Integrated Intermediation and Fintech Market Power

Stanford University

University of Geneva, Swiss Finance Institute, and CEPR

Boston College

Integrated Intermediation and

Fintech Market Power

∗

Greg Buchak N Vera Chau Adam Jørring

August 25, 2023

Abstract

We document that in the US residential mortgage market, the share of integrated

intermediaries acting as both originator and servicer has declined dramatically.

Exploiting a regulatory change, we show that borrowers with integrated servicers are

more likely to reﬁnance, and conditional on reﬁnance, are more likely to be recaptured

by their own servicer. Recaptured borrowers pay lower fees relative to other reﬁnancers.

This trend is partially oﬀset by a rise in integrated ﬁntech originator-servicers, who

recapture at higher frequency but at worse terms. We build and calibrate a dynamic

structural model to interpret these facts and quantify their impact on equilibrium

outcomes. Our model suggests that integreated intermediaries enjoy a marginal cost

advantage when reﬁnancing recaptured borrowers, and fully disintegrating them would

reduce reﬁnancing frequencies and increase fees. Fintechs use technology to reacquire

customers and reduce borrower inertia against reﬁnancing. This endogenously creates

market power, which ﬁntechs exploit through higher fees. Despite worse terms ex-post,

ﬁntechs increase consumer welfare ex-ante by increasing reﬁnancing frequencies. Taken

together, our results highlight the importance of intermediaries’ scope in consumer

ﬁnancial outcomes and highlight a novel, quantitatively important application of

ﬁntech: customer acquisition.

Keywords: Mortgage market structure, mortgage reﬁnance, ﬁntech, intermediation.

JEL Classiﬁcation Codes: G2, L5

∗

We thank Peter DeMarzo, Darrell Duﬃe, Steve Grenadier, Josh Rauh, Amit Seru, Jeﬀ Zwiebel, and other seminar partic-

ipants at Stanford for comments and suggestions. We also thank Yoonjoo Hwang for oustanding research assistance. Buchak is

at Stanford Graduate School of Business. Email: [email protected]. Chau is at Swiss Finance Institute & Geneva Finance

Research Institute. Email: v[email protected]h. Jørring is at Boston College. Email: adam.jorring@bc.edu

Electronic copy available at: https://ssrn.com/abstract=4552425

A loan origination creates two conceptually distinct assets: the right to the cashﬂows

paid by the borrower—the loan—and the right and obligation to collect payments from the

borrower and forward them to the loan’s owner in exchange for a fee—the servicing right. In a

traditional balance sheet model of banking, origination, servicing, and receiving cashﬂows are

integrated within one ﬁnancial intermediary. However, the modern industrial organization

of ﬁnancial intermediation has seen a striking trend of separating loan origination from loan

ownership, e.g., in the context of residential mortgage origination (Buchak et al., 2023) or

small business lending (Gopal and Schnabl, 2020). This paper documents and explores the

consequences of an additional margin of dis-integration: the increased separation of loan

originator and loan servicer.

We study this phenomenon in he context of the US residential mortgage market. As in

other contexts, the mortgage servicer occupies an important role in the mortgage interme-

diation chain. In particular, the servicer maintains a long-term, ongoing relationship with

the borrower: The servicer collects monthly payments and forwards them to the mortgage

owner. The servicer reminds the borrower when there are late payments. The servicer is

potentially responsible for ex-post modiﬁcations in the event of delinquency or default. In

short, unlike the originator or mortgage owner, the servicer plays an active role in the ﬁ-

nancial life of the mortgage borrower far after the point of origination. For this reason, it is

natural to expect that characteristics of the servicer will have an impact on the borrower’s

ex-post ﬁnancial decisions, such as reﬁnancing or default. This paper focuses on the the

reﬁnancing impliactions of mortgage servicers.

Our paper uncovers two key economic forces that distinguish combined originator-

servicers—which we term integrated intermediaries—and integrated ﬁntechs from other, dis-

integrated ﬁnancial intermediaries. First, we show that integrated intermediaries possess a

cost advantage in reﬁnancing their existing servicing customers. In particular, by virtue of

their long-term relationship with the borrower, the servicer already possesses documents,

credit information, and access that are key in the origination of a new loan. Second, we ﬁnd

that integrated “ﬁntech”

intermediaries appear to be able to use data and customer access

for the purposes of more eﬀective customer acquisition in a manner that non-ﬁntech lenders

lack the technical expertise to implement. In particular, ﬁntech lenders are able to reﬁnance

their “woodhead” servicing customers who would not otherwise be looking to reﬁnance, and

can exploit their resulting market power to reﬁnance the borrower at a higher markup.

We ﬁrst establish these facts in reduced form in three steps. After documenting a secular

We deﬁne a “Fintech” intermediary as in Buchak et al. (2018a), where a lender is a Fintech if the loan

application occurs entirely online and the potential borrower is able to obtain a ﬁrm, contractual rate quote

without interacting with a human loan oﬃcer.

Electronic copy available at: https://ssrn.com/abstract=4552425

decline in the market share of integrated intermediaries that is partially oﬀset by a rise in

integrated ﬁntech intermediaries, we provide identiﬁed evidence that integrated intermedi-

aries lead to more reﬁnancing using a regulatory shock to the capital treatment of mortgage

servicing rights on bank balance sheets. Next, using a novel forward-merged data set that

allows us to match mortgage originations to future reﬁnances of those same mortgages, we

show that reﬁnances originated by the existing servicer are executed with lower fees. This

suggests that these lenders possess a marginal cost advantage when reﬁnancing their own

customers. We ﬁnd, however, that ﬁntech intermediaries charge signiﬁcantly higher fees and

reﬁnance their servicing customers sooner and at lower rate beneﬁt to the borrower, suggest-

ing that they are targeting borrowers who would not otherwise be attempting to reﬁnance

and charging markups when doing so. Finally, we rule out that our results are driven by

alternate mechanisms, such as one intermediary type possessing superior underwriting skills.

After establishing these facts in reduced form, we build and calibrate a structural model

that allows us to quantify the size of these forces and consider equilibrium counterfactuals

where originators and servicers are fully dis-integrated or where ﬁntech never entered. We

ﬁnd that fully dis-integrating the servicing market would decrease the annual rate of mort-

gage reﬁnance by roughly 10% and increase prices by roughly 5%. This is primarily driven

by the loss of the marginal cost advantage in loan production. The implication from this

counterfactual is that the rapid dis-integration of the servicing market was likely responsi-

ble for a reduction in the number of borrowers who reﬁnanced their mortgages while rates

declined between 2010 and 2022. Additionally, we ﬁnd that the customer acquisition aspect

of ﬁnancial technology has had a large quantitative impact on mortgage reﬁnance propen-

sities and costs. Fintech customer acquisition ability increased the annual rate of mortgage

reﬁnancing by roughly 25% relative to the baseline. It did so by overcoming woodheaded

borrower inertia and splitting the resulting surplus by increasing fees by roughly 5%.

Our paper begins by documenting two new empirical trends. First, we document that

the share of agency mortgages where the originator is also the servicer—a structure of inter-

mediation we term integrated—has fallen sharply since 2010. Figure 1 Panel A shows that

integrated intermediation declined from a high of nearly 70% of originations in 2010 to a low

of roughly 30% in 2015 before recovering back to 50% in 2020. Second, we document that a

signiﬁcant portion of the subsequent increase in integrated intermediation was driven by the

entry of ﬁntech intermediaries, who nearly all operate as integrated originator-servicers. The

integrated ﬁntech share of originations rose from zero in 2010 to roughly 12% in 2020. As

we show, integreated non-ﬁntech and integrated ﬁntech intermediaries exhibit qualitatively

diﬀerent behavior.

To study this behavior, we ﬁrst establish two key ﬁndings around integrated intermedi-

Electronic copy available at: https://ssrn.com/abstract=4552425

aries in general. We ﬁrst exploit a change in the regulatory treatment of mortgage servicing

assets in 2013: Previously, risk-based capital requirements attached a risk weight of 100% to

these assets. In 2013, this risk weight was increased to 250%.

This change, as documented

in Mayock and Shi (2022), induced many banks (but not non-banks) to sell their mortgage

servicing assets. Thus, when considering mortgages originated before 2013, those originated

by banks (the treatment group) are less likely to be serviced by an intermediary who is also

a mortgage originator than are those originated by non-banks (the control group) following

the policy. A diﬀerence-in-diﬀerence analysis shows that after the risk weight changes, the

treated mortgages are signiﬁcantly less likely to reﬁnance than the control mortgages. No

such eﬀect exists among ex-ante bank-serviced mortgages where the servicer at origination

was not the originator, ruling out that this eﬀect is merely driven by the shifting of servicing

assets from banks to non-banks. This analysis thus provides direct evidence that mortgages

serviced by an originator are more likely to reﬁnance than other mortgages.

After showing that whether the servicer is integrated or not impacts the quantity of re-

ﬁnancing, we next condition on reﬁnancing and ask whether mortgages reﬁnanced by the

servicer, recaptured reﬁnances, diﬀer from those reﬁnanced by a third party. This analysis

requires linking the ﬁrst origination, which provides the identity of the servicer, to the sub-

sequent reﬁnancing mortgage, which provides the identity of the subsequent originator, as

well as the new interest rate, timing, and fees. To do this, we undertake a merge, which to

our knowledge is novel in the literature, of merged GSE/HMDA originations with forward-

matched HMDA originations. With this data, we ﬁrst conﬁrm the own-servicer reﬁnancing

channel described above by directly showing that these reﬁnancing borrowers tend to reﬁ-

nance with their own servicer in particular. In terms of pricing, we next show that recaptured

reﬁnances have roughly a $150 lower fee, or 7 basis points as a fraction of the principal. This

suggests that reﬁnancing an existing customer has a lower marginal cost as compared to an

outsider and that some of these marginal cost savings are passed on to the borrower.

We next show that integrated ﬁntech intermediaries in particular appear to possess an

advantage in customer acquisition gives them access to otherwise uncontested reﬁnancing

customers. With our forward-matched reﬁnancing dataset, we ﬁrst document that ﬁntech

intermediaries focus on recapturing their own servicing customers. Next, we use merged

HMDA application data to study borrower search and shopping behavior around reﬁnanc-

ing. We ﬁnd that ﬁntech customers exert signiﬁcantly less eﬀort in searching and shopping

as compared to other lenders’ customers: Fintech customers have 10% fewer reﬁnance appli-

cations to other competing lenders than non-ﬁntech customers, and customers who receive

Given a capital requirement of 6%, this change means that having $1 worth of mortgage servicing rights

went from requiring $0.06 inequity ﬁnancing to $0.15 in equity ﬁnancing.

Electronic copy available at: https://ssrn.com/abstract=4552425

ﬁntech loan oﬀers are nearly 50% less likely to reject the oﬀer. In consequence, ﬁntech lenders

face less eﬀective competition and oﬀer their borrowers signiﬁcantly worse loan terms. Fin-

tech recaptures occur roughly one year earlier in the life of the loan, oﬀer a 20 basis point

smaller rate improvement to the borrower, and have a $300 higher fee, or roughly 17 basis

points as a fraction of the principal.

Motivated by these facts, we build and calibrate a quantitative structural model of mort-

gage reﬁnancing and servicing. We do this for three reasons: First, the mortgage reﬁnancing

decision is a complicated dynamic problem and in particular is model dependent. Formal-

izing the reﬁnancing decision into a structural model allows us to properly account for the

dynamic relationships of all agents involved and in particular to discuss notions of welfare

and optimal reﬁnancing, which is not possible with the reduced form approach. Second,

we use the model to back out the key economic forces we identiﬁed above—marginal cost

advantage and customer acquisition advantage—and to discuss the quantitative relevance of

both of them. Third, we use the model to examine the equilibrium price and quantity im-

pact of several policy counterfactuals, such as the impact of further servicer dis-integration

or ﬁntech entry.

Our model features a long-lived borrower with a 30-year ﬁxed rate mortgage. As interest

rates evolve stochastically, the borrower may want to reﬁnance her mortgage into a lower

rate. Consistent with a wealth of empirical literature, the borrower faces signiﬁcant inertia

when reﬁnancing her mortgage, which we operationalize as a ﬁxed cost she must pay in order

to obtain mortgage oﬀers. If she decides to reﬁnance her mortgage, she receives oﬀers from

imperfectly competitive lenders. To match the empirical facts, we model the lender price

as an origination fee that is a markup over their marginal cost of originating the mortgage.

If she accepts the oﬀer from an integrated intermediary, a servicing relationship is formed

between her and the intermediary. This relationship intermediary then possesses a marginal

cost advantage in subsequent reﬁnances to the same borrower.

In the empirical results, we showed that loan characteristics for recaptured reﬁnancing

by ﬁntech lenders is diﬀerent from other integrated lenders. We therefore introduce into

the model an integrated ﬁntech lender who, in addition to possessing the same marginal

cost advantage in originating loans to its servicing customers, possesses a customer acquisi-

tion technology that applies to its servicing customers. We operationalize this by allowing

borrowers with ﬁntech servicers to receive a reﬁnancing oﬀer before soliciting oﬀers from

other lenders at a signiﬁcantly lower inertia cost. We interpret this as thet ﬁntech lender

exploiting servicer customer data to make targeted reﬁnance oﬀers that allow the borrower

to partially avoid the mental cost of seeking oﬀers. While seemingly simple, legacy lenders

Electronic copy available at: https://ssrn.com/abstract=4552425

have long struggled to implement similar systems.

This targeted oﬀer gives the ﬁntech

lender additional market power over other oﬀers, allowing the lender to endogenously set

higher markups.

We calibrate the model’s parameters chieﬂy through the simulated method of moments

(SMM). The key parameters underlying our mechanisms are (1) the marginal costs of orig-

inating loans for both insider and outsider originators, (2) the mental cost associated with

obtaining reﬁnance oﬀers, and (3) the reduction in mental cost from reﬁnancing through

the ﬁntech’s targeted oﬀer. We calibrate these parameters in an over-identiﬁed SMM by

matching model-predicted moments to those detailed in the reduced form section of the

paper. In particular, we utilize the overall level of fees, the changes in fees for ﬁntech and

non-ﬁntech integrated lenders, the overall reﬁnance rate, and intermediary market shares.

Despite targeting seven moments with four parameters, our model comes reasonably close

in matching these key moments in the data.

With the calibrated model, we ﬁrst explore the economic magnitudes of our two channels.

We ﬁnd that the marginal cost advantage of an originator reﬁnancing its own servicing

customers is roughly $250, or roughly 14% of the overall origination cost. Given the observed

fee diﬀerences, this number suggests that integrated originators realize this marginal cost

advantage and pass only some on to borrowers. This cost is not fully passed through because

only the relationship intermediary possesses the lending advantage. Further, we ﬁnd that

the ﬁntech customer acquisition technology reduces the mental cost of reﬁnancing by roughly

$1,588. The value of this reduction is high, but the ﬁntech captures only some of it through

higher fees because it must still compete with both the outside option of other reﬁnance

oﬀers and also its own targeted oﬀers in the following period.

We next consider two counterfactuals. We ﬁrst examine how reﬁnancing quantities and

prices would change following full dis-integration of the market. We ﬁnd that by turning

oﬀ the marginal cost advantage, the quantity of reﬁnances drops by roughly 10% relative

to the baseline and fees increase by roughly 5% relative to the baseline. These quantities

show that even though the marginal cost saving of $250 is relatively modest, it has large

quantitative implications for reﬁnancing. We next examine a counterfactual world with no

ﬁntech targeted oﬀers. Without ﬁntech customer acquisition technology, reﬁnance quantities

Legacy lenders’ desires and diﬃculties in acquiring customers is perhaps best exempliﬁed by a fraud in

which JPMorgan Chase paid $175m to acquire student borrower data that was later found to be wholly fab-

ricated. https://www.nytimes.com/2023/01/21/business/jpmorgan-chase-charlie-javice-fraud.

html. More generally, many commentators have observed the diﬃculties banks face in even sim-

ple technological innovation. Documented reasons include, e.g., a heightened concern over the risk of

upgrading legacy systems. https://www2.deloitte.com/us/en/pages/financial-services/articles/

modernizing-legacy-systems-in-banking.html. Third-party apps providing such services to banks and

non-ﬁntech lenders, e.g., Blend (https://blend.com/company/about/) came online only in the 2020s.

Electronic copy available at: https://ssrn.com/abstract=4552425

drop by roughly 25% and prices fall by roughly 5%. Quantities fall because there are many

borrowers on the extensive margin of reﬁnancing who would not reﬁnance without the lower

mental-cost ﬁntech option. Prices fall because without these targeted oﬀers, there are fewer

high-markup ﬁntech reﬁnances in equilibrium. This ﬁnal result highlights that while the

ﬁnancial technology is used to create and exploit market power, on net this has positive

welfare beneﬁts relative to a world where there are fewer people reﬁnancing.

Our paper has several important implications. First, we show that the industrial or-

ganization of the servicing market has large quantitative consequences for policy-relevant

ex-post borrower outcomes like mortgage reﬁnancing. Because it is the largest consumer

ﬁnance market in the United States, the reﬁnancing of ﬁxed rate residential mortgages has

large direct welfare implications for households and for the passthrough of monetary policy,

as well as having pricing implications for the RMBS secondary markets. Di Maggio et al.

(2017) showed the eﬀect of mortgage interest rates on other household consumption deci-

sions using adjustable rate mortgages. Our paper highlights that institutional or regulatory

changes that impact servicing are likely to have spillover eﬀects into the macroeconomy. Sec-

ond, our paper highlights important aspects around the contract design of mortgages and

mortgage servicing rights. For example, contract provisions that restrict which parties are

allowed to reﬁnance mortgages would likely have large quantitative impacts on reﬁnancing

propensities.

Finally, our paper emphasizes a key role for ﬁnancial technology beyond the most com-

monly cited application: underwriting. For institutional reasons, our setting largely takes

the underwriting implications of ﬁntech oﬀ the table, and yet, ﬁntech lenders have a large and

growing market share. Instead, our paper highlights a customer acquisition application of

ﬁntech, where data allows the ﬁntech to source otherwise uncontested customers. By sourc-

ing these underserved customers, ﬁntechs in our setting are able to both bring in customers

on the extensive margin and charge them higher markups for the services they provide. An

important takeaway from our paper is that in the largest consumer ﬁnance setting, the chief

application of ﬁntech is a marketing application, not an underwriting application.

Related Literature

Recent changes in the composition of institutions in the ﬁnancial sector have prompted

increased interest in the eﬀect of intermediaries. In the mortgage market, Buchak et al.

(2018a) and Fuster et al. (2022) study the role of shadow banks and ﬁntech lenders, and

Fuster, Lo and Willen (2017) and Bhutta, Fuster and Hizmo (2020) study the impact of in-

termediation on mortgage interest rates. A related set of papers following the ﬁnancial crisis

Electronic copy available at: https://ssrn.com/abstract=4552425

of 2008 focused on how institutions and market structure impacted household outcomes. For

example, Keys et al. (2010) and Purnanandam (2011) study the originate-to-distribute mar-

ket. We contribute a brand new intermediary type to this discussion. We deﬁne integrated

lenders who have a close, existing relationship with borrowers through their servicing role.

A related strand of literature studies so-called “Fintech” intermediaries that harness

technology in some way to eﬀect lending and other aspects of intermediation. Berg, Fuster

and Puri (2022) provides a useful survey of the recent developments in analyzing ﬁntech

lenders in a number of consumer ﬁnance markets. Fuster et al. (2022) and Buchak et al.

(2018a) are two such examples in the mortgage market. Broadly, the ﬁntech literature

tends to focus on these intermediaries mustering new data or models to better underwrite

or otherwise value underlying assets. See, e.g., Di Maggio and Yao (2019), Buchak et al.

(2020). Babina, Buchak and Gornall (2022) highlight additional applications such as product

improvement. Our contribution to this literature is to introduce and identify a novel ﬁntech

mechanism, namely, their ability to mobilize data in order to acquire customers and market

power.

Our paper also draws on recent advances in using structural models to study industrial

organization in household ﬁnance applications. For example, Buchak et al. (2018b), Egan,

Horta¸csu and Matvos (2008), and Egan, Lewellen and Sunderam (2022). Allen, Clark and

Houde (2019) structurally estimates the impact of search frictions in the Canadian mortgage

market. Matteo (2020) analyzes the impact of bank regulation on the UK residential mort-

gage market. Our paper builds a dynamic, structural model of borrower reﬁnancing which

incorporates a marginal cost advantage and superior matching ability for a speciﬁc set of

lenders corresponding to ﬁntech and integrated institutions.

In the mortgage market, our paper relates to important papers about the structure of

mortgage origination. Many of these papers were a response to the subprime mortgage crisis

and highlight the importance of intermediaries and market structure in household ﬁnance.

For example, see Mian and Suﬁ (2009), Acharya et al. (2011), and Elenev, Landvoigt and

Van Nieuwerburgh (2016). In the mortgage space, our paper also contributes to an ongoing

discussion regarding the perplexingly low rates of reﬁnancing in the market. There are

several potential causes to this rigidity. There can be real or behavioral costs (Andersen

et al., 2020), lack of ﬁnancial sophistication (Jørring, 2020), or matching frictions (Agarwal

et al., 2020) which render borrowers inert. Other papers documenting reﬁnancing rates and

comparing them to models of reﬁnancing include Keys, Pope and Pope (2016) and Agarwal,

Rosen and Yao (2016). Our paper considers the role of integrated and ﬁntech lenders in

alleviating the matching frictions which may lead to low rates of reﬁnancing.

More broadly, our paper is related to a line of literature around search costs, competition,

Electronic copy available at: https://ssrn.com/abstract=4552425

and customer acquisition technology. In addition to the aforementioned Allen, Clark and

Houde (2019), applications of search to mortgage markets include Agarwal et al. (2020)

and Gurun, Matvos and Seru (2016). Egan (2019) studies costly search in the context of

consumer bond investments. We also add to a broad literature on competition in consumer

ﬁnance. In the mortgage space, Scharfstein and Sunderam (2016) and Buchak and Jørring

(2021) study the eﬀect of competition on interest rates and fees, respectively. Yannelis and

Zhang (2021) studies competition in a diﬀerent setting, the market for auto-loans, and Jiang

et al. (2023) study the role of technology in mortgage origination. In our setting, ﬁntech

lenders can alleviate some of the search costs associated with reﬁnancing. If they can target

borrowers who have high search thresholds, they can capture a portion of the market and

accumulate market power.

Finally, our paper has implications for monetary policy transmission. Related papers

which consider how intermediary types can inﬂuence monetary policy transmission include

Drechsler, Savov and Schnabl (2017), Scharfstein and Sunderam (2016) and Xiao (2020).

The intermediary mechanisms studied in this paper has the potential to imapct the eﬃcient

pass-through of monetary policy. Borrowers who don’t reﬁnance when interest rates are low

can dampen the monetary policy eﬀect.

Our paper proceeds as follows. In Section 1 we describe the institutional setting and

data. In Section 3 we document empirical facts motivating our paper. In Section 4 we

describe the model. In Section 5 we show estimation results and counterfactuals from the

model. Section 6 provides a brief discussion of our results and concludes.

1 Data and Institutional Details

We focus on the agency-backed single-family mortgage market. In 2020, agency-backed

mortgages comprised approximately 60% of mortgage originations and 72% of reﬁnancings.

The overwhelming majority of these loans are 30-year fully amortizing ﬁxed rate mortgages

that are prepayable at no cost. The pre-payment feature means that borrowers can beneﬁt

from reﬁnancing their mortgage into a lower rate if prevailing mortgage rates fall.

“Agency-backed” loans are originated and sold to the Fannie Mae or Freddie Mac (the

GSEs). The GSEs then pool these loans into securities that are tradeable on the secondary

market. In exchange for a GSE-set insurance fee, often referred to as a “g-fee,” the GSEs

guarantee the payments to the mortgage pool against delinquency and default. Importantly

for our setting, as a result of these guarantees, originators and mortgage holders bear es-

sentially no credit risk from borrower delinquency and default. However, mortgage holders

still face interest rate risk arising from (1) standard duration risk, where the present value

Electronic copy available at: https://ssrn.com/abstract=4552425

of long-dated cashﬂows declines as rates rise, and (2) prepayment risk, where borrowers may

reﬁnance into a lower rate and prepay their existing mortgage, thus cutting the life of the

mortgage short.

In this paper, we study two parts of the mortgage life-cycle: servicing and reﬁnancing. A

mortgage origination creates an auxiliary asset called a mortgage servicing right (“MSR”).

This servicing right gives its owner the right and obligation to act as the mortgage’s servicer.

Servicers interact with borrowers on behalf of the owner of the mortgage, which in our

case are the GSEs and owners of the securitized mortgage pools. Servicers collect monthly

payments, ensure the loan is properly insured, and handle delinquencies. In exchange, the

MSR owner collects a fee, typically a percentage of the monthly payments collected by the

servicer. The serving fee is set, essentially by statute, by the GSEs. While lenders can

and often do have mortgage servicing businesses and service their own loans, MSRs can be

sold at origination to another institution to conduct the actual servicing of a loan. Thus,

while related, the mortgage itself and the MSR are conceptually separate assets owned and

managed by potentially diﬀerent parties. Finally, while servicing rights occasionally trade

in a secondary market post origination, the market is illiquid.

A mortgage reﬁnancing is a prepayment of an existing loan, ﬁnanced by taking out a

new loan with diﬀerent terms. The proceeds from the new loan are used to pay down

the remaining principal of the existing mortgage, and because the mortgages in our setting

have a prepayment option, this terminates all rights and obligations of the initial mortgage.

Mortgages are typically reﬁnanced either to obtain better terms, e.g., a lower rate or longer

maturity, or to increase leverage on the underlying asset in order to extract equity (a so-

called “cash-out” reﬁnance). In this paper we focus only on the former. The contractual

design of agency mortgages dictates that mortgages can be prepaid at no cost, but there

are, of course, costs associated with taking out the new loan. For example, the cost of a

loan oﬃcer or appraisal, or simply a markup that the new lender charges at origination.

Importantly, because the reﬁnancing loan is a new mortgage, the borrower is not limited to

their existing lender and can choose to reﬁnance with any lender.

1.1 Data

We draw from two main sources of data, which are common in the literature: The Fannie

Mae and Freddie Mac origination and performance data (the “GSE data”), and the Home

Mortgage Disclosure Act (“HMDA”) origination data.

GSE data: Fannie Mae and Freddie Mac provide information on the GSEs’ portfolios of

30-year single-family conforming ﬁxed-rate mortgages. This data provides information about

Electronic copy available at: https://ssrn.com/abstract=4552425

the loan at origination, including the names of the lender and servicer at origination, the

interest rate, the term, and the principal balance. The origination dataset includes all new

loans; including reﬁnancing loans. The monthly performance updates provide information

on the loans delinquency status, as well as the date at which a loan is terminated. This can

occur if a loan is paid oﬀ through a prepayment, end of the term, or reﬁnancing.

In order to study comparable mortgages, we follow the literature and restrict our sample

to 30-year mortgages originated for purchases of primary residence homes. These loans are

fully amortizing and have full documentation. The loan-level origination dataset further

provides interest rates, FICO scores, LTVs, and DTIs, as well as loan size, type, purpose,

and location. It also identiﬁes the originator that sold the loan to the GSE in cases where

the originator had suﬃciently high origination market share in the reporting period.

HMDA: HMDA data covers is a mortgage-application-level dataset covering the near

universe of U.S. mortgage applications and originations. This data, used extensively in the

literature, includes lender identiﬁcation, application outcome, and loan type, purpose, size,

year of origination, and location at the census tract level.

Since 2018, HMDA has further recorded several further variables that are key in our

analysis: loan interest rate, non-interest-rate charges (including origination charges, discount

points, and lender credits), loan-to-value (LTV) ratio, and debt-to-income (DTI) ratio. Con-

sequently, while our main sample includes loans since 2005, we do several test including these

additional variables in the 2018–2021 sample.

Merging procedure: There is no direct crosswalk linking the GSE data to the HMDA

data. In addition to the two underlying data sources, we create a new merged dataset that

links mortgages to their new, reﬁnanced loan. To create this match, we ﬁrst take GSE loans

to their HMDA counterparts. The match is based on year, zip code, loan amount, loan

purpose, occupancy type, and the original lender name. Then, for these matched loans, we

take the sample which is exiting through a reﬁnancing in the month that the loan is recorded

as leaving the dataset. We forward match to executed reﬁnancings in the HMDA dataset.

Importantly, the substantial information regarding borrowers available in the HMDA dataset

facilitates this match. Rather than use loan characteristics, we use borrower characteristics

(year, census tract, race, ethnicity, loan amount) to match reﬁnancings with their new loans.

To summarize, the empirical analysis for this project consists of several datasets: a random

sample from GSE-backed origination and monthly performance data, a random sample of

HMDA applications, and forward-linked reﬁnanced loans.

Electronic copy available at: https://ssrn.com/abstract=4552425

2 Motivation

We document two observable trends motivating our focus on integration and ﬁntech

lenders. First, the share of integrated originations has declined signiﬁcantly since 2010.

Second, this has been partially oﬀset by a rise in integrated ﬁntech originations beginning

in 2012.

2.1 The declining share of integrated originations

We deﬁne an integrated origination (or integrated loans) to be a mortgage origination

where the originator and servicer at origination are the same entity. We will often inter-

changeably refer to originators or servicers as integrated. In principal, an originator could

retain the servicing rights for only a proportion of the loans they originate each month, but

in practice, lenders tend to retain the vast majority of their servicing rights or sell most of

them.

Additionally, while integration is deﬁned at origination, MSR owners can sell the

servicing rights at a later date. While this is rare except in some special cases (e.g., after the

implementation of the MSR capital rule that we discuss below), the upshot of this potential

mismeasurement is that our results will tend to overstate the fraction of loans currently

integrated and understate the eﬀects of integration.

Figure 1 plots the monthly share of loans that are integrated at origination. The share of

integrated originations is roughly 50% at the beginning of our sample in 2002, rises gradually

through the ﬁnancial crisis before hitting a high of roughly 70% in 2009, before declining

steadily to a low of 30% in 2015. Starting from 2015, the share of integrated originations

again begins to rise, hitting roughly 50% by the end of our sample in 2020. These aggregate

trends are only partially coincident with the rise of non-bank (“shadow bank”) originators

in the agency mortgage sector discussed in Buchak et al. (2018a). While traditional bank

originators are somewhat more likely than shadow bank originators to retain servicing rights,

there is signiﬁcant time-series variation within both types, and as we show in Appendix Table

A2, the aggregate shadow bank origination share explains less than 30% of the variation in

integrated share.

Figure A1 in the appendix shows the histogram of retained ratios for each lender-month in a sample of

FNMA and FHLMC originated loans data. As the ﬁgure shows, these shares cluster at 0% and 100%, with

little mass in between.

See Appendix Table A2 Panel A for the aggregate time series. In Panel B, we regress aggregate

origination share on shadow bank origination share, bank integration share, and shadow bank integration

share. We ﬁnd that shadow bank origination share explains roughly 30% of the variation in integration

share, while within-lender-type variation explains roughly 70% of the variation. Hence, while the broad shift

to shadow bank originations plays a large role in the aggregate trends highlighted here, it explains far less

than half of the variation.

Electronic copy available at: https://ssrn.com/abstract=4552425

2.2 The rise of integrated ﬁntech originations

We deﬁne ﬁntech lenders as in Buchak et al. (2018a), which broadly captures the idea

of utilizing IT in the mortgage origination process. Under this classiﬁcation, a lender is

a ﬁntech lender if the consumer can receive a ﬁrm rate quote without human interaction.

Figure 1 Panel B plots the share of loans originated by integrated ﬁntech lenders. Fintech

lenders largely entered the market around 2013, at which point their market share beings to

grow dramatically. At their height in 2019, integrated ﬁntech lenders originated almost 12%

of all agency loans.

Critically for our paper, nearly all ﬁntech lenders are integrated. In our sample of agency

loans, 88% of ﬁntech originations are integrated, as compared to 50% or less for other non-

bank intermediaries. Thus, as ﬁntech lenders gain market share in originations, the aggregate

share of integrated originations also rises. As shown in Figure 1 Panel A, the rise of ﬁntech

thus partially oﬀsets the steep secular decline in integrated origination highlighted above.

3 Facts about integrated intermediaries

Having established the decline in integrated lenders and the rise of integrated ﬁntech

lenders, we turn to studying mortgage characteristics related to these intermediary types.

We document a set of key facts around (1) integrated intermediaries in general, and (2)

ﬁntech integrated intermediaries speciﬁcally, which motivate our model in section 4. First,

having an integrated servicer causes borrowers to reﬁnance at signiﬁcantly higher frequen-

cies, with these borrowers often reﬁnancing directly with their servicer. These recaptured

borrowers pay signiﬁcantly lower fees as compared to non-recaptured borrowers. Second, ﬁn-

tech intermediaries focus on recapturing borrowers, their customers are less likely to search

for other oﬀers and are more likely to accept ﬁntech oﬀers. On origination, these recaptured

ﬁntech borrowers realize a less advantageous rate improvement from reﬁnancing and pay

signiﬁcantly higher fees.

We interpret the evidence at the conclusion of this section. To brieﬂy anticipate our

proposed mechanism, our evidence on integrated intermediaries in general suggests that in-

termediaries possess a marginal cost advantage in reﬁnancing their own servicing customers

that they partially pass through as lower fees. With respect to ﬁntech integrated inter-

mediaries, our evidence is consistent with ﬁntechs utilizing their servicing pool to acquire

customers who consider few alternatives, thus creating and exploiting market power and

charging high markups. While we do not take a stand on the exact source for these advan-

tages, an illustrative example of the former is lenders already possessing identity and credit

Electronic copy available at: https://ssrn.com/abstract=4552425

information about their servicing customers. An illustrative example of the latter is ﬁntech

lenders integrating borrower-facing servicing and origination interfaces to reduce consumer

cognitive burdens around reﬁnancing. In the model of section 4 we explore this mechanism

formally.

3.1 Integrated lenders and borrower recapture

Our ﬁrst fact in this subsection provides causal evidence that integration of a borrower’s

servicer and originator increases the borrower’s reﬁnancing propensity. We then show that

this is likely driven by lenders directly reﬁnancing their own servicing customers, and that

these recaptured borrowers pay lower fees to reﬁnance.

3.1.1 Servicing customers of integrated borrowers reﬁnance more frequently

We present causal evidence linking the servicer’s status as an originator to borrower reﬁ-

nancing propensity, and show that borrowers with loans from integreated intermediaries are

more likely to reﬁnance. We posit a connection between servicer, originator, and borrower

reﬁnance for several reasons. First, an originator may possess a cost advantage in underwrit-

ing loans to its own servicing customers because it already possesses borrower information,

paperwork, and, importantly, contact information. Servicers are in the best position to know

when their servicing customers would most beneﬁt from a reﬁnance, and can supply their

information to their aﬃliated loan-origination division, thus obviating customer acquisition

frictions. Third, incentives for reﬁnancing are most aligned when the servicer does the re-

ﬁnancing. When executed with an outside originator, the reﬁnancing means the servicer

would lose future servicing income. For an aﬃliated originator, reﬁnancing the loan elimi-

nates the old servicing right but creates a new servicing right that immediately ﬂows back

to the servicer. Thus integrated servicers have an incentive to encourage their customers to

reﬁnance, so long as they reﬁnance with the aﬃliated originator. Finally, customers have a

slight preference for their existing servicers with whom they already have a relationship. In

a recent U.S. Mortgage Servicer Satisfaction Study survey,

customers were more likely to

reconsider an originator in the future if that originator had retained the servicing rights.

To identify this channel, we exploit a regulatory change around the capital treatment of

mortgage servicing rights (MSRs) that impacted banks but not non-banks. In particular,

the Basel III framework, approved in 2013, increased regulatory capital weights on MSRs

from 100 to 250 percent. That is, given a 6% capital requirement, a bank pre-reform was

required to ﬁnance an MSR asset with 6% equity. Following the reform, a bank was required

https://www.housingwire.com/articles/how-retaining-servicing-provides-a-competitive-advantage/

Electronic copy available at: https://ssrn.com/abstract=4552425

to ﬁnance the MSR with 15% equity. In consequence, the reform made holding MSR assets

on balance sheet signiﬁcantly more expensive for depository institutions, and as shown in

Mayock and Shi (2022), depository institutions sold oﬀ these assets in large numbers in early

2013.

This regulatory shock provides a natural experiment we use to evaluate the importance of

the servicer-originator connection in mortgage reﬁnancing. We consider all agency mortgages

originated prior to 2013 where the originator retained the servicing rights. Our treatment

group consists of all mortgages initially serviced (and thus originated) by a traditional bank,

and our control group consists of all mortgages initially serviced (and thus originated) by

non-banks. At origination, having retained the servicing rights, both groups are on equal

footing with respect to reﬁnancing their servicing customers. However, following the approval

of the more onerous MSR capital rule in 2013, the bank mortgage servicers that comprise the

treatment group receive a negative shock in continuing to hold the servicing rights, and are

therefore induced to sell these servicing rights to a third party.

While the original servicer

was by deﬁnition also an originator, the third party buyers may or may not be. Thus,

mortgages in the treatment group are less likely to be serviced by a mortgage originator

after 2013. In contrast, the non-bank servicer-originator mortgages in our control group did

not face such a regulatory shock and therefore had no incremental incentive to dispose of

their MSRs.

Our hypothesis is that reﬁnancing propensity for the treatment group should

decline after the capital rule change disintermediates them.

This motivates the following loan-month-level diﬀerence-in-diﬀerence and dynamic

diﬀerence-in-diﬀerence:

Refi

= βP ost

× Bank

+ γ

+ 

(1)

Refi

2015 Q4

τ =2011 Q1

t=τ

× Bank

+ γ

+ 

(2)

As discussed above, the sample consists of agency mortgages

originated between January

2000 and December 2012 where the originator was also the servicer. Here Refi

is a zero-one

indicator for whether loan i prepays at time t. β and β

are the diﬀerence-in-diﬀerence and

dynamic diﬀerence-in-diﬀerence coeﬃcients of interest. P ost

is a zero-one indicator for the

observation occurring in January 2013 or later. Bank

is an indicator for whether mortgage

Although these sales are not observable in our data, Mayock and Shi (2022) shows strong evidnece for

what is eﬀectively our ﬁrst stage.

Note that because we cannot directly measure MSR sales, we are estimating an intent to treat eﬀect.

For computationally eﬃciency we run the experiment on a random 1% sample of loans.

Electronic copy available at: https://ssrn.com/abstract=4552425

i was originated (and hence serviced at origination) by a depository institution.

In the ﬁxed eﬀects γ

and γ

, c refers to interacted binned FICO at origination, binned

loan-to-value (“LTV”) at origination, 3-digit zip code, and binned current loan size. t refers

to time, and b refers to originator type. Thus γ

are characteristic times time ﬁxed eﬀect and

account for time-and-characteristic varying changes in reﬁnance propensities. These ﬁxed

eﬀects absorb, for instance, a possible cofounding eﬀect where banks target diﬀerent customer

types from banks who happened to change their reﬁnancing behavior around 2013. γ

are

characteristic times originator type ﬁxed eﬀects and capture originator-and-characteristic

time-invariant diﬀerences in reﬁnance propensity.

Our mechanism predicts that β < 0. Following the MSR rule change that caused banks

to sell servicing rights, the reﬁ-propensity on bank-originated mortgages decreases because

the servicer is less likely to be the originator. Table 2 Panel A shows that this is indeed

the case. Column (1) is the baseline speciﬁcation, with ﬁnds that treated loans are roughly

1.6 percentage points less likely to reﬁnance on a given month. Column (2) adds a control

for the diﬀerence between the loan’s original interest rate and prevailing rate (measured in

percentage points), and ﬁnds little change in the coeﬃcient. Column (3) restricts the sample

to those loans that appear clearly in the money: loans with a rate diﬀerence of at least 50

basis points. Again, the results are roughly unchanged. These eﬀects are large relative to

the mean reﬁnance propensity of roughly 1.7 percent per month.

Next, we show the results of the dynamic diﬀerence-in-diﬀerence in Figure 2 Panel A.

First, the ﬁgure shows that bank and non-bank reﬁnance rates are on parallel trends pre-

treatment. Upon impact, however, bank-originated mortgages saw a clear downward shift

in their reﬁnance propensity.

An alternative mechanism is that non-bank servicers induce a lower rate of reﬁnancing

in their servicing customers than bank servicers do, and the Basel III rule shift merely shifts

servicing customers from banks to non-banks, thus reducing their reﬁnancing rates. There is

a natural placebo test to rule out such a mechanism. In particular, by rerunning our analysis

on the sample of loans where the originator was not the servicer at origination, but where

Basel III nevertheless induced a shift of the servicing rights from banks to non-banks, we

can construct a sample of loans that is not aﬀected by the severing of the originator-servicer

relationship but is aﬀected by the bank-to-non-bank shift. If the eﬀect comes merely from

the bank or non-bank status of the current servicer, we should observe a similar drop in the

reﬁnancing rate among non-integrated mortages previously serviced by banks. In contrast,

if the eﬀect comes speciﬁcally from severing the originator-servicer relationship, we should

see no eﬀect in this sample. Table 2 Panel B shows that there is no eﬀect among the non-

integrated loans, which rules out that the eﬀect occurs merely through the shift of servicing

Electronic copy available at: https://ssrn.com/abstract=4552425

rights from banks to non-banks.

3.1.2 Integrated originators recapture their customers

Next, we provide direct evidence that borrowers are more likely to reﬁnance with their

servicer if possible, i.e., if the servicer is also an originator. Our empirical approach is to

regress a lender’s market share among a population of reﬁnancing borrowers on whether the

lender is the population’s servicer. A positive coeﬃcient in this regression suggests that

borrowers are more likely to choose their own servicer when reﬁnancing.

To operationalize this approach, we begin with the forward-matched reﬁnance sample.

We aggregate this loan-level data to the lender-market level, where a market is deﬁned as a

year × zip × servicer. We then calculate each lender’s market share in each market. That

is, for a given lender, we measure the share of reﬁnances among borrowers in a given year

and zip code sharing a servicer. For each of these markets, the given lender may also be the

servicer. We then run the following regression at the lender-market level:

ltzs

= βOriginatorIsServicer

+ γ

tzs

+ γ

ltz

+ 

ltzs

(3)

ltzs

is the lender’s market share (between 0 and 1) among reﬁnancing borrowers

at year t, zip z, with servicer s. OriginatorIsServicer

is a zero-one indicator for whether

lender l is the same entity as servicer s. The coeﬃcient of interest is β. All speciﬁcations

include market (year × zip × servicer) ﬁxed eﬀects, which absorb market-level demand

shocks (and in particular absorb the outside option of not reﬁnancing). We vary the lender

ﬁxed eﬀects as described below. Standard errors are clustered at the lender level to allow

for correlated shocks within lender, e.g., an unobserved lender quality shock.

Table 3 shows the results. Column (1) includes no lender ﬁxed eﬀects. This column shows

that originators have a 21% higher market share among their own servicing customers. It is

possible that this is partly driven by unobserved lender quality diﬀerences, for instance, that

originators who are also servicers are of greater quality than originators who are pure-play

originators. Thus, in column (2), we include lender ﬁxed eﬀects, and ﬁnd a qualitatively

similar result of 18.4%. Importantly, including lender ﬁxed eﬀects absorbs diﬀerences be-

tween banks and shadow banks, and the small change to the estimate suggests that this

distinction is not driving our results. Finally, in column (3) we include lender × year × zip

ﬁxed eﬀects, which absorbs unobserved time- and location-varying quality diﬀerences at the

lender level, and ﬁnd essentially no change in the estimate. Taken together, these estimates

suggest a direct connection between servicing and reﬁnancing, where servicing customers are

Electronic copy available at: https://ssrn.com/abstract=4552425

signiﬁcantly more likely to reﬁnance with their own servicer, when possible.

3.1.3 Recaptured non-ﬁntech borrowers pay lower fees

Having shown a high degree of recapture—originators reﬁnancing their servicing

customers—our second fact now explores the characteristics and terms of these recaptured

reﬁnances in comparison to non-recaptured reﬁnances. For this and the following subsec-

tions, we run the following regression in the forward-matched reﬁnance sample at the loan

level:

Characteristic

T ype

+ F E + 

(4)

Here, Characteristic

is a characteristic of reﬁnancing loan i. We examine three variables

pertaining to the timing of the reﬁnance: the years between origination and reﬁnance, the

rate improvement, deﬁned as the diﬀerence between the old and the new rate (where a higher

positive number means a more favorable rate improvement), and the payment improvement,

deﬁned as the annual change in mortgage payments (again, a positive number means more

favorable terms for the borrower). We further examine three variables pertaining to the

cost of the reﬁnance: the dollar origination charges, origination charges as a percentage of

loan principal, and breakeven years, deﬁned as the dollar origination charges divided by the

annual rate increase. This last variable measures how many years it takes for the payment

improvement to equal the cost of reﬁnance; a larger number indicates that it takes longer

for the borrower to break even in payment terms in NPV-unadjusted terms.

J indexes characteristics of the reﬁnancing: Fintech recapture, Fintech new borrower, and

Non-ﬁntech recapture. Fintech recapture denotes recaptured loans where the original lender

was a ﬁntech company. Fintech new borrower is a loan reﬁnanced by a ﬁntech lender where

the original loan was not a ﬁntech loan. Non-ﬁntech recapture denotes loans recaptured by

non-ﬁntech lenders. We reserve discussion of the two ﬁntech categories to section 3.2.3 below

when we discuss ﬁntech reﬁnancing characteristics. The omitted category is non-ﬁntech non-

recaptures, i.e., a reﬁnance done by a non-ﬁntech lender who is not the borrower’s servicer.

All speciﬁcations include lender type ﬁxed eﬀects (bank vs. non-bank), loan controls (for

the reﬁnancied loans), loan controls (for the original loan), and market (year × three-digit

zip) ﬁxed eﬀects. The results are shown in Table 4.

We begin with the non-ﬁntech recapture results in Panel A. This table shows that while

non-ﬁntech recaptured reﬁs occur slightly earlier than non-recaptured reﬁs, there is no statis-

tically or economically signiﬁcant diﬀerent in terms of interest rate improvement or payment

Electronic copy available at: https://ssrn.com/abstract=4552425

improvement. The rate improvement is only 0.9 basis points higher than the baseline and is

not statistically signiﬁcant. In Appendix A3, we decompose the rate improvement into dif-

ferences into the rate on the old loan and the rate on the reﬁnance. We ﬁnd that non-ﬁntech

recaptures reﬁnance at rates slightly above baseline, but also reﬁnance loans whose existing

rates are slightly above baseline, and on net these cancel out. This 0.9 net basis point dif-

ference translates to an insigniﬁcant $35 payment diﬀerence on average over the course of

a year. These magnitudes are small in an absolute sense, and are also small relative to the

means of these variables: The mean rate improvement for a reﬁnance is 1.14% (114 basis

points) and the mean annual payment improvement is nearly $3,000. The diﬀerences for

non-ﬁntech recaptures represent a de minimis change relative to this baseline.

Much larger diﬀerences emerge when examining fees in Panel B. We ﬁnd that non-ﬁntech

recaptures have $140 lower fees in dollar terms, or 6.5 basis points lower as a fraction of

loan principal. These quantities are statistically and economically signiﬁcant, representing

more than a 5% diﬀerence relative to baseline. The diﬀerence is even more stark in the

context of how long a borrower would take to break even after reﬁnancing, as shown in

column 3. We ﬁnd that a borrower reﬁnancing with their non-ﬁntech servicer breaks even

0.29 years, or 3.5 months earlier relative to baseline. Economically, this is a large eﬀect, and

is equal to more than 20% of the mean break-even time of roughly 15 months. To summarize,

while borrowers reﬁnancing with their own servicer appear to have no advantage in terms

of rate improvement, they do enjoy a signiﬁcant cost advantage in doing so, which at least

partially explains previous results showing borrower’s higher propensity to reﬁnance with

their servicer.

3.2 The special role of integrated ﬁntech lenders

We next study ﬁntech intermediaries in particular. Recall that ﬁntech servicers are

overwhelmingly integrated, and a signiﬁcant portion in the increase in integrated lender

market share since 2015 has coincided with the rise of ﬁntech intermediaries. While this

suggests that integrating lending and servicing is likely to play a large role in ﬁntechs’

intermediation behavior, in this section we show that the consequences of ﬁntechs’ integration

is qualitatively diﬀerent from that of other lenders. We ﬁrst show that even relative to other

integrated lenders, ﬁntech lenders recapture at a signiﬁcantly higher rate. We next show

that ﬁntech reﬁnancing applicants are less likely to reject ﬁntech oﬀers or shop elsewhere.

Finally, we show that conditional on recapturing, ﬁntech recaptures occur at signiﬁcantly

worse terms for borrowers.

Electronic copy available at: https://ssrn.com/abstract=4552425

3.2.1 Fintechs recapture at higher frequencies

Using our forward matched reﬁnance sample, we ﬁrst examine whether ﬁntech reﬁnances

are more likely to be recaptures than other lenders’ reﬁnances. At the loan level, we run the

following regression:

Recaptured

= βF intech

+ F E + 

(5)

Recaptured is a zero-one indicator for whether the originator is the same entity as the

previous servicer. Fintech is a zero-one indicator for whether the originator is a ﬁntech lender,

as classifed in Buchak et al. (2018a). We include lender type, borrower new loan controls

(measured for the newly reﬁnanced loan), original borrower controls (measured with the

original loan), and market (year times three-digit zip) ﬁxed eﬀects. The results are shown

in Table 5.

Columns (1) and (2) use the full sample of forward-matched loans; column (2) includes

additional borrower controls. Across both speciﬁcations, we ﬁnd that ﬁntech originations are

signiﬁcantly more likely to be borrower recaptures. While unconditionally, roughly 14% of

the reﬁnances in our sample are recaptures, ﬁntech reﬁnances are roughly 12% more likely to

be recaptures. This estimate is both statistically and economically signiﬁcant and represents

nearly a 100% greater likelihood of recapturing as compared to the base rate. Columns (3)

and (4) limit the sample to only reﬁnances by integrated lenders. We impose this restriction

because mechanically, only integrated lenders can recapture servicing customers, and in

principle the ﬁrst two columns could simply have been picking up the fact that ﬁntech

lenders tend to be integrated. In this restricted sample, the unconditional recapture rate is

30%, with a ﬁntech reﬁnances having an incremental 20% likelihood of being a recapture.

These results point to the recapture of servicing customers to be an important source

of ﬁntech customers. While recaptures constitute only 14% of a typical lender’s reﬁnance

portfolio, they constitute nearly 30% of a ﬁntech lender’s reﬁnance portfolio. We now turn

to have the terms on these ﬁntech recaptures diﬀer from other reﬁnances.

3.2.2 Fintech lenders face less competition in reﬁnancing

We next show that ﬁntech lenders appear to operate in a less competitive consumer

environment than other lenders. In particular, we show that ﬁntech applicants are less likely

to shop for other reﬁnance oﬀers, and that once an oﬀer is made, ﬁntech applicants are more

likely to accept it.

Electronic copy available at: https://ssrn.com/abstract=4552425

Customer Search Intensity and Fintech Lenders : We examine whether borrow-

ers applying for ﬁntech reﬁnances apply to fewer other lenders. We use this as a proxy for the

search intensity of borrowers. We link loan applications to each other based on immutable

borrower characteristics such as the borrower’s location, race, income, and property type to

form application “chains.” A detailed description of this match process is available in the

appendix. At the application chain (index b) level, we run the following regression:

Applications

= βF intech

+ F E + X

controls

+ 

Applications

is the number of applications for borrower b. This value is one for loans that

were origianted on the ﬁrst try. It also includes denials that were not ultimately successful.

F intech

is an indicator for whether the set of matched applications includes an application

to a ﬁntech lender or alternatively an originated loan with a ﬁntech lender. We run the

analysis on a 5% sample of all application chains between 2018 and 2021.

Table 6 reports these results. Column (1) shows that application spells that include

ﬁntech applications have fewer applications overall. A similar result holds for spells that

include ﬁntech originations shown in Column (2). The result in Column (2) indicates that

successful ﬁntech originations were executed with fewer applications. However, Column (1)

shows that multiple ﬁntech attempts which were unsuccessful were also rare. Together,

ﬁntech lenders are more successful at matching overall. All application spells have, by

deﬁnition, at least one application, but the mean number of applications is 1.05. To interpret

quantities, our measured coeﬃcients, which range from -0.004 to -0.006, represent roughly a

10% increase in the probability of having more than one application.

Fintech Loan Takeup Rates : In a similar vein, we next examine whether, once

approved for a loan, ﬁntech borrowers are more likely to actually complete the borrowing

process through origination. Using a random sample of all approved HMDA reﬁnance ap-

plications between 2018 and 2021 , we run the following regression at the loan application

level,

Originated

= βF intech

+ F E + 

(6)

Originated

is a zero-one indicator for whether the approved loan is originated. F intech

is a zero-one indicator for whether the lender is a ﬁntech lender. F E is a collection of lender

type, binned borrower characteristics, and market ﬁxed eﬀects. The results are shown in

Electronic copy available at: https://ssrn.com/abstract=4552425

Table 6 Panel B Columns (1) and (2). Column (2), which is the most saturated speciﬁcation,

shows that approved ﬁntech reﬁnance loans have a 2.7% greater likelihood of being accepted,

or conversely, a 2.7% lower likelihood of being turned down by the borrower. The overall

rate at which borrowers decline lender oﬀers is roughly 5%, so the ﬁntech eﬀect represents

more than a 50% decrease versus the baseline value.

We run an additional speciﬁcation examining a conceptually similar but slightly diﬀerent

outcome. Among all HMDA applications not rejected by the lender, we examine whether

the borrower withdraws her application (or turns down an approved loan). Note that in

comparison to the previous speciﬁcation, this speciﬁcation includes some loans that are

withdrawn by the borrower before the lender makes a credit decision. In consequence, the

sample is slightly larger. We show these results in Table 6 Panel B Columns (3) and (4). Here,

we ﬁnd that relative to a mean withdrawl rate of 16%, ﬁntech applications are withdrawn 10%

less frequently, again representing an economically and statistically signiﬁcant diﬀerence.

3.2.3 Fintech lenders charge high fees to recaptured borrowers

Finally, we return to the analysis based on regression (4) with results shown in Table 4.

Recall, Panel A shows diﬀerences in the timing of reﬁnance among various reﬁnance types,

and Panel B shows diﬀerences in fees. Column (1) in Panel A shows that compared to

the base group of non-ﬁntech non-recaptures, recaptured ﬁntech reﬁnances occur nearly 0.9

years, or 10.5 months, earlier. This represents a large eﬀect—nearly 30% faster—relative

to the typical time between origination and reﬁnance of 2.8 years. Column (2) shows that

this faster reﬁnance translates into a much smaller rate improvement, with the realized rate

improvement on recaptured ﬁntech loans being roughly 16 basis points smaller than average.

Again, this is a large eﬀect—roughly 14% smaller—relative to the mean improvment of

1.14%. In consequence, the realized annual payment improvement is more than $300 lower,

or 10% relative to baseline, for ﬁntech recaptures.

The story is quite diﬀerent when ﬁntechs reﬁnance borrowers who are not their servicing

customers. Column (1) shows that ﬁntechs reﬁnance new customers only modestly early,

roughly 0.20 years faster relative to baseline. Interestingly, these reﬁnances are modestly

more advantageous to borrowers than the baseline arms-length non-ﬁntech reﬁnances, re-

alizing a rate improvement that is 6 basis points better for borrowers. This translates to

a payment improvement that is roughly $200 greater. In Appendix A3, we decompose the

rate improvement into diﬀerences into the rate on the old loan and the rate on the reﬁ-

nance. We ﬁnd that both recaptured and arms-length ﬁntech reﬁnances occur at roughly

market rates, but recaptured ﬁntech reﬁnances target borrowers with lower-than-baseline pre-

existing rates, while arms-length ﬁntech reﬁnances capture borrowers with slightly higher-

Electronic copy available at: https://ssrn.com/abstract=4552425

than-baseline pre-existing rates.

Turning to fees, Panel B shows that both ﬁntech recaptures and ﬁntech new borrowers

charge signiﬁcantly higher fees compared to baseline reﬁnances. Fintechs charge recaptures

$315, or 15 basis points per loan principal more, while the corresponding premia are $400, 20

basis points, for new customers. These higher fees, combined with the diﬀerential payment

improvements shown in Panel A, imply that the break-even time for a ﬁntech recapture is

nearly one quarter year longer, or 20% greater than the baseline time. Similarly, the break-

even time for a ﬁntech new borrower is .14 years longer, or 10% longer than the baseline

time.

To summarize the ﬁndings in this section, we ﬁnd that Fintech lenders focus much more

on recapturing their servicing customers than other lenders do. These recaptured reﬁnances

occur much sooner and realize much smaller rate improvements than similar reﬁnances ef-

fected by arms-length, non-ﬁntech lenders. Despite the fact that these recaptures oﬀer

smaller economic beneﬁts to customers, ﬁntechs charge signiﬁcantly higher fees, with the net

eﬀect being that it takes much longer for recaptured ﬁntech reﬁs to oﬀer a net beneﬁt to the

borrower.

3.3 Discussion and interpretation

We conclude the empirical section with a brief discussion and interpretation of these facts.

The market has been shifting away from a model of mostly integrated originator-servicers

towards one where these activites occur in diﬀerent entities. This is partially driven by mar-

ket share gains in non-bank originators, who are on average less likely to be integrated, but

is largely driven by within-lender type shifts. The rise of ﬁntech intermediaries, which are

overwhelmingly integrated, somewhat oﬀsets this secular decline in integrated origination-

servicing. Our diﬀerence-in-diﬀerence evidence suggests that there is an important causal

connection between servicer integration and reﬁnance rates, driven by lenders directly reﬁ-

nancing their own servicing customers.

Our results highlight two mechanisms for this channel. First, lenders appear to have a

cost advantage in reﬁnancing their own servicing customers. The empirical evidence for this

mechanism comes through the lower fees paid by recaptured borrowers compared to arms-

length borrowers. While we formalize this idea subsequently in our structural model, these

lower fees suggest that originators face lower marginal costs which, through competition, are

(partially) passed through to borrowers.

Second, ﬁntech lenders in particular appear to attract reﬁnance borrowers who are not

considering other oﬀers, and they draw heavily from their pool of servicing customers in order

Electronic copy available at: https://ssrn.com/abstract=4552425

to do so. To contextualize this mechanism, it is important to consider the broader context in

which the overall reﬁnancing rate is “low,” with a signiﬁcant portion of this low rate being

explained by search costs or other borrower inertia against reﬁnancing.

Our results are

consistent with ﬁntech lenders being able to partially overcome that intertia for their servicing

customers in particular. In doing so, they obtain local market power over these borrowers

and exploit it by charging higher fees. The empirical evidence for this mechanism comes

from several results. Fintechs’ own share of recaptured reﬁnancings is much higher than

other lenders’. Their recaptured customers reﬁnance much earlier and at less advantageous

rates as compared to other borrowers, suggesting that they are reaching these customers

earlier and before they look to reﬁnance on their own. Finally, once recaptured, these ﬁntech

borrowers pay much higher fees, suggesting that ﬁntechs are charging a higher markup in

response to their market power.

Finally, we discuss and rule out an alternate ﬁntech lending channel, which is better

underwriting. Broadly, much of the ﬁntech literature has examined whether ﬁntech uses

technology to better screen borrowers. We note ﬁrst, for institutional reasons, that lenders

originating GSE mortgages have little incentive to invest in more predictive underwriting

technology for the simple fact that the GSEs guarantee risk on the loans and that the price

of these guarantees is set by a simple and pre-determined FICO × LTV grid. Additionally,

in Appendix A.3, we examine whether ﬁntech loans perform better, and ﬁnd little evidence

in support of this mechanism.

Together, the empirical facts point towards two related mechanism. The ﬁrst is a margianl

cost advantage for integrated lenders. The second is a customer acquisition advantage for

ﬁntech lenders. We note that these two advantages appear to complement each other for

the ﬁntech case. In the model section below, this is a point we study more closely. Before

concluding this empirical section, we discuss supporting evidence and mechanism results

regarding the ﬁntech lenders.

4 A Model of Mortgage Reﬁnancing and Servicing

Motivated by our empirical results, we develop a dynamic quantitative model of

mortgage reﬁnancing featuring various servicers in a central role. Our model provides an

economic framework to help understand our reduced form results. We calibrate the model

using several empirical moments from our reduced-form analysis in Section 3 in order to oﬀer

a quantitatively realistic picture of the market. Our calibrated model oﬀers an interpretable

See Keys, Pope and Pope (2016), Agarwal, Rosen and Yao (2016), and Agarwal et al. (2020). See

Appendix A.1 for our own analogous presentation of these facts.

Electronic copy available at: https://ssrn.com/abstract=4552425

quantiﬁcation of key economic forces, namely, the implied cost to a borrower of reﬁnancing

her mortgage, the marginal cost beneﬁt accruing to an integrated originator-servicer

when reﬁnancing its own servicing customers, and the technological advantage in cus-

tomer acquisition enjoyed by ﬁntech servicers. Finally, we undertake several counterfactuals

in Section 5 to explore the quantity and price impact of servicer integration and ﬁntech entry.

4.1 Model setting

We describe the model setting. We begin by detailing the contracts and agents before

explaining the model’s timeline and agents’ decisions.

4.1.1 Contracts and agents

Mortgage contracts: A mortgage contract indexed by (r

, t, M) has a (reﬁnanceable)

ﬁxed rate r

, years to maturity t, and initial principal balance M. At origination, the time to

maturity is T . Mortgages are fully amortizing and payments are ﬁxed over the life of the loan.

We deﬁne the (constant given the interest rate) period payment as equal to r

× A(M ; T ).

Simplifying notation, we ﬁx M and T and suppress them in the notation, so that a mortgage

contract uniquely identiﬁed by (r

, t) has period payment r

× A. Upon reﬁnancing to a new

rate r, nothing is changed about the loan except that its period payment becomes r × A.

Servicing contracts: Implicitly associated with each mortgage contract is a servicing

contract. So long as the mortgage contract exists, the servicing contract pays its owner a

fee proportional to the mortgage payment less servicing costs. We denote this net fee, as

a fraction of the mortgage payment, as s, so that each period the owner of the servicing

contract receives a ﬂow payment of s × r

× A. We take s to be set exogenously by the

GSEs. Reﬂecting our institutional setting, this fee is paid by the GSEs, who are outside of

our model.

Borrowers: Borrower i has an existing mortgage contract indexed by (r

, t) faces a

prevailing interest rate r, and has a pre-existing servicing relationship with servicer j. Each

period, the borrower makes her mortgage interest payment, r

× A, and decides whether to

reﬁnance her mortgage. Except in the case where she reﬁnances with her ﬁntech servicer,

the borrower must pay a ﬁxed cost c in order to obtain oﬀers from lenders. This ﬁxed cost

c represents the cost of time spent and mental eﬀort of deciding whether to reﬁnance and

getting oﬀers and captures the “inert” nature of households in reﬁnancing their mortgages.

Borrowers discount the future by a factor β < 1 and exit the model at time t = 0, when

their loan is paid oﬀ.

Electronic copy available at: https://ssrn.com/abstract=4552425

Intermediaries: Intermediary j originates and services mortgage reﬁnances for borrow-

ers. When originating a mortgage for consumer i, intermediary j charges a fee f

and pays

a marginal cost of origination equal to mc

. As we discuss shortly, intermediaries possess

market power and in equilibrium will earn an origination markup f

− mc

> 0. When

in possession of a servicing contract S(r

, t) the intermediary earns the net servicing fee

s × r

× A. To simplify the model, we assume that the mortgage (the stream of mortgage

payments, as opposed to the origination charge or servicing fees) are sold into a competitive

secondary market so that originators only earn proﬁts from origination and servicing fees.

There are three types of intermediaries:

1. Disintegrated intermediaries, D: the originator is not a servicer. Upon reﬁnancing

a mortgage, the disintegrated originator earns the origination markup and sells the

servicing contract to a disintegrated servicer at the expected discounted value of the

servicing cashﬂows. There are N

disintegrated intermediaries.

2. Integrated non-ﬁntech intermediaries, I: the originator is also a servicer. Upon reﬁ-

nancing a mortgage, the integrated non-ﬁntech originator retains the servicing contract

and collects the servicing cashﬂows period-by-period. There are N

integrated inter-

mediaries.

3. Integrated ﬁntech intermediaries, F : additionally, the ﬁntech intermediary possesses

an advantage in customer acquisition that we detail shortly. There is one ﬁntech

intermediary.

Interest rates: The reﬁnance interest rate available to a borrower evolves exogenously

as r

∼ F (r; θ) where r

is next period’s interest rate, r is this period’s interest rate, and θ

parameterize the distribution F (·).

4.1.2 Borrower’s timeline

Borrower i and her associated servicer j enter the period with the state variables

, r, t, j), where r

is the ﬁxed rate on the mortgage, r is the rate at which the borrower

can reﬁnance, t is the years to maturity on the mortgage, and j is the identity of the ser-

vicer. At the start of the period, the borrower makes her mortgage payment r

× A and the

servicer receives its servicing fee s × r

× A. If t = 0, the game ends. Let B(r

, r, t, j) denote

the borrower’s value function and let S(r

, r, t, j) denote the servicer’s value function. The

mortgage matures at t = 0, so that B(r

, r, t, j)|

t<0

= S(r

, r, t, j)

t<0

= 0

Standard reﬁnance market: For t > 0, the borrower now decides whether to begin

the reﬁnancing process. Unless she has a ﬁntech servicer, the borrower must pay a cost

Electronic copy available at: https://ssrn.com/abstract=4552425

c to obtain mortgage oﬀers from intermediaries. For intermediary k’s oﬀer to borrower i,

borrower i obtains the following utility:

= −f

+ βE [B(r, r

, t − 1, k)] + 

(7)

= βE [B(r

, r

, t − 1, j)] + 

(8)

Where u

represents the outside option of not reﬁnancing after seeing the oﬀers. Intu-

itively, the cost of reﬁnancing is the fee paid to the intermediary, and the beneﬁt is changing

the ﬁxed rate from r

to r (as well as changing the servicer from j to k, which may be

indirectly beneﬁcial). The outside option is to not pay the fee and keep the status quo mort-

gage rate and servicer. 

and 

are horizontal taste shocks that realize after the borrower

chooses to obtain mortgage oﬀers (capturing, e.g., non-price attributes of intermediaries spe-

ciﬁc to the borrower). We later specialize as type-1 extreme value shocks for computational

simplicity.

Among the oﬀers and outside option k ∈ {0, 1, . . . , N

+ N

+ 1} the borrower chooses

that which maximizes her utility. Her decision problem and ex-ante expected utility from

choosing among these oﬀers is as follows:

Refi

= arg max

} (9)

EU[Standard Refi

] = E

max

}

(10)

Where the expected value is taken over realizations of the 

’s.

The borrower only makes this choice if she chooses to obtain oﬀers. In deciding whether

to obtain oﬀers, she considers the expected utility from each choice:

i,get offers

= −c + EU [Standard Refi

] + 

i,get offers

(11)

i,no offers

= βE [B(r

, r

, t − 1, j)] + 

i,no offers

(12)

c is the cost of obtaining an oﬀer. If she obtains oﬀers, she receives in expectation the

value of considering a standard reﬁnance, deﬁned above. If she does not obtain oﬀers, she

proceeds in the status quo with her current mortgage rate and servicer. Finally, 

i,get offers

and 

i,no offers

are horizontal taste shocks, parameterized again as type-1 extreme value

shocks. Her decision problem and ex-ante expected utility from the choice of whether to

Electronic copy available at: https://ssrn.com/abstract=4552425

obtain oﬀers is as follows:

Offer

= arg max {u

i,get offers

, u

i,no offers

} (13)

EU[Offer

] = E [max {u

i,get offers

, u

i,no offers

}] (14)

Targeted ﬁntech market: When the borrower’s existing lender is a ﬁntech lender, the

ﬁntech lender can make a special targeted oﬀer to the borrower before the borrower decides

to obtain other oﬀers. The ﬁntech’s fee is f

. The advantage of the ﬁntech’s targeted oﬀer

is that its targeted nature reduces the borrower’s mental cost of reﬁnancing her mortgage.

In particular, rather than costing her c, reﬁnancing in this manner costs only c − ∆c

. The

borrower’s utility from reﬁnancing in the targeted ﬁntech market with the ﬁntech lender j

or not is as follows:

i,F

= −f

− (c − ∆c

) + βE [B(r, r

, t − 1, j)] + 

(15)

im∼F

= EU [Of fer

] + 

(16)

That is, if the borrower reﬁnances, she pays the fee, pays the mental reﬁnancing cost,

and updates her mortgage rate and servicer. If she does not reﬁnance, she moves on to the

decision, deﬁned earlier, of whether or not to obtain broader oﬀers. Note that even if she

rejects the ﬁntech’s targeted oﬀer at this stage, she can still accept the ﬁntech’s oﬀer in the

standard reﬁnance market. As before, the borrower’s expected utility from this decision,

with expectation taken over the type-1 shocks is:

EU[F intech

] = E [max {u

i,F

, u

i,∼F

}] (17)

Combining these equations, we are ready to deﬁne the borrower’s value function in re-

cursive form:

B(r

, r, t, j) = −r

× A + EU[F intech

] (j is ﬁntech)

B(r

, r, t, j) = −r

× A + EU[Offer

] (j is not ﬁntech)

Observe that EU[F intech

] nests EU[Offer

] and that both nest the discounted contin-

uation value of reﬁnancing or not. The decisions and implicit market shares that arise from

Electronic copy available at: https://ssrn.com/abstract=4552425

the solution to borrower’s problem are as follows:

• p

(f, r

, r, t, j), the probability that the borrower accepts the targeted ﬁntech oﬀer

given the state variables and oﬀered fee.

• p

, r, t, j), the probability that the borrower chooses to obtain standard reﬁ oﬀers.

• p

({f}, r

, r, t, j), the probability that the borrower accepts intermediary k’s oﬀer given

the state variables and menu of oﬀered fees by intermediary k and others.

4.1.3 Intermediaries’ timeline

Intermediaries follow an analogous timeline. The current servicer has a value function

S(r

, r, t, j) and at the start of the period collects the servicing fee s × r

× A.

Standard reﬁnance market: If the borrower chooses to get reﬁnance oﬀers in the

standard reﬁnance market, lenders compete in diﬀerentiated Bertrand competition, simulta-

neously setting fees. Each intermediary type, (D,I,F ) has a diﬀerent dynamic problem.

Disintegrated intermediary’s problem: A disintermediated intermediary maintains no ser-

vicing relationship with a reﬁnance origination customer and instead immediately sells the

servicing rights. The origination problem for disintermediated intermediary j is as follows:

= max

, f

−j

, r

, r, t, j)



− mc

+ βE



(r, r

, t − 1, j)



(18)

Here, p

, f

−j

, r

, r, t, j) is the (endogenous) probability that the borrower reﬁnances,

is the fee, and mc

is the marginal cost. If the borrower accepts the intermediary’s oﬀer,

the intermediary receives the fee, pays the marginal origination cost, and sells the servicing

relationship at its discounted expected value. That value is given recursively as follows:

, r, t − 1, j) = s × A × r

1 − p

, r, t, j)

({f}, r

, r, t, k)

βE [S(r

, r

, t − 1, j)] (19)

That is, the servicing relationship pays the servicing fee and survives so long as no other

intermediaries reﬁnance the borrower. The borrower reﬁnances if she chooses to obtain oﬀers

and accepts one of the oﬀers.

Integreated non-ﬁntech’s problem: The problem for an integrated non-ﬁntech intermedi-

ary j with no prior servicing relationship is as follows:

Electronic copy available at: https://ssrn.com/abstract=4552425

= max

, f

−j

, r

, r, t, j)



− mc

+ βE



(r, r

, t − 1, j)



(20)

Here, p

, f

−j

, r

, r, t, j) is the (endogenous) probability that the borrower reﬁnances,

is the fee, and mc

is the marginal cost. If the borrower accepts the intermediary’s oﬀer,

the intermediary receives the fee, pays the marginal origination cost, and gains the servicing

relationship. The problem for intermediary j with a prior servicing relationship is slightly

diﬀerent:

= max

, f

−j

, r

, r, t, j)



− mc

+ βE



(r, r

, t − 1, j)



1 −

({f}, r

, r, t, k)

βE



, r

, t − 1, j)



(21)

In this case, (1) the intermediary possesses a marginal cost advantage in reﬁnancing its

own servicing customer, and pays mc

< mc

. Additionally, the second term incorporates

that if the borrower accepts no oﬀers then the status-quo servicer maintains the servicing

relationship through the next period. Unlike the disintegrated intermediary, the integrated

intermediary with a relationship takes the fact that it may lose its valuable servicing rela-

tionship into account when setting prices. The value function for the servicing relationship

is given recursively as follows:

, r, t − 1, j) = s × A × r

+ p

, r, t, j)Π

+ (1 − p

, r, t, j)) βE



, r

, t − 1, j)



(22)

That is, the servicing relationship pays the servicing fee. If the borrower chooses to get

oﬀers, the intermediary obtains the insider expected proﬁt P i

, and if the borrower does

not choose to get oﬀers, the intermediary continues with the existing servicing relationship

through to the next period.

Fintech intermediary’s problem: The ﬁntech intermediary’s problem in the standard

reﬁnance market exactly mirrors that of the integrated non-ﬁntech problem, except that

it has a diﬀerent value function, S

, r, t, j), which we deﬁne shortly.

Targeted ﬁntech market: When there is an existing ﬁntech servicing relationship, the

ﬁntech lender can make a special targeted oﬀer to the borrower before the borrower decides

Electronic copy available at: https://ssrn.com/abstract=4552425

whether to obtain other oﬀers. In that case, the ﬁntech’s problem is as follows:

, r, t, j) = s × A × r

+ max

(f, r

, r, t, j)



f − mc

+ βE



(r, r

, t − 1, j)



+ (1 − p

(f, r

, r, t, j))



, r, t, j)Π

+ (1 − p

, r, t, j)) βE



, r

, t − 1, j)



(23)

In this case, the intermediary ﬁrst obtains the service ﬂow. Second, the intermediary sets

a fee f . In setting the fee, the intermediary trades oﬀ a higher fee versus a lower chance

of reﬁnancing the borrower. If the borrower reﬁnances with the ﬁntech, it pays the fee and

maintains the servicing relationship. If the borrower does not reﬁnance with the ﬁntech,

it still maintains the servicing relationship with the ﬁntech unless she both (1) chooses to

obtain oﬀers and (2) accepts one of the non-ﬁntech oﬀers.

Finally, Figure 3 provides a graphical schematic for the model’s timeline. The black

boxes correspond to the borrower; the red boxes correspond to the intermediary. Borrowers

begin each period with a rate, time to maturity, potential reﬁnance rate, and a relationship.

Borrowers then pay interest on their loans. If the borrower has a ﬁntech servicer, the borrower

can reﬁnance in the targeted ﬁntech reﬁnance market. If no reﬁnance occurs, the borrower

can then choose to pay a cost to obtain reﬁnance oﬀers in a monopolistically competitive

market. If she reﬁnances, her rate updates and the cycle repeats. If not, she maintains her

rate and servicer relationship. Likewise, the servicer’s within-period problem tracks that of

the borrower where the servicer sets reﬁnance fees to maximize its lifetime expected income.

4.2 Equilibrium

Equilibrium in the model is deﬁned as follows. First, given fees, borrowers make optimal

choices over reﬁnancing with the ﬁntech, obtaining oﬀers, and accepting oﬀers if obtained.

These choices are summarized in p

(f, r

, r, t, j), p

, r, t, j), and p

({f}, r

, r, t, j).

Second, taking borrower and other intermediaries’ behavior as given, intermediaries set

optimal fees. These are: (f

, f

), which denote the ﬁntech fee in the targeted

ﬁntech market, the ﬁntech fee in the non-targeted market, the non-ﬁntech integrated no-

relationship fee, the non-ﬁntech integrated fee relationship fee, and the disintegrated fee.

Optimal fees are deﬁned for each point in the state space.

Electronic copy available at: https://ssrn.com/abstract=4552425

4.3 Model calibration

We calibrate the model using a combination of time-series estimates, institutional pa-

rameters or parameters drawn directly from the literature, and the simulated method of

moments. In this section, we describe the calibration process. Table 7 Panel A shows the

moments we target in the simulated method of moments; Panel B shows the calibrated

parameters and the source of the calibration.

Parameters set using institutional details or existing literature: We set the

initial loan principal M to be equal to $250k, which is roughly the median mortgage size

during our time period. The modal mortgage in our data set is a 30-year ﬁxed rate loan,

and so we set the initial loan length T = 30. We set the servicing fee, expressed in an annual

percentage of total loan payments, equal to 50 basis points, which is roughly in line with

GSE servicing guidelines. Finally, in line with the literature, we set the subjective discount

factor β = 0.95 Table 7 Panel B summarizes these details.

Parameters set with time-series estimates: We calibrate the interest rate process,

= F (r, θ) using a simple autoregressive process in levels. In particular, we assume that

= ρr + (1 − ρ)µ

+ σ

,  ∼ 0, ∞, where ρ is the AR parameter, µ

is the long-run mean,

and σ

is the volatility of the shocks. We obtain these parameters by regressing the 30-year

mortgage interest rate on its one-year lagged value. These estimates are reported in Table 7

Panel B.

Parameters calibrated through SMM: Finally, we have four remaining parameters to

calibrate. These are: mc

, the marginal cost of origination for a non-relationship originator,

, the marginal cost of origination for a relationship originator, c, the implied mental cost

for a borrower to obtain reﬁnance oﬀers, and ∆c

, the mental cost advantage of reﬁnancing

with a targeted ﬁntech oﬀer. To calibrate these parameters, we use an over-identiﬁed simu-

lated method of moments approach. Table 7 Panel A shows the target moments in the data

and those that obtain from the calibration.

While the parameters are identiﬁed jointly, we provide a brief intuitive discussion high-

light what moments are most informative about our model parameters. Marginal costs,

and mc

are most directly informed by the fees observed in equilibrium in the data.

In particular, our model imposes a parametric structure between fees and marginal costs,

where observed fees reﬂect the combination of a markup (reﬂecting market power) and a

markdown (reﬂecting some intermediaries’ willingness to oﬀer lower fees to win valuable fu-

ture servicing ﬂows). The overall level of fees implies a level of marginal costs, which we

ﬁnd to be roughly $1,775. Additionally, recall that fees charged by relationship non-ﬁntech

originator-servicers are lower by approximately $150. This diﬀerence is informative about

the marginal cost diﬀerences for outsiders, mc

, and insiders, mc

. Our calibration estimates

Electronic copy available at: https://ssrn.com/abstract=4552425

that this diﬀerence is roughly $250, meaning that originators with a servicing relationship

possess a cost advantage that they only partially pass through to borrowers.

The mental, or inertia cost c of obtaining reﬁnance oﬀers is identiﬁed, to a ﬁrst order,

by the overall frequency of reﬁnance. A higher frequency of reﬁnance implies that borrowers

are obtaining oﬀers more often, which in turn implies that c is low. In contrast, a lower

frequency of reﬁnance implies that borrowers are obtaining oﬀers more rarely, which in turn

implies that c is high. We ﬁnd that matching the (low) overall share of reﬁnancing implies

a high mental cost of roughly $11k.

Finally, the mental cost advantage for ﬁntech lenders is identiﬁed jointly through a num-

ber of moments, including the ﬁntech lending share, the ﬁntech markup over non-ﬁntech fees,

and the share of ﬁntech reﬁnances that are recaptures as opposed to new clients. Broadly,

a higher ﬁntech lending share, a higher ﬁntech markup, and a higher share of ﬁntech recap-

tures is consistent with a greater ﬁntech mental cost advantage. Our calibration estimates

this advantage to be roughly $1,588. Despite our SMM calibration being overidentiﬁed, the

model is fairly successful in hitting the target moments in the data.

5 Model Analysis and Counterfactuals

We conclude by discussing key economic forces in the model and running counterfactuals

to understand the importance of servicer intermediation and disintermediation and the role

of ﬁntech customer acquisition technology.

5.1 Model analysis

We begin by exploring key economic forces in the model and connecting thtem to our

reduced form results by examining some of the model’s policy functions. These are shown in

Figure 4. Policy functions in this model are a four-dimensional object because they depend

in (r

, t, r, j), the current ﬁxed rate, the mortgage’s time to maturity, the potential reﬁnance

rate, and the identity of the current servicer. To collapse dimensionality, we ﬁx t = 5 and

r = 6%, that is, when there are ﬁve years left until maturity of the mortgage and prevailing

interest rates are 6%. In panels (a), (c), and (e), we plot market shares (conditional on

reﬁnancing), and in panels (b), (d), and (f), we plot reﬁnance origination fees. In each case,

we plot these outcomes versus the reﬁnance rate gap, r

− r, i.e., the rate improvement a

reﬁnancing borrower would obtain. Values to the left represent less improvement; values to

the right represent more improvement.

Panels (a) and (b) show shares and fees when the incumbent servicer is a disintegrated

Electronic copy available at: https://ssrn.com/abstract=4552425

servicer—that is, the servicer is unable to reﬁnance the borrower itself. Panel (a) shows that

in this case, non-ﬁntech originators have more market share than ﬁntech originators and

that these market shares do not vary with the potential rate improvement. Here, non-ﬁntech

market share exceeds ﬁntech market share simply because there are many more non-ﬁntech

originators than there are ﬁntech originators, and in the case where the ﬁntech intermediary

has no special access through being the servicer, ﬁntechs possess no special advantage.

Panel (b) shows fees. Observe that both ﬁntech and non-ﬁntech fees are increasing

through the region where reﬁnance becomes beneﬁcial to the borrower, moving from roughly

2.20 to 2.40 as r

− r increases from -0.01 to 0.02. Intuitively, as the reﬁnance becomes more

beneﬁcial to the borrower, they are willing to pay more for it, and because the market is

not perfectly competitive, lenders can charge higher prices as the reﬁnance becomes more

valuable to the borrower. Interesting, although the lenders in this scenario have no marginal

cost diﬀerences, ﬁntech lenders charge slightly lower fees than non-ﬁntech lenders. This is

because, as we will see, the ﬁntech-borrower relationship is more valuable than the non-

ﬁntech-borrower relationship and so ﬁntechs are willing to charge a lower price upfront in

order to have a greater chance of building the relationship.

Panels (c) and (d) show shares and fees when there is a non-ﬁntech integrated servicer

with a preexisting servicing relationship, whose policy function is now shown in green. Inter-

estingly, the relationship servicer charges much higher fees when reﬁnancing has a relatively

low value, and decreases its fees when reﬁnancing has a relatively high value. These dif-

ferences reﬂect (1) the servicer’s incentives to maintain the servicing relationship, and (2)

the passthrough of its marginal cost advantage. First, when the value of a reﬁnance is low

to a borrower, it is very unlikely that the borrower will reﬁnance with any lender. Hence,

the relationship lender does not need to worry about losing the relationship, and instead

can charge a high fee in hopes that in the highly unlikely case that the borrower chooses to

reﬁnance, it will earn signiﬁcant money on the fee.

In contrast, when the value of a reﬁnance is high to the borrower, the relationship servicer

dramatically reduces prices. It reduces prices because when reﬁnancing is very valuable to

the borrower, the is likely to obtain reﬁnancing oﬀers and reﬁnance. The borrower’s more

aggressive eﬀort to obtain reﬁnancing oﬀers endogenously increases the level of competition

in the lending market, and in order to maintain its valuable servicing relationship, the

relationship servicer reduces prices. Due to its marginal cost advantage, it is able to reduce

prices far below those of other lenders. This gap between the relationship servicer’s rates

and rates from others neatly reﬂects the rate diﬀerences shown in Table 4.

Finally, panels (e) and (f) show shares and fees when there is a ﬁntech servicer with a

preexisting servicing relationship. In this case, the ﬁntech’s ability to send targeted rate

Electronic copy available at: https://ssrn.com/abstract=4552425

oﬀers at a lower cost to the borrower gives the ﬁntech signiﬁcant market power, particualrly

in states of the world where the borrower is unlikely to obtain oﬀers itself. This occurs

primarily in the region of the state space where reﬁnancing has a relatively low value. Here,

panel (e) shows that the ﬁntech has a signiﬁcant market share advantage. This is precisely

consistent with our ﬁndings on borrower terms in Table 4, showing that ﬁntech reﬁnances

tend to occur earlier in the loan’s life and at less advantageous rate gaps for the borrower.

Panel (f) shows that the ﬁntech lender exploits this market power to charge signiﬁcant

markups. As the value of reﬁnancing increases to just about 0, this market power advantage

evaporates as the ﬁntech expects borrowers to begin obtaining outside reﬁnancing oﬀers.

The ﬁntech loses market share and partially responds by decreasing prices. Again, these

ﬁndings are consistent with 4 showing that ﬁntech recaptures have much higher fees than

other origination types despite being relatively disadvantagous to borrowers.

Interestingly, as the value of reﬁnancing becomes extremely high—above 2% in terms of

rate diﬀerence—ﬁntech fees again rise. This occurs because the dynamic value of having

a ﬁntech servicer to oﬀer easier reﬁnances makes the ﬁntech oﬀer more desirable to the

borrower even if the original oﬀer is obtained in the standard reﬁnance market. In other

words, in anticipation of easier reﬁnances in the future, borrowers are wiling to pay more for

ﬁntech services today.

5.2 Counterfactual I: Intermediation and disintermediation

We now turn to our ﬁrst counterfactual. Here, we quantify the importance of the

originator-servicer integration in the quantities and costs of reﬁnancing by eliminating the

marginal cost advantage of the insider-servicer. In other words, we set mc

≡ mc

, so that

the insider servicer and the outsider servicer are on equal footing in terms of future reﬁnanc-

ing costs. We show the results in Figure 5 Panels (a), (c), and (e). In each panel, we again ﬁx

t = 5 years and r = 6%, and plot outcomes against the rate gap r

− r. We show the results

by the identity of the incumbent servicer—a disintegrated servicer, a ﬁntech servicer, and

an incumbent servicer. Panel (a) shows the percentage change in total reﬁnancing compared

to the baseline calibration. Panel (c) shows this in absolute, rather than percentage terms,

and Panel (e) shows the percentage change in fees.

Panel (a) shows dramatic percentage reductions in the quantity of reﬁnancing: Around a

50 basis point rate gap, the reﬁnancing rate decreases by roughly 7.5% relative to the baseline

when the incumbent servicer is a ﬁntech, and by roughly 5% relative to the baseline when

the incumbent servicer is a non-ﬁntech integrated servicer. In terms of annual quantities,

panel (c) shows that the fraction of the population reﬁnancing declines by roughly 50 and 25

Electronic copy available at: https://ssrn.com/abstract=4552425

basis points, respectively. Panel (e) shows how this change is mediated through increases in

fees of between 2 and 4 percent relative to the baseline. The eﬀect is greater when there is

a ﬁntech incumbent because, due to the ﬁntech’s advantage in targeting their own servicing

customers, ﬁntechs are the most likely to reﬁnance their own servicing customers and thus

when own-servicer reﬁnancing costs see a relative increase, ﬁntechs are the most aﬀected.

This counterfactual illustrates that the cost advantage coming from originator-servicer

integration leads to an economically meaningful change in the quantity of reﬁnancing activ-

ity. Consistent with our reduced form results, eliminating that connection—as many trends

such as increased MSR capital requirements have done—likely have lead to economically

meaningful reductions in household reﬁnancing.

5.3 Counterfactual II: Fintech customer acquisition technology

Finally, we examine the importance of ﬁntech’s ability to use technology to target their

servicing customers for reﬁnancing. We run this counterfactual by sending ∆c

→ −∞,

eﬀectively making it impossible for the ﬁntech lender to reﬁnance any customers outside of

the standard reﬁnance market. The results, presented analogously to the previous counter-

factual, are shown in Figure 5 panels (b), (d), and (f).

Panel (b) shows a dramatic percentage reduction in reﬁnance activity relative to the

baseline where ﬁntechs are able to target their servicing customers. These changes are par-

ticularly stark when reﬁnancing is relatively less valuable. In these cases, it is highly unlikely

that borrowers would pay the cost to obtain oﬀers, and hence the ﬁntech lender’s targeting

is essentially their only chance at reﬁnancing. However, even for positive and signiﬁcant re-

ﬁnance values, eliminating ﬁntech’s customer acquisition technlogy has quantitatively large

eﬀects. When the reﬁnance gap is 50 basis points, the percentage of reﬁnances decreases by

32% relative to the baseline, and even when the reﬁnance gap is 1%, the decrease is roughly

25%.

These ﬁndings, where the relative impact of ﬁntech is largest among borrowers who ben-

eﬁt the least from reﬁnancing, are largely consistent with our reduced form ﬁndings. In

particular, our results in Table 5 showed that overall, ﬁntech reﬁnancings tended to be re-

captures. In Appendix Table A7, we replicate this analysis on subsamples of the borrower

population formed by their ex-ante potential rate improvement. We ﬁnd that ﬁntech reﬁ-

nances with the smallest rate improvements are the most likely to be recaptures, whereas

ﬁntech reﬁnances with rate improvements with greater than 2% are not more likely than

other reﬁnances to be recaptures.

We next move to absolute changes in reﬁnancing quantities, as shown in Panel (c).

Electronic copy available at: https://ssrn.com/abstract=4552425

Observe that almost no change in absolute reﬁnancing quantity is detectable when the re-

ﬁnancing gap is less than 50 basis points. This is because reﬁnancing at this place in the

state space—through a ﬁntech relationship or not—is so unlikely that absolute changes are

vanishingly small. However, as the reﬁnance gap increases to 1% and the overall rate of reﬁ-

nancing increases, the eﬀect of eliminating ﬁntech customer acqusition technology becomes

economically meaningful: the annual rate of reﬁnancing decreases by 6% in absolute terms,

which, as discussed above, constitutes roughly 25% of all reﬁnancing.

Panel (f) shows the impact on fees. Recall that one eﬀect of the ﬁntech’s ability to

target customers was to give it signiﬁcant market power over those customers. Eliminating

that ability reduces the ﬁntech’s ability to exploit its market power, and so reﬁnancing fees

shrink. These results highlight an apparent tradeoﬀ of the ﬁntech’s customer acquisition

technology: On one hand, the technology appears to give the ﬁntech market power, and

it exploits that market power by charging higher fees. However, that market power exists

because it is oﬀering borrowers a signiﬁcantly cheaper way to reﬁnance their mortgages,

while still preserving their existing options. Thus, ﬁntech’s customer acquisition technology

increases the total surplus, and the ﬁntech captures a portion of that surplus through higher

markups. Unlike in a standard market power setting, however, the ﬁntech’s higher markups

are actually accompanied by a quantity increase through oﬀering a more convenient product.

6 Conclusion

Our paper documented two dramatic changes in the structure of loan servicing in the US

residential mortgage market. First, the share of mortgages where the originator is also the

servicer has declined dramatically from 70% in 2010 to 30% in 2015, recovering modestly since

then. Second, a large share of this recovery was driven by the rise of “Fintech” intermediaries,

who both originate and service loans. We explored the consequences for these major shifts

for household mortgage reﬁnancing prices and quantities.

We emphasize two economic forces for integrated ﬁntech and non-ﬁntech intermediaries.

First, integrated intermediaries have a marginal cost advantage in reﬁnancing their servicing

customers. This cost advantage is partially passed through to borrowers in the form of lower

fees. Second, ﬁntech intermediaries in particular use data and customer access for more

eﬀective customer acquisition. The exploit this unique customer access ability to extract

higher markups. We established these facts in reduced form using a change to capital risk

weights on mortgage servicing assets that impacted banks and not non banks to show that

integrated servicers reﬁnance their servicing portfolio at a higher rate. We also used a

novel forward-merged data set to show that reﬁnances originated by the existing servicer

Electronic copy available at: https://ssrn.com/abstract=4552425

are done with lower fees for non-ﬁntechs, and with higher fees and at less advantageous

terms for ﬁntechs. We then built and calibrated a quantitative structural model to quantify

the importance of these economic forces and run policy counterfactuals. We showed that a

fully disintegrated servicing sector would see roughly 10% fewer mortgage reﬁnances at 5%

higher prices. Additionally, the ﬁntech customer acquisition channel increases the annual

rate of mortgage reﬁnancing by roughly 25% relative to the baseline and an increase in fees

by roughly 5% relative to the baseline.

We conclude with several important implications from our paper. First, the industrial

organization of the servicing market has large quantitative consequences for policy-relevant

ex-post borrower outcomes like mortgage reﬁnancing. Because it is the largest consumer

ﬁnance market in the United States, the reﬁnancing of ﬁxed rate residential mortgages has

large direct welfare implications for households and for the passthrough of monetary pol-

icy. Second, Second, there are imporant aspects for the contract design of mortgages and

mortgage servicing rights that would have large impacts on reﬁnancing prices and quanti-

ties. Finally, we emphasized a novel role of ﬁntech: customer acquisition, as opposed to

underwriting. In our setting, ﬁntechs source otherwise uncontested customers and charge

higher markups. An important takeaway from our paper is that in the largest consumer ﬁ-

nance setting, the chief application of ﬁntech is a marketing application, not an underwriting

application.

Electronic copy available at: https://ssrn.com/abstract=4552425

Figure 1: Mortgage Market Recent Trends

Note: The ﬁrst panel of the ﬁgure plots the monthly share of originated loans where the lender is

an integrated lender-servicer. The sample is a random draw from the full set of single-family FNMA

and FHLMC loan originations. We identify integrated loans by matching lender and servicer names.

Loans where the lender or servicer is too small to be named in the data ﬁles are removed. The second

panel shows the monthly share originated by integrated ﬁntech lenders. The sample is identiﬁcal to

panel A but the numerator includes loans that are not only integrated but the lender is identiﬁed as

a ﬁntech lender using the deﬁnition in Buchak et al. (2018a).

Panel A: Integrated Origination Share

Panel B: Integrated Fintech Origination Share

Electronic copy available at: https://ssrn.com/abstract=4552425

Figure 2: Integrated Servicers Refinancing: Event Study

Note: The ﬁgure plots the coeﬃcients from the dynamic diﬀerence-in-diﬀerence test given by

Refi

2015 Q4

τ =2011 Q1

t=τ

× Bank

+ γ

+ 

The sample consists of all agency loans originated between January 2000 and Decmeber 2012

with an integrated lender. The dependent variable is an indicator equal to “1

if the loan exits

through a reﬁnancing at time t. The event depicted in the event study is January 2013 when

the Basel III framework changed the costs of holding on to MSRs as an asset for bank lenders.

Following this implementation, banks were less likely to also be servicers and thus, became

unintegrated.

The ﬁxed eﬀects γ

are interacted, binned FICO at origination, binned LTV, 3-digit zip code,

and binned current loan size. γ

refers to the originator type.

The plot shows a signiﬁcant decline in the reﬁnancing rates of loans originated by bank lenders

following the 2013 change in the MSR capital rule.

Electronic copy available at: https://ssrn.com/abstract=4552425

Figure 3: Model timeline

Note: This ﬁgure provides a representation of the within-period timeline of the quantitative structural model. The black boxes correspond to the borrower; the

red boxes correspond to the intermediary. Borrowers begin each period with a rate, time to maturity, potential reﬁnance rate, and a relationship. Borrowers

then pay interest on their loans. If the borrower has a ﬁntech servicer, the borrower can reﬁnance in the targeted ﬁntech reﬁnance market. If no reﬁnance

occurs, the borrower can then choose to pay a cost to obtain reﬁnance oﬀers in a monopolistically competitive market. If she reﬁnances, her rate updates and

the cycle repeats. If not, she maintains her rate and servicer relationship. Likewise, the servicer’s within-period problem tracks that of the borrower where

the servicer sets reﬁnance fees to maximize its lifetime expected income.

Electronic copy available at: https://ssrn.com/abstract=4552425

Figure 4: Policy functions

Note: This ﬁgure shows originator policy functions from the calibrated structural model. All ﬁgures

show optimal policies versus the rate gap between the current ﬁxed rate and the rate at which borrowers

could reﬁnance for t = 5 years remaining and r = 6%. Panels (a), (c), and (e) show origination shares;

panels (b), (d), and (f) show fees. Shares are expressed in decimals and fees are expressed in thousands

of dollars. Panels (a) and (b); (c) and (d); (e) and (f) show results for when the incumbent servicer

is dis-integrated, integrated non-ﬁntech, and ﬁntech, respectively. The red, blue, and green lines are

the policy functions of ﬁntechs, non-ﬁntech outsiders, and non-ﬁntech insiders, respectively.

(a) Shares, disintegrated servicer (b) Fees, disintegrated servicer

(e) Shares, ﬁntech servicer (f) Fees, ﬁntech servicer

Electronic copy available at: https://ssrn.com/abstract=4552425

Figure 5: Counterfactuals

Note: This ﬁgure shows counterfactuals from the calibrated structural model. All ﬁgures show optimal

policies versus the rate gap between the current ﬁxed rate and the rate at which borrowers could

reﬁnance for t = 5 years remaining and r = 6%. Panels (a), (c), and (e) show the counterfactual

that eliminates the marginal cost advantage of insider servicers; Panels (b), (d), and (f) show the

counterfactual that eliminates customer access advantage of ﬁntechs. Panels (a) and (b) show changes

in reﬁnancing propensities in percentage terms relative to the baseline. Panels (c) and (d) show the

same changes in absolute changes. Panels (e) and (f) show changes in reﬁnancing fees in percentage

terms relative to the baseline.

(a) Share percent changes, MC counterfactual (b) Share percent changes, ﬁntech counterfactual

(e) Fee % changes, MC counterfactual (f) Fee % changes, ﬁntech counterfactual

Electronic copy available at: https://ssrn.com/abstract=4552425

Table 1: Summary Statistics

Note: Panel A shows summary statistics for the random sample of FNMA and FHLMC loans used

in the MSR capital event study shown in table 2. The sample consists of loans at origination

and the summary statistics are separated by bank vs. non-bank lenders. Panel B shows summary

statistics for the forward-linked reﬁnancing sample. These are FNMA and FHLMC loans exiting

through a reﬁnancing which we have forward linked to their new mortgage using HMDA data. The

summary statistics show loans originated by ﬁntech lenders as compared to the full set of integrated

lender-servicer loans. Note that integrated ﬁntech loans falls under both categories.

Panel A: GSE sample

Unique Loans 79,593

Bank Lenders 66,759

Non-bank Lenders 12,834

Mean Std

Bank Non-bank Bank Non-bank

Interest Rate 5.11 4.22 0.96 0.76

Reﬁ 0.02 0.02 0.15 0.15

DTI 33 31 12 10

Loan Amount 200,395 208,114 118,889 128,796

Rate gap 1.21 0.37 1.00 0.82

Panel B: Linked Reﬁnances sample

Unique Loans 78,484

Integrated 36,643

Fintech 20,258

Mean Std

Fintech Integrated Fintech Integrated

DTI 35 35 9.32 9.16

Loan Amount 307,764 155,079 151,298 155,079

Reﬁ Amount 289,462 292,152 149,093 154,379

Rate Improvement 1.14 1.06 0.64 0.61

Electronic copy available at: https://ssrn.com/abstract=4552425

Table 2: Post-MSR capital rule event study

Note: Panel A shows the baseline results from the diﬀerence-in-diﬀerences MSR capital rule change. The

speciﬁcation is,

Refi

= βP ost

× Bank

+ γ

+ 

P ost

is an indicator for months after January 2013 when the Basel III rule changed the cost to banks of holding

MSR assets. The sample consists of a random draw from the FNMA and FHLMC monthly performance dataset

and observations are at the loan-month level. The summary statistics for these loans at origination is reported

in table 1. The results show the coeﬃcient β for speciﬁcations without rate-gap controls in the ﬁrst column, with

rate-gap controls, and in-the-money loans. The rate gap is deﬁned as r

old

− r

prevailing

. In the money loans are

those for whom the rate gap is higher than 50 basis points. Loan controls × servicer are ﬁxed eﬀects interacting

bins of borrower FICO scores, debt-to-income ratios, loan-to-value, loan size, and zip codes. Loan controls ×

date are the same borrower controls interacted with report date instead of servicer. Standard errors, clustered

at the report date level, are shown in parentheses. Panel B shows the results for the placebo test which limits

the sample to non-integrated originators only.

Panel A: Event study

Dependent variable:

% Reﬁ

(1) (2) (3)

Post x Bank servicer at origination −1.635

∗∗∗

−1.504

∗∗∗

−1.557

∗∗∗

(0.196) (0.196) (0.296)

E[Y] 1.69 1.69 1.95

Loan controls x servicer Y Y Y

Loan controls x date Y Y Y

Rate gap control N Y Y

Sample All All ITM

Observations 2,714,576 2,714,576 2,029,899

0.487 0.488 0.536

Note:

∗

p<0.1;

∗∗

p<0.05;

∗∗∗

p<0.01

Panel B: Placebo (non-integrated originators)

Dependent variable:

% Reﬁ

(1) (2) (3)

Post x Bank servicer at origination −0.226 −0.179 0.131

(0.156) (0.156) (0.169)

E[Y] 1.68 1.68 1.90

Loan controls x servicer Y Y Y

Loan controls x date Y Y Y

Rate gap control N Y Y

Sample All All ITM

Observations 3,439,660 3,439,660 2,951,084

0.718 0.718 0.747

Note:

∗

p<0.1;

∗∗

p<0.05;

∗∗∗

p<0.01

Electronic copy available at: https://ssrn.com/abstract=4552425

Table 3: Borrower recapture and lender market share

Note: This table shows how originator market share varies with whether the lender is the borrower’s servicer.

Data are constructed from the forward-matched reﬁnance data. Observations are at the lender-market level,

where a market is a year × three-digit zip × servicer. The left-hand side variable is market share. Originator is

servicer is an indicator for whether the lender is the servicer for the current market. All columns include market

(year × three-digit zip × servicer) ﬁxed eﬀects. Column (2) includes lender ﬁxed eﬀects. Column (3) includes

lender × year × three-digit zip ﬁxed eﬀects. Standard errors, in parentheses, are clustered at the lender level.

Dependent variable:

Originator market share

(1) (2) (3)

Originator is servicer 0.217

∗∗∗

0.184

∗∗∗

0.186

∗∗∗

(0.015) (0.010) (0.011)

Year × Zip × Servicer (Market) FE X X X

Lender FE X

Lender × Year × Zip FE X

Observations 47,603 47,603 47,603

Adjusted R

0.742 0.818 0.793

Note:

∗

p<0.1;

∗∗

p<0.05;

∗∗∗

p<0.01

Electronic copy available at: https://ssrn.com/abstract=4552425

Table 4: Lending Consequences of Recapture

Note: The two panels of this table relate loan characteristics to the reﬁnancing type. The sample is the forward-linked

reﬁnance data. We distinguish lender types based on “Fintech recapture” (ﬁntech loans which are recaptured), “Fintech new

borrower” (ﬁntech reﬁnancings where the previous lender was diﬀerent), and “Non-ﬁntech recapture” (recaptured loans where

the lender was not a ﬁntech lender). Panel A shows outcomes around the timing of the reﬁnance: Years to reﬁ, which is the

number of years between origination and reﬁnance, Rate improvement, which is r

old

− r

new

(in percentage points) between

the old rate and the new rate (a positive number means a larger improvement from the borrower’s perspective), and payment

improvement, which is the approximate change in annual payment in dollar terms after the reﬁnance (New Loan Amt ×

old

− r

new

)

100

). Panel B shows outcomes around the cost of the reﬁnance: The fee, both in dollar terms and in percentage

points of the loan principal, and breakeven years, which is the fee divided by the annual payment improvement. It measures

how many years have to pass before the payment improvement exceeds the origination charge. E[Y ] is the mean of the left-

hand side variable. All columns include lender type (bank versus shadow bank) ﬁxed eﬀects, market ﬁxed eﬀects, ﬁxed eﬀects

associated with loan charactersitics of the new reﬁnancing loan, and continuous borrower controls for the loan at origination

of the original loan. Note that the coeﬃcients of interest are robust to which set (origination or reﬁnancing) borrower controls

are used. The new loan ﬁxed eﬀects are interactions of HMDA variables: debt-to-income ratio bins, combined loan-to-value

bins, bins of borrower income, and applicant age. Continuous origination controls are debt-to-income, CLTV, original loan

size, and original loan term. Finally, market ﬁxed eﬀects are 3-digit zip code × year. Standard errors, clustered at the

lender-county level, are shown in parentheses.

Panel A: Timing of reﬁnance

Dependent variable:

Years to reﬁ Rate improvement Payment improvement

(1) (2) (3)

Fintech recapture −0.889

∗∗∗

−0.156

∗∗∗

−322.418

∗∗∗

(0.053) (0.017) (39.105)

Fintech new borrower −0.197

∗∗∗

0.062

∗∗∗

198.765

∗∗∗

(0.035) (0.008) (19.741)

Non-ﬁntech recapture −0.690

∗∗∗

0.009 35.544

(0.043) (0.012) (35.286)

E[Y] 2.84 1.14 2987

Lender Type FE Y Y Y

Borrower New Loan FE Y Y Y

Borrower Controls Y Y Y

Market FE Y Y Y

Observations 78,373 78,373 78,373

Adjusted R

0.280 0.240 0.403

Note:

∗

p<0.1;

∗∗

p<0.05;

∗∗∗

p<0.01

Panel B: Fees on recaptured Loans

Dependent variable:

Fee ($) Fee ( % of principal) Breakeven years

(1) (2) (3)

Fintech recapture 315.219

∗∗∗

0.154

∗∗∗

0.233

∗∗∗

(37.847) (0.015) (0.040)

Fintech new borrower 401.427

∗∗∗

0.211

∗∗∗

0.142

∗∗∗

(28.613) (0.013) (0.022)

Non-ﬁntech recapture −140.046

∗∗∗

−0.065

∗∗∗

−0.290

∗∗∗

(25.303) (0.009) (0.022)

E[Y] 2,201 1.05 1.29

Lender Type FE Y Y Y

Borrower New Loan FE Y Y Y

Borrower Controls Y Y Y

Market FE Y Y Y

Observations 78,373 78,373 78,373

Adjusted R

0.113 0.350 0.205

Note:

∗

p<0.1;

∗∗

p<0.05;

∗∗∗

p<0.01

Electronic copy available at: https://ssrn.com/abstract=4552425

Table 5: Fintech Borrower Recapture

Note: The table shows results from regressions of recaptured status on a ﬁntech lender indicator. The sample

consists of all loans in the forward-linked reﬁnancing dataset. The dependent variable is one if the reﬁnanced loan

was recaptured and zero otherwise. The ﬁrst two columns show results using the full sample while the third and

fourth columns show results for the sub-sample of reﬁnanced loans that were originated by integrated lenders.

The table indicates whether or not the results include lender type (bank versus shadow bank) ﬁxed eﬀects, market

ﬁxed eﬀects, ﬁxed eﬀects associated with loan charactersitics of the new reﬁnancing loan, and continous borrower

controls for the loan at origination of the original loan. Note that the coeﬃcients of interest are robust to which

set (origination or reﬁnancing) borrower controls are used. The new loan ﬁxed eﬀects are interactions of HMDA

variables: debt-to-income ratio bins, combined loan-to-value bins, bins of borrower income, and applicant age.

Continuous origination controls are debt-to-income, CLTV, original loan size, and original loan term. Finally,

market ﬁxed eﬀects are 3-digit zip code × year. Standard errors, clustered at the lender-county level, are shown

in parentheses.

Panel A: Fintech and borrower recapture

Dependent variable:

Recaptured reﬁnance

(1) (2) (3) (4)

Fintech 0.125

∗∗∗

0.123

∗∗∗

0.207

∗∗∗

0.204

∗∗∗

(0.010) (0.010) (0.014) (0.014)

Sample All Integrated only

E[Y] 0.14 0.14 0.30 0.30

Lender Type FE Y Y Y Y

Borrower New Loan FE Y Y Y Y

Borrower Controls N Y N Y

Market FE Y Y Y Y

Observations 78,484 78,373 36,643 36,603

Adjusted R

0.086 0.101 0.129 0.133

Note:

∗

p<0.1;

∗∗

p<0.05;

∗∗∗

p<0.01

Electronic copy available at: https://ssrn.com/abstract=4552425

Table 6: Fintech borrower shopping

Note: This table examines the shopping behavior of ﬁntech borrowers. Panel A examines borrower search behavior by

regressing the number of mortgage applications in a given application spell on whether the spell includes a ﬁntech search or

origination. Applications counts the number of searches. Fintech application and Fintech lender are indicators for whether

there is a ﬁntech application or a ﬁntech origination in the spell, respectively. Panel B examines the likelihood that borrowers

accepts a loan oﬀer made by a ﬁntech lender as compared to all other lenders. In Columns (1) and (2), the sample is all

approved reﬁnance applications in HMDA between 2018 and 2021. A loan is approved if the loan was eventually originated

or it was withdrawn by the borrower. The left-hand side variable is a zero-one indicator for whether the approved loan

is originated. In Columns (3) and (4), the sample is all reﬁnancing applications. This includes loans that were eventually

originated, withdrawn by the borrowers, or any applications that were unsuccessful for any reason including incompleteness.

The left-hand side variable is an indicator equal to one if the borrower either withdrew an application before it was complete

or did not proceed with the loan after the oﬀer was approved. Standard errors are clustered at the lender-county level. Colums

in both panels include, as indicated, lender type ﬁxed eﬀects, borrower controls, continuous borrower controls, and market

ﬁxed eﬀects. Lender type is an indicator variable for bank vs. non-bank lenders. Borrower controls are interacted ﬁxed

eﬀects of debt-to-income ratios, combined loan-to-value ratios, applicant income, and age. Continuous origination controls

are debt-to-income, CLTV, original loan size, and original loan term. Market ﬁxed eﬀects are three-digit zip × year of the

HMDA record.

Panel A: Fintech borrower search behavior

Dependent variable:

Applications

(1) (2)

Fintech application −0.004

∗∗∗

(0.001)

Fintech lender −0.004

∗∗∗

(0.001)

E[Y] 1.05 1.05

Lender Type FE X X

Borrower Controls X X

Market FE X X

Observations 1,010,492 1,010,492

Adjusted R

0.008 0.008

Note:

∗

p<0.1;

∗∗

p<0.05;

∗∗∗

p<0.01

Panel B: Fintech oﬀer acceptance rates

Dependent variable:

Originated Borrower withdrawn

(1) (2) (3) (4)

Fintech 0.023

∗∗∗

0.027

∗∗∗

−0.074

∗∗∗

−0.102

∗∗∗

(0.002) (0.002) (0.002) (0.003)

E[Y] 0.95 0.95 0.16 0.16

Lender Type FE N Y N Y

Borrower Controls N Y N Y

Market FE N Y N Y

Observations 94,515 94,515 134,483 134,483

Adjusted R

0.002 0.257 0.007 0.446

Note:

∗

p<0.1;

∗∗

p<0.05;

∗∗∗

p<0.01

Electronic copy available at: https://ssrn.com/abstract=4552425

Table 7: Model calibration

Note: This table shows the moments used in the calibration (Panel A) and the calibrated parameters values (Panel B). Panel

A shows the moment description, the value of the moment in the data, the value of the moment in the calibration, and their

diﬀerence. Note that the calibration is signiﬁcantly overidentiﬁed, using seven moments to calibrate four parameters. Panel

B shows the parameter values, a brief description, the calibrated value, and the source for the calibration.

Panel A: Targeted Moments

Moment Data Model Diﬀerence

Reﬁnance probability 0.203 0.225 0.023

Reﬁnance fee 2.291 2.284 -0.007

Non-ﬁntech recapture ∆ price -0.150 -0.204 -0.054

Fintech recapture ∆ price 0.300 0.212 -0.087

Recapture share of ﬁntech 0.310 0.340 0.030

Recapture share of non-ﬁntech 0.140 0.065 -0.074

Fintech share 0.120 0.266 0.146

Panel B: Calibrated Parameters

Parameter Description Value Source

ρ Rate AR coeﬃcient 0.850 30-year mortgage rate time series

Long-term rate mean 0.060 30-year mortgage rate time series

Rate shock volatility 0.006 30-year mortgage rate time series

M Initial loan principal ($k) 250 Median mortgage size

T Loan length (years) 30 30-year FRM

s Servicing fee (annual % of principal) 0.0050 GSE policy

β Subjective discount factor 0.95 Standard assumption

Outsider marginal cost 1.775 SMM

Insider marginal cost 1.539 SMM

c Inertia cost ($k) 11.337 SMM

∆c

Eﬀective ﬁntech inertia advantage ($k) 1.588 SMM

Electronic copy available at: https://ssrn.com/abstract=4552425

References

Acharya, Viral V, Matthew Richardson, Stijn Van Nieuwerburgh, and

Lawrence J White. 2011. “Guaranteed to fail: Fannie Mae, Freddie Mac, and the

debacle of mortgage ﬁnance.” Princeton University Press.

Agarwal, Sumit, John Grigsby, Ali Hortascu, Gregor Matvos, Amit Seru, and

Vincent Yao. 2020. “Searching for Approval.” NBER working paper, 27341.

Agarwal, Sumit, Richard J Rosen, and Vincent Yao. 2016. “Why do borrowers make

mortgage reﬁnancing mistakes?” Management Science, 62(12): 3494–3509.

Allen, Jason, Robert Clark, and Jean-Fran¸cois Houde. 2019. “Search friction

and market power in negotiated-price markets.” Journal of Political Economy, 127(Au-

gust): 1550–1598.

Andersen, Steﬀen, John Y Campbell, kasper Meisner Nielsen, and Tarun Ra-

madorai. 2020. “Sources of Inaction in Household Finance: Evidence from the Danish

Mortgage Market.” American Economic Review, 110(10): 3184–3230.

Babina, Tania, Greg Buchak, and Will Gornall. 2022. “Customer data access and

ﬁntech entry: Early evidence from open banking.”

Berg, Tobias, Andreas Fuster, and Manju Puri. 2022. “FinTech Lending.” Annual

Review of Financial Economics, 14: 187–207.

Bhutta, Neil, Andreas Fuster, and Aurel Hizmo. 2020. “Paying too much? Price

dispersion in the US mortgage market.”

Buchak, Greg, and Adam Jørring. 2021. “Do Mortgage Lenders Compete Locally?

Evidence Beyond Interest Rates.” Working Paper.

Buchak, Greg, Gregor Matvos, Tomasz Piskorski, and Amit Seru. 2018a. “Fintech,

regulatory arbitrage, and the rise of shadow banks.” Journal of Financial Economics,

130(3): 453–483.

Buchak, Greg, Gregor Matvos, Tomasz Piskorski, and Amit Seru. 2018b. “The

limits of shadow banks.” National Bureau of Economic Research.

Buchak, Greg, Gregor Matvos, Tomasz Piskorski, and Amit Seru. 2020. “Why is

intermediating houses so diﬃcult? evidence from ibuyers.” National Bureau of Economic

Research.

Electronic copy available at: https://ssrn.com/abstract=4552425

Buchak, Greg, Gregor Matvos, Tomasz Piskorski, and Amit Seru. 2023. “MacroPru

in the Shadow of Declining Balance Sheet Models of Banking.” NBER Chapters.

Di Maggio, Marco, Amir Kermani, Benjamin J Keys, Tomasz Piskorski, Rod-

ney Ramcharan, Amit Seru, and Vincent Yao. 2017. “Interest Rate pass-through:

Mortgage rates, household consumption, and voluntary deleveraging.” American Economic

Review, 107(11): 3550–88.

Di Maggio, Marco, and Vincent Yao. 2019. “Fintech borrowers: Lax-screening or

Cream-Skinning?” Review of Financial Studies, 34(10): 4546–4618.

Drechsler, Itamar, Alexi Savov, and Philipp Schnabl. 2017. “The deposits channel

of monetary policy.” Quarterly Journal of Economics, 132: 1819–1876.

Egan, Mark. 2019. “Brokers versus Retail Investors: Conﬂicting Interests and Dominated

Products.” The Journal of Finance, 74(3): 1217–1260.

Egan, Mark, Horta¸csu, and Gregor Matvos. 2008. “Deposit competition and ﬁnancial

fragility: evidence from the US banking sector.” American Economic Review, 107: 169–

216.

Egan, Mark, Stefan Lewellen, and Adi Sunderam. 2022. “The cross section of bank

value.” The Review of ﬁnancial studies, 35(May): 2101–2142.

Elenev, Vadim, Tim Landvoigt, and Stijn Van Nieuwerburgh. 2016. “Phasing otu

the GSEs.” Journal of Monetary Economics, 81: 111–132.

Fuster, Andreas, Matthew C Plosser, Philip Schnabl, and James I Vickery. 2022.

“The role of technology in mortgage lending.” The Review of Financial Studies, 32: 1854–

99.

Fuster, Andreas, Stephanie Lo, and Paul Willen. 2017. “The Time-Varying Price of

Financial Intermediation in the Mortgage Market.” National Bureau of Economic Research

w23706, Cambridge, MA.

Gopal, Manasa, and Philipp Schnabl. 2020. “The Rise of Finance Companies and

FinTech Lenders in Small Business Lending.” Available at SSRN.

Gurun, Umit G., Gregor Matvos, and Amit Seru. 2016. “Advertising Expensive

Mortgages.” The Journal of Finance, 71(5): 2371–2416.

Electronic copy available at: https://ssrn.com/abstract=4552425

Jiang, Sheila, Jose Ignacio Cuesta, Adam Jørring, and Douglas Xu. 2023. “Bank

Technology Adoption and Loan Production in the U.S Mortgage Market.” Working Paper.

Jørring, Adam. 2020. “Financial Sophistication and Consumer Spending.” Working Paper.

Keys, Benjamin J, Devin G Pope, and Jaren C Pope. 2016. “Failure to Reﬁnance.”

Journal of Financial Economics, 122: 482–499.

Keys, Benjamin J, Tanmoy Mukherjee, Amit Seru, and Vikrant Vig. 2010. “Did

securitization lead to lax screening? Evidence from subprime loans.” Quarterly Journal of

Economics, 125(Feb): 307–362.

Matteo, Benetton. 2020. “Leverage regulation and market structure: A structural model

of the UK mortgage market.” The Journal of ﬁnance, 76(Dec): 2997–3053.

Mayock, Tom, and Lan Shi. 2022. “Adverse selection in the market for mortgage servic-

ing.” Journal of Housing Economics, 58(101858).

Mian, Atif, and Amir Suﬁ. 2009. “The consequences of mortgage credit expan-

sion: Evidence from the U.S. mortgage default crisis.” Quarterly Journal of Economics,

124(Nov): 1449–96.

Purnanandam, Amiyatosh. 2011. “Originate-to-distribute model and the subprime mort-

gage crisis.” The Review of Financial Studies, 24: 1881–1915.

Scharfstein, David, and Adi Sunderam. 2016. “Market power in mortgage lending and

the transmission of monetary policy.” Unpublished working paper., Harvard University.

Xiao, Kairong. 2020. “Monetary Transmission through Shadow Banks.” The Review of

Financial Studies, 33(6): 2379–2420.

Yannelis, Constantine, and Anthony Lee Zhang. 2021. “Competition and selection in

credit markets.” National Bureau of Economic Research.

Electronic copy available at: https://ssrn.com/abstract=4552425

A Appendix

A.1 Low rates of mortgage reﬁnancing

We adopt a simple strategy to highlight the low level of reﬁnancing. We begin following a

cohort of loans the ﬁrst time that it is suﬃciently “deep in the money” for reﬁnancing: when

its current rate r

old

is more than 50 basis points above the prevailing new origination rate

prevailing,t

. For example, the January 2019 cohort consists of all loans whose rate exceeds

the prevailing rate by 50 basis points for the ﬁrst time on January 2019. We then ask what

fraction of loans in that cohort reﬁnance within three years. We show these reﬁnancing rates

in Figure A4 Panel A. This ﬁgure shows that for the majority of the sample, the reﬁnancing

share is below 10%. That is, even for mortgages that are deeply in-the-money to reﬁnance,

only a small fraction actually reﬁnance within a reasonable time frame. Panel B of that same

ﬁgure shows a diﬀerent slightly diﬀerent version of this analysis. For each month, the sample

consists of all loans whose interest rates are 50 basis points higher than the prevailing. The

y-axis then shows what proportion of those loans actually reﬁnance in that month. If they

don’t choose to exit, they reappear in the following month as long as their interest rate is

still at least 50 basis points above the prevailing. The highest reﬁnancing share achieved in

this measure is 30%.

A.2 Fintech Reﬁnance Share

While the discussion above indicated low reﬁnance rates on average, we supplement this

analysis with results on ﬁntech lenders. Table A5 shows results of a cox proportional hazard

model of exits through reﬁnancing on a zero-one ﬁntech indicators. Age is deﬁned as months

since origination and a death even is deﬁned as exit through reﬁnancing. Panel A shows the

results of the model with borrower controls. The coeﬃcient of 0.064 implies exp(0.064) = 1.06

indicating that borrowers with ﬁntech lenders are 6% more likely to reﬁnance. The results

in panel A includes borrower controls which are not available in the kaplan-meier survival

probability curves shown in panel B. On average, ﬁntech borrowers diverge in their survival

likelihood at around 30 months.

A.3 Fintech and borrower delinquency

Here, we rule out an alternate force that might lead ﬁntech lenders to behave diﬀerently:

diﬀerent underwriting ability. The idea, popular in the literature, is that ﬁntechs possess a

technological advantage in underwriting and are thus better able to predict borrower default.

We note ﬁrst, for institutional reasons, that lenders originating GSE mortgages have little

Electronic copy available at: https://ssrn.com/abstract=4552425

incentive to invest in more predictive underwriting technology for the simple fact that the

GSEs guarantee risk on the loans and that the price of these guarantees is set by a simple

and pre-determined FICO × LTV grid.

Nevertheless, there are some marginal reasons that originators may have incentives to

detect borrower risk beyond the GSEs’ pricing grid. First, there is a short window after

origination but before loans are sold to GSEs where the originator still owns the loan and

is consequently exposed to borrower credit risk. Second, there is the risk that loans are put

back from the GSEs, which occurs if the borrower defaults very early in the life of the loan

(in which case default is taken as per se evidence that the underwriting was fraudulent), or if

as a result of default later on, the GSEs allege and prove fraud in the underwriting process.

Therefore, we provide evidence around how borrower performance varies with whether

the originator was a ﬁntech lender. We run the following loan-level regression:

CreditEvent

= βF intech

+ F E + 

(24)

(25)

Here, CreditEvent

is an indicator for a given credit event. We test (1) whether the

loan becomes 30-days delinquent within the ﬁrst three years of origination, (2) whether the

loan becomes 60-days delinquent within the ﬁrst three years of origination, (3) whether the

loan becomes 90-days delinquent within the ﬁrst three years of origination, (4) whether the

loan is put back to the originator within the ﬁrst three years of origination, and (5) whether

there is a foreclosure on the borrower within the ﬁrst three years of origination. F intech

a zero-one indicator for whether the lender is a ﬁntech lender. F E is a collection of lender

type, binned borrower characteristics, and market ﬁxed eﬀects. The results are shown in

Table A6.

As shown in Columns (1)–(3), we ﬁnd no statistically signiﬁcant relationship between

ﬁntech and 30-, 60-, and 90-day delinquency. In columns (4) and (5), we ﬁnd a small

negative relationship but the results are again not statistically signiﬁcant. Putbacks shown

in column (4) are of potential consequence to orignators and servicers because it imposes

a cost on the originator and terminates the servicing contract. Here, we ﬁnd that ﬁntech

technology makes no diﬀerence for this economically relevant outcome for both servicer

and originator. Foreclosures are also potentially important to servicers as it terminates the

servicing contract. Together, the results suggest that ﬁntech lenders do not possess any

underwriting advantage in originating loans.

Electronic copy available at: https://ssrn.com/abstract=4552425

A.4 Search Intensity Match Details

This section details the search intensity match used in table 6. We start by splitting each

HMDA yearly ﬁle into “denied” and “executed” datasets where executed denotes loans that

were originated. We limit the sample to applications and loans for reﬁnancing only. In each

dataset, we create a matchkey consisting of combinations of applicant characteristics. To

identify borrowers, we use a set of characteristics that are immutable between application

attempts. These characteristics include,

• loan type: Indicator for FHA or conventional loan

• derived dwelling category: single family or multifamily home type

• business or commercial purpose

• occupancy type: owner occupied or investor

• loan amount: Amount of the loan (Note that these are rounded in the data. It’s

possible that borrowers re-apply for a diﬀerenet loan amount but because the value

is already an approximation, we use it as is. We assume that it is unlikely for loan

amounts to change signiﬁcantly between applications. For reﬁnancings where the loan

needs to replace an existing balance, this is even less likely to change.)

• census tract

• county code

• applicant ethnicity

• co applicant ethnicity (note that if there is no co-applicant, this received a speciﬁc

encoding)

• applicant race 1

• co applicant race 1

• applicant sex

• co applicant sex

• income

• lien status: First lien or subordinate lien

Electronic copy available at: https://ssrn.com/abstract=4552425

For observations in the denied set, these characteristics deﬁne a chained set. For observa-

tions in the executed set, any observations which are not unique along these characteristics

are removed. These keys are then used to conduct a one-to-one match between denials and

executed observations. By deﬁnition, this method creates a matched set of loans with at

least one documented failure. However, there are a large number of unmatched applications

remaining. For our analysis, we treat these loans as loans which were originated on the ﬁrst

try or loans which were never successfully executed. For the ﬁnal dataset, we take a 5%

random sample of the full data including unmatched orinated loans, unmatched denials, and

matched chains.

To facilitate analyses, we deﬁne several variables which classify the type of search expe-

rience. Applications denotes the total number of applications associated with a chain. This

includes the originated loan and any matched denials. For unmatched denials, this includes

the total number of applications and for unmatched originated loans, the applications num-

ber is one. This is our main dependent variable of interest. On the right side, we deﬁne a

Fintech indicator for any application with a Fintech lender. This includes applications with

Fintech lenders which were not executed. The second variable of interest is an indicator for

Fintech Executed. By contrast to the Fintech indicator, this variable is only set to one for

originated Fintech loans, all Fintech denials are set to zero.

For the analyses in the main paper, we use the full dataset including unmatched loans.

However, one might be concerned that some loans are unmatched due to the algorithm

implemented. In appendix table A8, we include a column that restricts the sample to the

matched set. These are loans with at least one failed attempt before being executed. This

is a smaller, more restrictive sample but it does not suﬀer from the potential concern with

the unmatched loans.

Note that the randomization is conducted at the chain level, not the individual HMDA observation

level.

Electronic copy available at: https://ssrn.com/abstract=4552425

Table A1: Within lender-month integration shares

Note: The ﬁgure plots lender-month integration shares. The sample is a 1% random draw from the FNMA and FHLMC

origination-level dataset. For each lender-month, the plot shows the proportion of all of their originated loans for which they

are also the servicer.

Electronic copy available at: https://ssrn.com/abstract=4552425

Table A2: Does shadow bank share drive integration share?

Note: This table shows descriptively what explains the aggregate integration share. Panel A shows the share of traditional

bank originations that are integrated (blue), the share of shadow bank originations that are integrated (red), and the shadow

bank market share (black, dashed line). Panel B shows a regression of aggregate integration share on shadow bank origination

share (Column (1)), bank integration share and shadow bank integration share (Column (2)), and all three (Column (3)).

Panel A: Lender type integration share and shadow bank origination share

Panel B: Explanatory power regression

Dependent variable:

Aggregate integration share

(1) (2) (3)

Shadow bank origination share −0.265

∗∗∗

−0.288

∗∗∗

(0.028) (0.007)

Bank integration share 0.986

∗∗∗

0.774

∗∗∗

(0.042) (0.015)

Shadow bank integration share 0.053

∗∗

0.243

∗∗∗

(0.022) (0.009)

Observations 216 216 216

0.292 0.736 0.971

Adjusted R

0.288 0.734 0.970

Note:

∗

p<0.1;

∗∗

p<0.05;

∗∗∗

p<0.01

Electronic copy available at: https://ssrn.com/abstract=4552425

Table A3: Recaptured rate decomposition

Note: This table examines existing and new interest rates for recaptured loan reﬁnances. The sample is the forward-linked

set of reﬁnancings. We distinguish lenders based on “Fintech recapture” (ﬁntech loans which are recaptured), “Fintech new

borrower” (ﬁntech reﬁnancings where the previous lender was diﬀerent), and “Non-ﬁntech recapture” (recaptured loans where

the lender was not a ﬁntech lender). In Column (1), the left-hand side variable is the rate of the loan that is reﬁnanced. In

Column (2), the left-hand side variable is the new rate at which the loan is reﬁnanced. E[Y ] is the mean of the left-hand

side variable. All columns include lender type (bank versus shadow bank) ﬁxed eﬀects, new loan characteristic ﬁxed eﬀects,

borrower controls, and market (3-digit zip code × year) ﬁxed eﬀects. The new loan ﬁxed eﬀects are interactions of HMDA

variables: debt-to-income ratio bins, combined loan-to-value bins, bins of borrower income, and applicant age. Continuous

origination controls are debt-to-income, CLTV, original loan size, and original loan term. Standard errors, clustered at the

lender-county level, are shown in parentheses.

Dependent variable:

Old rate New rate

(1) (2)

Fintech recapture −0.168

∗∗∗

−0.012

(0.018) (0.012)

Fintech new borrower 0.020

∗∗∗

−0.042

∗∗∗

(0.006) (0.004)

Non-ﬁntech recapture 0.058

∗∗∗

0.049

∗∗∗

(0.012) (0.015)

E[Y] 4.16 3.01

Lender Type FE Y Y

Borrower New Loan FE Y Y

Borrower Controls Y Y

Market FE Y Y

Observations 78,373 78,373

Adjusted R

0.429 0.599

Note:

∗

p<0.1;

∗∗

p<0.05;

∗∗∗

p<0.01

Electronic copy available at: https://ssrn.com/abstract=4552425

Table A4: Refinance rates

Note: Panel A depicts cohort-level reﬁnancing rates over time. This is a 1% sample taken from the monthly loan-level

performance dataset provided by FNMA and FMHLC. Each month, we take the diﬀerence between a loan’s current interest

rate and the prevailing mortgage rate (we use the average of all mortgages originated in the economy that month). Loans

who are within three years of their potential rate improvement exceeding 50 basis points (r

old

− r

prevailing

> 0.5) is included

in the sample. Each cohort depicted on the x-axis consists of loans which ﬁrst became in-the-money that quarter. Then, the

y-axis plots the share which has reﬁnanced within 3 years of that date. Panel B shows a diﬀerent version of these reﬁnancing

rates. Rather than cohorts, the x-axis shows all loans still active in that month for which the in-the-money criterion holds.

These are loans that would beneﬁt from reﬁnancing. The y-axis shows the proportion of loans which do reﬁnance that month.

Loans that do not reﬁnance reappear the following month if they are still in-the-money.

Panel A: Reﬁnance propensity (by cohorts)

Panel B: Reﬁnance Propensity (monthly)

Electronic copy available at: https://ssrn.com/abstract=4552425

Table A5: Proportional Hazard Results

Note: Panel A below shows results from a cox proportional hazard analysis of ﬁntech lenders. Age is deﬁned as months since

the loan was originated. A “death” is deﬁned as exit through reﬁnancing while “Fintech” is a zero-one indicator for whether

or not the loan is originated by a ﬁntech lender. Panel B shows kaplan-meier plots comparing survival probabilities for ﬁntech

lenders with non-ﬁntech ledners. We include the results in panel A to show the survival probability of an exit event with the

inclusion of borrower controls. The coeﬃcient on Fintech is 0.064. Taking the exponential, exp(0.064) = 1.06 indicating that

borrowers with ﬁntech lenders are 6% more likely to reﬁnance. The controls include lender type (bank versus shadow bank)

ﬁxed eﬀects, market ﬁxed eﬀects and continuous borrower controls, debt-to-income, CLTV, original loan size, and original

loan term. The sample is limited to loans for whom r

old

− r

prevailing

> 0.5 and loans that are younger than 45 months.

Panel A: Proportional Hazard Results

Dependent variable:

age

Fintech 0.664

∗∗∗

(0.025)

Observations 596,821

0.014

Max. Possible R

0.553

Log Likelihood −236,049.300

Wald Test 11,143.830

∗∗∗

(df = 7)

LR Test 8,521.517

∗∗∗

(df = 7)

Score (Logrank) Test 10,830.420

∗∗∗

(df = 7)

Note:

∗

p<0.1;

∗∗

p<0.05;

∗∗∗

p<0.01

Panel B: Reﬁnance hazard rates

Electronic copy available at: https://ssrn.com/abstract=4552425

Table A6: Delinquency and default Rates

Note: This table shows delinquency rates for integrated and ﬁntech lenders. We examine a 1% random draw of GSE loans

originated between 2013 and 2021. The sample is limited to 2013 and onwards because this is when Fintech sellers begin

to gain market share. The left-hand side variables are (1) whether the loan becomes 30-days delinquent within the ﬁrst

three years of origination, (2) whether the loan becomes 60-days delinquent within the ﬁrst three years of origination, (3)

whether the loan becomes 90-days delinquent within the ﬁrst three years of origination, (4) whether the loan is put back to

the originator within the ﬁrst three years of origination, and (5) whether there is a foreclosure on the borrower within the

ﬁrst three years of origination. F intech

is a zero-one indicator for whether the lender is a ﬁntech lender. Each regression

includes lender type (bank vs. non-bank) ﬁxed eﬀects, continuous borrower controls, and market F E is a collection of lender

type, binned borrower characteristics, and market ﬁxed eﬀects. Standard errors, clustered at the zip-code level are shown in

parentheses.

Dependent variable:

30 DPD 60 DPD 90 DPD putback foreclose

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

Fintech 0.0002 0.004 0.003 −0.00003 −0.00001

(0.004) (0.003) (0.003) (0.0001) (0.00002)

E[Y] 0.013 0.007 0.005 0.000 0.000

Lender Type FE Y Y Y Y Y

Borrower Controls Y Y Y Y Y

Market FE Y Y Y Y Y

Observations 11,029,224 11,029,224 11,029,224 4,579,875 4,579,875

0.038 0.035 0.030 0.028 0.050

Adjusted R

0.029 0.026 0.021 0.006 0.028

Note:

∗

p<0.1;

∗∗

p<0.05;

∗∗∗

p<0.01

Electronic copy available at: https://ssrn.com/abstract=4552425

Table A7: Fintech Recapture over Targets

Note: The table shows results from regressions of recaptured status on a ﬁntech lender indicator. The sample consists of

all loans in the forward-linked reﬁnancing dataset. The dependent variable is one if the reﬁnanced loan was recaptured and

zero otherwise. The columns show results for diﬀerent subsets of borrowers based on their potential rate improvement at

the time of the reﬁnancing. The table indicates whether or not the results include lender type (bank versus shadow bank)

ﬁxed eﬀects, market ﬁxed eﬀects, ﬁxed eﬀects associated with loan charactersitics of the new reﬁnancing loan, and continous

borrower controls for the loan at origination of the original loan. Note that the coeﬃcients of interest are robust to which

set (origination or reﬁnancing) borrower controls are used. The new loan ﬁxed eﬀects are interactions of HMDA variables:

debt-to-income ratio bins, combined loan-to-value bins, bins of borrower income, and applicant age. Continuous origination

controls are debt-to-income, CLTV, original loan size, and original loan term. Finally, market ﬁxed eﬀects are 3-digit zip

code × year. Standard errors, clustered at the lender-county level, are shown in parentheses.

Dependent variable:

Recaptured reﬁnance

(1) 0.5 − 0.75 (2) 0.75 − 1 (3) 1 − 2 (4) > 2

Fintech 0.138

∗∗∗

0.156

∗∗∗

0.089

∗∗∗

0.017

(0.024) (0.026) (0.009) (0.031)

E[Y] 0.16 0.15 0.13 0.10

Lender Type FE Y Y Y Y

Borrower New Loan FE Y Y Y Y

Borrower Controls Y Y Y Y

Market FE Y Y Y Y

Observations 11,443 11,830 30,230 6,229

Adjusted R

−0.035 0.057 0.133 0.030

Note:

∗

p<0.1;

∗∗

p<0.05;

∗∗∗

p<0.01

Electronic copy available at: https://ssrn.com/abstract=4552425

Table A8: Fintech borrower shopping

Note: This table examines the shopping behavior of ﬁntech borrowers. Panel A examines borrower search behavior by

regressing the number of mortgage applications in a given application spell on whether the spell includes a ﬁntech search or

origination. Applications counts the number of searches. Fintech application and Fintech lender are indicators for whether

there is a ﬁntech application or a ﬁntech origination in the spell, respectively. Columns (1)-(2) include all loan spells, Columns

(3) restrict the sample to those with an origination. Colums in both panels include, as indicated, lender type ﬁxed eﬀects,

borrower controls, continuous borrower controls, and market ﬁxed eﬀects. Lender type is an indicator variable for bank

vs. non-bank lenders. Borrower controls are interacted ﬁxed eﬀects of debt-to-income ratios, combined loan-to-value ratios,

applicant income, and age. Continuous origination controls are debt-to-income, CLTV, original loan size, and original loan

term. Market ﬁxed eﬀects are three-digit zip × year of the HMDA record.

Panel A: Fintech borrower search behavior

Dependent variable:

Applications

(1) (2)

Fintech application −0.004

∗∗∗

(0.001)

Fintech lender −0.004

∗∗∗

(0.001)

E[Y] 1.05 1.05

Lender Type FE X X

Borrower Controls X X

Market FE X X

Observations 1,010,492 1,010,492

Adjusted R

0.008 0.008

Note:

∗

p<0.1;

∗∗

p<0.05;

∗∗∗

p<0.01

Electronic copy available at: https://ssrn.com/abstract=4552425

c/o University of Geneva, Bd. Du Pont d'Arve 42, CH-1211 Geneva 4

T +41 22 379 84 71, rps@sﬁ.ch, www.sﬁ.ch

Swiss Finance Institute

Swiss Finance Institute (SFI) is the national center for fundamental

research, doctoral training, knowledge exchange, and continuing

education in the ﬁelds of banking and ﬁnance. SFI’s mission is to

grow knowledge capital for the Swiss ﬁnancial marketplace. Created

in 2006 as a public–private partnership, SFI is a common initiative

of the Swiss ﬁnance industry, leading Swiss universities, and the

Swiss Confederation.