics. For purchase mortgages the advantage of FinTech lenders comes primarily from

the right tail (i.e., there are few loans with very long processing times), while for reﬁnances

the entire distribution is shifted to the left. This again suggests that for reﬁnances, it is

more easily possible for FinTech lenders to realize eﬃciency gains, while for purchase loans

In unreported results we also condition on whether the loans are FHA or VA insured loans, since

anecdotally, underwriting rules are less ﬂexible for these loans, possibly constraining the advantages of

FinTech lenders. Indeed, we ﬁnd for reﬁnances that the FinTech lender advantage is lower by 3 days

(relative to a sample of non-government or “conventional” loans). However, we detect no corresponding

diﬀerence in FinTech lenders’ processing time advantage among new purchase loans.

the scope may be more limited.

B. Additional analysis

One potential concern is that our processing time results are aﬀected by endogenous match-

ing between borrowers and lenders. For instance, if younger borrowers are more likely to

use FinTech lenders and also tend to submit their paperwork faster, FinTech lenders would

appear to process mortgages more quickly, even if they do not have an inherent techno-

logical advantage. Alternatively, FinTech lenders may attract the most complex mortgage

applications, which would attenuate the estimated FinTech processing time advantage.

We emphasize that the coeﬃcient on FinTech lenders is robust across speciﬁcations and

samples. If FinTech lenders matched with borrowers or loan types that are easier to process,

then adding the control variables should attenuate the estimated coeﬃcient; instead, the

coeﬃcient tends to get larger with additional controls. To the extent that unobservable

factors that make some borrowers faster than others are also correlated with observables,

this is a ﬁrst piece of evidence that our results are unlikely to be driven by endogenous

matching or other unobserved variables, but instead represent the direct eﬀect of FinTech

lending on processing times.

To investigate further, we examine whether the FinTech processing time advantage is

driven by “fast borrowers” migrating to FinTech lenders. We implement this test in two

stages. In the ﬁrst stage, we predict the probability that each loan is originated by a

FinTech lender as a function of loan and borrower characteristics. We then take this predicted

probability and use it as an explanatory variable in a second stage analysis of processing

times among non-FinTech mortgages. If non-FinTech lenders lose their faster customers to

FinTech lenders, non-FinTech processing times should have increased disproportionately for

borrower and loan types with high FinTech penetration (as measured by a high ﬁrst-stage

probability).

The second stage results are shown in Table A.1 in the Internet Appendix. In our baseline

speciﬁcation, we ﬁnd a positive eﬀect of the predicted FinTech probability on non-Fintech

processing times. This is consistent with selection, although the coeﬃcient is not nearly

large enough to explain our earlier processing time results.

In addition, the coeﬃcient of

interest ﬂips sign once we control for lender-by-census tract ﬁxed eﬀects to allow for the

possibility that FinTech lenders have a high market share in areas where traditional lenders

are slow (col. 2 and 4 of the table). In sum, it does not appear that selection eﬀects could

easily explain the large processing time diﬀerences we document.

Furthermore, as a direct test of whether FinTech lenders match with “fast” borrowers,

we study whether FinTech originators have gained the highest market share in geographic

locations where processing times were shortest ex ante, measured in 2010 prior to the growth

in FinTech. These results are presented in Section VII. To preview the key result, we in fact

ﬁnd the opposite; FinTech lenders have become popular in locations where processing times

were originally slow conditional on observables. This is inconsistent with an “endogenous

selection” interpretation of our processing time estimates, and in fact suggests that slow

processing by traditional lenders may be a driver of the growth in FinTech lending.

Summing up, our results suggest that FinTech mortgage lenders are roughly 20% faster

at processing mortgage originations than other lenders; the estimated eﬀects range from 7.5-

9.4 days for purchase mortgages and 9.3-14.6 days for reﬁnances. Several pieces of evidence

suggest that this ﬁnding is not due to endogenous borrower-lender matching or other omitted

variable biases.

IV Is FinTech More Eﬃcient or Just Less Careful?

The faster processing speeds of FinTech lenders could simply be a product of less careful

screening of borrowers, rather than greater eﬃciency.

We test this “lax screening” hypoth-

For instance, the coeﬃcient of 2.5 in column (1) of Table A.1 means that moving from the 1st to the

99th percentile in predicted FinTech propensity, corresponding to a diﬀerence of 0.335, increases expected

processing time by 0.85 days. This is only about one-tenth of the processing time advantage of FinTech

lenders as estimated in Table 3. Magnitudes are similar in column (3), which limits the sample to reﬁnances.

As a “reality check”, our estimates also appear roughly comparable to industry-based estimates of the

processing-time advantage of technology-based lending. In particular, Quicken Loans (2017b) claims that

importing income and asset information through their online platform reduces client mortgage processing by

12 days on average. Although it is not clear exactly how this statistic is calculated, it is interesting that it is

in the same ballpark as our estimate of a 8-14 day diﬀerence in processing times between FinTech originators

and other lenders.

For instance, using proprietary lender data, LaCour-Little (2007) documents that prior to the ﬁnancial

crisis, processing times were shortest for non-agency non-prime mortgages. This category of loans subse-

esis by studying the ex-post performance of FinTech loans compared to similar mortgages

from other lenders. We focus on FHA lending, which has been the riskiest segment of the

mortgage market in recent years and where we are therefore most likely to detect diﬀerences

in loan risk.

We use two separate sources of publicly available data on FHA mortgage

defaults: segment level data extracted from the FHA Neighborhood Watch Early Warning

System (“FHA NW data”) and FHA loan-level data from Ginnie Mae (“FHA Ginnie Mae

data”). To our knowledge, this is the ﬁrst academic study to make systematic use of either

of these data sources.

A. Analysis of default rates in FHA NW data

We start by analyzing default rates on FHA loans using FHA NW data. The data contains

origination volume and default rates for each lender at the national level and by state and

metropolitan statistical area (MSA). The data are available for all FHA loans as well as cer-

tain subcategories including home purchase mortgages, reﬁnances, and mortgages originated

in underserved census tracts.

The data generally covers the period 2015:Q3 to 2017:Q3,

although state and national data for all loans (not broken down by loan type) are available

over a longer sample period from 2012:Q3 to 2017:Q3.

Default rates are calculated as the share of loans that become at least 90 days delinquent

or are the subject of an FHA insurance claim within a speciﬁc time horizon after origination.

The data include rates at one-year (“1 Year Default”) and two-year (“2 Year Default”)

quently experienced extremely high default rates during the crisis.

FHA mortgages require a down payment of as little as 3.5% and are generally made to borrowers with

low credit scores who do not qualify for a prime conforming loan. FHA loans are government-guaranteed,

which limits the credit risk for the lender. However the lender is not fully indemniﬁed against risk since

the FHA can refuse to compensate the lender for credit losses in cases of fraud or other defects in mortgage

underwriting. FHA lenders have also paid out large legal settlements on FHA loans due to breaches of the

False Claims Act and other laws. As a result of these risks, many large bank lenders have withdrawn from

FHA lending or wound back their participation in the market (see e.g., Wall Street Journal, 2015).

The FHA Ginnie Mae data are similar to the loan-level data made available by government-sponsored

enterprises (GSEs) Fannie Mae and Freddie Mac. These are analyzed by Buchak et al. (2017), who ﬁnd little

diﬀerence in default probabilities between FinTech and other lenders (for origination vintages 2010-2013).

The main drawback of the GSE data is that these prime agency mortgages have experienced very low default

rates for recent vintages (as they are signiﬁcantly less risky than FHA loans) so that it may be diﬃcult to

detect diﬀerences across lenders.

A census tract is considered underserved by the FHA based on an administrative classiﬁcation derived

from median income and the share of minority households.

horizons. In order to control for geographic variation in default rates we scale a lender’s

default rate in each location by the overall default rate in that area. As an alternative to

raw default rates, the data also contain the “Supplementary Performance Metric” (SPM),

which scales a lender’s default rate by a benchmark default rate deﬁned based on the credit

score distribution of the underlying mortgages. Again, we then take the ratio of the lender’s

SPM to the overall SPM in the area. The SPM is only available at the state and national

level and at a two-year horizon after origination (“Mix-Adjusted 2 Year Default”).

Our analysis focuses on the diﬀerence in default rates between FinTech lenders and

other lenders. We compute the diﬀerence by taking by taking the weighted average of

FinTech relative default rates using origination volume by region and lender as weights and

subtracting one. This measure yields zero if there no diﬀerences in default rates between

FinTech lenders and other lenders. We use a diﬀerence-in-means test to examine the null

hypothesis that FinTech lender default rates are the same as other lenders.

Table 4 reports the results. Column (1) presents the relative diﬀerence in default rates

for FinTech lenders using 1-year default as the default measure. In Panel A, we ﬁnd that

loans originated by FinTech lenders are 35% less likely to default than comparable loans

originated by non-FinTech lenders. The coeﬃcient is almost unchanged when using MSA-

level data instead of state-level data and when using the 2-year default rate instead of

the 1-year default rate (col. 2). The coeﬃcient remains statistically signiﬁcant, albeit the

eﬀect is smaller (-25.5%) when using the mix-adjusted default rate, based on the SPM, as

the outcome variable (col. 3). We ﬁnd quantitatively similar results when restricting the

sample to high-market share regions (Panel B), when considering home purchase loans or

reﬁnances separately (Panel C), for loans to underserved neighborhoods (Panel D), and when

considering a longer sample period (Panel E). Overall, we ﬁnd no evidence that FinTech loans

are risker than non-FinTech loans; in fact, they appear to default less often.

B. Loan-level analysis of FHA default rates

We complement this evidence with a loan-level analysis of data on FHA mortgages securitized

into Ginnie Mae MBS. The main advantage of the Ginnie Mae data relative to the FHA NW

data is that they include a rich set of loan and borrower characteristics (e.g., the borrower’s

credit score and the loan-to-value ratio). This allow us to investigate whether FinTech lenders

target speciﬁc borrower types based on their riskiness and whether diﬀerences in default

rates can be explained by diﬀerences in observable characteristics. A disadvantage of the

Ginnie Mae data is that they only include the identity of the MBS issuer, not the mortgage

originator. Hence, the data do not perfectly identify which loans come from FinTech lenders.

However, the issuer and originator are typically the same and a comparison to HMDA

suggests mismeasurement is concentrated among small lenders.

Our sample consists of data from September 2013 (when the Ginnie Mae data ﬁrst become

available) until May 2017. We restrict the sample to 30-year ﬁxed-rate mortgages, which are

by far the most common FHA loan type. We estimate the following OLS regression:

Default

ijst

= α + βFinTech

+ γControls

ijst

+ 

ijst

(2)

where Default

ijst

on loan i by lender j in state s originated in month t is an indicator variable

equal to one if a loan ever becomes delinquent for 90 days or longer over our observation

period, FinTech

is an indicator variable equal to one for FinTech issuers, and Controls

ijst

a broad set of control variables such as origination month or state-by-origination month ﬁxed

eﬀects, loan purpose ﬁxed eﬀects, and other loan controls including borrower FICO score,

loan-to-value ratio (LTV) and debt-to-income ratio (DTI).

We cluster standard errors at

the issuer-origination month level.

Table 5 presents the results. Column (1) controls for origination-month ﬁxed eﬀects

only and ﬁnds that FinTech borrowers are 1.29 percentage points less likely to default than

non-FinTech borrowers, equivalent to 35% of the overall default rate of 3.65%. This result

is very similar to the estimates based on FHA NW data. Column (2) adds loan purpose

ﬁxed eﬀects. The eﬀect declines to 0.91 percentage points, or 27%, but remains statistically

For some small FinTech lenders, the number of MBS-issued loans is substantially smaller than their

number of originated loans in HMDA, implying that they sell a signiﬁcant portion of their loans to other

ﬁrms before issuance. The eﬀect on identifying FinTech loans should be limited given that this issue primarily

aﬀects smaller lenders.

Other loan controls include the log of the loan amount and indicators for the number of borrowers, the

property type, whether the borrower received down payment assistance, and for whether a loan’s FICO,

LTV, or DTI are missing.

signiﬁcant. This result reﬂects the fact that FinTech lenders issue more reﬁnance mortgages

than home purchase mortgages, and that reﬁnances tend to be less risky (especially those

not involving cash-out). In column (3), we add further loan level controls such as FICO

and LTV. We see that this has only a small incremental eﬀect on the coeﬃcient of interest,

implying that FinTech lenders do not originate loans that are less risky based on these

observable characteristics. Columns (4) and (5) split the sample by loan purpose; the eﬀect

is slightly larger for home purchase mortgages, but is also sizable for reﬁnances.

C. Are FinTech lenders cream skimming?

Our analysis of default rates ﬁnds no evidence that FinTech lenders originate riskier mortgages—

in fact, in the FHA market we ﬁnd the opposite result. The diﬀerence in default rates varies

across speciﬁcations but is statistically signiﬁcant in almost all of them and the magnitude

is economically large—default rates for FinTech-originated loans are about 25% lower in

column (3) of Table 5, which includes the largest set of controls, and ranges between 10-40

percent in the other speciﬁcations. The results are robust to using two diﬀerent datasets

(FHA Neighborhood data and Ginnie Mae loan-level data) and to diﬀerent sets of controls

for loan, borrower and location characteristics.

Our ﬁndings speak directly against the “lax screening” hypothesis. If anything, they sug-

gest that the automated technologies used by FinTech lenders may screen borrowers more

eﬀectively than the more labor-intensive methods used by other lenders (e.g., because the

automated systems directly check databases of original source documents, reducing the pos-

sibility of fraud). This reasoning has been emphasized by industry experts (e.g., Goodman,

2016), and to our knowledge we provide the ﬁrst systematic evidence to support it.

Although superior screening of credit risk can be viewed as an advantage of FinTech

lending, it may also have negative consequences for some borrowers, or for the government,

due to “cream skimming” of the highest value customers. For example, cream skimming

could lead to ex ante credit rationing by weakening the credit quality of the remaining

In further (unreported) regressions, we have found that the relative eﬀect size is fairly stable if we repeat

the regressions for each loan origination year 2013-2017. Furthermore, the Ginnie Mae data also contain

mortgages guaranteed by the Department of Veterans Aﬀairs (VA); for those loans, which default at lower

rates than FHA ones, the relative decrease in default hazard for FinTech-originated loans is again similar.

borrower pool—this mechanism is explored by Mayer et al. (2013) in the context of private

subprime mortgage lending. Alternatively, it could shift costs to the government if private

and public lenders compete for borrowers, an argument that has been made in the context

of FinTech lenders like SoFi in the student loan market

In the context of mortgage lending in the current environment, it is unlikely that cream

skimming by FinTech lenders has economically signiﬁcant eﬀects. The reason is that during

our analysis period the vast majority of all risky mortgages in the U.S. are government insured

at a pre-set price, either by the FHA or other government agencies such as the Department of

Veterans Aﬀairs. Consequently, cream skimming by FinTech lenders is unlikely to materially

aﬀect credit access for remaining borrowers, who will still qualify for government insurance.

Even so, we estimate two speciﬁcations to investigate possible cream-skimming eﬀects.

First, we examine whether a higher FinTech market share in a location helps to reduce

overall mortgage default risk in that location, as opposed to FinTech lenders just selecting

the lowest-default borrowers from a ﬁxed pool. We also test whether the default advantage

of FinTech lenders diminishes as their market share increases. If the distribution of risky

borrowers is unchanged by the presence of FinTech lenders, then as their market share

increases in an area their performance advantage will diminish, as they expand their lending

to the more risky borrowers.

We present the results, based on the Ginnie Mae data from the previous subsection, in

columns (6) and (7) of Table 5. Column (6) estimates the direct eﬀect of FinTech state-level

market share on default. Although the point estimate is negative, the eﬀect is economically

small and statistically insigniﬁcant. This result suggests that there is no discernable eﬀect

of an increased FinTech footprint on the overall default risk of borrowers receiving mortgage

credit. We note that the estimate has large standard errors; it would be interesting to revisit

this analysis in the future when the market share of FinTech lending is larger.

Column (7) adds the interaction of the FinTech lender indicator variable and local Fin-

Tech market share and ﬁnds that the coeﬃcient on the interaction is negative and marginally

signiﬁcant. This result suggests that the better default performance of FinTech mortgages

See e.g. https://www.bloomberg.com/news/articles/2015-06-10/student-loan-refinancing-

boom-could-cost-u-s-taxpayers-billions.

in fact tends to be more pronounced in regions where FinTech has a larger market share. On

the other hand, however, in Panel B of Table 4, we ﬁnd a lower FinTech default advantage in

markets where the lender has a high market share (although the diﬀerence is not statistically

signiﬁcant, and even in these locations, FinTech default rates remain lower than the market

as a whole).

While somewhat mixed, none of the results suggest a robust positive relation between

market share and risk. In addition, we ﬁnd no evidence that the lower default rate of FinTech

lenders disappears in locations where their market share is high. In sum, the ﬁndings suggest

that the lower default rates associated with FinTech lending is not simply due to positive

selection of low-risk borrowers.

D. Are FinTech lenders charging higher interest rates?

Related, we can also use FHA loan-level data from Ginnie Mae to test whether FinTech

lenders charge higher or lower mortgage interest rates conditional on observables. Results

are shown in Table A.2 in the Internet Appendix. We ﬁnd that FinTech lenders oﬀer interest

rates which are 2.3bp lower overall—splitting the sample by loan purpose, the eﬀect is 7.5bp

for purchase mortgages and eﬀectively zero for reﬁnances.

Although these diﬀerences

are small in magnitude, the direction of the eﬀect is consistent with the Buchak et al.

(2017) estimates for FHA loans (although they are cautious in drawing inferences from their

results because their FHA dataset includes fewer loan-level controls than the data used here).

However, it contrasts with Buchak et al.’s ﬁnding that FinTech lenders charge higher rates

for GSE mortgages. One possible explanation that could account for both sets of results,

and is in line with some of Buchak et al.’s other evidence, is that lower-income borrowers,

who are more likely to obtain FHA loans, are more price sensitive and less willing to pay a

premium for convenience.

We note that these coeﬃcients are not particularly stable if we allow them to vary over time — in some

time periods the FinTech coeﬃcient is positive and signiﬁcant, but over others it is negative and signiﬁcant.

A potential explanation is that movements in market interest rates may be reﬂected at diﬀerent times on

rates on originated loans between FinTech and other lenders, due to diﬀerences in processing times. The

Ginnie Mae data, or the GSE data used by Buchak et al. (2017), does not easily allow one to cleanly control

for this.

V Is FinTech Lending More Elastic?

We next study whether FinTech lenders are better able to accommodate shocks to mort-

gage demand. Mortgage application volumes in the U.S. ﬂuctuate enormously over time,

primarily due to movements in interest rates that can lead to “reﬁnancing waves.” There is

also substantial cross-sectional variation in demand for new mortgages, for example due to

diﬀerential housing market trends.

Managing volatility in mortgage applications is a key challenge for lenders. If a lender

receives more applications than their underwriting process can accommodate, their process-

ing cycle times increase and they risk losing money (and future business) due to loans not

closing in a timely manner. Figure 2, which is similar to evidence in Fuster et al. (2017b),

illustrates two main points. First, as shown in panel (a), there is large variation in the level

of monthly applications, with the peak level being almost three times as high as the trough.

Application volume co-moves closely with borrowers’ average incentive to reﬁnance, here

proxied by the diﬀerence between the average coupon rate on outstanding mortgages and

the 10-year Treasury yield. Second, panel (b) shows that ﬂuctuations in median processing

times are sizable (from a low of 37 days to a high of 51 days), and that processing times are

strongly positively correlated with total mortgage applications.

By automating, centralizing and standardizing much of the underwriting process, FinTech

lenders may conceivably increase the short-run elasticity of lending supply in response to

demand shocks. However, testing whether capacity constraints are less binding for FinTech

lenders presents a clear empirical challenge: the volume of applications a lender receives

is endogenous. For example, lenders may solicit applications when processing constraints

are slack and discourage applications when processing times are expected to be long. Both

behaviors would attenuate the relationship between applications and processing time and

obfuscate diﬀerences across lenders.

One exception: between October and December 2015, processing times increase even though applications

decrease. This is most likely due to the implementation of new loan disclosure rules (“TILA-RESPA Inte-

grated Disclosure,” or TRID) on October 3, 2015. These new rules required many lenders to adjust their un-

derwriting processes, resulting in delays. For more details, see e.g. https://www.wsj.com/articles/new-

mortgage-rules-may-spark-delays-frustration-1443519000.

A. Demand shocks and processing time

We identify diﬀerences in supply elasticity by exploiting demand shocks that vary application

volumes independent of ﬁrm-speciﬁc conditions. We use time-series variation in total ap-

plications, which is primarily determined by macroeconomic factors, in particular long-term

interest rates, and is plausibly exogenous to the capacity contraints facing any individual

lender. We test whether FinTech mortgage processing times are less sensitive to variation in

total application volume by estimating the following regression using loan-level HMDA data

from 2010 through June of 2016:

Processing Time

ijct

= γApplications

+ βApplications

× FinTech

+ α

+ δ

+ θControls

+ 

ijct

(3)

where Processing Time

ijct

is the number of days between application and closing for mortgage

i from lender j in census tract c and application month t, Applications

is the log of aggregate

mortgage applications in month t, FinTech

is an indicator variable equal to one for FinTech

lenders, α

and δ

are vectors of lender and census-tract ﬁxed eﬀects, and Controls

includes

borrower and loan controls similar to Table 3 as well as calendar month dummies to account

for seasonality and dummies for loan purpose (purchase versus reﬁnancing). Standard errors

are clustered by lender-month.

Table 6 presents the results. The ﬁrst two columns consider all originated loans; column

(1) controls only for lender dummies, while column (2) includes additional controls for loan

and borrower characteristics, location, and month. We ﬁnd that FinTech lenders are about

half as sensitive to aggregate mortgage application volumes as other lenders. Quantitatively,

a 10% rise in overall application volume increases processing time by 1.3 days for non-FinTech

lenders but only 0.7 days for FinTech ﬁrms (based on column 2). Column (3) restricts the

sample to reﬁnances, the market where FinTech lenders specialize and where interest rates

matter most for demand. We ﬁnd that processing times for reﬁnances are more sensitive

to aggregate volumes, but again FinTech lenders are only half as sensitive. Column (4)

considers all applications, including denied and withdrawn applications; again, the results

are similar. All results are statistically signiﬁcant at the 1% level. Columns (5)-(7) repeat the

prior three speciﬁcations but restrict the sample to nonbanks. The degree to which FinTech

lenders are less sensitive to aggregate applications is not as large in this sample but the

magnitudes are still economically meaningful. FinTech lenders are 20-40% less sensitive to

aggregate volumes relative to nonbanks, again statistically signiﬁcant at the 1% level except

for column (5), where p = 0.14.

The diﬀerential sensitivity of FinTech lender processing times to application volume is

also illustrated visually in Figure 3. This binned scatter plot conﬁrms that FinTech lenders

have shorter processing times on average, as already shown in Section III. More impor-

tantly, processing time for Fintech lenders is also less sensitive to demand for new mortgages

compared to banks and (to a lesser extent) other non-bank lenders. This lower sensitivity is

particularly apparent at the highest levels of application volume (when aggregate application

volume exceeds 1.2 million mortgages per month).

B. Alternative demand shocks and processing time

We repeat the analysis using the weighted average coupon on the universe of ﬁxed-rate MBS

less the 10-year Treasury yield (“Reﬁ Incentive”) as our measure of mortgage demand, in-

stead of log application volume.

Our ﬁndings, presented in columns (1)-(3) of Table A.3 in

the Internet Appendix, are similar to those discussed above. A higher reﬁnancing incentive

is signiﬁcantly correlated with longer processing times across speciﬁcations, but processing

times for FinTech lenders are signiﬁcantly less sensitive, if anything by a larger proportion

than in our main results. The consistency with our earlier ﬁndings is sensible given that

we show in Figure 2 that reﬁnancing incentives are the key determinant of mortgage appli-

cation volume. The result does, however, address any concerns that our earlier results are

aﬀected by idiosyncratic shocks to individual large lenders that are large enough to inﬂuence

aggregate applications.

As alternative approach, we also construct a “Bartik-style” index of exposure to local

ﬂuctuations in mortgage application volume based on the geography of lender activity. The

In unreported results, we consider speciﬁcations with time ﬁxed eﬀects and draw similar conclusions.

While it absorbs all time-series variation, this alternative speciﬁcation does not allow us to observe the

uninteracted coeﬃcient on aggregate application volume.

As in Fuster et al. (2017b), we use the 10-year Treasury rate rather than a market rate on new mortgages

in order to prevent endogeneity to concurrent supply conditions in the mortgage market.

index is calculated as the log of the weighted sum of county-level mortgage applications in

month t, where the weights are the lender’s average market shares in each county measured

over the entire sample period. The identiﬁcation assumption is that application volumes in

a geographic area are exogenous to any given market participant’s share. We present the

results in in columns (4)-(6) of Table A.3 in the Internet Appendix. Processing times are

positively correlated with the proxy, although once again, less so for FinTech lenders when

we consider reﬁnancing loans and all applications. There is no statistically signiﬁcant eﬀect

for the sample of all originated loans (col. 4).

Taking together, results based on two alternative measures of loan demand indicate that

FinTech lenders are less sensitive to exogenous demand shocks than other lenders, supporting

our main ﬁndings in Table 6.

C. Demand shocks and application denial rates

A possible concern is that our results may reﬂect credit rationing. If FinTech lenders avoid

capacity constraints by becoming more selective and rationing credit when total mortgage

demand rises, their processing times may seem less sensitive to demand even if they are

not actually more elastic. We test this hypothesis by examining whether denial rates for

FinTech lenders are diﬀerentially sensitive to aggregate application volume. The regression

speciﬁcation is identical to Eq. (3) above, except that the left-hand side variable is an

indicator variable equal to one if a loan application was denied and zero otherwise. The

sample includes all applications that were either approved or denied.

Table 7 presents the results. We ﬁnd that FinTech lenders reduce denial rates by 1.1%

percentage points for each 10% increase in application volume (col. 1). The eﬀect is similar

when focusing on reﬁnance mortgages (col. 2), restricting the sample to nonbanks (col. 3),

and focusing on reﬁnance mortgages among nonbanks (col. 4). These results are inconsistent

with credit rationing and instead provide further evidence that FinTech lenders’ credit supply

is more elastic than those of other lenders.

In unreported results, we ﬁnd that FinTech lenders on average have a roughly 2.5 percentage point higher

denial rate than banks (though the diﬀerence is statistically insigniﬁcant), and a 3.5 percentage higher denial

rate than other nonbank lenders (signiﬁcant at p < 0.1), conditional on our typical set of controls. This

could reﬂect more stringent screening, or alternatively that with online applications, there is no “ﬁltering”

We point out that the direct eﬀect of application volume on denial rates is negative

across all speciﬁcations. This may seem counter-intuitive, although it likely reﬂects the fact

that when applications rise due to changes in interest rates, the average credit quality of

applicants improves.

D. Demand shocks and origination volumes

We also analyze whether mortgage origination volumes for FinTech lenders respond diﬀer-

entially to changes in total applications. Analysis of quantities over our short sample period

is diﬃcult because there are diﬀerential trends in application volumes across lender types

and across individual ﬁrms within a type. We estimate a model in ﬁrst diﬀerences to partial

out these trends:

∆Originations

= γ∆Applications

+ β∆Applications

× FinTech

+ α

+ 

where Originations

is the log of originated applications (by lender j for applications in

month t) and Applications

is the log of aggregate application volume. Lender origination

changes are winsorized at the top and bottom 1% to mitigate the impact of extreme outliers.

We include lender ﬁxed eﬀects, α

, to allow for lender-speciﬁc diﬀerences in the average

growth rate over the analysis period. We restrict the sample to lenders who rank in the top

500 in volume at some point during the sample period.

We ﬁnd no meaningful diﬀerence in origination sensitivity for FinTech lenders. As shown

in Table 8, FinTech origination volume appears equally sensitive to changes in aggregate

application volumes as those of all other lenders (col. 1 and 2) and nonbanks (cols. 3 and

4). Hence, similar to our results on denials, we ﬁnd no evidence that the lower sensitivity

of FinTech lender processing times comes at the expense of lower originations due to credit

rationing; conversely, though, we do not see an obvious increase in origination growth for

by a loan oﬃcer that may discourage borrowers from applying when their chances of approval are low.

In line with this interpretation, Fuster and Willen (2010) show that denial probabilities fell for all

income levels in the wake of the ﬁrst MBS purchase announcement by the Federal Reserve in late 2008

(when application volumes surged), and that average FICO scores (which are not in HMDA) increased

sharply.

FinTech lenders when application volume rises. Overall, we are cautious about drawing

strong conclusions from this analysis as it is quite diﬃcult to establish lender-type speciﬁc

eﬀects given the strong and nonlinear upward trend in the FinTech lender market share

during this period.

VI FinTech and Mortgage Reﬁnancing

This section examines whether the presence of FinTech lenders aﬀects mortgage reﬁnancing

behavior by borrowers. Prior work has shown that many borrowers do not reﬁnance their

ﬁxed-rate mortgages optimally; they commit errors either by failing to reﬁnance when it is

in their ﬁnancial interest to do so, or by reﬁnancing even though the costs of doing so exceed

the beneﬁts (e.g., Campbell, 2006; Agarwal et al., 2015; Andersen et al., 2015; Keys et al.,

2016). In addition to behavioral factors, intermediation frictions in the mortgage market

also contribute to ineﬃcient reﬁnancing patterns (e.g., Agarwal et al., 2017; Bond et al.,

2017; Scharfstein and Sunderam, 2016). These frictions weaken the “reﬁnancing channel” of

monetary policy (e.g., Beraja et al., 2017; Di Maggio et al., 2016; Wong, 2016). Examining

the eﬀect of FinTech on reﬁnancing is thus important, since this is one channel through

which technological progress in the mortgage industry may have real eﬀects on the economy.

Industry reports and academic research indicate that mortgage-backed securities backed

by FinTech loans do exhibit faster prepayment speeds than pools from other lenders, con-

sistent with an eﬀect of FinTech on the speed of reﬁnancing (e.g., Goldman Sachs Research,

2016; Buchak et al., 2017). However, it is unclear whether this fact reﬂects faster-prepaying

borrowers selecting into mortgages from FinTech lenders, or whether FinTech lending directly

aﬀects the likelihood of reﬁnancing, thereby potentially aﬀecting aggregate reﬁnancing be-

havior. If FinTech mortgage lending does aﬀect the market-wide propensity to reﬁnance,

an important follow-up question is whether this is due to a reduction in errors of omission

(meaning that more borrowers who should reﬁnance do so), or instead reﬂects an increase

in errors of commission (more borrowers reﬁnance even when they should not). Below, we

assess this based on the optimal reﬁnancing decision rule of Agarwal et al. (2013).

To measure reﬁnancing behavior, we use data from Equifax’s Credit Risk Insight Servicing

McDash (CRISM) dataset, which merges McDash mortgage servicing records from Black

Knight with credit bureau data from Equifax. The sample period is January 2010 through

June 2016. The CRISM dataset provides information on loan and borrower characteristics

such as FICO score, CLTV, interest rate, and loan term, and features borrower identiﬁers

that allow us to track a borrower across loans and thereby identify mortgage reﬁnancing.

We focus on the 500 largest counties by loan volume. Details on the sample construction and

how we measure reﬁnancing are provided in Section D of the Internet Appendix, where we

also conﬁrm that the average reﬁnance propensity we measure in our data is closely aligned

with variation over time in the volume of of reﬁnancing loans in HMDA.

A. Reﬁnancing propensity

We measure the eﬀect of FinTech lending on monthly reﬁnancing propensities using the

following OLS regression:

Reﬁ Propensity

c,t

= α

+ α

+ β · FinTechShare

c,t−s

+ Γ · X

c,t

+ 

c,t

(4)

where Reﬁ Propensity

c,t

is the share of mortgages in county c in month t that are reﬁnanced

and FinTechShare

c,t−s

is the one-quarter-lagged four-quarter moving average market share

of FinTech mortgage lenders among reﬁnance loans in a county. We include county ﬁxed

eﬀects, α

, to control for ﬁxed unobservable diﬀerences in reﬁnancing speeds across counties

and month ﬁxed eﬀects, α

, to control for aggregate conditions.

The time-varying county-

level controls X

c,t

include average FICO score, average CLTV, the average interest rate on

outstanding loans, and the share of outstanding loans that are FHA or VA loans. We run

the regressions separately for the sample of all outstanding loans, and restricting to 30-year

CRISM has previously been used to study reﬁnancing by Beraja et al. (2017) and Di Maggio et al.

(2016).

Market shares are calculated based on HMDA. Results are similar if we use all loans, rather than just

reﬁnances, to calculate the market share (although we view this alternative approach as less conceptually

appealing, because composition eﬀects imply that the overall FinTech market share will be aﬀected by the

relative volume of purchase loans versus reﬁnances, which in turn is related to our outcome variable). Market

shares are calculated on a volume-weighted basis, although results are similar if we use loan-count weighted

shares instead. Similarly, our results are robust to using alternative timing conventions, such as the share

over the previous calendar year.

FRMs only. We cluster standard errors by county.

Table 9 shows that there is a positive and strongly statistically signiﬁcant association

between reﬁnancing propensities and FinTech market share. The estimate of 0.689 in column

(2) implies that an 8 percentage point increase in FinTech market share (corresonding to

moving from the 10th to 90th percentile of county averages in 2015) is associated with a

0.055 percentage point increase in the reﬁnancing propensity, about a 10% increase relative

to the average monthly reﬁnancing propensity over this time period of 0.54%. Thus, the

magnitude of the eﬀect is economically meaningful.

Figure 4 further illustrates the eﬀect. Here we sort the counties into ﬁxed terciles based

on their FinTech market share in the middle of our sample period (between mid-2012 to

mid-2013). We then plot the average reﬁnance propensity in each tercile over time. We see

that in 2010, before the growth in FinTech lending, the tercile of counties where FinTech

lenders subsequently gained the most market share has the slowest reﬁnancing speeds. The

reﬁnancing propensity across the three terciles converge over time, however, coincident with

the growth in FinTech lending. This suggests that FinTech mortgage lenders have helped

“slow” reﬁnancing counties to “catch up.”

In summary, our results show that the faster prepayment speeds on FinTech mortgages

are also reﬂected in overall market-wide local reﬁnancing propensities, rather than just being

due to a selection of “fast” borrowers into FinTech loans. As a caveat on our results, we note

that, although our regressions condition on county and time ﬁxed eﬀects, the diﬀerential

growth in FinTech market share across counties may not be exogenous with respect to

reﬁnancing propensities. For instance, it is possible that FinTech lenders predicted correctly

which geographic regions still had the most potential for higher reﬁnancing volumes, and

advertised and expanded most intensively there. At the least, however, our results suggest

that a higher presence of FinTech lending leads to faster mortgage reﬁnancing, perhaps by

reducing the transaction or time costs of reﬁnancing.

We have also replicated these regressions using HMDA data, where we use the log of the number of

reﬁnance loans originated in a county-month as the dependent variable. The lagged FinTech market share

(controlling for county and month ﬁxed eﬀects) is again positively and signiﬁcantly associated with reﬁnance

originations, and the magnitude of the eﬀect is comparable to the estimates in Table 9.

Our ﬁndings here have an interesting parallel to earlier work by Bennett et al. (2001), who ﬁnd evidence

that technological innovation in the 1990s reduced reﬁnancing frictions and increased mortgage reﬁnancing

speeds.

B. Reﬁnancing optimality

We next examine whether the increase in reﬁnancing speeds documented above is associated

with more optimal reﬁnancing decisions. In other words, is it driven by higher reﬁnancing

by borrowers with a high mortgage coupon rate relative to the available rate on new loans,

and who thus realize large savings in interest costs by reﬁnancing (fewer errors of omission)?

Or is it due in large part to borrowers who reﬁnanced more frequently but should not have

done so because the interest savings were small and outweighed by the costs of reﬁnancing

(more errors of commission)?

To determine the breakeven interest rate diﬀerential beyond which a borrower should

reﬁnance, we use the “square-root rule” of Agarwal, Driscoll, and Laibson (2013), henceforth

ADL.

For this analysis, we restrict the sample to 30-year FRMs, since the ADL optimality

calculation was derived for FRMs; it does not apply to adjustable-rate loans. We note

that the ADL calculation embeds a number of assumptions about the costs and beneﬁts of

reﬁnancing; among these, it does not account for other common reﬁnancing motives, such

as cashing out home equity, shortening the term of the loan, or reﬁnancing from a mortgage

type that requires mortgage insurance (such as FHA loans) to one that does not. While the

ADL benchmark is very useful, it is ultimately a simpliﬁcation.

Based on the ADL rule, we ﬁnd, similar to Keys et al. (2016), that at certain points over

our sample period, a lot of borrowers should reﬁnance. For instance, in late 2012, about 60%

of all 30-year FRM borrowers in our sample should have reﬁnanced according to the ADL

benchmark, yet only a signiﬁcantly smaller percentage did so. However, similar to Agarwal

et al. (2015), we ﬁnd that among reﬁnances that do occur, more than half are executed at a

rate diﬀerential that is too small when assessed against the ADL rule. Thus, at least when

compared to this benchmark, there is substantial scope for enhanced reﬁnancing eﬃciency.

In Table 10, we estimate loan-level regressions similar to the county-level speciﬁcations

from above, but now separating outstanding mortgages into diﬀerent bins depending on

how “in the money” the reﬁnancing option is. Speciﬁcally, the “reﬁ incentive” shown in

The square-root rule is a second-order approximation that comes close to the (more complicated) optimal

decision rule derived by ADL. As inputs to the calculation, we require assumptions on discount rates, tax

rates, moving probabilities, interest rate volatility, and (most importantly) the upfront costs associated with

reﬁnancing; we use the same parameter values as ADL’s baseline calibration (following also Keys et al. 2016).

the table is equal to the diﬀerence between the outstanding mortgage rate and current

market rate, minus the optimal reﬁnancing diﬀerential according to ADL.

For instance, if

a borrower is currently paying 6.5%, the market interest rate is at 4.5%, and the optimal

reﬁnancing diﬀerential is 1.5% based on the ADL formula, the borrower would have a 0.5%

positive reﬁnancing incentive. If the market rate increases to 5.5%, the reﬁnancing incentive

would become negative. Even though a reﬁnancing in this situation would still reduce the

borrower’s monthly payments, the savings would not be suﬃcient to outweigh the ﬁxed costs

of reﬁnancing and the loss of option value.

Column (1) shows that for borrowers where the reﬁnancing option is more than 1 per-

centage point out of the money, a higher local FinTech share has a marginally negative eﬀect

on the likelihood of reﬁnancing. The eﬀect of a higher FinTech share then becomes positive,

and is highest for borrowers that are within 50 basis points of their optimal reﬁnancing

diﬀerential (columns 3 and 4). The eﬀect size then decreases again somewhat for borrowers

that have a large reﬁnancing incentive according to ADL; these borrowers in many cases

have suboptimally failed to reﬁnance for a long period of time (given that the reﬁnancing in-

centive is driven by market interest rates, which drift only slowly through time). In industry

jargon, these borrowers are sometimes called “woodheads” (Deng and Quigley, 2012).

The results imply that an increased presence of FinTech lenders is most strongly asso-

ciated with higher reﬁnancing when the borrower’s incentive to reﬁnance is either “at the

money” or just “in the money.” It does not appear that FinTech lenders induce an increase

in grossly ineﬃcient reﬁnancings (if anything, the reverse is true), although they also do not

spur a large increase in reﬁnancing for the borrowers who would gain the most from doing

so.

We present additional evidence in Table A.4 of the Internet Appendix. Speciﬁcally we

examine reﬁnances from 30-year FRMs to new 30-year FRMs, and ﬁnd that a higher local

FinTech share is associated with a higher fraction of reﬁnances that would be classiﬁed as

optimal; larger decreases in the interest rate a borrower is paying; and a higher fraction of

cash-out reﬁnancings.

The FRM market rate is measured using the 30-year FRM rate from Freddie Mac’s Primary Mortgage

Market Survey, which is a standard source for academic research and policy analysis.

In sum, the results suggest that a higher share of FinTech lending is associated not

just with faster reﬁnancing, but also more optimal reﬁnancing decisions. However, this

eﬀect appears somewhat weaker for the borrowers that would most beneﬁt from additional

reﬁnancing. Such borrowers may be less ﬁnancially literate; we show in the next section

that proxies for ﬁnancial literacy are negatively correlated with takeup of mortgages from

FinTech lenders. In some cases seemingly “woodhead” borrowers may face other obstacles

that prevent reﬁnancing (e.g., a signiﬁcant decline in income since the original mortgage was

received). To the extent that an increased FinTech share in the market overall continues to

lead to faster reﬁnancing, it is in fact possible that borrowers who themselves are limited

in their sophistication or are otherwise unable to reﬁnance may be worse oﬀ: in equilibrium

faster prepayment speeds will be priced in MBS valuations, which could feed through to

higher market mortgage rates.

VII Who Borrows From FinTech Lenders?

The market share of FinTech lenders varies signiﬁcantly by geography and across segments

of the mortgage market.

In this section we estimate a simple model of the likelihood of

borrowing from a FinTech mortgage lender as a function of loan, borrower and location

characteristics. We then compare the cross-sectional patterns in the data to a number of

hypotheses about the drivers of the growth in FinTech mortgage borrowing.

We estimate a loan-level linear probability model using pooled HMDA data on mortgage

originations from 2010 to 2016:

FinTech

i,c,t

= α

+ β · Controls

i,c,t

+ Γ · location

+ 

l,c,t

(5)

where FinTech

i,c,t

is equal to 100 if mortgage i in census tract c originated at time t was orig-

Table 2 provides univariate summary statistics about the characteristics of mortgages from FinTech

lenders. Figure E in the Internet Appendix maps the FinTech market share of mortgage originations in

2010 and 2016. This map highlights the substantial geographic variation in the market share of FinTech

lenders, as well as the widespread growth in technology-based lending over our sample period. One limitation

of HMDA is that market coverage outside of MSAs may be more limited, because very small lenders and

lenders that do not operate in MSAs do not need to report. However, our regression results in this section

are robust to restricting the sample to MSAs.

inated by a FinTech lender and zero otherwise, α

is a vector of time dummies, Controls

i,c,t

is a set of loan and borrower characteristics drawn from HMDA, and location

is a set of

local geographic and socioeconomic variables measured at the census tract or county level,

drawn from a variety of sources including the U.S. Census, American Community Survey

and the FRBNY consumer credit panel. These location variables are measured in 2010, or

otherwise as early in time as possible, to minimize any concerns about reverse causality.

Data deﬁnitions and sources are provided in the Data Appendix.

We estimate this model separately for purchase mortgages and reﬁnances, because the

determinants of demand may be quite diﬀerent between the two, and because the market

share of FinTech lenders is signiﬁcantly higher for reﬁnances (Table 2). Since diﬀerences in

borrower characteristics between FinTech lenders and banks may be driven by regulatory

factors, we present each speciﬁcation both including and excluding mortgages from banks.

In the speciﬁcations excluding banks, FinTech lenders are compared to other nonbanks, who

are regulated similarly and have the same funding model.

A. Results

Estimates are presented in Table 11. Each continuous right-hand side variable is normalized

to have a standard deviation of one, so that magnitudes can be compared across variables.

Below, we discuss the cross-sectional patterns relative to the predictions of four sets of

potential drivers of the growth in FinTech mortgage lending: (i) access to ﬁnance, (ii)

technology adoption and ﬁnancial literacy, (iii) Internet access, and (iv) local mortgage and

housing market conditions.

Access to ﬁnance. We ﬁnd no strong evidence that FinTech lenders disproportionately

cater to ﬁnancially constrained borrowers (e.g., borrowers with low incomes or poor credit

Beyond the main variables of interest reported in the table, regressions also control for loan size (in logs),

dummies for loan type (jumbo, FHA, VA), dummies for coapplicant and investor loan, additional individual-

level race dummies, state dummies, and dummies for missing values for each variable with incomplete data

coverage. See Internet Appendix F for a full table of results including these variables. The table in the

Internet Appendix also presents results of univariate regressions in which the FinTech dummy is regressed

individually on each right-hand-side variable. The only additional control included in these univariate re-

gressions is the vector of time dummies. Comparing the univariate and multivariate results helps to show

which results are robust to the inclusion or exclusion of other variables.

histories), particularly compared to other nonbanks. Although the FinTech market share

is higher in census tracts where mortgage borrowers have lower average credit scores, the

results for income and the share of FHA/VA-insured mortgages are mixed depending on

the type of loan and comparison group.

As shown in Section IV, FinTech mortgages also

have comparatively lower default rates on FHA loans, suggesting they are not targeting the

riskiest borrowers. Nonbank lenders overall, including FinTech lenders, originate a signiﬁ-

cantly higher share of FHA and VA loans than banks, however (see Table 2 or the Internet

Appendix). Buchak et al. (2017) attribute the low levels of bank FHA lending to regulatory

and legal factors. FinTech lenders attract a higher share of female borrowers, although the

share of black or hispanic borrowers is lower in most speciﬁcations.

Also related to access to ﬁnance, we test whether FinTech borrowing is higher in census

tracts with few physical bank branches. We measure branch access as the number of bank

branches within a 25 mile radius of the geographic midpoint of the census tract, based on

FDIC Summary of Deposits data. Even though the FinTech business model is focused on

online applications, we ﬁnd that the share of FinTech borrowing is increasing in bank branch

density.

Technology adoption and ﬁnancial literacy. Early technology adoption is often

concentrated in urban areas as well as among younger and more educated consumers. Indeed,

we ﬁnd that the FinTech market share is higher in more urban neighborhoods, measured by

population density, although only for purchase mortgages (for reﬁnances, the coeﬃcient ﬂips

sign across speciﬁcations).

Examining education and age directly, we ﬁnd that the FinTech market share is increasing

in the fraction of adults with at least a bachelor’s degree. Interestingly, however, the share

of FinTech borrowing is increasing with average mortgage borrower age. Although this may

Among purchase mortgages, FinTech borrowers have higher incomes either compared to all lenders or

just nonbanks, and are also less likely to be FHA/VA guaranteed than loans from other nonbanks. For

reﬁnances, FinTech borrowers have lower incomes and the fraction of FHA/VA insured loans is generally

higher. –see Internet Appendix F for the FHA and VA coeﬃcients

Our estimates with regard to borrower gender and race should be treated with caution, because as

discussed earlier, a signiﬁcant fraction of race and gender ﬁelds in HMDA are coded as missing or “NA”

for FinTech lenders. As a result, the measured shares of female and minority borrowers are likely to be a

lower bound. Using census tract variation in minority population from the 2010 Census, we ﬁnd that the

FinTech share is lower in tracts with a high share of minority borrowers, although this is not always true in

the univariate speciﬁcations in the Internet Appendix.

seem counterintuitive, we have been told by industry practitioners that ﬁrst-time mortgage

borrowers often prefer to interact face-to-face with a mortgage broker or loan oﬃcer, rather

than applying online, because they are less familiar with the steps involved; in other words

they are less ﬁnancially experienced and literate (consistent with Agarwal et al., 2009, who

ﬁnd that ﬁnancial literacy increases with age up to individuals’ mid-50’s).

Internet access. As online services become more ubiquitous, there is growing concern

about the “digital divide”, meaning that inequality in access to Internet services may be ex-

acerbating income and wealth inequality. We examine whether the availability of high-speed

Internet is a constraint on FinTech mortgage borrowing (where applications are generally

completed online), using two data sources: ﬁrst, the fraction of households with high-speed

Internet access from the Census Bureau American Community Survey (available from 2013

by county); second, census-block data from the FCC and NTIA for the ten largest states

by population on the fraction of households with the option to connect to ﬁber or cable

Internet, which we aggregate by census tract.

Empirically, these two variables have opposite signs, and the coeﬃcients are generally

small in magnitude. We conclude that lack of access to adequate Internet does not ap-

pear to be a signiﬁcant constraint on the diﬀusion of technology-based mortgage lending.

This interpretation is also consistent with more detailed analysis described below about the

staggered rollout of Google Fiber in one local market.

Local mortgage and housing market conditions. As FinTech lending becomes

more widely available, it may be particularly beneﬁcial in areas with long processing times.

Indeed, we ﬁnd that a higher share of FinTech mortgage borrowing in census tracts where

mortgage processing times were slow ex ante, measured in 2010 prior to the growth in

FinTech lending.

This result suggests that borrowers have turned to FinTech lenders in

part to alleviate bottlenecks in mortgage origination associated with “traditional” mortgage

lenders.

We measure average processing time in 2010 conditional on borrower characteristics. Using 2010 HMDA

data, we regress processing time on loan and borrower characteristics. We then take the residuals from this

regression and aggregate them to the census tract level.

As discussed in Section III, the sign of this correlation also speaks against concerns that our earlier

processing time results are due to selection eﬀects. If “fast processor” borrowers (conditional on observables)

are attracted to FinTech lenders, we would have expected a higher ex post FinTech share in neighborhoods

where 2010 processing times were faster than would be expected based on observables. In fact, however, we

We also examine whether FinTech mortgage borrowing is more prevalent in “hot” real

estate markets where prices are rising rapidly and quick closing may be more important.

Anecdotal evidence suggests that borrowers may be particuarly attracted to the convenience

and fast processing speeds of FinTech lenders in such markets.

We ﬁnd no empirical

support for this hypothesis; in fact, home price growth is negatively correlated with the

market share of FinTech lending for purchase mortgages (which are the relevant group to

consider for this hypothesis).

We examine whether the FinTech share is lower in neighborhoods with high average

home prices (measured in 2010). This hypothesis is motivated by the fact that FinTech

lenders rely on securitization for funding mortgages; as a result these lenders originate few

jumbo loans, which are diﬃcult to securitize. This in turn means that FinTech lenders may

advertise less in high-home-price markets where jumbo mortgages predominate. There also

may be less social learning about the FinTech mortgage lending model in such markets. We

do indeed ﬁnd that the likelihood of borrowing from a FinTech lender is lower in high home

price areas, conditional on loan size and other observables.

B. Interpretation

We interpret our evidence as supporting the view that FinTech mortgage borrowers are

attracted to the faster processing times and greater convenience involved with online appli-

cations and partial automation of mortgage underwriting. This is consistent with the faster

growth of FinTech in census tracts with previously long mortgage processing cycle times and

the higher incomes and education of FinTech borrowers. It is also consistent with the high

share of reﬁnances for FinTech lenders.

We ﬁnd no empirical support for the hypothesis

that FinTech lenders have grown by disproportionately targeting risky, marginal borrowers.

ﬁnd that the opposite correlation is true in the data.

For example, a September 2015 The Street article titled “Online Mortgage Lenders Are Beating Tra-

ditional Bank Loans” highlights the shorter closing times of online lenders, and includes the following

quote from the CEO of the lender Bank of the Internet “We have very short underwriting term times and

that’s a plus for our purchase oriented borrowers – we give quick answers,” Garrabrants said. ”In a really

hot market, that’s important.” (see https://www.thestreet.com/story/13282079/1/online-mortgage-

lenders-are-beating-traditional-bank-loans.html).

The more standardized set of tasks involved in reﬁnancing a mortgage may be the best ﬁt for FinTech

lending at the current state of technology.

Despite the emphasis of the FinTech lending model on online applications and interactions,

we also ﬁnd no evidence that younger borrowers or borrowers located in census tracts with

better Internet access are more likely to borrow from FinTech lenders.

We emphasize that we do not attach a strong causal interpretation to our results, given

the reduced-form nature of our analysis. Yet, we believe that the empirical relationships

documented here are a useful benchmark for further analysis.

C. Evidence from Google Fiber rollout

In addition to our reduced-form analysis, we also conduct an in-depth empirical analysis

of the potential causal eﬀect of improved Internet access on FinTech lending. We exploit

the staggered entry of a new high-speed Internet provider in a single market, namely the

entry of Google Fiber in Kansas City starting in late 2012. Google Fiber is a large-scale

initiative by Google to establish a new Internet service provider. The ﬁrst metro area that

Google Fiber entered was the Kansas City area, consisting of Johnson, Leavenworth, and

Wyandotte counties in Kansas and Cass, Clay, Jackson, and Platte counties in Missouri.

A major factor behind the selection of Kansas City was that households had poor access

to high-speed Internet prior to the entry of Google Fiber. High-speed Internet was only

available to a relatively small fraction of households over this period (cable and ﬁber Internet

from providers other than Google was limited), and there was a rapid expansion in high-

speed Internet access over this time period as Google Fiber became available broadly across

the Kansas City area.

Using HMDA data, we analyze how the market share of FinTech lenders across diﬀerent

neighborhoods in the Kansas City MSA evolved over this period, exploiting the staggered

rollout of Google Fiber across census tracts and controlling for time and census-tract ﬁxed

eﬀects. Results are presented in Internet Appendix G. Our main ﬁnding is that the discrete

improvements in Internet access generated by the entry of Google Fiber did not induce a

We note that, where comparable, our results are generally consistent with Buchak et al. (2017), who

estimate similar reduced-form regressions of the determinants of borrowing from FinTech lenders and other

nonbanks. For example, Buchak et al. (2017) also ﬁnd that FinTech lenders originate a signiﬁcantly higher

share of reﬁnances compared to purchase mortgages, and have a high incidence of missing race and gender.

Besides diﬀerences in modelling choices, we also examine a number of determinants of FinTech demand

which Buchak et al. (2017) do not (e.g., bank branch density, borrower age, or Internet access).

higher share of FinTech borrowing. The results are often not statistically signiﬁcant; in

the cases where they are signiﬁcant they have the opposite sign to that predicted, due to a

migration in mortgage lending from nonbanks to banks.

Consistent with our earlier evidence, these results indicate that adequate Internet access is

unlikely to be a signiﬁcant constraint on the diﬀusion of online mortgage lending, mitigating

concerns about the “digital divide” in this setting.

VIII Conclusion

This paper presents new evidence on how technology is reshaping the U.S. mortgage mar-

ket by studying the vanguard of technology-based lenders. Our results show that FinTech

lenders oﬀer a faster origination process that is less sensitive to ﬂuctuations in demand than

traditional lenders. These improvements are associated with an increase in the propensity to

reﬁnance, especially among borrowers that are likely to beneﬁt from it. We ﬁnd no evidence

that FinTech lending is more risky.

Going forward, we expect that other lenders will seek to replicate the “FinTech model”

characterized by electronic application processes with centralized, semi-automated under-

writing operations. However, it is unclear whether traditional lenders or small institutions

will all be able to adopt these practices as these innovations require signiﬁcant reorganiza-

tion and sizable investments. The end result could be a more concentrated mortgage market

dominated by those ﬁrms that can aﬀord to innovate. From a consumer perspective, we

believe our results shed light on how mortgage credit supply is likely to evolve in the fu-

ture. Speciﬁcally, technology will allow the origination process to be faster and to more

easily accommodate changes in interest rates, leading to greater transmission of monetary

policy to households via the mortgage market. Our ﬁndings also imply that technological

diﬀusion may reduce ineﬃciencies in reﬁnancing decisions, with signiﬁcant beneﬁts to U.S.

households.

Our results have to be considered in the prevailing institutional context of the U.S.

mortgage market. Speciﬁcally, at the time of our study FinTech lenders are non-banks that

securitize their mortgages and do not take deposits. It remains to be seen whether we ﬁnd

the same beneﬁts of FinTech lending as the model spreads to deposit-taking banks and their

borrowers. Changes in banking regulation or the housing ﬁnance system may aﬀect FinTech

lenders going forward. Also, the beneﬁts we document stem from innovations that rely on

hard information; as these innovations spread, they may aﬀect access to credit for those

borrowers with applications that require soft information or borrowers that require direct

communication with a loan oﬃcer. We leave these issues for future research.

References

Acharya, V. V., P. Schnabl, and G. Suarez (2013): “Securitization without risk

transfer,” Journal of Financial Economics, 107, 515–536.

Agarwal, S., G. Amromin, S. Chomsisengphet, T. Landvoigt, T. Piskorski,

A. Seru, and V. Yao (2017): “Mortgage Reﬁnancing, Consumer Spending, and Com-

petition: Evidence From the Home Aﬀordable Reﬁnance Program,” Working Paper 21512,

National Bureau of Economic Research.

Agarwal, S., J. Driscoll, D. Laibson, and X. Gabaix (2009): “The Age of Reason:

Financial Decisions over the Life Cycle and Implications for Regulation,” Brookings Papers

on Economic Activity, 30, 51–101.

Agarwal, S., J. C. Driscoll, and D. I. Laibson (2013): “Optimal Mortgage Reﬁ-

nancing: A Closed-Form Solution,” Journal of Money, Credit and Banking, 45, 591–622.

Agarwal, S., R. J. Rosen, and V. Yao (2015): “Why Do Borrowers Make Mortgage

Reﬁnancing Mistakes?” Management Science, 62, 3494–3509.

Andersen, S., J. Y. Campbell, K. M. Nielsen, and T. Ramadorai (2015): “Inat-

tention and Inertia in Household Finance: Evidence from the Danish Mortgage Market,”

Working Paper 21386, National Bureau of Economic Research.

Badarinza, C., J. Y. Campbell, and T. Ramadorai (2016): “International compar-

ative household ﬁnance,” Annual Review of Economics, 8, 111 – 144.

Bartlett, R., A. Morse, R. Stanton, and N. Wallace (2017): “Consumer Lending

Discrimination in the FinTech Era,” Working paper, UC Berkeley.

Bennett, P., R. Peach, and S. Peristiani (2001): “Structural Change in the Mortgage

Market and the Propensity to Reﬁnance,” Journal of Money, Credit and Banking, 33, 955–

975.

Beraja, M., A. Fuster, E. Hurst, and J. Vavra (2017): “Regional Heterogeneity

and Monetary Policy,” Staﬀ Report No. 731, Federal Reserve Bank of New York.

Bond, P., R. Elul, S. Garyn-Tal, and D. K. Musto (2017): “Does Junior Inherit?

Reﬁnancing and the Blocking Power of Second Mortgages,” Review of Financial Studies,

30, 211–244.

Buchak, G., G. Matvos, T. Piskorski, and A. Seru (2017): “Fintech, Regulatory

Arbitrage, and the Rise of Shadow Banks,” Working Paper 23288, National Bureau of

Economic Research.

Campbell, J. Y. (2006): “Household Finance,” Journal of Finance, 61, 1553 – 1604.

——— (2013): “Mortgage Market Design,” Review of Finance, 17, 1 – 33.

Choi, D. B., H.-S. Choi, and J.-E. Kim (2017): “Clogged Intermediation: Were Home

Buyers Crowded Out?” Working paper, Federal Reserve Bank of New York.

Collard-Wexler, A. and J. De Loecker (2015): “Reallocation and Technology: Ev-

idence from the US Steel Industry,” American Economic Review, 105, 131–71.

D’Acunto, F. and A. G. Rossi (2017): “Regressive Mortgage Credit Redistribution in

the Post-crisis Era,” Working paper, University of Maryland.

DeFusco, A. A., S. Johnson, and J. Mondragon (2017): “Regulating Household

Leverage,” Working paper, Northwestern University.

Deng, Y. and J. M. Quigley (2012): “Woodhead Behavior and the Pricing of Residential

Mortgages,” Working Paper IRES2012-025, NUS Institute of Real Estate Studies.

Di Maggio, M., A. Kermani, B. J. Keys, T. Piskorski, R. Ramcharan, A. Seru,

and V. Yao (2017): “Interest rate pass-through: Mortgage rates, household consumption,

and voluntary deleveraging,” American Economic Review, 107, 3550–3588.

Di Maggio, M., A. Kermani, and C. Palmer (2016): “How Quantitative Easing Works:

Evidence on the Reﬁnancing Channel,” Working Paper No. 22638, National Bureau of

Economic Research.

Drechsler, I., A. Savov, and P. Schnabl (2017): “The deposits channel of monetary

policy,” Quarterly Journal of Economics, 132, 1819–1876.

Freddie Mac (2016): “Your Step-by-Step Mortgage Guide: From Application to

Closing,” http://www.freddiemac.com/singlefamily/docs/Step_by_Step_Mortgage_

Guide_English.pdf, accessed 2017-08-15.

Fuster, A., P. Goldsmith-Pinkham, T. Ramadorai, and A. Walther (2017a):

“Predictably Unequal? The Eﬀects of Machine Learning on Credit Markets,” Discussion

Paper 12448, CEPR.

Fuster, A., S. H. Lo, and P. S. Willen (2017b): “The Time-Varying Price of Financial

Intermediation in the Mortgage Market,” Staﬀ Report 805, Federal Reserve Bank of New

York.

Fuster, A. and P. S. Willen (2010): “$1.25 Trillion is Still Real Money: Some Facts

About the Eﬀects of the Federal Reserve’s Mortgage Market Investments,” Public Policy

Discussion Paper 10-4, Federal Reserve Bank of Boston.

Gete, P. and M. Reher (2017): “Mortgage Supply and Housing Rents,” Review of

Financial Studies, forthcoming.

Goldman Sachs Research (2015): “The Future of Finance,” https://www.planet-

fintech.com/file/167061, accessed 2018-01-15.

——— (2016): “The Mortgage Analyst: Prepays catching up to Quicken? Not

so quick,” http://www.thebondbeat.com/bondbeat/wp-content/uploads/2016/02/

gs-TMA.pdf, accessed 2017-03-15.

Goodman, L. (2016): “Why Rocket Mortgage Won’t Start Another Housing Crisis,” Urban

Institute Urban Wire Blog, February 12. Accessed 2017-03-15.

Housing Wire (2015): “Guaranteed Rate Purchases Remains of Discover Home Loans,”

https://www.housingwire.com/articles/34986-guaranteed-rate-purchases-

remains-of-discover-home-loans, accessed 2018-01-15.

——— (2017): “Quicken Loans’ Rocket Mortgage Finally a Fully Digital Mortgage Thanks

to Pavaso Partnership,” https://www.housingwire.com/articles/41498-quicken-

loans-rocket-mortgage-finally-a-fully-digital-mortgage-thanks-to-pavaso-

partnership, accessed 2018-01-15.

Jiang, W., A. A. Nelson, and E. Vytlacil (2014): “Securitization and Loan Perfor-

mance: Ex Ante and Ex Post Relations in the Mortgage Market,” Review of Financial

Studies, 27, 454–483.

Keys, B., D. Pope, and J. Pope (2016): “Failure to Reﬁnance,” Journal of Financial

Economics, 122, 482 – 499.

Keys, B. J., T. Mukherjee, A. Seru, and V. Vig (2010): “Did securitization lead

to lax screening? Evidence from subprime loans,” Quarterly Journal of Economics, 125,

307–362.

LaCour-Little, M. (2000): “The Evolving Role of Technology in Mortgage Finance,”

Journal of Housing Research, 11, 173–205.

——— (2007): “The Home Purchase Mortgage Preferences of Low- and Moderate-Income

Households,” Real Estate Economics, 35, 265–290.

Mayer, C., T. Piskorski, and A. Tchistyi (2013): “The ineﬃciency of reﬁnancing:

Why prepayment penalties are good for risky borrowers,” Journal of Financial Economics,

107, 694 – 714.

Mian, A. and A. Sufi (2009): “The consequences of mortgage credit expansion: Evidence

from the US mortgage default crisis,” Quarterly Journal of Economics, 124, 1449–1496.

Oliver Wyman (2016): “The Future of Technology in Mortgage Originations,” White

paper.

Petersen, M. and R. Rajan (2002): “Does distance still matter? The information

revolution in small business lending,” Journal of Finance, 57, 2533 – 2570.

Philippon, T. (2016): “The FinTech Opportunity,” Working Paper 22476, National Bureau

of Economic Research.

Purnanandam, A. (2010): “Originate-to-distribute model and the subprime mortgage

crisis,” Review of Financial Studies, 24, 1881–1915.

Quicken Loans (2017a): “Quicken Loans and eOriginal Partner on Next Mortgage

Revolution,” http://www.quickenloans.com/press-room/2017/10/17/quicken-

loans-eoriginal-partner-next-phase-digital-mortgage-revolution/, accessed

2018-01-08.

——— (2017b): “Quicken Loans Continues Simplifying Mortgage Process, Allows

More Financial Information in Fewer Steps,” http://www.quickenloans.com/press-

room/2017/10/23/quicken-loans-continues-simplifying-mortgage-process-

allows-clients-share-financial-information-fewer-steps/, accessed 2018-01-06.

Scharfstein, D. and A. Sunderam (2016): “Market Power in Mortgage Lending and

the Transmission of Monetary Policy,” Working Paper, Harvard University.

Sharpe, S. A. and S. M. Sherlund (2016): “Crowding Out Eﬀects of Reﬁnancing on

New Purchase Mortgages,” Review of Industrial Organization, 48, 209–239.

Stein, J. (2002): “Information production and capital allocation: Decentralized versus

hierarchical ﬁrms,” Journal of Finance, 57, 1891 – 1921.

Syverson, C. (2011): “What Determines Productivity?” Journal of Economic Literature,

49, 326–65.

The Economist (2016): “A New Foundation: One of America’s Biggest Mortgage Lenders

is Not Like the Others,” May 26.

Wall Street Journal (2015): “FHA Oﬀers Olive Branch to Hesitant Lenders,” Septem-

ber 1.

Wong, A. (2016): “Population Aging and the Transmission of Monetary Policy to Con-

sumption,” Working paper, Northwestern University.

Appendix: Data Sources and Variable Deﬁnitions

Variable Deﬁnition Level of

Aggregation

HMDA

Log(income) Log of borrower income Individual

Log(loan size) Log of loan amount Individual

Loan-to-income (LTI) Loan amount divided by borrower income Individual

Jumbo Loan [0,1] Indicator variable equal to one if loan amount exceeds FHFA con-

forming loan limit for the month of origination

Individual

Loan Type [0,1] Indicator variables for conventional, FHA, and VA loans. Con-

ventional is omitted category in regressions.

Individual

Loan purpose [0,1] Indicator variables for whether whether loan is a home purchase

loan or a reﬁnancing

Individual

No Coapplicant [0,1] Indicator variable equal to one if no coapplicant on the mortgage

application

Individual

Owner Occupied [0,1] Indicator for the property being the borrower’s principal dwelling Individual

Gender [0,1] Indicators for borrower gender being Female, Male, or Unknown

(unreported). Male is omitted category in regressions.

Individual

Race & Ethnicity [0,1] Indicators for race & ethnicity. Non-hispanic white is omitted

category in regressions.

Individual

Processing Time Number of calendar days between the application date and action

date of a loan (based on restricted-use version of HMDA data).

Individual

Log(application volume) Log of aggregate application volume National

U.S. Census

% Black or Hispanic Percent of population identifying as Black, Hispanic, or both in

2010

Tract

Population density Thousands of residents in thousands per square mile in 2010 Tract

American Community Survey

% bachelor degree Percent of population 25 years or older with a bachelor’s or higher

degree in 2010

Tract

% with broadband sub-

scription

Percent of population with an Internet subscription other than

dial-up (including DSL, cable, ﬁber optic, mobile broadband,

satellite, or ﬁxed wireless) in 2010

County

Equifax

Credit score Mean credit score of all individuals with a positive mortgage bal-

ance (measured as of 2014Q3)

Tract

Age Median age of individuals with a positive mortgage balance (mea-

sured as of 2014Q3)

Tract

Zillow

Home price appreciation Home price appreciation over the 12 months up to the month of

loan origination

County

Log(2010 home price) Log of average home price as of January 2010 County

NTIA/FCC

High speed internet cov-

erage

Percent of households with access to either cable or ﬁber services,

or Google Fiber, measured at a half-yearly frequency. Data is

collected from the NTIA from 2011 to mid-2014 and from the

FCC from end-2014 to end-2016. Data only available for the ten

largest US states.

Tract

Variable Deﬁnition Level of

Aggregation

FDIC Summary of Deposits

Bank branch density Number of bank branches (in 000s) within a 25 mile radius of the

center of census tract in 2010

Tract

FHA Neighb orhood Watch Early Warning System

One- and two-year com-

parative default rate

Default rate (90+ days delinquent or teriminated in a claim as % of

loans) for FinTech lender i measured as the percent deviation from

the default rate for all FHA loans in the same geography and time

period. 1 year default rate considers only defaults occurring in the

ﬁrst year of the mortgage life. For most analysis, performance is

measured over the sample period 2015:Q3 to 2017:Q3.

MSA,

state or

national

Mix-adjusted 2-year de-

fault rate

Based on the FHA supplementary performance measure (SPM,

equal to the ratio of the lender’s two-year default rate to a bench-

mark default rate for a portfolio of loans with the same mix of

credit scores.) Mix-adjusted default rate is the ratio of the lender’s

SPM to the SPM for all FHA loans in the same geography and

time period.

MSA,

state or

national

Ginnie Mae loan-level data

Default Indicator variable for whether FHA mortgage enters 90+ days

delinquency

Individual

FICO Borrower credit score at origination Individual

LTV Ratio of loan amount to appraised property value Individual

DTI Ratio of total debt payment to borrower income Individual

Equifax CRISM borrower level data

Coupon minus 10-year

Treasury yield

Diﬀerence between average coupon rate on outstanding mortgages

and 10-year Treasury bond yield (used in Figure 2)

National

Reﬁnancing propensity Share of outstanding mortgages in month t that are reﬁnanced in

month t+1

County

FICO Borrower credit score (updated monthly), county average County

CLTV Combined loan-to-value ratio (sum of all mortgage liens divided

by updated home value)

County

Average current rate Current mortgage coupon rate, county average County

FHA/VA share Share of FHA/VA mortgages in county County

Author derived variables

2010 conditional process-

ing time

Average processing time in 2010 (census tract average, residual

after regressing processing time in days on HMDA borrower &

loan characteristics)

Tract

Reﬁnancing incentive Diﬀerence between coupon rate and rate at which borrower would

be indiﬀerent between reﬁnancing and not reﬁnancing based on

the ‘square root rule’ of Agarwal et al. (2013)

Individual

Figures and Tables

Figure 1: Number of FinTech mortgage lenders (according to our classiﬁcation) over time

Figure 2: Application volume, interest rate, and processing time

(a) Mortgage application volume and interest rates

(b) Mortgage application volume and processing time

Figure shows the evolution of the number of applications for new loans in HMDA that result in loan origina-

tions (divided by the number of business days in a given month), plotted against (a) a proxy for borrowers’

reﬁnancing incentive (the diﬀerence between the average coupon on outstanding mortgages and the 10-year

Treasury bond yield), and (b) the median processing time for new loan applications that result in loan

originations and were submitted in the same month.

Figure 3: Illustrating diﬀerences in processing times and elasticity to demand across lender

types.

Processing time (days)

800 900 1000 1100 1200 1300 1400

Aggregate loan application volume (monthly, in 000s)

Banks FinTech Other

Figure shows binned scatter plot (with linear ﬁt) of processing times of originated loans against the volume

of aggregate loan applications by lender type, all measured in HMDA 2010-2016:Q2. Processing times and

application volume are ﬁrst residualized with respect to the following variables: census tract indicators,

calendar month indicators, the log of applicant income, the log of the loan amount, indicators for FHA

loans, VA loans and jumbo loans, applicant race, gender, whether the loan is a reﬁnance, whether the loan

has a coapplicant, whether a pre-approval was requested, the occupancy and lien status of the loan, the

property type, and a dummy indicating whether income is missing. Application volumes are then grouped

into 10 bins, and for each bin the mean of the residualized processing time is calculated and the mean

processing time is added, separately for the three lender types in our analysis.

Figure 4: Reﬁnancing propensities across counties with diﬀerent levels of FinTech market

1.5

Monthly Refinance Propensity (group average, %)

2010m1 2011m1 2012m1 2013m1 2014m1 2015m1 2016m1

Lowest FinTech Tercile Middle Tercile Highest FinTech Tercile

Figure shows average monthly reﬁnance propensities across three groups of counties, sorted based on county-

level FinTech market shares (among reﬁnance loans) over mid-2012 to mid-2013. Data source: CRISM.

Table 1: FinTech and Top 20 Mortgage Originators in 2016, based on Home Mortgage

Disclosure Act (HMDA)

Rank Type of

Lender

Lender Name Volume

(Bn)

Market

Share (%)

FinTech

Since

1 Bank Wells Fargo 132.58 6.62

2 FinTech Quicken Loans 90.55 4.52 2010

3 Bank JPMorgan Chase 75.52 3.77

4 Bank Bank Of America 60.24 3.01

5 FinTech Loandepot.com 35.94 1.80 2016

6 Mtg Freedom Mortgage 32.17 1.61

7 Bank US Bank 30.69 1.53

8 Mtg Caliber Home Loans 27.94 1.40

9 Bank Flagstar 27.31 1.36

10 Mtg United Shore 22.93 1.15

11 Bank Citibank 21.73 1.09

12 FinTech Guaranteed Rate 18.44 0.92 2010

13 Mtg Finance of America 17.63 0.88

14 Mtg Fairway Independent 16.10 0.80

15 Bank USAA Federal Savings 15.52 0.78

16 Mtg Guild Mortgage 15.07 0.75

17 Mtg Stearns Lending 14.93 0.75

18 Bank Suntrust Mortgage 14.77 0.74

19 Bank Primelending 13.87 0.69

20 Mtg Nationstar Mortgage 13.50 0.67

...

23 FinTech Movement Mortgage 11.61 0.58 2014

39 FinTech Everett Financial

(Supreme)

7.62 0.38 2016

534 FinTech Avex Funding

(Better.com)

0.49 0.02 2016

Bank = Depository Institution, Mtg = Non-bank Mortgage Lender, FinTech = FinTech Lender

Market share is based on dollar volume of originations in HMDA.

Table 2: Descriptive statistics by lender type, HMDA data 2010 - mid-2016

Banks Non-bank lenders All Lenders

Non-FinTech FinTech

Mean p50 Mean p50 Mean p50 Mean p50

Originated Mortgages

Applicant Income 121 86.00 102 82.00 102 84.00 115 84.00

Loan amount / income 1.96 1.80 2.46 2.40 2.34 2.19 2.13 2.00

Home Purchase 0.34 0 0.52 1 0.22 0 0.38 0

Reﬁnancing 0.66 1 0.48 0 0.78 1 0.62 1

Jumbo 0.05 0 0.02 0 0.02 0 0.04 0

Loan Type:

Conventional 0.86 1 0.61 1 0.71 1 0.78 1

FHA 0.09 0 0.28 0 0.20 0 0.15 0

VA 0.05 0 0.11 0 0.09 0 0.07 0

Owner Occupied 0.88 1 0.92 1 0.92 1 0.89 1

Male 0.67 1 0.69 1 0.59 1 0.68 1

Female 0.25 0 0.27 0 0.26 0 0.26 0

No Coapplicant 0.45 0 0.52 1 0.50 0 0.48 0

Race:

White 0.79 1 0.78 1 0.68 1 0.78 1

Black/African American 0.04 0 0.06 0 0.05 0 0.05 0

Asian 0.05 0 0.07 0 0.04 0 0.06 0

Other 0.01 0 0.01 0 0.01 0 0.01 0

Unknown 0.11 0 0.09 0 0.22 0 0.11 0

Processing Time 53.04 45.00 50.20 40.00 42.58 37.00 51.71 43.00

Observations 32,751,662 14,742,227 2,306,237 49,800,126

All Applications

Loan Outcome

Originated 0.64 1 0.58 1 0.66 1 0.62 1

Approved, Not Accepted 0.04 0 0.05 0 0.03 0 0.04 0

Denied 0.20 0 0.16 0 0.27 0 0.19 0

Withdrawn 0.09 0 0.15 0 0.03 0 0.11 0

Closed for Incompleteness 0.03 0 0.06 0 0.01 0 0.04 0

Processing Time 47.16 40.00 46.98 35.00 40.11 35.00 46.80 38.00

Observations 51,448,444 25,604,501 3,473,506 80,526,451

Table contains summary statistics of HMDA data by lender type, for loan applications from January 2010 through June 2016. “Banks” are

depository institutions, “Non-bank lenders” are non-bank mortgage lenders, and “FinTech” lenders are classiﬁed according to Section II.

In the ﬁrst part of the table, summary statistics are calculated for originated mortgages only. In the second part of the table, statistics are

calculated for all applications, which include applications that ended up being originated, approved by the lender but not accepted by the

borrower, denied, withdrawn by the applicant before a decision was made, or closed for incompleteness. “Applicant Income” is in thousands

of USD and does not include missing values. ”Loan amount / income“ (LTI) is loan amount divided by applicant income; LTI is winsorized

at the 0.5% level and does not include missing values. “Jumbo” is an indicator for the loan amount being greater than the applicable FHFA

Conforming Loan Limit. “Owner Occupied” is an indicator for the property being the borrower’s principal dwelling. “No Coapplicant” is

an indicator for no coapplicant provided for the loan. Race: “Other” is an indicator for applicant race being American Indian, Alaskan,

Hawaiian, or Paciﬁc Islander. “Unknown” is an indicator for race being unreported or “Not Applicable”. “Processing Time” is the number

of days between application date and action date of a loan. Loan outcomes: “Originated” are applications that were successfully originated.

“Approved, Not Accepted” are applications where the application was approved, but not accepted by applicant. “Denied” are applications

that were denied by originator. “Withdrawn” are applications that were withdrawn by the applicant before a credit decision was made.

“Closed for Incompleteness” are applications where the application ﬁle was closed for incompleteness.

Table 3: FinTech lenders and processing times: Loan-level results

Panel A: Purchase Loans (1) (2) (3) (4) (5)

FinTech Lender -7.93*** -9.43*** -8.33*** -9.24*** -7.46***

(0.52) (0.61) (0.43) (0.48) (0.45)

Log(Applicant Income) -0.55*** -1.00*** -0.44***

(0.04) (0.03) (0.06)

Log(Loan Amount) 4.47*** 4.90*** 6.09***

(0.08) (0.05) (0.13)

FHA 0.65*** 0.27*** -0.34***

(0.15) (0.10) (0.09)

VA 1.68*** 1.51*** 1.91***

(0.22) (0.20) (0.30)

Jumbo 3.17*** 5.29*** 5.90***

(0.22) (0.15) (0.28)

Observations 19,159,345 19,159,345 18,551,855 18,551,855 7,185,042

0.00 0.02 0.23 0.24 0.34

Census Tract-Month No No Yes Yes Yes

Loan Controls No Yes No Yes Yes

Sample All Lenders All Lenders All Lenders All Lenders Nonbanks

Panel B: Reﬁnance Loans (1) (2) (3) (4) (5)

FinTech Lender -9.99*** -13.64*** -10.82*** -14.61*** -9.32***

(0.59) (0.57) (0.79) (0.71) (0.53)

Log(Applicant Income) 0.04 -0.17*** -0.25**

(0.09) (0.06) (0.12)

Log(Loan Amount) 4.74*** 4.60*** 1.24***

(0.10) (0.09) (0.14)

FHA 5.83*** 5.67*** 5.48***

(0.46) (0.40) (0.29)

VA 1.70*** 2.04*** 1.42***

(0.53) (0.44) (0.41)

Jumbo 7.06*** 7.23*** 9.71***

(0.23) (0.21) (0.17)

Observations 30,616,247 30,616,247 30,169,300 30,169,300 8,041,746

0.01 0.10 0.18 0.24 0.29

Census Tract-Month No No Yes Yes Yes

Loan Controls No Yes No Yes Yes

Sample All Lenders All Lenders All Lenders All Lenders Nonbanks

Table reports regressions of loan processing time (in days) on an indicator variable identifying FinTech

lenders, census tract-month ﬁxed eﬀects, loan controls and borrower controls. Data source: HMDA. The

sample consists of originated purchase loans in Panel A and reﬁnancing loans in Panel B with application

dates from 2010 to 2016Q2. Displayed loan controls include the log of applicant income, the log of the loan

amount, indicators for FHA loans, VA loans and jumbo loans. Suppressed loan controls include applicant

race, gender, whether the loan has a coapplicant, whether a preapproval was obtained, the occupancy and

lien status of the loan, the property type, and a dummy indicating whether income is missing. Column (5)

excludes bank lenders. Standard errors reported in parentheses are clustered by lender-month. ***, **, *

indicate statistical signiﬁcance at 1%, 5%, and 10%, respectively.

Table 4: Default rate for FinTech lenders, FHA loans (% deviation from region)

Default deﬁnition

1 Year 2 Year Mix-Adjusted

Default Default 2 Year Default

A. All FHA loans

State level -35.0*** -35.8*** -25.5***

(3.8) (3.9) (4.5)

MSA level -35.3*** -36.2***

(2.4) (2.6)

B. High market share regions only (loans in top quartile markets by lender):

State level -24.7*** -23.4*** -11.7

(8.4) (8.6) (9.9)

MSA level -19.8*** -20.5***

(6.0) (6.1)

C. Disaggregated by purchase versus reﬁnancing (all loans, state level)

Purchase -14.8*** -17.1*** -10.1*

(3.6) (4.3) (5.9)

Reﬁnancing -30.6*** -27.5*** -40.3***

(2.6) (3.5) (2.4)

D. Disaggregated by neighborhood socioeconomic status (all loans, state level)

Underserved (low income/minority) -33.5*** -32.4*** -25.3***

(4.5) (4.4) (4.4)

Not Underserved -36.8*** -36.5*** -25.4***

(3.6) (3.9) (4.3)

E. All FHA loans: longer time series

National level -44.7*** -45.6*** -32.7***

(4.8) (4.5) (5.7)

State level -45.4*** -46.1*** -33.4***

(3.2) (3.1) (3.9)

Table reports weighted average percent diﬀerence in default rate between mortgages from

FinTech lenders and all FHA mortgages originated in same time period and market (ei-

ther MSA, state or national market). Values less than zero indicate lower default rates

for FinTech lenders. These statistics are calculated as the weighted average of compare

(default rate

/default rate

)−1 across FinTech lenders i and regions r, weighting by lender

origination volume. In practice we calculate this weighted average by regressing compare

on a constant term using weighted least squares. Default deﬁnition is either default within

ﬁrst year, default within ﬁrst two years, or the ‘mix-adjusted’ default rate which is based on

the supplementary performance metric, an adjusted default rate which takes into account

the credit score distribution of originations (see text for details). Sample period 2015:Q3 to

2017:Q3 except for panel E, where sample period is 2012:Q3 to 2017:Q3. Data extracted

from the FHA Early Warning System portal in December 2017. Robust standard errors

in parentheses; standard errors clustered by state in state-level regression using longer time

series (section E). ***, **, * indicate statistical signiﬁcance at 1%, 5%, and 10%, respectively.

Table 5: FHA mortgage default regressions based on Ginnie Mae data.

(1) (2) (3) (4) (5) (6) (7)

FinTech -1.29*** -0.97*** -0.93*** -1.51*** -0.79*** -0.91***

(0.13) (0.13) (0.14) (0.13) (0.11) (0.14)

FinTech Share -0.10

(1.00)

FT X FT share -1.70*

(0.98)

FICO -0.04*** -0.05*** -0.03*** -0.04*** -0.04***

(0.00) (0.00) (0.00) (0.00) (0.00)

LTV 0.04*** 0.07*** 0.03*** 0.04*** 0.04***

(0.00) (0.00) (0.00) (0.00) (0.00)

DTI 0.06*** 0.07*** 0.04*** 0.06*** 0.06***

(0.00) (0.00) (0.00) (0.00) (0.00)

Purpose FE No Yes Yes Yes Yes Yes Yes

Month FE Yes Yes No No No Yes No

State FE No No No No No Yes No

MonthXState FE No No Yes Yes Yes No Yes

Loan Controls No No Yes Yes Yes Yes Yes

Mean Y 3.65 3.65 3.65 4.00 2.73 3.65 3.65

R2 0.02 0.02 0.04 0.05 0.03 0.04 0.04

Observations 4097569 4097568 4097544 2966644 1130881 4097548 4097544

Loan Sample All All All Purch. Reﬁ All All

Table reports regressions of indicator for a loan ever entering 90+ day delinquency on an indicator variable

identifying FinTech issuers (or the state-level FinTech market share, demeaned by month; or the interaction

of the FinTech indicator with FinTech market share), state-by-origination month ﬁxed eﬀects, loan controls

and borrower controls. The sample consists of FHA-insured 30-year ﬁxed-rate mortgages originated over

June 2013 to May 2017, obtained from Ginnie Mae MBS monthly loan-level disclosures. Displayed loan

controls include the borrower FICO score, the loan-to-value ratio (LTV) and the debt-to-income ratio

(DTI). Suppressed loans controls include loan purpose type, the log of the loan amount, and indicators

for the number of borrowers, the property type, whether the borrower received down payment assistance,

and for whether FICO, LTV, or DTI are missing. Standard errors reported in parentheses are clustered by

issuer-origination month. ***, **, * indicate statistical signiﬁcance at 1%, 5%, and 10%, respectively.

Table 6: Elasticity of processing time with respect to aggregate application volume: FinTech

vs. other lenders

(1) (2) (3) (4) (5) (6) (7)

ln(App. Vol.) 11.76*** 13.48*** 18.88*** 13.43*** 8.85*** 13.60*** 10.55***

(0.52) (0.47) (0.67) (0.47) (0.45) (0.81) (0.79)

ln(App. Vol.) × FinTech -7.55*** -6.15*** -9.57*** -7.46*** -2.06 -4.45*** -4.47***

(1.46) (1.51) (1.80) (1.50) (1.40) (1.67) (1.56)

Observations 49,775,550 49,775,312 30,615,852 80,495,817 17,024,138 8,927,175 29,048,184

0.14 0.20 0.25 0.17 0.20 0.16 0.16

Loan Controls No Yes Yes Yes Yes Yes Yes

Lender FE Yes Yes Yes Yes Yes Yes Yes

Census Tract FE No Yes Yes Yes Yes Yes Yes

Month FE No Yes Yes Yes Yes Yes Yes

Application Sample Originated Originated Reﬁ All Originated Reﬁ All

Lender Sample All All All All Nonbanks Nonbanks Nonbanks

Table reports regressions of loan processing time (in days) on the log of aggregate monthly application

volume, an interaction with the FinTech indicator, loan controls, lender ﬁxed eﬀects, census-tract ﬁxed

eﬀects and calendar month ﬁxed eﬀects. Data source: HMDA. The sample is restricted to application

dates from 2010 to 2016:Q2. Columns (1), (2), and (5) include all originated loans; Columns (3) and (6)

include originated reﬁnancing loans; and Columns (4) and (7) include all applications (including denied

and withdrawn applications). The sample of lenders includes all lender types in Columns (1)-(4) and

nonbanks only in Columns (5)-(7). Loan controls include the log of applicant income, the log of the loan

amount, indicators for FHA loans, VA loans and jumbo loans, applicant race, gender, whether the loan has

a coapplicant, whether a preapproval was obtained, the occupancy and lien status of the loan, the property

type, and a dummy indicating whether income is missing. Columns (4) and (7) also include indicators

for whether a loan was denied or withdrawn. Standard errors reported in parentheses are clustered by

lender-month. ***, **, * indicate statistical signiﬁcance at 1%, 5%, and 10%, respectively.

Table 7: Elasticity of mortgage application denial probabilities with respect to variation in

aggregate demand for loans — FinTech vs. other lenders.

(1) (2) (3) (4)

ln(App. Vol.) -0.068*** -0.107*** -0.067*** -0.124***

(0.003) (0.004) (0.006) (0.009)

ln(App. Vol.) × FinTech -0.108*** -0.087*** -0.108*** -0.072***

(0.016) (0.015) (0.016) (0.016)

Loan controls Yes Yes Yes Yes

Lender FE Yes Yes Yes Yes

Census Tract FE Yes Yes Yes Yes

R2 0.18 0.18 0.27 0.30

Observations 68,793,269 44,728,223 23,538,398 13,328,381

Application Sample All Reﬁ All Reﬁ

Lender Sample All All Nonbanks Nonbanks

Table reports regressions of indicator for loan application denial on the log of aggregate application volume,

an interaction with the FinTech indicator, loan controls, lender ﬁxed eﬀects, census-tract ﬁxed eﬀects and

calendar month ﬁxed eﬀects. Data source: HMDA. The sample is restricted to application dates from 2010

to 2016:Q2. Applications are included if they result in either a loan origination, in the application being

approved by the lender but not accepted by the borrower, or an application denial. The sample of lenders

includes all lender types in Columns (1)-(2) and nonbanks only in Columns (3)-(4). Loan controls include

the log of applicant income, the log of the loan amount, indicators for FHA loans, VA loans and jumbo

loans, applicant race, gender, whether the loan has a coapplicant, whether a preapproval was obtained,

the occupancy and lien status of the loan, the property type, and a dummy indicating whether income is

missing. Standard errors reported in parentheses are clustered by lender-month. ***, **, * indicate statistical

signiﬁcance at 1%, 5%, and 10%, respectively.

Table 8: Elasticity of originations with respect to changes in aggregate volume: FinTech vs.

other lenders

(1) (2) (3) (4)

∆ ln(App. Vol.) 1.17*** 1.57*** 1.17*** 1.71***

(0.01) (0.01) (0.01) (0.02)

∆ ln(App. Vol.) × FinTech -0.06 0.06 -0.06 -0.08

(0.07) (0.12) (0.07) (0.12)

Observations 52,030 51,311 24,450 23,831

Adjusted R

0.35 0.35 0.32 0.33

Month FE Yes Yes Yes Yes

Application Sample Originated Reﬁ Originated Reﬁ

Lender Sample All All Nonbanks Nonbanks

Table reports regressions of the log change in lender-level originated loans on the log change in aggregate

application volumes, an interaction with the FinTech indicator, loan controls, lender ﬁxed eﬀects, census-

tract ﬁxed eﬀects and calendar month ﬁxed eﬀects. The unit of observation is lender-month. Data source:

HMDA. The sample is restricted to 2010 through 2016:Q2. Columns (1) and (3) include all originated loans;

columns (2) and (4) included originated reﬁnancing loans. The sample of lenders includes all lender types in

columns (1) and (2) and nonbanks only in columns (3) and (4). Standard errors reported in parentheses are

White-Huber standard errors. ***, **, * indicate statistical signiﬁcance at 1%, 5%, and 10%, respectively.

Table 9: FinTech market share and reﬁnancing propensities: county-level results.

(1) (2) (3) (4)

All All 30yr FRM 30yr FRM

FinTech Share

Q−1

(MA) 1.121

∗∗∗

0.689

∗∗∗

1.195

∗∗∗

0.706

∗∗∗

(0.204) (0.142) (0.223) (0.157)

Average FICO/10 0.067

∗∗∗

0.071

∗∗∗

(0.012) (0.013)

Average CLTV/10 -0.094

∗∗∗

-0.104

∗∗∗

(0.007) (0.008)

Average current rate 1.135

∗∗∗

1.202

∗∗∗

(0.059) (0.062)

FHA/VA share 0.190 0.185

(0.315) (0.332)

County ﬁxed eﬀects Yes Yes Yes Yes

Month ﬁxed eﬀects Yes Yes Yes Yes

Mean of dep var 0.54 0.54 0.59 0.59

R2 0.78 0.81 0.76 0.79

Obs. 39000 39000 39000 39000

Table regresses county-level monthly reﬁnancing propensities (deﬁned as the share of outstanding mortgages

in month t − 1 that are reﬁnanced in month t, in percentage points) on the FinTech share in a county

(4-quarter rolling average, lagged one quarter; range [0,1]), county ﬁxed eﬀects, month ﬁxed eﬀects, and

average characteristics of outstanding loans in the county: FICO, updated combined loan-to-value ratios

(CLTV), average coupon rate, and the share of FHA/VA mortgages. Data sources: CRISM for reﬁnancing

propensities and loan characteristics, HMDA for FinTech market shares. In columns (1) and (2), reﬁnancing

propensities are calculated based on all loans; in columns (3) and (4), based on 30-year ﬁxed-rate mortgages

only. Sample covers January 2010 through June 2016 for the largest 500 counties by count of outstanding

mortgages in December 2013 (see Appendix D for details). Standard errors reported in parentheses are

clustered by county. ***, **, * indicate statistical signiﬁcance at 1%, 5%, and 10%, respectively.

Table 10: FinTech market share and reﬁnancing propensities, by reﬁnancing incentive bins.

(1) (2) (3) (4) (5) (6) (7)

Reﬁ incentive (ADL) < −1 [−1, −0.5) [−0.5, 0) [0, 0.5) [0.5, 1) ≥ 1 All

FT Share

Q−1

(MA) -0.140* 1.028*** 2.008*** 1.985*** 1.444*** 0.507* 1.436***

(0.073) (0.200) (0.304) (0.353) (0.347) (0.267) (0.229)

County FE Yes Yes Yes Yes Yes Yes Yes

Month FE Yes Yes Yes Yes Yes Yes Yes

Loan controls Yes Yes Yes Yes Yes Yes Yes

Mean Y 0.12 0.46 0.85 1.04 1.05 0.78 0.59

R2 0.00 0.00 0.01 0.01 0.01 0.01 0.00

Obs. 64,866,392 42,085,823 38,988,748 29,249,088 19,039,098 20,745,039 214,996,787

Table regresses indicator for whether a borrower reﬁnanced their loan in a given month on the FinTech share

in a county (4-quarter rolling average, lagged one quarter), county ﬁxed eﬀects, month ﬁxed eﬀects, and the

following loan controls: 5-point bins of CLTV, 20-point bins of FICO, a cubic function in the age of the

reﬁnanced loan, the log of the balance of the reﬁnanced loan, and an indicator for whether the reﬁnanced loan

was an FHA/VA loan. Data sources: CRISM for reﬁnancing propensities and loan characteristics, HMDA

for FinTech market shares. For columns (1)-(6), borrowers are separated into 6 bins depending on their

reﬁnancing incentive based on the Agarwal et al. (2013) (ADL) calculation. Negative incentives (expressed

in percentage points of interest rates) mean that according to ADL a borrower should not reﬁnance; positive

incentives mean they should reﬁnance. The ﬁnal column (7) pools all bins. Sample includes 30-year ﬁxed-

rate mortgages only. Standard errors reported in parentheses are clustered by county. ***, **, * indicate

statistical signiﬁcance at 1%, 5%, and 10%, respectively.

Table 11: Who Borrows from FinTech Mortgage Lenders?

Dependent variable: = 100 if originator is FinTech lender, = 0 otherwise

Purchases Reﬁnances

All Nonbanks All Nonbanks

Borrower income and demography

Log(income) 0.104

∗∗∗

0.701

∗∗∗

-0.833

∗∗∗

-0.159

∗∗∗

(0.00650) (0.0173) (0.00725) (0.0275)

Gender:

Female 0.0592

∗∗∗

0.184

∗∗∗

0.756

∗∗∗

3.056

∗∗∗

(0.00947) (0.0208) (0.0126) (0.0379)

Unknown 2.887

∗∗∗

10.13

∗∗∗

6.728

∗∗∗

24.99

∗∗∗

(0.0421) (0.117) (0.0437) (0.100)

Race and ethnicity:

Black -0.306

∗∗∗

-0.387

∗∗∗

-0.415

∗∗∗

1.166

∗∗∗

(0.0254) (0.0495) (0.0291) (0.0814)

Hispanic -0.880

∗∗∗

-1.577

∗∗∗

-1.432

∗∗∗

-1.982

∗∗∗

(0.0200) (0.0391) (0.0250) (0.0629)

Unknown 1.551

∗∗∗

3.220

∗∗∗

3.632

∗∗∗

6.540

∗∗∗

(0.0294) (0.0658) (0.0310) (0.0710)

% black or hispanic

TRACT

-0.228

∗∗∗

-1.064

∗∗∗

-0.256

∗∗∗

-2.273

∗∗∗

(0.0166) (0.0394) (0.0165) (0.0501)

Access to ﬁnance

Credit score

TRACT

-0.279

∗∗∗

-0.731

∗∗∗

-1.068

∗∗∗

-3.002

∗∗∗

(0.0192) (0.0468) (0.0193) (0.0618)

Bank branch density

TRACT

0.467

∗∗∗

0.954

∗∗∗

0.275

∗∗∗

0.479

∗∗∗

(0.0262) (0.0574) (0.0201) (0.0530)

Technology diﬀusion and adoption

Population density

TRACT

0.141

∗∗∗

0.920

∗∗∗

-0.0691

∗∗∗

0.421

∗∗∗

(0.0275) (0.0697) (0.0236) (0.0607)

Borrower age

TRACT

0.119

∗∗∗

0.340

∗∗∗

0.263

∗∗∗

0.869

∗∗∗

(0.0168) (0.0400) (0.0169) (0.0502)

% bachelor degree

TRACT

0.307

∗∗∗

0.920

∗∗∗

0.262

∗∗∗

0.690

∗∗∗

(0.0213) (0.0529) (0.0180) (0.0553)

Internet access

% high speed coverage

TRACT

0.101

∗∗∗

0.255

∗∗∗

0.0689

∗∗∗

0.371

∗∗∗

(0.0127) (0.0316) (0.0127) (0.0461)

% with broadband subscription

CTY

-0.132

∗∗∗

-0.487

∗∗∗

-0.0344

∗∗

-0.0551

(0.0179) (0.0460) (0.0167) (0.0555)

Local housing market conditions

% home price appreciation

CTY

-0.0362

∗∗∗

-0.836

∗∗∗

0.277

∗∗∗

-1.258

∗∗∗

(0.0114) (0.0258) (0.0132) (0.0382)

Processing time coeﬃcients

TRACT

0.0182 0.205

∗∗∗

0.588

∗∗∗

1.599

∗∗∗

(0.0111) (0.0269) (0.0119) (0.0397)

Log(2010 home price)

CTY

-0.127

∗∗∗

-0.688

∗∗∗

-0.812

∗∗∗

-2.993

∗∗∗

(0.0188) (0.0471) (0.0213) (0.0675)

Additional controls Yes Yes Yes Yes

Observations 20790255 8901875 32936746 9888845

Mean of Dependent Variable 2.888 6.745 6.129 20.41

Linear probability model based on HMDA data from 2010-16. All continuous right-hand size variables

normalized to have mean of zero and standard deviation of one. Superscripts

TRACT

and

CTY

indicate

variable is measured at the census tract or county level of aggregation, respectively, rather than at the

loan level. Robust standard errors in parentheses, clustered by census tract. Regressions include controls

for loan size, loan type, dummies for jumbo loan, coapplicant, owner occupied, other race categories, and

missing values for any variable with positive incidence of missing values. See Internet Appendix for full

results including coeﬃcients on these variables as well as univariate regressions. See Data Appendix for

variable deﬁnitions and sources.

∗

p < 0.10,

∗∗

p < 0.05,

∗∗∗

p < 0.01

Internet Appendix for

“The Role of Technology in Mortgage Lending”

Andreas Fuster, Matthew Plosser, Philipp Schnabl, and James Vickery

A Processing time: additional analysis

Figure A.1: Distribution of processing times by lender type. (These are residuals after

controlling for loan characteristics and census tract × month ﬁxed eﬀects as in Table 3.)

(a) Purchase mortgages

(b) Reﬁnance mortgages

Table A.1: Testing whether high FinTech probability is associated with slower processing

time for non-FinTech lenders.

(1) (2) (3) (4)

FinTech 2.547*** -0.808*** 3.109*** -0.559***

(0.163) (0.144) (0.216) (0.183)

Loan controls Yes Yes Yes Yes

Lender-Month FE Yes Yes Yes Yes

Lender-Census Tract FE No Yes No Yes

R2 0.22 0.27 0.30 0.33

Observations 47180463 44210993 28616765 26127401

Loan type All All Reﬁ Reﬁ

Sample Non-Fintech Non-Fintech Non-Fintech Non-Fintech

Table regresses loan processing time (in days) for non-FinTech lenders on the predicted probability that an

application would go to a FinTech lender (

FinTech), lender-month ﬁxed eﬀects, lender-census tract ﬁxed

eﬀects, and loan controls.

FinTech comes from an unreported ﬁrst-stage OLS regression where, in the full

sample including all lender types, an indicator for a loan being originated by a FinTech lender is regressed on

census tract-month ﬁxed eﬀects and loan controls. In both stages, loan controls include the log of applicant

income, the log of the loan amount, indicators for FHA loans, VA loans and jumbo loans, applicant race,

gender, loan purpose (purchase or reﬁnancing), whether the loan has a coapplicant, whether a preapproval

was obtained, the occupancy and lien status of the loan, the property type, and a dummy indicating whether

income is missing. Both purchase and reﬁnance loans are included in columns (1)-(2), while only reﬁnance

loans are included in columns (3)-(4). Standard errors reported in parentheses are clustered by lender-month.

***, **, * indicate statistical signiﬁcance at 1%, 5%, and 10%, respectively.

B Is FinTech lending cheaper?

Table A.2: FHA mortgage interest rate regressions based on Ginnie Mae data. Includes

30-year ﬁxed-rate mortgages originated 2013-2017.

(1) (2) (3) (4)

FinTech 0.000 -0.023** -0.075*** 0.002

(0.009) (0.010) (0.010) (0.008)

FICO -0.002*** -0.002*** -0.001***

(0.000) (0.000) (0.000)

LTV 0.000*** 0.003*** -0.001***

(0.000) (0.000) (0.000)

DTI 0.000*** 0.001*** -0.001***

(0.000) (0.000) (0.000)

Sample All All Purch. Reﬁ

Purpose FE? No Yes Yes Yes

Month FE? Yes No No No

MonthXState FE? No Yes Yes Yes

Loan cont.? No Yes Yes Yes

Mean Y 4.00 4.00 4.01 3.96

R2 0.31 0.41 0.42 0.46

Observations 4097569 4097544 2966644 1130881

Table regresses mortgage interest rate on an indicator variable identifying FinTech issuers, state-by-

origination month ﬁxed eﬀects, loan controls and borrower controls. The sample consists of FHA-insured

30-year ﬁxed-rate mortgages originated over June 2013 to June 2017, obtained from Ginnie Mae MBS

monthly loan-level disclosures. Displayed loan controls include the borrower FICO score, the loan-to-value

ratio (LTV) and the debt-to-income ratio (DTI). Suppressed loans controls include loan purpose type, the

log of the loan amount, and indicators for the number of borrowers, the property type, whether the bor-

rower received down payment assistance, and for whether FICO, LTV, or DTI are missing. Standard errors

reported in parentheses are clustered by issuer-origination month. ***, **, * indicate statistical signiﬁcance

at 1%, 5%, and 10%, respectively.

C Is FinTech lending more elastic? Additional results

Table A.3: Elasticity of processing time with respect to demand proxies: FinTech vs. other

lenders

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

Reﬁ Incentive 4.79*** 6.72*** 5.14***

(0.20) (0.28) (0.19)

Reﬁ Inc. × FinTech -3.95*** -5.45*** -4.56***

(0.64) (0.74) (0.65)

Bartik App. × FinTech 9.73*** 13.78*** 8.97***

(0.34) (0.50) (0.35)

Bartik App. × FinTech 0.05 -3.27*** -2.51***

(0.82) (0.82) (0.67)

Observations 49,775,312 30,615,852 80,495,817 49,775,312 30,615,852 80,495,817

0.20 0.25 0.17 0.19 0.24 0.17

Loan Controls Yes Yes Yes Yes Yes Yes

Lender FE Yes Yes Yes Yes Yes Yes

Census Tract FE Yes Yes Yes Yes Yes Yes

Month FE Yes Yes Yes Yes Yes Yes

Application Sample Originated Reﬁ All Originated Reﬁ All

Lender Sample All All All All All All

Table A.3 regresses loan processing time on two proxies for aggregate mortgage demand: the average

outstanding coupon less the 10-yr Treasury yield (Reﬁ Incentive) and the log of the weighted sum of

county-level applications where weights are the unconditional market share of applications received in

the county (Bartik Applications). Regressions include an interaction between the proxy and the FinTech

indicator, loan controls, lender ﬁxed eﬀects, census-tract ﬁxed eﬀects and calendar month ﬁxed eﬀects. The

sample is restricted to application dates from 2010 to 2016Q2. Columns 1 and 4 include all originated

loans; Columns 2 and 5 included originated reﬁnancing loans; and Columns 3 and 6 include all applications

(including denied applications). The sample of lenders includes all lender types. Loan controls include the

log of applicant income, the log of the loan amount, indicators for FHA loans, VA loans and jumbo loans,

applicant race, gender, whether the loan has a coapplicant, whether the application was a preapproval,

the occupancy and lien status of the loan, the property type, and a dummy indicating whether income is

missing. Columns 3 and 6 also include indicators for whether a loan was denied or withdrawn. Standard

errors reported in parentheses are clustered by lender-month. ***, **, * indicate statistical signiﬁcance at

1%, 5%, and 10%, respectively.

D FinTech & reﬁnancing: additional analysis

A.1 Sample construction

We pull all active loans in CRISM in December 2013 and select the 500 counties with the

highest number of loans. To limit the sample size while still having suﬃcient data coverage

across the counties, we take 12,000 loans from each county (roughly the number of loans in

the smallest county in the top 500). We then take the individual CRISM identiﬁers that were

associated with these loans, and pull all records associated with those individuals from 2010

through 2016. By restricting to the largest counties, we are able to get accurate reﬁnance

propensities for a cross-section of the country at the county level while limiting our sample

size for computational reasons. This sample selection procedure gives us a sample of over

325 million loan-month observations, made up of 7.2 million distinct loans from 5.1 million

distinct borrowers.

We identify reﬁnances and calculate reﬁnance propensities and cashouts at the county

level following the same procedure in Beraja et al. (2017). Reﬁnance propensities at the

county level are deﬁned as the percentage of loans from month t − 1 that are reﬁnanced

in month t. We create panels both at the county and individual level with these identiﬁed

reﬁnances.

Figure A.2 shows the average reﬁnance propensity over time as well as the number of

originated reﬁnance loans in the same counties, as recorded in HMDA (where we sum loans

by application month). We track the evolution of originations fairly closely.

A.2 Additional results

In Table A.4 we complement the ﬁndings in Section VI by studying the properties of 30-

year FRMs that were reﬁnanced into new 30-year FRMs over our sample period. The ﬁrst

two columns of the table study whether a reﬁnance was optimal (i.e. whether the interest

rate saving was large enough) according to the ADL rule. In column (1), we do this based

on comparing the rate on the old mortgage to the market rate at the time the reﬁnance

happened (similar to how we deﬁne reﬁnancing incentives in the main text). In column (2),

we instead directly use the rate on the new (reﬁnance) mortgage. We see that in both cases,

a higher local FinTech market share increases the probability that a reﬁnance is classiﬁed

as optimal. Interestingly, the association is stronger when we use the actual mortgage rate

rather than the market rate, even though based on that metric, actually fewer reﬁnances

Note that our CRISM sample design (explained above) over-samples the relatively smaller counties

among the top 500; if we weight counties similarly in HMDA, the two lines become even closer.

1.2

Aggregate Refinance Propensity (%)

100

200

300

400

500

600

HMDA Refinances (000s)

2010m1 2011m1 2012m1 2013m1 2014m1 2015m1 2016m1

HMDA Refinances (000s)

Aggregate Refinance Propensity (from CRISM)

Figure A.2: Reﬁnance propensity over time: comparing CRISM-derived measure to number

of reﬁnance mortgages in top 500 counties recorded in HMDA.

(only 41%) are classiﬁed as optimal.

In column (3), instead of relying on the ADL calculation, we directly look at the gap

between the old mortgage rate and the new mortgage rate, which averages 1.35%. There,

again, a higher local FinTech share is associated with a larger gap. Finally, the last columns

shows that in places with higher FinTech shares, borrowers were more likely to also withdraw

some home equity when reﬁnancing.

This reﬂects the fact that, on average, rates on actually originated mortgages tend to be somewhat

higher than the rate reported in the Freddie Mac Primarly Mortgage Market Survey, which applies to the

highest credit quality borrowers.

The cash out indicator that is used as the left-hand side variable here is equal to 1 if, after subtracting 2

percent from the new loan to cover closing costs, the new mortgage is at least $5,000 above the old mortgage

(including junior liens) that is being paid oﬀ.

Table A.4: Testing for link between local FinTech share and properties of realized reﬁnances

of 30-year ﬁxed-rate mortgages.

(1) (2) (3) (4)

Opt. reﬁ? Opt. reﬁ? Rate gap Cash out?

(mkt rate) (actual rate) (old−new)

FT Share

Q−1

(MA) 0.266*** 0.610*** 0.939*** 0.175**

(0.083) (0.092) (0.122) (0.081)

County FEs Yes Yes Yes Yes

Month FEs Yes Yes Yes Yes

Loan controls Yes Yes Yes Yes

Mean Y 0.55 0.41 1.34 0.17

R2 0.35 0.25 0.42 0.13

Obs. 666,070 666,070 666,070 666,072

Table shows results of four diﬀerent regressions of characteristics of reﬁnance loans in CRISM where both

the old and new mortgage are 30-year FRMs. The left-hand side variables are, by column, 1) an indicator

for whether a reﬁnancing occurred at a time where the market interest rate was below the rate at which

the Agarwal et al. (2013) (ADL) rule would prescribe that the borrower reﬁnance (so “1” would mean the

reﬁnancing was “optimal” in that sense); (2) an indicator of whether the mortgage rate on the new (reﬁnance)

loan is below the ADL rate; (3) the diﬀerence between the old mortgage rate and the new mortgage rate

(winsorized at 1%); (4) an indicator variable for the reﬁnance involving “cashing out” home equity (set equal

to 1 if the balance of the new loan exceeds the balance of the old loan by more than $5000 plus closing costs

(assumed to correspond to 2 percent of the loan amount). Independent variables in each case include the

one-quarter-lagged four-quarter county-level FinTech market share, county ﬁxed eﬀects, month ﬁxed eﬀects,

and the following loan controls: 5-point bins of CLTV, 20-point bins of FICO, a cubic function in the age of

the reﬁnanced loan, the log of the balance of the reﬁnanced loan, and an indicator for whether the reﬁnanced

loan was an FHA/VA loan. Standard errors reported in parentheses are clustered by county. ***, **, *

indicate statistical signiﬁcance at 1%, 5%, and 10%, respectively.

E Spatial Variation in FinTech Mortgage Lending

Figure A.3: Market share of FinTech lenders by county

Calendar year 2010

Calendar year 2016

FinTech market share by county in 2010 and 2016. Figure reﬂects all lender types and both purchase

mortgages and reﬁnancings. FinTech lenders classiﬁed using the procedure described in Section II. Data

source: HMDA.

F Cross-sectional regressions: Additional results

Dependent variable: = 100 if ﬁntech lender, = 0 otherwise

Purchases Reﬁnances

All Nonbanks All Nonbanks

Univariate Multivariate Univariate Multivariate Univariate Multivariate Univariate Multivariate

Borrower income and demography

Log(income) -0.0932

∗∗∗

0.104

∗∗∗

0.761

∗∗∗

0.701

∗∗∗

-0.549

∗∗∗

-0.833

∗∗∗

-2.677

∗∗∗

-0.159

∗∗∗

(0.00777) (0.00650) (0.0242) (0.0173) (0.00877) (0.00725) (0.0321) (0.0275)

Gender:

Female 0.00683 0.0592

∗∗∗

-0.502

∗∗∗

0.184

∗∗∗

-0.130

∗∗∗

0.756

∗∗∗

0.199

∗∗∗

3.056

∗∗∗

(0.00973) (0.00947) (0.0218) (0.0208) (0.0119) (0.0126) (0.0380) (0.0379)

Unknown 3.027

∗∗∗

2.887

∗∗∗

13.09

∗∗∗

10.13

∗∗∗

8.712

∗∗∗

6.728

∗∗∗

30.88

∗∗∗

24.99

∗∗∗

(0.0334) (0.0421) (0.120) (0.117) (0.0384) (0.0437) (0.0990) (0.100)

Race and ethnicity:

Black 0.0808

∗∗∗

-0.306

∗∗∗

-1.181

∗∗∗

-0.387

∗∗∗

-0.218

∗∗∗

-0.415

∗∗∗

-2.862

∗∗∗

1.166

∗∗∗

(0.0276) (0.0254) (0.0568) (0.0495) (0.0298) (0.0291) (0.0877) (0.0814)

Hispanic -0.729

∗∗∗

-0.880

∗∗∗

-3.314

∗∗∗

-1.577

∗∗∗

-1.542

∗∗∗

-1.432

∗∗∗

-7.162

∗∗∗

-1.982

∗∗∗

(0.0180) (0.0200) (0.0370) (0.0391) (0.0253) (0.0250) (0.0759) (0.0629)

Unknown 2.594

∗∗∗

1.551

∗∗∗

8.604

∗∗∗

3.220

∗∗∗

7.206

∗∗∗

3.632

∗∗∗

19.53

∗∗∗

6.540

∗∗∗

(0.0262) (0.0294) (0.0796) (0.0658) (0.0303) (0.0310) (0.0814) (0.0710)

% black or hispanic

TRACT

0.0449

∗∗∗

-0.228

∗∗∗

-0.816

∗∗∗

-1.064

∗∗∗

0.288

∗∗∗

-0.256

∗∗∗

-1.452

∗∗∗

-2.273

∗∗∗

(0.0102) (0.0166) (0.0224) (0.0394) (0.0117) (0.0165) (0.0393) (0.0501)

Access to ﬁnance

Credit score

TRACT

-0.0777

∗∗∗

-0.279

∗∗∗

0.408

∗∗∗

-0.731

∗∗∗

-0.532

∗∗∗

-1.068

∗∗∗

-2.523

∗∗∗

-3.002

∗∗∗

(0.0123) (0.0192) (0.0315) (0.0468) (0.0120) (0.0193) (0.0423) (0.0618)

Bank branch density

TRACT

0.523

∗∗∗

0.467

∗∗∗

1.040

∗∗∗

0.954

∗∗∗

0.266

∗∗∗

0.275

∗∗∗

-1.382

∗∗∗

0.479

∗∗∗

(0.0239) (0.0262) (0.0604) (0.0574) (0.0186) (0.0201) (0.0623) (0.0530)

Technology diﬀusion and adoption

Population density

TRACT

0.269

∗∗∗

0.141

∗∗∗

0.672

∗∗∗

0.920

∗∗∗

-0.000996 -0.0691

∗∗∗

-1.538

∗∗∗

0.421

∗∗∗

(0.0237) (0.0275) (0.0669) (0.0697) (0.0194) (0.0236) (0.0714) (0.0607)

Borrower age

TRACT

0.0400

∗∗∗

0.119

∗∗∗

0.673

∗∗∗

0.340

∗∗∗

-0.0186 0.263

∗∗∗

1.680

∗∗∗

0.869

∗∗∗

(0.0154) (0.0168) (0.0390) (0.0400) (0.0161) (0.0169) (0.0538) (0.0502)

% bachelor degree

TRACT

0.116

∗∗∗

0.307

∗∗∗

0.940

∗∗∗

0.920

∗∗∗

-0.199

∗∗∗

0.262

∗∗∗

-1.388

∗∗∗

0.690

∗∗∗

(0.0175) (0.0213) (0.0476) (0.0529) (0.0143) (0.0180) (0.0489) (0.0553)

Internet access

% high speed 0.192

∗∗∗

0.101

∗∗∗

0.294

∗∗∗

0.255

∗∗∗

0.120

∗∗∗

0.0689

∗∗∗

-0.611

∗∗∗

0.371

∗∗∗

coverage

TRACT

(0.0118) (0.0127) (0.0316) (0.0316) (0.0130) (0.0127) (0.0575) (0.0461)

% with broadband -0.0924

∗∗∗

-0.132

∗∗∗

-0.466

∗∗∗

-0.487

∗∗∗

-0.279

∗∗∗

-0.0344

∗∗

-2.864

∗∗∗

-0.0551

subscription

CTY

(0.0131) (0.0179) (0.0341) (0.0460) (0.0138) (0.0167) (0.0462) (0.0555)

Local housing market conditions

% home price -0.0522

∗∗∗

-0.0362

∗∗∗

-0.971

∗∗∗

-0.836

∗∗∗

0.315

∗∗∗

0.277

∗∗∗

-1.999

∗∗∗

-1.258

∗∗∗

appreciation

CTY

(0.0112) (0.0114) (0.0271) (0.0258) (0.0137) (0.0132) (0.0443) (0.0382)

Processing time 0.0961

∗∗∗

0.0182 0.204

∗∗∗

0.205

∗∗∗

0.760

∗∗∗

0.588

∗∗∗

1.561

∗∗∗

1.599

∗∗∗

coeﬃcients

TRACT

(0.0108) (0.0111) (0.0290) (0.0269) (0.0133) (0.0119) (0.0502) (0.0397)

Log(2010 home price)

CTY

-0.150

∗∗∗

-0.127

∗∗∗

-0.628

∗∗∗

-0.688

∗∗∗

-0.440

∗∗∗

-0.812

∗∗∗

-4.321

∗∗∗

-2.993

∗∗∗

(0.0111) (0.0188) (0.0284) (0.0471) (0.0139) (0.0213) (0.0411) (0.0675)

Observations 20790255 20790255 8901875 8901875 32936746 32936746 9888845 9888845

Mean Dependent Var 2.888 2.888 6.745 6.745 6.129 6.129 20.41 20.41

Dependent variable: = 100 if ﬁntech lender, = 0 otherwise

Purchases Reﬁnances

All Nonbanks All Nonbanks

Univariate Multivariate Univariate Multivariate Univariate Multivariate Univariate Multivariate

Additional race variables

American Indian/Alaska Native -0.605

∗∗∗

-0.351

∗∗∗

-1.837

∗∗∗

-1.165

∗∗∗

0.185

∗∗∗

0.873

∗∗∗

0.431

∗∗

1.100

∗∗∗

(0.0471) (0.0469) (0.100) (0.103) (0.0697) (0.0681) (0.200) (0.199)

Asian -0.401

∗∗∗

-0.223

∗∗∗

-0.820

∗∗∗

-0.573

∗∗∗

-1.673

∗∗∗

-0.582

∗∗∗

-8.575

∗∗∗

-2.263

∗∗∗

(0.0289) (0.0309) (0.0679) (0.0698) (0.0403) (0.0445) (0.105) (0.123)

Hawaiian/Paciﬁc Islander -0.198

∗∗∗

-0.0433 -1.731

∗∗∗

-0.307

∗∗∗

-0.815

∗∗∗

-0.316

∗∗∗

-5.427

∗∗∗

0.309

(0.0611) (0.0614) (0.114) (0.113) (0.0809) (0.0916) (0.212) (0.220)

Missing variable indicators

Missing Log(income) -3.037

∗∗∗

-5.187

∗∗∗

-7.263

∗∗∗

-11.55

∗∗∗

-2.545

∗∗∗

-9.490

∗∗∗

-14.23

∗∗∗

-18.42

∗∗∗

(0.0144) (0.0297) (0.0526) (0.151) (0.0207) (0.0290) (0.0609) (0.0693)

Missing % black or hispanic

TRACT

4.111

∗∗∗

5.285

∗∗∗

2.929

∗∗∗

7.582

∗∗∗

-0.0931 3.341 -5.258 4.777

(1.214) (1.383) (0.556) (1.971) (2.551) (2.698) (6.680) (5.378)

Missing Credit score

TRACT

-0.916

∗∗∗

-0.589

∗∗∗

-2.476

∗∗∗

-0.948 -1.354

∗∗∗

-0.312 -4.354

∗∗∗

-1.116

(0.191) (0.198) (0.462) (0.636) (0.401) (0.397) (1.191) (1.318)

Missing Bank branch density

TRACT

0.156

∗∗∗

0.129

∗∗∗

-0.213

∗∗∗

0.0722 0.615

∗∗∗

0.234

∗∗∗

0.468

∗∗∗

0.469

∗∗∗

(0.0272) (0.0318) (0.0670) (0.0771) (0.0294) (0.0305) (0.108) (0.101)

Missing Population density

TRACT

1.412 -1.939

∗∗∗

1.100 -3.873

∗∗∗

-1.702 -2.794 -6.470

∗

-1.920

(1.632) (0.432) (1.476) (0.996) (1.100) (1.920) (3.674) (5.038)

Missing Borrower age

TRACT

-0.724

∗∗

0.176 -2.392

∗∗∗

-0.343 -1.070 -0.336 -4.657

∗∗

-1.447

(0.363) (0.429) (0.706) (0.998) (0.767) (0.707) (2.145) (2.225)

Missing % bachelor degree

TRACT

2.803

∗

0.992 1.941

∗

0.544 -1.160 -0.289 -6.413 -3.266

(1.535) (0.726) (1.098) (1.945) (1.560) (1.629) (4.561) (5.304)

Missing % high speed -0.947

∗∗∗

-0.886

∗∗∗

-1.532

∗∗∗

-1.917

∗∗∗

-0.718

∗∗∗

-0.652

∗∗∗

-0.575

∗∗∗

-2.222

∗∗∗

coverage

TRACT

(0.0251) (0.0250) (0.0628) (0.0604) (0.0268) (0.0243) (0.0979) (0.0766)

Missing % with broadband -0.222

∗∗∗

0.613

∗∗∗

2.194

∗∗∗

3.157

∗∗∗

-0.482

∗∗∗

0.0861

∗∗

6.369

∗∗∗

2.676

∗∗∗

subscription

CTY

(0.0280) (0.0362) (0.0968) (0.116) (0.0326) (0.0355) (0.142) (0.134)

Missing % home price -0.675

∗∗∗

0.109

∗∗

-0.190

∗∗

0.150 -0.395

∗∗∗

0.0437 6.062

∗∗∗

0.391

appreciation

CTY

(0.0241) (0.0518) (0.0835) (0.172) (0.0315) (0.0747) (0.123) (0.270)

Missing Processing time 0.179

∗∗∗

0.0334 -0.594

∗∗∗

-0.314

∗∗∗

0.733

∗∗∗

0.135

∗∗∗

-0.398

∗∗∗

-0.387

∗∗∗

coeﬃcients

TRACT

(0.0426) (0.0475) (0.0962) (0.106) (0.0455) (0.0475) (0.140) (0.136)

Missing Log(2010 home price)

CTY

-0.704

∗∗∗

-0.728

∗∗∗

0.0130 -0.906

∗∗∗

-0.415

∗∗∗

0.384

∗∗∗

5.403

∗∗∗

2.375

∗∗∗

(0.0237) (0.0536) (0.0790) (0.171) (0.0306) (0.0762) (0.118) (0.273)

Other loan controls

Log(loan size) 0.0252

∗∗∗

0.0909

∗∗∗

-0.138

∗∗∗

-0.494

∗∗∗

1.295

∗∗∗

2.435

∗∗∗

-5.397

∗∗∗

-1.665

∗∗∗

(0.00930) (0.00931) (0.0264) (0.0240) (0.0108) (0.00762) (0.0487) (0.0626)

Jumbo Loans -1.951

∗∗∗

-2.599

∗∗∗

0.605

∗∗∗

-0.524

∗∗∗

-4.578

∗∗∗

-6.870

∗∗∗

-6.122

∗∗∗

-0.401

∗∗∗

(0.0226) (0.0314) (0.116) (0.100) (0.0272) (0.0358) (0.132) (0.129)

Loan Type: FHA 1.078

∗∗∗

1.223

∗∗∗

-1.124

∗∗∗

-0.379

∗∗∗

5.864

∗∗∗

9.225

∗∗∗

-2.041

∗∗∗

2.884

∗∗∗

(0.0165) (0.0149) (0.0358) (0.0286) (0.0341) (0.0362) (0.0585) (0.0593)

Loan Type: VA 0.0610

∗∗∗

0.497

∗∗∗

-1.282

∗∗∗

-0.889

∗∗∗

2.990

∗∗∗

7.633

∗∗∗

-3.873

∗∗∗

3.893

∗∗∗

(0.0229) (0.0222) (0.0458) (0.0444) (0.0494) (0.0486) (0.0846) (0.0806)

No Coapplicant 0.533

∗∗∗

0.469

∗∗∗

0.498

∗∗∗

0.860

∗∗∗

0.451

∗∗∗

0.138

∗∗∗

-0.755

∗∗∗

-1.694

∗∗∗

(0.00977) (0.00945) (0.0230) (0.0215) (0.0121) (0.0121) (0.0368) (0.0343)

Owner Occupied 0.543

∗∗∗

0.0652

∗∗∗

-1.652

∗∗∗

-0.589

∗∗∗

1.881

∗∗∗

0.908

∗∗∗

3.194

∗∗∗

3.855

∗∗∗

(0.0176) (0.0193) (0.0619) (0.0593) (0.0203) (0.0189) (0.0694) (0.0658)

Observations 20790255 20790255 8901875 8901875 32936746 32936746 9888845 9888845

Mean Dependent Var 2.888 2.888 6.745 6.745 6.129 6.129 20.41 20.41

Linear probability model based on HMDA data from 2010-16. All continuous right-hand size variables normalized to have mean of zero and standard deviation of one.

TRACT

and

CTY

indicate variable is measured at the census tract or county level of aggregation, respectively, rather than at the loan level. Robust standard errors in

parentheses, clustered by census tract. Regressions include controls for loan size, loan type, dummies for jumbo loan, coapplicant, owner occupied, other race categories,

and missing values for any variable with positive incidence of missing values. See Internet Appendix for full results including coeﬃcients on these variables as well as

univariate regressions. See Data Appendix for variable deﬁnitions and sources.

∗

p < 0.10,

∗∗

p < 0.05,

∗∗∗

p < 0.01

G Diﬀusion of Google Fiber in Kansas City

Table A.5: Summary Statistics for Kansas City Regression Variables

mean sd min p50 max

% with Google Fiber

CTY

0.07 0.24 0.00 0.00 1.00

Log(income) 4.41 0.61 0.00 4.41 9.21

Log(loan size) 4.96 0.72 0.00 5.02 10.77

Female 0.26 0.44 0.00 0.00 1.00

Unknown 0.07 0.26 0.00 0.00 1.00

Black 0.04 0.19 0.00 0.00 1.00

Hispanic 0.03 0.17 0.00 0.00 1.00

Unknown 0.10 0.30 0.00 0.00 1.00

American Indian/Alaska Native 0.01 0.07 0.00 0.00 1.00

Asian 0.02 0.15 0.00 0.00 1.00

Hawaiian/Paciﬁc Islander 0.00 0.04 0.00 0.00 1.00

Jumbo Loans 0.02 0.15 0.00 0.00 1.00

Loan Type: FHA 0.19 0.39 0.00 0.00 1.00

Loan Type: VA 0.07 0.26 0.00 0.00 1.00

No Coapplicant 0.47 0.50 0.00 0.00 1.00

Owner Occupied 0.92 0.28 0.00 1.00 1.00

Table A.6: Fintech Mortgage Share & Google Fiber Access

Dependent variable: = 100 if ﬁntech lender, = 0 otherwise

Purchases Reﬁnances

All Nonbank All Nonbank

% with Google Fiber

TRACT

-0.800

∗∗∗

-0.738

∗∗∗

-1.260 -0.903 -0.487 -0.417 0.0186 0.595

(0.244) (0.243) (0.935) (0.926) (0.357) (0.341) (1.023) (0.996)

Year-Month FEs Y Y Y Y Y Y Y Y

Census Tract FEs Y Y Y Y Y Y Y Y

Borrower & Loan Controls N Y N Y N Y N Y

Observations 138306 138306 34796 34796 180777 180777 51890 51890

Mean Dependent Var 2.147 2.147 8.535 8.535 5.189 5.189 18.08 18.08

Linear probability model of borrowing from a FinTech lender on Google Fiber access, based on HMDA data from 2011-16.

Robust standard errors in parentheses, clustered by census tract. Borrower and loan characteristics include applicant in-

come, indicator for missing applicant income, loan size, borrower gender, race, & ethnicity indicators, loan type, coapplicant

indicator, and owner occupied indicator.

∗

p < 0.10,

∗∗

p < 0.05,

∗∗∗

p < 0.01

Figure A.6: Staggered Entry of Google Fiber

Google Fiber Availability in December 2011

No Google Fiber

Google Fiber Availability in December 2015

No Google Fiber

<75%

75% - 95%

≥95%

Figure shows the share of the population for each census tract that lives in a census block with Google

Fiber in Kansas City. Source: NTIA and FCC data on Internet coverage by census block, provider, and

technology in December 2011 and 2015.