Finance and Economics Discussion Series
Divisions of Research & Statistics and Monetary Affairs
Federal Reserve Board, Washington, D.C.
Paying Too Much? Price Dispersion in the US Mortgage Market
Neil Bhutta, Andreas Fuster, Aurel Hizmo
2020-062
Please cite this paper as:
Bhutta, Neil, Andreas Fuster, and Aurel Hizmo (2020). “Paying Too Much? Price
Dispersion in the US Mortgage Market,” Finance and Economics Discussion Se-
ries 2020-062. Washington: Board of Governors of the Federal Reserve System,
https://doi.org/10.17016/FEDS.2020.062.
NOTE: Staff working papers in the Finance and Economics Discussion Series (FEDS) are preliminary
materials circulated to stimulate discussion and critical comment. The analysis and conclusions set forth
are those of the authors and do not indicate concurrence by other members of the research staff or the
Board of Governors. References in publications to the Finance and Economics Discussion Series (other than
acknowledgement) should be cleared with the author(s) to protect the tentative character of these papers.
Paying Too Much? Price Dispersion in the US Mortgage Market
Neil Bhutta Andreas Fuster Aurel Hizmo
June 30, 2020
Abstract
We document wide dispersion in the mortgage rates that households pay on identical loans,
and show that borrowers’ financial sophistication is an important determinant of the rates ob-
tained. We estimate a gap between the 10th and 90th percentile mortgage rate that borrowers
with the same characteristics obtain for identical loans, in the same market, on the same day,
of 54 basis points—equivalent to about $6,500 in upfront costs (points) for the average loan.
Time-invariant lender attributes explain little of this rate dispersion, and considerable disper-
sion remains even within loan officer. Comparing the rates borrowers obtain to the real-time
distribution of rates that lenders could offer for the same loan and borrower type, we find that
borrowers who are likely to be the least financially savvy tend to substantially overpay relative
to the rates available in the market. In the time series, the average overpayment decreases when
overall market interest rates rise, suggesting that a rising level of borrowing costs encourages
more search and negotiation. Furthermore, new survey data provide direct evidence that finan-
cial knowledge and shopping both affect the mortgage rates borrowers get, and that shopping
activity increases with the level of market rates.
We thank Jason Allen, Robert Avery, John Campbell, Nick Embrey, Serafin Grundl, Katherine Guthrie, Haj
Hadeishe, Michael Haliassos, Chris Hansman, Gregor Matvos, Raven Molloy, John Mondragon, Christopher Palmer,
Saty Patrabansh, Chad Redmer, James Rowe, David Zhang, as well as seminar and conference participants at the
American University (Kodog), Arizona State University, Baruch College (Zicklin), EPFL, Federal Reserve Bank
of Atlanta, Federal Reserve Board, Freddie Mac, Norges Bank, NYU Stern, Oxford Sa
¨
ıd, AFA Annual Meeting,
AREUEA National Conference, CEPR European Conference on Household Finance (Ortygia), Cherry Blossom Fi-
nancial Education Institute, FCA-Imperial Household Finance Conference, FDIC Consumer Research Conference,
and the NBER Summer Institute (Real Estate) for helpful comments. The views expressed are those of the authors
and do not necessarily reflect those of the Federal Reserve Board, the Federal Reserve System, or the Swiss National
Bank.
Bhutta and Hizmo are at the Federal Reserve Board; Fuster is at the Swiss National Bank. Emails:
1 Introduction
Recent survey data indicate that half of the borrowers taking out a mortgage in the US in 2016
only seriously considered one lender, and just three percent of the borrowers considered more than
three lenders.
1
Ninety-six percent of the respondents reported that they were satisfied that they
received the lowest interest rate for which they could qualify. Taking these facts at face value, one
might be led to conclude that either there is not much price dispersion in the mortgage market,
or that borrowers are very efficient at searching and finding the most competitive lenders. This
might seem a reasonable conclusion especially when considering that the mortgage market appears
highly competitive: the majority of mortgages in the US are very standardized and guaranteed by
the government (through the GSEs and FHA/VA), and in our data there are over one hundred
different lenders offering mortgages in a local market in any given day. However, in contrast to
borrowers’ perceptions, we document a striking amount of variation in the prices consumers pay
for mortgages, and that overpayment varies systematically across borrower types and over time.
Dispersion in consumer interest rates, particularly mortgage rates, has substantial implications
for households’ financial well-being. Policies to improve market functioning require an understand-
ing of why dispersion occurs in equilibrium. A standard explanation for price dispersion is the
existence of search costs due to the opportunity cost of time. But other factors such as financial
literacy and a willingness/ability to negotiate could also play important roles. In this paper, we ar-
gue that the standard search cost explanation for price dispersion does not fit many of the patterns
we see in the data. We provide several pieces of evidence suggesting that financial acumen is an
important determinant of the rates that people get, and that shopping activity is not determined
purely by rational factors—as hinted at by the survey data above.
Identifying dispersion in mortgage rates is challenging. Even in a perfectly competitive, friction-
less market, we would expect to see considerable variation across consumers in their interest rates
arising from several factors such as differences in credit risk, day-to-day fluctuations in market rates,
and heterogeneity in risk and time preferences that can affect borrowers’ choices of various contract
terms. To address this challenge, we draw on a unique source of data—an industry platform used
by lenders to price mortgages, initiate rate locks, manage pipeline risk, and sell mortgages to in-
vestors. The platform provides data on both available rates—the rates that lenders could offer for
specific mortgages in each market and each day—and data on the mortgages locked, or obtained,
by consumers. The data on locked mortgages include key variables for evaluating mortgage pricing,
including several that are unavailable in any other dataset, such as “discount points”, exact time of
rate lock (as opposed to the closing date), and the lock period (e.g. 30 or 60 days).
Turning first to the data on mortgage interest rate locks, we document wide interest rate
dispersion. We find that the difference between the 90th and 10th percentile interest rate that
identical borrowers lock in for the same (30-year fixed-rate fully-documented) loan in the same
market, on the same day, and paying the same points, is 54 basis points (bp). Using the average
1
Statistics are from the National Survey of Mortgage Originations, which is conducted jointly by the Federal
Housing Finance Agency (FHFA) and the Consumer Financial Protection Bureau (CFPB).
1
point-rate trade-off in our data, a 54bp rate difference is equivalent to an upfront payment of 2.6
points, or about $6,500 for an average loan of $250,000. Notably, large dispersion exists even for
federally-insured loans where lenders take little to no credit risk, and for loans that meet the credit
standards for purchase by Fannie Mae and Freddie Mac. As such, it is unlikely that unobserved
risk factors could explain this dispersion.
Allowing for lender-specific time-varying pricing and branch-by-month fixed effects cuts the
residual dispersion by almost one-half. In other words, different lenders, as well as branches within
lenders, set different prices, and the cheapest lenders and branches change over time. Still, sig-
nificant dispersion remains within branch and even within loan officer, with the largest residual
dispersion among borrowers who are likely to be the least financially sophisticated (e.g. low credit
score and inexperienced home buyers). Overall, getting a low rate is not simply about “going to the
right lender.” Instead, it appears that in order to get a low rate, borrowers must be knowledgeable
and able to negotiate no matter which lender they choose.
2
Next, we draw on the real-time distribution of available interest rates that lenders could offer
to borrowers. These “offer” rates are specific to loan and borrower characteristics (loan-to-value
ratio [LTV], credit score [FICO], loan amount, points, etc.) in a given market on a given day.
Importantly, while these rates are available to borrowers (since otherwise we would not observe them
on the platform), they are not necessarily advertised by lenders or easily observed by borrowers.
3
For a given borrower, we compute the difference between the rate they locked and the median offer
rate available for the same type of loan and borrower on the same day in the same market. We
find that this “locked-offer rate gap” is positive on average, meaning that borrowers tend to get
mortgage rates that are somewhat higher than the median available rate for an identical mortgage.
4
More importantly, the locked-offer rate gap varies substantially across borrower types. For
example, “jumbo” borrowers, who tend to have relatively high incomes, on average obtain a rate
that is 21bp below the median of their offer distribution, suggesting that such borrowers are able
to find relatively good deals. In contrast, FHA borrowers, who tend to have lower income, wealth,
and credit scores, on average pay 25bp more than what the median lender could offer for their loan.
Remarkably, one-quarter of FHA borrowers pay in excess of 45bp more than the median available
rate for their exact same loan.
5
2
Notably, including lender, branch and loan officer fixed effects arguably accounts for differences in service quality
(and aspects such as convenience of office location) that might have helped explain residual price dispersion. Also,
the lending platform providing our data is mostly used by monoline nonbank mortgage originators, and our results
are unchanged when we limit the sample exclusively to such lenders. Thus we can rule out explanations related to
cross-selling or bundling of other services (Horta¸csu and Syverson, 2004) as factors that could explain observed price
dispersion in this market.
3
See Duncan (2019) for a discussion of what makes comparison shopping in the mortgage market more complicated
and time-consuming than shopping for ordinary goods.
4
Because the most popular lenders may be relatively expensive, and because there is substantial within-lender
dispersion in rates as noted above, it is not necessarily surprising that the locked-offer gap would be positive. As
we document in the Appendix, our results are qualitatively identical when we consider the expected gains from one
extra search as an alternative to the locked-offer rate gaps.
5
Looking across locations, locked-offer rate gaps are highest in ZIP codes with low median household incomes,
fewer college-educated households, and high minority shares.
2
We also document that low-FICO and high-LTV borrowers have relatively high locked-offer
rate gaps even within the same branch and controlling for loan amount. One possible reason such
borrowers might pay more would be if they tend to require more attention and service from loan
officers. However, using data on loan officer compensation, which we observe for a subset of lenders,
we do not find support for such a story. Overall, these results suggest that low-FICO and high-
LTV borrowers pay more for mortgages not simply because they present more credit risk, but also
because they search and negotiate less effectively.
If driven simply by borrowers’ time-invariant search costs, the gap between locked and available
rates should not vary over time. However, we find that average overpayment declines when the
level of market interest rates rises. This may partly reflect affordability constraints becoming more
binding as rates rise; however, we show that even borrowers that appear unconstrained (based
on their debt-to-income ratio) exhibit the same relationship. Thus, we conclude that behavioral
factors, such as feeling less of a need to shop or negotiate when rates are already low, likely influence
search effort.
6
Finally, we provide direct support for the importance of borrower sophistication using new
data from the National Survey of Mortgage Originations (NSMO). The NSMO combines detailed
administrative records on recent mortgage originations with survey data on the individuals who
took out those mortgages. The survey component focuses on borrowers’ shopping behavior and
their knowledge of mortgages and interest rates. Using these data, we show that shopping and
knowledge are predictive of borrowers getting lower mortgage rates, controlling for an array of
credit risk variables and other individual characteristics.
7
We construct a composite measure of
the rate component attributable to shopping/knowledge and—despite the coarseness of the survey
questions—find a sizeable 26bp gap between borrowers at the 90th percentile of this measure and
those at the 10th percentile. FHA, low-income, less-educated, and low-FICO borrowers tend to
do particularly poorly based on this measure. Lastly, we show that shopping activity is elevated
in higher interest rate environments, consistent with our conjecture that a rise in rates encourages
people to shop more.
8
Overall, our empirical results suggest that a large fraction of the borrower population in the US
overpays for mortgages, and a key reason for this seems to be a lack of financial sophistication. The
borrowers that fare the worst often get government-guaranteed loans through the FHA program,
which is aimed at lowering the cost of homeownership for lower-income households. Our results
suggest that government entities such as the FHA might consider ways to reduce price dispersion
6
In line with this, we also find evidence suggesting that FHA and jumbo borrowers may be “anchoring” to the
average prime conforming rate, which is the rate most often advertised and reported in the media: they obtain
relatively better rates (a lower locked-offer rate gap) when the difference between their offer rate and the prime
conforming offer rate increases.
7
In independent work made public after our initial draft, Malliaris et al. (2020) document similar patterns in the
NSMO data.
8
We further provide complementary evidence using the 2016 Survey of Consumer Finances (SCF). We find that
higher financial literacy, as gauged by the Lusardi-Mitchell financial literacy “test”, is associated with significantly
lower interest rates. We also find in the SCF data that borrowers who report shopping intensely for credit end up
with substantially lower rates.
3
and excessive markups to help fulfill their policy objectives. Our findings also suggest that the lack
of consumer search is important for the pass-through of monetary policy to the mortgage market:
reduced search effort appears to prevent borrowers’ rates from falling as much as they could when
market rates decrease, thereby weakening the pass-through of expansive policy.
Previous literature has documented price dispersion in several consumer finance markets, in-
cluding mutual funds (Horta¸csu and Syverson, 2004; Choi et al., 2010), auto loans (Argyle et al.,
2017), and credit cards (Stango and Zinman, 2016). We study price dispersion in mortgages—by
far the largest financial liability of households. Careful measurement of dispersion in this market
is important not only because of the significance of mortgages for households’ finances, but also
because this market is highly subsidized and dispersion has important policy implications for the
incidence of those subsidies. In addition, because of the widespread demand for homeownership,
participation in mortgages spans the income, wealth, and financial sophistication spectrums. Cross-
sectional variation in outcomes can help shed light on the factors driving dispersion in consumer
financial markets.
While standard search costs undoubtedly play a role in generating dispersion, as shown recently
by Argyle et al. (2017), a number of other factors could contribute to dispersion, including various
behavioral factors (see Zinman, 2015, for a review). The degree of dispersion we find, along with
the cross-sectional and time-series patterns of overpayment, suggest that factors beyond standard
search costs play a key role. We further support this conclusion by the direct survey evidence
relating obtained rates to shopping behavior and market knowledge—variables that are generally
unobserved in other settings where price dispersion has been studied. Thus, we view our results as
being in line with the growing literature pointing at financial literacy/sophistication as a key driver
of differential outcomes in household finance (e.g., Hastings et al., 2013; Gomes et al., 2020).
We are not the first to study dispersion in the US mortgage market. In particular, recent work
by Alexandrov and Koulayev (2017) and McManus et al. (2018) shows dispersion in lenders’ offer
rates, while Gurun et al. (2016) and Agarwal et al. (2019) study transacted rates.
9
Our results on
dispersion in offers are similar to those in the earlier work. However, when studying transacted
rates, we believe that our dataset allows improved measurement relative to previous papers, as we
are able to control for several variables not previously observable (e.g. discount points, rate lock
date, and rate lock period). Indeed, while we find wide dispersion in locked rates, it is considerably
narrower than the dispersion found in earlier work. Furthermore, we unpack the dispersion in
transacted rates by assessing the relative explanatory power of (time-varying) controls for lenders,
branches, and loan officers.
Most importantly, we advance the literature by comparing the rates that borrowers actually
obtain to the ones they could obtain in the same market at the same time. Our ability to compare
9
Some work also exists outside the US. Allen et al. (2014) study the Canadian market, where there is no dispersion
in posted rates, but large dispersion in contracted rates, which they argue arises due to differences in bargaining
leverage across consumers. In the UK market, Iscenko (2018) finds that many borrowers choose products that are
dominated in cost terms by other available alternatives, while Liu (2019) shows that many borrowers appear to neglect
non-salient fees and that lenders exploit this in their price setting. Damen and Buyst (2017) provide evidence that
mortgage borrowers in Belgium who shop more achieve substantial savings.
4
transacted rates to offers is key to understanding who overpays. Studying price dispersion (the
second moment) alone does not allow one to assess that question; similarly, differences in mean
transacted rates alone would not allow the separation of overpayment from credit risk premia. Fi-
nally, to our knowledge this is the first paper documenting how a measure of borrower overpayment
changes with market rates over time.
Agarwal et al. (2019) argue that overpayment by certain groups need not imply that they
are unsophisticated (or have high search costs), but could be a rational response of relatively
risky borrowers who fear being rejected. These authors document that the relationship between
contracted rate and the number of “inquiries” recorded by credit bureaus—their proxy for borrower
search—is U-shaped. This suggests that borrowers that search a lot may do so because their
application gets rejected, which in turn may lead these borrowers to accept relatively worse offers.
This channel may contribute to some of the overpayment we document. At the same time, however,
we find considerable overpayment even among a sizeable fraction of well-qualified borrowers, and
provide evidence from NSMO and SCF data suggesting that variation in sophistication is important
to understand cross-sectional dispersion.
10
Finally, in a well-known paper, Woodward and Hall (2012) document dispersion in the fees paid
by borrowers to mortgage brokers who arrange loans between borrowers and lenders. Woodward
and Hall use data from 2001, a time when brokers and loan officers were able to increase their
commission from lenders by getting borrowers to accept a high interest rate (as we discuss in
Section 2). However, since 2011 new regulations prohibit brokers and loan officers from earning
more in this way, which could in theory significantly reduce dispersion. Yet we continue to find
wide dispersion in our data covering recent years. Furthermore, while Woodward and Hall were
limited to a sample of 1500 FHA loans, our much larger dataset enables us to examine dispersion
and overpayment across market segments and borrower types, as well as within lender and even
loan officer.
The rest of the paper is organized as follows. In the next section, we provide some institutional
detail that will be important for the rest of the paper. Section 3 describes the Optimal Blue
data on rate locks and mortgage offers. Section 4 measures and unpacks price dispersion in the
offer data and the lock data. Section 5 explores how locked rates on average compare to the offer
distribution, and how this varies across borrowers with different characteristics. Section 6 studies
how these patterns evolve over time as market rates change. Section 7 introduces survey data from
the NSMO and presents direct evidence on the connection between shopping, mortgage knowledge,
and interest rate outcomes. Finally, Section 8 concludes with some potential policy implications.
10
Over the period we study, underwriting standards in the GSE and FHA segments of the market are largely
dictated directly by these agencies. Thus, for the vast majority of borrowers that get approved for a loan, it should
also be easy to get a loan from a different lender. That said, the perception that other lenders are unlikely to accept
one’s application may be sufficient to induce a borrower to accept a relatively “bad” offer.
5
2 Mortgage Pricing and Originations in the US
In this section, we provide a brief overview of some of the institutional details that will be important
for the rest of the paper.
11
In the US, there are multiple channels through which a borrower can obtain a loan. One of
them is to go directly to a bank or credit union. An alternative is to obtain a loan through a
specialized mortgage originator, a so-called “mortgage bank.” These lenders, contrary to what the
name suggests, are not depository institutions, and typically do not keep any of the mortgages on
their own balance sheet. Finally, it is also possible to go through a mortgage broker, who may have
relationships with both bank and nonbank originators, and acts as an intermediary connecting
borrowers to those institutions. When a loan is originated directly by a lender who will either
retain the loan in portfolio or sell it directly in the secondary (mortgage-backed securities, or MBS)
market, this is called a “retail loan”; if a loan is originated via a nonbank entity that originates the
loan for another lender, this is called “wholesale.”
Regardless of the channel, a borrower will generally interact (in person or just by phone/online)
with a loan officer or broker (henceforth LO) who will have access to various “rate sheets” that
provide the detailed pricing available at a given point in time (generally updated at least once a
day). Importantly, for any loan type and combination of characteristics, there is no single interest
rate—instead, the rate sheet shows a combination of note rates and “(discount) points”. To obtain
a low note rate, a borrower can pay points (where 1 point = 1 percent of the loan amount). If the
borrower is willing to take a higher rate, they can receive points (often called rebates or credits)
which in turn can be used toward the origination costs.
In the case of a retail loan, the available pricing will come directly from the lender’s pricing
desk; in the case of wholesale lending, the rate sheets can come from several different lenders (often
referred to as “investors”). Each rate sheet will provide pricing for different loan programs (e.g. GSE
loans, FHA, or jumbos) with adjustments depending on a few loan and borrower characteristics,
typically FICO, LTV, loan amount, geographic region, loan purpose and property type. Pricing
depends on the value that a lender assigns to the loan—often based on the current value of such
a loan in the MBS market, where most loans are ultimately sold.
12
Prices also take into account
required “guarantee fees” set by the agencies that securitize the loans and insure the credit risk,
namely the GSEs and Ginnie Mae (for FHA/VA loans).
13
Furthermore, lenders will add a margin
that may depend, among other things, on the level of demand for loans (Fuster et al., 2017).
On top of the prices from the rate sheet, the costs to the borrower include compensation of
11
For additional discussion, see e.g. Fuster et al. (2013) or https://files.consumerfinance.gov/f/201301_cfpb_
final-rule_loan-originator-compensation.pdf.
12
Generally, prices in the MBS market depend on the yields on alternative investments (especially Treasuries) as
well as investors’ projections of future prepayments of the underlying mortgages (since mortgage borrowers have a
free prepayment option). Prepayments are in turn affected by factors such as the volatility of interest rates, home
price growth, or relevant policies by the GSEs and FHA (e.g. streamlined refinance programs).
13
In addition to the guarantee fee, which is a flow insurance premium over the life of a mortgage, the GSEs
charge upfront “loan-level price adjustments” that depend on borrower and loan characteristics—see e.g. https:
//www.fanniemae.com/content/pricing/llpa-matrix.pdf.
6
the LO and/or their employer (e.g., the mortgage bank). This compensation may be explicit (via
upfront origination fees) or implicit (via lender profit margins on rate sheets). Historically, LOs
had strong incentives to sell loans with higher interest rates, all else equal, and thereby generate
more compensation not only for the lender but also for themselves (often called the “yield spread
premium”). However, in the wake of the financial crisis, new regulations were imposed so that LO
compensation may no longer vary with the interest rate and other terms of the loan. But lenders,
of course, still profit when borrowers take higher interest rates.
14
Importantly, this does not imply
that all LOs in a firm simply get paid an identical, fixed amount for each loan they originate. In
fact, LOs are frequently given a choice between different possible compensation plans, for example
trading off fixed salary for higher commission rates per dollar of originated loans.
Finally, it is not the case that the combination of rate sheets and a specific LO’s compensation
plan in all cases determine the final rate and points/fees that given borrower is offered: there may
be “exceptions” granted, for instance to meet a competitive outside offer. Lenders generally have
specific procedures for these exceptions, since they want to avoid violating fair lending laws.
15
An important step in the origination process is the mortgage rate lock. A lock is a guarantee
that the borrower will be issued a mortgage with a specific combination of interest rate and points if
the mortgage closes by a specific date. Borrowers typically lock their mortgage rates as a protection
against rate increases between the time of the lock and the time when the mortgage closes. A lock
can occur at the same time a borrower submits a loan application with a lender, but can also
happen at a later time. Not all rate locks ultimately lead to originated mortgages, since the loan
application can still be rejected afterwards (e.g. because the appraisal of the home comes in lower
than expected) or the borrower could renege. However, the lock is binding on the lender, as long as
the characteristics of the loan and borrower (such as the loan amount or the credit score) remain
as specified at the time of the lock. Lenders typically do not charge an explicit fee for a rate lock,
though there are generally loan application fees. Also, if a loan does not close by the time the lock
period expires, extending the lock typically requires a fee.
16
3 Optimal Blue Data
Our main data comes from an industry platform called Optimal Blue that connects over 600 mort-
gage lenders with more than 200 whole loan investors. Through the platform, mortgage originators
can gather information on mortgage pricing, initiate rate locks, manage pipeline risk, and sell mort-
gages to investors. Over forty thousand unique users access the system each month to search loan
programs and lock in consumer mortgages. More than $500bn of mortgages were processed through
14
These rules were first changed in 2011 as part of the Truth in Lending Act; the Consumer Financial Protection
Bureau published its final rule on LO compensation requirements in January 2013.
15
See e.g. https://www.crai.com/sites/default/files/publications/Managing-the-Fair-Lending-Risk-of%
2DPricing-Discretion-Whitepaper-Oct-2014.pdf or https://www.mortech.com/mortechblog/
pricing-discretion-fair-lending-risk.
16
For more information on rate locks, see e.g. https://www.bankrate.com/finance/mortgages/
questions-rate-lock-answered.aspx.
7
this system in 2017, thus accounting for about 25% of loan originations nationally.
The lenders using the platform tend to be nonbank monoline mortgage lenders (which have
gained substantial market share in the post-crisis period, see e.g. Buchak et al., 2018) and smaller
community banks or credit unions. That said, many institutions on this platform act as correspon-
dent lenders, meaning that they originate loans intended to be sold to other financial institutions
such as a large bank like JP Morgan or Wells Fargo.
For this study we use two components of the data generated by the platform: a) data on
mortgage products and mortgage prices actually accepted by consumers, and b) data on mortgage
products available and mortgage prices offered by lenders.
3.1 Mortgage Rate Lock Data
The first source of data is the universe of “rate lock”agreements for the mortgages processed through
the Optimal Blue platform. We have access to all the mortgage locks generated by the platform
since late 2013. Since the market coverage increases over the course of 2013-2014, we start using
the data from January 2015; we end in December 2019. The data have wide geographical coverage
of about 280 metropolitan areas as well as rural areas. All of the standard loan characteristics used
for underwriting are included: loan-to-value (LTV) ratio, FICO score, debt-to-income (DTI) ratio,
loan amount, loan program, loan purpose (purchase or refinancing), asset documentation, income
documentation, employment status, occupancy status, property type, ZIP code location etc.
There are a number of unique features of the data relative to servicing data that are typically
used in mortgage research. First, it includes not only the contracted mortgage rate, but also the
discount points or credits associated with that rate (meaning additional upfront payments made
or received by the borrower). Second, we observe the exact time-stamp of when the lock occurred,
while in most other datasets only the closing date is recorded, which generally differs from the
pricing-relevant lock date by several weeks or even months. Finally, we have unique identifiers for
the lender, branch, and loan officer that processes each mortgage. For some lenders we can also
observe loan officer compensation, expressed as a percentage of the loan amount.
17
While the lock data features numeric lender identifiers, it does not directly provide us with
information on the lenders. However, we are able to classify a subset of lenders into whether
they are an independent nonbank or not by relying on a match between Optimal Blue locks and
administrative FHA data used in Bhutta and Hizmo (2020). This will be useful later to assess
whether lender type and cross-selling might be driving the patterns in the data that we observe.
We restrict the sample in various ways to ensure that we study a relatively uniform set of loans
that is representative of the type of mortgages originated in recent years. For instance, we only
keep 30-year fixed-rate mortgages on owner-occupied single-unit properties, with full documentation
of assets and income, and drop self-employed borrowers. We also drop loans for amounts under
$100,000, and those with implausible values for LTV, DTI, or points/credits. Finally, we drop VA
17
Some lenders process compensation outside of the Optimal Blue system, or do not compensate loan officers
directly on a per-loan basis.
8
loans and streamline refinances (which are a small part of the sample). This leaves us with 3.6
million observations. For the analysis in Sections 5 and 6 we will further restrict the sample in order
to match the locked mortgages to offers for identical characteristics, as will be described there.
Table 1 presents some summary statistics from the lock data sample that we use for the analysis
in this paper, separating between the four loan programs in the data, since they differ substantially
in terms of borrower and loan characteristics. The four programs are: conforming (so they are
typically securitized through Fannie Mae or Freddie Mac), super-conforming (with loan amounts
above the national conforming limit but below the local limit, so that Fannie Mae or Freddie Mac
can still securitize the loan, but potentially at slightly worse prices), jumbo (loan amount above
the local conforming limit, meaning the loan cannot be securitized through the government-backed
entities), and FHA loans (which require mortgage insurance from the FHA and are securitized
through the government entity Ginnie Mae).
The table shows that FHA loans are most likely to go to first-time homebuyers with low FICO
scores and high LTV and DTI. Jumbo loans, the only loan type where the credit risk is not
guaranteed by the government, tend to go to the most creditworthy borrowers and feature relatively
low LTVs. They only constitute about 2% of our sample. The table also shows that FHA borrowers
on average pay fewer discount points than borrowers in the other programs; Appendix Figure A-1
displays the cumulative distribution of points paid (or received) by program.
As noted above, not all lenders use the Optimal Blue platform, and not all rate locks necessarily
result in an originated mortgage. Thus, there is a concern that the distribution of interest rates
recorded in our rate lock data may not accurately represent the rates that borrowers ultimately
end up with. However, in Appendix A.1, we show that the interest rates observed in the rate lock
data mirror the interest rates observed in the well-known McDash mortgage servicing dataset on
originated mortgage loans, both in terms of averages and dispersion. Furthermore, loan/borrower
characteristics in Optimal Blue locks also look very similar to those in data on originated loans.
18
3.2 Mortgage Offers Data
As our second source of data, we collect data on the menu of mortgage products and mortgage rates
that lenders offer through the platform’s pricing engine. Optimal Blue’s “Pricing Insight” allows
users to retrieve the real-time distribution of offers for a loan with certain characteristics in a given
local market (where an offer consists of a combination of a note rate and upfront fees and points
that the borrower pays or receives with this rate). The Insight interface is designed for lenders to
compare their pricing against that of peers.
For any combination of day, MSA, and loan/borrower characteristics, we measure an “offer” rate
for each lender on the platform. This offer rate reflects the interest rate (with zero points) that
18
See Appendix Table A-1 and Figure A-3 for details. For jumbo mortgages, the locked interest rates in Optimal
Blue tend to be higher than those in McDash, which could reflect that the relatively smaller lenders that use the
Optimal Blue platform may not be as competitive for these types of loans as for FHA and conforming loans. It is also
the case that average jumbo loan amounts are somewhat smaller in Optimal Blue locks than in McDash originations,
which could reflect some differential selection of borrowers. The dispersion of rates is still very similar, however.
9
the lender could offer a prospective borrower, including fees under the assumption that the loan is
originated by the loan officer (LO) that has locked the most loans for that lender in that market.
19
If a lender represents multiple different investors, the offer we observe is based on the most
competitive investor offer. Thus, a borrower locking a loan with this lender would not necessarily
get exactly the observed offer rate for three reasons. First, the locked rate can vary depending on
which LO the borrower goes through, since different LOs can charge different markups. Second,
the LO may offer a loan that is not based on the rate sheet of the most competitive investor, but
on one from a different investor.
20
Third, as noted earlier, borrowers may be able to negotiate and
get an “exception” or a lower rate from the lender.
We conduct daily searches in one local market (Los Angeles), twice-weekly searches in four
markets, and weekly searches for 15 additional markets.
21
We collect offer distributions for 100
different loan types, differing across the following dimensions: FICO score, LTV ratio, loan program,
loan purpose (purchase or cash-out refinance), occupancy (owner-occupied or investor), rate type
(30-year fixed or 5/1 adjustable), and loan amount. The mortgages require full documentation of
income, assets and employment, and are used to finance single-unit homes.
An important limitation of the offers data is that we are not able to track institutions over time
or match them directly to the lenders in the lock data, since there is no fixed lender identifier. The
time series is also slightly shorter than for the locks data, as we started systematically tracking
offers in April 2016.
4 Dispersion in Mortgage Rates
4.1 Dispersion in Offer Rates
We begin by briefly presenting some findings from the Optimal Blue Insight data on offer rate
dispersion. Our analysis here, along with additional findings presented in Appendix A.2, adds to
recent work looking at offer rate dispersion using other sources of data in Alexandrov and Koulayev
(2017) and McManus et al. (2018).
Figure 1 shows the dispersion in mortgage rates available from different lenders, pooling data
over time and across all of the 20 metropolitan areas for which we obtained data. To make distri-
butions comparable across time and locations, we demean the offer rates for each mortgage type in
each market and day. Figure 1 indicates wide dispersion in offer rates. There is a 53bp difference
between the 10th and 90th percentile offers, which is similar to what Alexandrov and Koulayev
(2017) and McManus et al. (2018) have documented.
In Appendix A.2, we additionally show that the degree of offer rate dispersion is quite similar
19
As explained further in Appendix A.2, we observe a distribution of prices (points) for a given note rate, which
we transform into a distribution of rates for zero points.
20
One reason why an LO might want to do this is to maintain active relationships with multiple investors.
21
The markets with twice-weekly searches are New York City, Chicago, Denver, and Miami. The markets with
weekly searches are Atlanta, Boston, Charlotte, Cleveland, Dallas, Detroit, Las Vegas, Minneapolis, Phoenix, Port-
land, San Diego, San Francisco, Seattle, Tampa, and Washington DC.
10
across different types of loans, different types of borrowers, and across all 20 cities in our sample.
Finally, it worth noting that it is not necessarily the case that a given lender occupies the same spot
in the offer distribution over time. Lenders could move around in the distribution if pricing does
not simply reflect time-invariant cost factors. Unfortunately, since we cannot follow lenders over
time in the Insights data, we cannot assess this directly in the offer data. However, the analysis in
the next subsection will shed some light on whether lenders’ relative pricing changes over time.
4.2 Dispersion in Locked Rates
In the previous subsection we observed wide variation in mortgage rates available from different
lenders for identical borrowers on the same day and in the same market. In this section we aim
to investigate whether identical borrowers who choose the same mortgage product, in the same
market, and at the same time, actually lock in different interest rates. If many borrowers shop
around, we may observe less dispersion in locked rates than we observe in offer rates.
To investigate dispersion in locked mortgage rates, we regress locked rates on borrower and
loan characteristics, as well as time effects, and then add an increasingly fine set of fixed effects.
Our outcome of interest is the remaining dispersion in the residual, which we measure in terms of
standard deviations, as well as the gap between 75th-25th or 90th-10th percentiles. Comparing the
residual dispersion across specifications allows us to “unpack” the relative importance of different
drivers of price dispersion in this market.
Table 2 shows the results from various specifications, estimated on the same set of 2.96 million
loans locked over the five-year period 2015-2019.
22
In the first column, as a benchmark, we include
only lock date-by-MSA fixed effects, in order to document the amount of overall interest rate
dispersion within the same MSA on the same day. These day-by-MSA fixed effects explain just
under 60 percent of the total variation in rates, and the standard deviation of the residual is 33bp.
In column (2), we add our baseline set of controls: an extensive set of underwriting variables,
which consist of fully interacted bins of values for FICO, LTV, and loan program, interacted with
lock month to allow for time-variation in risk pricing.
23
We also include borrower ZIP code fixed
effects, lock period fixed effects, property type fixed effects, cubic functions of loan amount and DTI,
as well as linear controls for FICO and LTV (to allow for within-bin variation).
24
This specification
is similar to regressions one could typically run with a mortgage servicing dataset.
25
We see that
22
The estimation drops “singleton” observations that are completely determined by the set of fixed effect. There
are more such singletons as we add more fixed effects; to ensure that our results are not driven by changing samples,
we use the remaining sample from the most restrictive specification (10) in all specifications. However, using the
largest possible sample for each specification instead does not materially affect the results.
23
We include 13 FICO bins, 9 LTV bins, and 12 dummies for the four loan programs interacted with three loan
purposes (purchase, rate refinance, and cash-out refinance). The choice of FICO and LTV bins is motivated by the
loan-level price adjustments set by the GSEs.
24
The lock period typically varies from 15 to 90 days, with 30 and 45 days being the most common choices. A
longer lock period leads to a slight increase in the fee (or equivalently the interest rate).
25
It is already somewhat more precise, since here we control for the date in which a loan is locked, along with
the length of the lock period, while in typical dataset loans originated in the same month may have been locked in
different months.
11
the controls explain a sizable share of the raw variation in interest rates—the adjusted R-squared
is 0.75—but that substantial dispersion remains: the standard deviation in residuals is 0.26, and
the borrower at the 90th percentile of the residual distribution pays 58bp more than the borrower
at the 10th percentile.
Column (3) adds bins for the points paid or received by the borrower (interacted with program
by lock month).
26
This (usually unobserved) variable indeed explains some of the rate differences
across borrowers, but substantial dispersion remains—e.g. the 90th-10th percentile difference is
still 54bp, which is almost identical to the corresponding dispersion in offer rates shown in the
previous subsection.
Based on the regression coefficient on discount points (not shown in the table), we can translate
interest rates to upfront points.
27
This coefficient implies that 1 discount point changes the interest
rate by about 21bp on average. Therefore, 54bp in rate is approximately equivalent to 2.6 upfront
discount points or 2.6% of the mortgage balance. In other words, our results imply that a borrower
with a $250k mortgage borrowing at the 90
th
percentile interest rate should be getting—but in fact
is not getting—a lender credit of about $6,500 relative to someone borrowing at the 10
th
percentile
interest rate.
Thus, observably identical borrowers within the same market, on the same day, getting the
same loan can pay dramatically different prices. Table 3 shows how the residual dispersion in
interest rates varies across different loan programs and characteristics. The middle column of the
table uses the residuals from specification (3). We see an extreme amount of dispersion for the two
lowest FICO groups. We also see substantial dispersion for FHA-insured loans, despite the fact
that these loans are fully insured by the government and thus lenders and investors take very little,
if any, credit risk. In other words, it seems unlikely that unobserved risk factors could explain the
wide dispersion in FHA interest rates. Along the same lines, we also find fairly wide dispersion
for conforming and super-conforming loans, which meet the credit standards of the GSEs and will
likely be purchased and fully guaranteed by these institutions.
28
Finally, we also see wide dispersion
even when we focus just on low-risk borrowers: those with prime FICO scores in excess of 680, and
those with LTVs of less than 75 percent.
Jumping back to Table 2, in column (4) we add lender fixed effects to allow for the possibility
that some of the price differences may reflect differences in lender characteristics such as service
quality or advertising costs. We find that the 90th-10th percentile difference decreases only slightly,
26
We include 8 point bins, as well as a linear function in points to allow for within-bin variation.
27
We estimate the relationship between discount points and interest rates in a regression specification identical to
column (10) of Table 2, with the only exception that discount points are allowed to only enter linearly.
28
One caveat here is that lenders may be worried about so-called “put-back” risk where loans in default must be
repurchased by the lender due to some defect in the underwriting found by the FHA or GSEs. However, at least in
the case of the GSEs, Goodman (2017) documents that put-back risk has been negligible since lenders have stopped
issuing low-documentation and other non-traditional loans. For FHA loans, perhaps the biggest concern for lenders
has been litigation risk under the False Claims Act, which allows the federal government to sue lenders that knowingly
submit false or fraudulent claims to the FHA. Under the Obama Administration, some of the largest lenders settled
with the government, paying fines close to $5 billion. That said, this risk is most salient for large banks with significant
capital at risk, unlike the nonbanks that dominate our data. Also, this risk has eased in recent years.
12
by 6bp. In columns (5) and (6), we further interact the lender fixed effects with lock day fixed
effects and other controls, to allow for the possibility that lenders’ (relative) pricing may change
over time, or may differ across loan types. Here the 90-10 gap drops more substantially, by 10bp (or
over 20 percent) from column (4). Overall, the results in columns (4)-(6) suggest that more so than
time-invariant differences in lender quality, price dispersion may reflect lender pricing strategies
that vary over time and across programs. Such variation would make it difficult for borrowers to
find low rates simply by following the recommendations of family, friends or real estate agents—yet
this is a common approach borrowers take to finding a mortgage.
In columns (7) and (8), we further allow for pricing to differ across different branches of a lender.
As discussed earlier, the lenders in our dataset tend to be nonbank monoline mortgage lenders and
community banks. For a typical lender in our data, in a given MSA, most loans are originated
through just 2 or 3 branches located within that MSA. Differential branch pricing could reflect
differences in convenience of the office location and/or costs (e.g. office rent). In addition, as noted
earlier, different branches can have different markups and pricing strategies.
The branch fixed effects in column (7) have noticeable incremental explanatory power, increasing
the adjusted R-squared from 0.85 to 0.88 and reducing the residual dispersion. Adding branch-
by-month fixed effects in column (8) further reduces residual dispersion—consistent again with
time-varying price strategies playing a role in the rates borrowers obtain, but in this case at the
branch level. Nevertheless, even in column (8), which should come close to looking at nearly-
identical borrowers getting a loan from the same branch at the same time, the 90-10 gap remains
at 31bp, and the interquartile range at 14bp.
Lastly, in columns (9) and (10), we further allow for pricing to vary across different loan officers
(LOs) in the same branch, which could reflect for instance differences across LOs in terms of
experience, compensation, or willingness/ability to negotiate. Which LO a borrower matches up
with (within a branch) does appear to matter somewhat for the rate they end up with, since the
adjusted R-squared further increases and the residual dispersion decreases in the last two columns.
Nevertheless, even after including LO fixed effects that are allowed to vary across time and programs,
the 90th-10th percentile difference remains at 26bp, and the interquartile range at 11bp.
The last column of Table 3 shows that the cross-sectional patterns in residual dispersion, al-
ready discussed above, remain similar in the most restrictive specification (10): the dispersion is
substantially larger for loan types and borrower characteristics that are associated with being more
financially constrained and potentially less sophisticated, such as FHA loans, low-FICO borrowers,
or first-time homebuyers.
The final rows of the table show that the residual dispersion is identical if we only consider loans
that were locked with lenders that we are able to classify as independent nonbanks (as discussed
in Section 3.1).
29
This suggests that the large dispersion is not driven by unobservable pricing
adjustments that banks or credit unions might make for customers that already have accounts or
29
The residual dispersion results are also essentially unchanged if we restrict the estimation sample to these inde-
pendent nonbanks only.
13
other business with them. The nonbank lenders are only in the business of originating mortgages.
To sum up the findings from this analysis, there is a large amount of dispersion in the rates that
observably identical mortgage borrowers pay, even after controlling for the exact timing and upfront
payments. Adding lender, branch and LO controls reduces the residual rate dispersion by about
half. However, substantial dispersion remains, implying that two observably identical borrowers
may get quite different deals from the same lender branch or even the same loan officer at the same
time. Furthermore, this appears to be more pronounced for financially less well-off borrowers or
those that are inexperienced in the market.
5 Comparing the Locked Rates to Offer Rates
The analysis so far has focused on dispersion, or “second moments.” We now turn to the question
of whether different types of borrowers get good or bad deals on average (i.e. the first moment),
relative to what is available in the market at the time they lock their mortgage. This will allow us
to assess more directly which types of borrowers tend to “overpay” for their loans, and test different
hypotheses for what is driving differences across borrowers and over time.
To do so, we use the data on lenders’ offer rates (described in Section 3.2) to compute median
offer rates by day, MSA, FICO, LTV, loan amount, and loan program (i.e. conforming, super-
conforming, jumbo, and FHA). We then match these benchmark median offer rates to observations
in the rate locks data with identical characteristics, and study the difference between the rate
obtained by consumers and the median rate available—the locked-offer rate gap.
30
We have fewer observations than in the previous analysis based on lock data only, since offer
rates are only available for a subset of loan types/characteristics, 20 MSAs, and a shorter time
period. In particular, this analysis is restricted to purchase mortgages, since we have the most
granular offers for them. Appendix A.3 provides additional detail on the matching.
In our main analysis, we focus on the distance between the rate locked by a borrower and
the rate available at the median lender, since we believe that this is a simple and transparent
benchmark. However, in Appendix A.4 we consider an alternative measure that is more directly
motivated based on search theory, namely the expected gain from obtaining one additional rate
quote from a different lender. As we show there, the main results from this section are qualitatively
identical when using this alternative measure.
5.1 Locked-Offer Rate Gaps by Borrower Type
The top panel of Figure 2 shows the distribution of the locked-offer rate gap for all mortgages in our
data. The dashed vertical line denotes the mean of the distribution. The locked-offer rate gap is
30
We use the rate at which the median lender offers a loan with zero points and fees. To compare to this offer,
we adjust the locked rate for points paid or received by the borrower based on the empirical relationship between
discount points and interest rates. We estimate this relationship in a regression specification identical to column (10)
of Table 2, with the only exception that discount points are allowed to only enter linearly (instead of entering in a
binned fashion as in Table 2).
14
positive on average (dashed line just to the right of the thick black line that denotes zero), meaning
that borrowers end up with mortgage rates that are more expensive than what the median lender
could offer for identical mortgages.
31
The bottom four panels of Figure 2 show the distribution of the locked-offer rate gap for various
sub-segments of the market. The summary statistics for these distributions are given in Table 4.
The locked-offer rate gap is largest for FHA loans, with an average of +25bp. This amounts to
about 1.2% of the mortgage balance in upfront points/fees, or $2,400 for a typical FHA loan of
$200k. Moreover, one-quarter of FHA borrowers overpay by 45bp or more relative to the median
offer. In contrast, the distributions for super-conforming and jumbo mortgages look very different:
the locked-offer rate gap is on average slightly negative at -4bp for super-conforming mortgages, and
even more negative at -21bp for jumbo mortgages. Thus, in these two market segments, borrowers
on average obtain relatively good deals, suggesting that they may be more sophisticated at shopping
and negotiating.
32
Note that the differences in average locked-offer gaps across market segments
generally follow a similar pattern as the differences in dispersion seen in Table 3, but in some cases
tell a more nuanced story: for instance, residual rate dispersion is identical in the conforming and
jumbo segments, but jumbo borrowers fare significantly better relative to offers. This illustrates
the value of studying the locked-offer rate gap, rather than relying on dispersion alone.
Table 4 further shows how the locked-offer rate gap distribution varies by FICO scores, LTV
ratios, whether the borrower is a first-time homebuyer, whether the borrower paid or received points
when taking out the loan, and whether we can classify the lender as independent nonbank or not.
On average, borrowers with a FICO larger than 740 lock in mortgage rates that are close to the
median offer, while borrowers with lower FICO scores lock in rates well above the median offer. For
instance, borrowers with FICO scores between 640 and 660 on average pay 23bp more than what
the median lender would offer for identical mortgages.
A similar pattern is evident when splitting the sample by LTV: borrowers with LTV less than
90% tend to obtain rates close to the median of the offer distribution, while higher LTV borrowers do
worse relative to the median offer. First-time homebuyers also tend to fare worse: on average, first-
time buyers pay 15bp more than what the median lender could offer them, while repeat homebuyers
pay only 7bp more.
Borrowers that pay discount points (positive values in the table) tend to end up with a higher
locked-offer rate gap than those who receive points (known as a rebate or credit) from the lender.
Note that since we adjusted the mortgage note rate for points paid, this relationship is not “me-
chanical.” Finally, the average rate gap is slightly higher when focusing on independent nonbanks
only; we return to discussing potential differences across lender types below.
It is worth noting that within each of the groups in Table 4, there is substantial dispersion in
31
In Figure A-2 in the appendix, we validate that our median offer rates derived from Optimal Blue Insights closely
track offer rates for comparable loans published by Mortgage News Daily, an industry website.
32
Consistent with our finding that jumbo borrowers tend to find good deals, Horta¸csu and Syverson (2004) find
that index fund investors, who may be relatively sophisticated given their participation in the stock market, place
the bulk of their assets in the the cheapest decile of funds.
15
the locked-offer rate gap, as shown in the table’s final three columns. Thus, even for high-FICO or
low-LTV borrowers, which on average have a gap close to zero, a non-trivial fraction of borrowers
lock rates well above what the median lender could offer them. However, dispersion tends to be
largest for the groups that on average fare the worst.
Table A-5 provides analogous summary statistics based on the median income, college education
share, minority share, and mortgage market concentration in a borrower’s location.
33
In the first
three cases, the observed differences between highest and lowest terciles are large: for instance, the
tercile of borrowers in ZIP codes with the lowest fraction of college educated residents on average
has a locked-offer rate gap of 16bp, while for the tercile with the highest fraction the average gap is
only 5bp. Similarly, average gaps are larger in areas with lower median incomes and higher minority
shares. In contrast, we do not find much evidence that average gaps increase in local mortgage
market concentration, although the dispersion of gaps is larger in more concentrated markets.
5.2 Regression Analysis
Next, we turn to a regression analysis to investigate whether the differences across FICO and LTV
groups in the locked-offer rate gap hold after controlling for certain loan characteristics, as well as
fixed effects for the particular lender and branch to which borrowers went. For a subsample of loans,
we can further control for loan officer compensation, which helps us assess whether differences in
locked-offer rate gaps may be driven by low-FICO or high-LTV borrowers being “more work” for
loan officers. Finally, we also test whether paying or receiving points is associated with getting a
worse deal on the loan.
One potential explanation for the results in Table 4 is that lower-FICO borrowers and higher-
LTV borrowers tend to have smaller loans and thus less of an incentive (in dollar terms) to shop
around. In columns (1) and (4) of Table 5, we regress locked-offer gaps on bins for different FICO
scores and LTV ratios, respectively, as well as fine loan amount bins and MSA-by-month fixed
effects. It is indeed the case that borrowers with the largest loan amounts pay substantially less
relative to their median offer rate (not shown in table). However, conditional on loan amount,
lower-FICO borrowers and higher-LTV borrowers continue to pay more, to a similar degree as
we observed in Table 4. Thus, such borrowers appear to obtain more expensive loans for reasons
beyond the differential monetary incentive to shop stemming from loan size variation.
Another potential explanation for why low-FICO and high-LTV borrowers are more likely to
pay too much is that they sort into more expensive lenders or branches. Borrowers might choose
expensive lenders because they offer better service or simply because they spend more on marketing
and are more visible. To investigate this explanation, we include branch fixed effects in columns (2)
and (5) of Table 5. In these columns, the R-squared jumps sharply to about 50 percent from less
than 20 percent, meaning that branch-specific pricing differences explain a fair amount of variation
33
Income, education, and minority shares are measured at the ZIP code level based on 2017 American Community
Survey data; mortgage market concentration is measured at the county level as the share of the largest four lenders
(following Scharfstein and Sunderam 2016) in the 2016 HMDA data.
16
in the locked-offer gap. Furthermore, the coefficients on FICO and LTV become slightly smaller in
magnitude, implying that sorting into lenders does explain some of the “overpayment” by low-FICO
and high-LTV borrowers, but the coefficients remain large.
Thus, it does not appear that, for example, lower-FICO borrowers end up with higher locked-
offer gaps just because they get their loans from more expensive lenders or branches. Even within
the same branch, low-FICO and high-LTV borrowers tend to pay more relative to their benchmark
median offer. One reason why this might occur is that (some of) these borrowers could be “more
work” for loan officers, who therefore require additional compensation through a higher rate (or
equivalently, more points upfront). Since by law the LO compensation can no longer depend on
the interest rate paid by the borrower, the postulated channel would have to work through low-
FICO and high-LTV borrowers being matched with LOs that “specialize in difficult cases” and
charge more. In order to test for this possibility, in columns (3) and (6) we directly control for LO
compensation (in % of the loan amount) for the subset of loans for which we observe it.
34
The
coefficients on this variable are strongly significant, and their magnitude of about +0.15 suggests
that higher LO compensation is reflected almost one-for-one in the rate the borrower pays (since
we earlier noted that one percent of the loan amount—one point—corresponds to about 0.2% in
rate terms). However, the coefficients on FICO and LTV remain similar, implying that low-FICO
and high-LTV borrowers do not pay higher rates simply because they match with expensive LOs.
The final two columns of the table test whether borrowers who pay or receive points get a worse
deal relative to the omitted category (those with points between -0.2 and +0.2).
35
Column (7)
reproduces the result seen in Table 4 that borrowers who pay (receive) points tend to pay high
(low) rates relative to what is available in the market. Column (8) shows that once we control for
lender/branch, the coefficients on the dummies for having paid or received points are close to zero;
this means that the overall relationship is driven by sorting into cheap/expensive lenders.
The main takeaway from this analysis is that low-FICO and high-LTV borrowers on average
tend to pay substantially higher rates not just due to credit risk premia embedded in lender offers,
but to a large extent due to the fact that they end up with worse rates relative to what is in principle
available in the market. This is illustrated in Figure 3, where the magnitude of the coefficients on
FICO and LTV bins from columns (1) and (4) of Table 5 are compared to coefficients from a similar
regression where we use the offer rates as dependent variable. We see that for FICO, the locked-
offer rate gap is about one-third as large as the offer differences.
36
For LTV, it is in fact not the
case that lender offers for high-LTV loans on average feature worse rates; if anything, the reverse
is true. This may be surprising, but is mostly due to the fact that in the conforming segment,
borrowers with LTVs above 80 are required to get private mortgage insurance, which effectively
34
We only observe loan officer compensation for a subset of lenders. LO compensation typically amounts to 1-2%
of the loan amount originated.
35
One reason why borrowers who pay/receive points might pay higher rates could be that lender offers become
more difficult to compare than at zero points, so that suboptimal decisions become more likely.
36
One reason for the higher offer rates for low-FICO borrowers is that the GSEs charge additional loan-level price
adjustments for such loans.
17
reduces the risk to the lender/GSE (at least in terms of loss-given-default).
37
Thus, what this
analysis implies is that high-LTV borrowers only pay higher interest rates due to their less effective
search/negotiation process, rather than due to differences in offer rates.
Robustness. Table A-6 in the Appendix reproduces the same regressions for FHA loans only,
and obtains similar results. Thus, the previous findings are not due simply to sorting into different
loan programs, or driven by the different benchmark offer rates across programs.
In another robustness check, reported in Table A-7, we restrict the sample to lenders that we
can identify as independent nonbanks. Doing so leaves the coefficients from Table 5 essentially
unchanged. As noted earlier, the nonbank lenders that constitute the majority of our sample are
only in the business of originating mortgages. Thus, the results on differential locked-offer gaps
cannot be explained by potential price advantages that bank lenders might grant to financially
well-off (high FICO, low LTV) customers, for instance because they also have significant account
balances or other business with the bank.
38
As mentioned at the beginning of this section, in Appendix A.4 we reproduce the two tables
from this section using a borrower’s expected gain from additional search instead of the locked-offer
gap, and obtain qualitatively identical results.
Finally, it may be that most of the lenders making offers in our dataset are small and hard to
find. If that was the case, it would not be surprising that most borrowers pay more than what the
median lender is offering. To rule out this potential explanation, we replicate our analysis using
only offers from high-volume lenders, as designated on the Optimal Blue platform. Our results
remain qualitatively unchanged.
6 Time-series Movements in the Locked-Offer Rate Gap
The last section explored the cross-sectional patterns in the locked-offer rate gap. In this section,
we instead study how this gap moves over time, with a particular focus on how it responds to
changes in market interest rates. Are borrowers more likely to end up with worse rates (relative to
what the median lender could offer) when market rates are low, and more likely to get a good deal
as rates increase? If so, what might explain this relationship?
Figure 4 plots the average locked-offer rate gap against market interest rates, here measured by
the 10-year Treasury yield.
39
In the summer of 2016, the level of market interest rates as shown
by Treasury yields was very low. The locked-offer rate gap during this time was high, meaning
37
Reflecting this, the GSEs’ loan-level price adjustments tend to be lower at LTVs above 80 than at 80.
38
Table 4 showed that borrowers who obtain loans fron nonbanks tend to have higher locked-offer gaps; this could
either reflect overall advantageous pricing by banks/credit unions, or selection by borrowers. What we emphasize
here is that such differential pricing, if it exists, does not appear to vary with borrower creditworthiness.
39
For the average locked-offer gap, we use the estimated month fixed effects from a regression similar to those in
Table 5 but controlling simultaneously for FICO and LTV. We use the 10-year Treasury yield as our measure of
market rates since it is strongly correlated with the 30-year fixed mortgage rate, but avoids potential endogeneity
issues due to the measurement of the latter. However, using the mortgage rate or the current-coupon MBS yield
instead leaves our conclusions unchanged.
18
that borrowers were locking rates from the higher end of the offer rate distribution. As Treasury
yields increased, and as a result lenders increased their offer rates, the locked-offer gap shrunk,
indicating that borrowers moved toward the cheaper end of the offer distribution. When rates fell
again starting in late 2018, the inverse happened. Overall, the movements in the locked-offer gap
almost mirror movements in the Treasury yields.
We confirm the statistical significance of the relationship between the locked-offer rate gap
and market rates in Table 6.
40
The first column adds the 10-year Treasury yield as a control to
a regression similar to the ones estimated in Table 5 (controlling for all borrower characteristics
jointly, along with MSA fixed effects). The coefficient implies that as the 10-year Treasury yield
increases by 1 percentage point, the average locked-offer gap falls by about 6bp. This is sizable,
given that we saw earlier that over our sample as a whole, the gap averaged 11bp with a standard
deviation of 31bp.
In column (2) we add month fixed effects (interacted with MSA) and see that the relationship
between Treasury yields and the locked-offer rate gap is even stronger within-month—the magnitude
of the coefficient increases to 8.3bp. Column (3) further adds branch fixed effects, to see to what
extent the estimated relationship gets weaker once we control for potentially time-varying selection
of borrowers into expensive or cheap lenders/branches. The coefficient on the Treasury yield is
reduced (to 5.7bp), suggesting that some of the overall relationship may be due to borrowers
selecting cheaper lenders when rates are higher (consistent with additional shopping).
Next, we test the hypothesis that the relationship is driven purely by affordability constraints:
as market rates increase, the implied monthly mortgage payments increase, and more borrowers
may come up against DTI constraints embedded in mortgage underwriting.
41
To study whether this
is likely to be an important factor behind the relationship, we separate borrowers into those with
a DTI up to 36 percent (which are likely unconstrained by the payment burden) and those with a
higher DTI (for whom a higher rate may mean that they run up against underwriting constraints).
We thus repeat the same regressions, allowing for separate coefficients on the Treasury yield
depending on whether a borrower’s DTI is above 36 or not. Across columns (4) to (6), we see that
the estimated coefficient on the Treasury yield is indeed slightly more negative for the high-DTI
borrowers, suggesting that affordability constraints play some role in the relationship. However,
the coefficient on the Treasury yield remains sizeable even for those borrowers that are most likely
not constrained by the payment burden.
This suggests that the relationship may be driven at least partly by “behavioral” factors: for
instance, when the level of rates is already low, borrowers may feel less compelled to search for a good
40
Appendix Table A-10 repeats the analysis in this section with the expected gain from search as the dependent
variable, which leaves the qualitative results unchanged. This implies that the results cannot be explained by time
variation in the width of the offer distribution.
41
The relevant debt-to-income ratio in the US is usually the so-called “back-end” ratio, which divides the required
monthly payments on all debts (not just the mortgage) by the monthly income. Under the “qualified mortgage” rule
that has been in effect in the US since 2014, this back-end DTI ratio is supposed to be below 43 percent (see e.g.
DeFusco et al., 2020). However, conforming mortgages guaranteed by Fannie Mae and Freddie Mac are exempt from
this requirement; these entities therefore impose their own requirements, which in some cases can be higher.
19
deal or negotiate hard than when rates are higher, even though in dollar terms the consequences
are the same. This might be the case particularly after a recent drop in rates, as borrowers might
compare their offer to a higher reference level. In Section 7.3, we will show that according to survey
data, shopping effort does indeed increase when market rates are higher.
The hypothesis that borrowers anchor their beliefs about mortgage rates to a reference rate
also has implications for the cross-section of the locked-offer rate gap. If borrowers use a heavily
advertised rate, such as the prime conforming rate, as a reference rate, they should be willing to
search/negotiate more when the offer rates in their program are high relative to this reference rate.
Therefore, the locked-offer rate gap should be low when the gap between the offer rates in other
programs and the advertised prime conforming rate is high. We test this hypothesis in the last two
columns of Table 6.
We first compute a daily median offer rate for a typical borrower in each program.
42
Then,
we calculate the spread between that program-specific offer rate and the prime conforming offer
rate from Optimal Blue; by construction this spread is zero for borrowers in the conventional
conforming market.
43
Specification (7) regresses the locked-offer rate gap on the offer spread to
the prime conforming rate and shows that the coefficient of interest is negative and statistically
significant. Specification (8) also includes Treasury yields to control for the overall level of the
interest rates and the coefficient is unchanged. This mechanism may also partly explain why FHA
borrowers overpay more and jumbo borrowers underpay less than standard conforming borrowers:
the offer rates are typically lower for FHA than conforming (by 33bp on average over our sample)
but higher for jumbo than conforming (by 7bp on average).
44
Overall, these results support the
hypothesis the variation in the locked-offer rate gap is likely to be driven by behavioral phenomena,
such as anchoring to the level of a salient and observable mortgage rate.
7 Survey Evidence on Shopping, Knowledge, and Mortgage Rates
In this section, we use the National Survey of Mortgage Originations (NSMO) to document how
different measures of borrower shopping and financial literacy (in particular, knowledge about
mortgages) correlate with the mortgage rate a borrower obtains. We also document which borrower
types appear to overpay due to a lack of shopping and knowledge, and how shopping effort varies
with the level of market interest rates. In both cases, our findings align well with our earlier results.
The NSMO is a joint initiative of FHFA and CFPB as part of the “National Mortgage Database”
program. It surveys a nationally representative sample of borrowers with newly originated closed-
end first-lien residential mortgages in the US, focusing in particular on borrowers’ experiences
getting a mortgage, their perceptions of the mortgage market, and their future expectations. In
November 2018, micro level data for the first 15 survey waves were for the first time made public
42
For conforming, super-conforming, and jumbo loans, we compute the daily median offer rate for a borrower with
LTV=80, FICO=750, DTI=36. For FHA loans we compute it for a borrower with LTV=96, FICO=680, DTI=36.
43
Our results are robust to different choices for reference rate such as the Freddie PMMS rate, the Bankrate prime
rate, Mortgage News Daily 30-year fixed rate, or MBS yields.
44
FHA rates are lower because they do not include the insurance premium, which borrowers need to pay separately.
20
on the FHFA website, covering originations from January 2013 to December 2016.
45
The NSMO
contains a large number of questions, some of which were not asked in all waves, along with admin-
istrative information (from matched mortgage servicing and credit records) on borrower character-
istics such as FICO credit score at the time of origination, or the spread between a loan’s interest
rate and the market mortgage interest rate.
The full NSMO dataset contains 24,847 loans. For our analysis, we impose a number of sample
restrictions. The main ones are that we only consider mortgages on a household’s primary residence
and drop mobile/manufactured homes as well as 2-4 unit dwellings. In addition, we require the loan
term to be either 10, 15, 20, or 30 years, and drop construction loans or those obtained through a
builder, mortgages with an associated additional lien, and those with more than two borrowers on
the loan. Finally, we drop a few observations where the survey respondent was not a borrower on
the loan. This leaves us with 19,906 mortgages for the analysis.
Our analysis in this section will proceed in three parts: first, we estimate the relationship
between measures of borrower shopping or knowledge about the mortgage market and the rate
borrowers obtain on their loan, controlling for a rich set of borrower and loan characteristics.
Second, we study which borrower and loan attributes correlate with lower rate spreads solely due
to shopping and knowledge about the mortgage market. Third, we show that shopping effort
increases when market interest rates are higher.
7.1 The Relationship between Shopping, Knowledge, and Contract Rates
We estimate OLS regressions of the form
RateSpread
ijtw
= βX
i
+ ΓZ
ij
+ α
t
+ δ
w
+
ijtw
(1)
where RateSpread
ijtw
is the spread between the contract rate and the market mortgage rate prior
to origination, for borrower i with loan characteristics j, loan origination month t and responding
to survey wave w.
46
X
i
are different measures of borrower i’s shopping effort or knowledge about
the mortgage market, as described below. Z
ij
is a rich set of borrower and mortgage characteristics
that could influence the pricing of the loan. The full list of controls is provided in the note to
Table 7; it contains for instance flexible controls for FICO and LTV, fixed effects for MSA, loan
term, program (e.g. GSE or FHA) and purpose (purchase or refinance), as well as borrower income,
education, age, and race.
47
We further include origination month fixed effects α
t
and survey wave
45
See https://www.fhfa.gov/DataTools/Downloads/Pages/National-Survey-of-Mortgage-Originations-Public-Use-File.
aspx.
46
The market mortgage rate is measured through the Freddie Mac Primary Mortgage Market Survey (PMMS),
lagged by two weeks relative to the time of loan origination. In the public dataset, the gap is truncated at -1.5 and
+1.5 percentage points; however, we were able to run the analyses described in this section at FHFA on a version of
the data without truncation (and containing MSA indicators, which we are able to use as fixed effects). In earlier
drafts, we reported results based on the public version of the data (version as of February 12, 2019), with little
qualitative difference.
47
One limitation of the NSMO data is that it does not contain a direct measure of points paid or received by
the borrower. However, the controls for borrower wealth and expected time in the mortgage should help absorb
21
fixed effects δ
w
(since there were a few small changes to the wording of questions across waves). In
all our NSMO analyses, we use the provided analysis weights, which are based on sampling weights
and non-response adjustments.
We consider the following X
i
variables:
1. The answer to the question “How many different lenders/mortgage brokers did you seriously
consider before choosing where to apply for this mortgage?” 49.0% of respondents (weighted)
answer 1, 35.3% 2, 13.0% 3, 1.7% 4, and 1.0% 5 or more. We combine the last three groups
into “3+”.
2. The answer to “How many different lenders/mortgage brokers did you end up applying to?”
Here, 76.7% answer 1, 18.7% 2, 3.6% 3, 0.7% 4, and 0.3% 5 or more. We combine the last
four groups into “2+”.
3. Those who indicated that they applied to two or more lenders are asked which of four non-
exclusive reasons were driving the multiple applications. We create an indicator for those
who indicate that “searching for better loan terms” was a reason (81.4% of those that apply
to more than one lender, or 18.6% of the sample overall).
48
4. A series of questions are asked about nine different possible information sources the borrower
could use to get information about mortgages or mortgage lenders. For each of them, a
respondent can say they used a source “a lot”, “a little”, or “not at all”. We use the following,
which we think of as the best proxies for genuine search effort: “Other lenders or brokers”
(32.7% a little, 9.2% a lot); “Websites that provide information on getting a mortgage” (32.1%
a little, 22.2% a lot); and “Friends/relatives/co-workers” (32.0% a little, 15.1% a lot).
5. The answer to the question “When you began the process of getting this mortgage, how
familiar were you (and any co-signers) with [t]he mortgage interest rates available at that
time?” 61.7% respond “Very”, 32.9% “Somewhat”, and 5.5% “Not at all”.
6. An index of “mortgage knowledge” based on 6 responses to the questions “How well could you
explain to someone the... Process of taking out a mortgage / Difference between a fixed- and
an adjustable-rate mortgage / Difference between a prime and subprime loan / Difference
between a mortgage’s interest rate and its APR / Amortization of a loan / Consequences of
not making required mortgage payments”. In each case, the respondent picked from a three
point scale from “Not at all” (which we code as 1) to “Very” (3). We take the sum of the 6
responses and standardize it to have mean 0 and standard deviation 1.
differences in rates due to variation in points.
48
Overall, 4.6% of respondents stated that they applied to more than one lender because they got “turned down
on an earlier application”; 6.7% because of “concern over qualifying for a loan”; and 7.4% because of “information
learned from the ‘loan estimate’,” with overlap across these categories.
22
7. An indicator for whether a borrower agreed with the statement“Most mortgage lenders would
offer me roughly the same rates and fees.” This question was only added in Wave 7 and so we
only have responses for roughly half of the sample. Of those, 68.2% agree with the statement.
We think of the first four items as capturing shopping effort, while the remaining three capture
mortgage market knowledge. We first add these measures to the regression one at a time, and then
in a final specification jointly. The results are presented in Table 7. We see that most proxies for
intense shopping and better mortgage market knowledge are associated with lower mortgage rates:
for instance, considering 3+ lenders rather than just one lender is associated with a 9.5bp lower
rate, while applying to more than one lender in search of better loan terms is associated with a
7bp lower rate. Similarly, more intense use of other lenders/brokers and the web as info sources
predicts lower rates, while relying on friends, relatives and co-workers seems to have little effect.
A particularly strong predictor is familiarity with available mortgage rates at the beginning of the
process of getting the mortgage: those who state they were very familiar on average pay 20bp less
than those who say they were not at all familiar. A one-standard-deviation higher value in the
mortgage knowledge index is associated with a 6bp lower rate, while believing that all lenders offer
roughly the same rate is associated with a higher rate.
The final column controls for all X
i
jointly. As one might expect, some of the coefficients
are attenuated relative to the earlier columns, but many of them remain individually significant,
suggesting that there are different dimensions to shopping and knowledge that can contribute to
a borrower obtaining a low rate.
49
For instance, a borrower who is very familiar with market
conditions may not need to consider more than one lender, if they can negotiate a good rate purely
based on their knowledge. Conversely, shopping alone does not guarantee a good rate if a borrower’s
knowledge is low (see also Malliaris et al., 2020). Again, it is important to remember that all of
these regressions control finely for other factors that likely influence loan pricing, in order to rule
out to the extent possible that these correlations reflect omitted variables that affect loan pricing
due to default or prepayment risk.
In Appendix A.5, we provide a complementary analysis using data from the 2016 Survey of
Consumer Finances (SCF). Consistent with the NSMO results, we find that borrowers who report
shopping more, and borrowers with high financial literacy—based on their answers to the Lusardi-
Mitchell financial literacy questions—get significantly lower interest rates, even after controlling for
loan characteristics, borrower credit risk, and borrower demographics.
7.2 Who Pays More Because of a Lack of Shopping or Knowledge?
The previous subsection provides evidence that more intense mortgage shopping and more knowl-
edge about the mortgage market is associated with lower contracted rates. We next ask which
49
It is interesting to note that the coefficient on “applied to 2+ lenders” flips sign if we simultaneously control for
having applied to 2+ lenders in search of better loan terms. This likely reflects that those who applied to multiple
lenders but not in search of better terms got turned down on their previous application (or learned negative news in
the process), in line with the findings of Agarwal et al. (2019).
23
observable borrower and loan characteristics are associated with stronger reported shopping in-
tensity and mortgage knowledge, resulting in lower interest rates. To do this, we first isolate the
part of the interest rate spread that can be attributed solely to shopping and knowledge about the
mortgage market. Then, we study how this measure varies with observable characteristics.
We compute the predicted interest rate spread for each borrower using a regression almost
identical to the one in specification (10) of Table 7. The only changes we make are that we omit the
indicator for whether a borrower believed that most mortgage lenders would offer roughly the same
rates and fees (since that question is only asked in later waves), and instead of the “knowledge index”
we use each of the six underlying questions individually. All shopping and knowledge variables are
thus categorical, and for each of them we use as baseline/omitted value the one that corresponds
to the lowest level of shopping or knowledge. We thus compute for each borrower the predicted
rate spread relative to a hypothetical borrower that indicates that they did not engage in any
shopping-related activities and have the poorest possible understanding of the mortgage market.
We summarize this predicted rate spread in the top row of Table 8. Due to shopping and
mortgage knowledge, the average borrower pays 35bp less than the hypothetical non-shopping,
completely clueless borrower. Perhaps more interesting is the magnitude of the difference between
the 10th and 90th percentile, which is 26bp. This implies that there are substantive differences
across borrowers in shopping behavior and mortgage knowledge amounting to a 26bp difference in
rates paid.
If shopping and mortgage knowledge are correlated with borrower and loan characteristics, then
interest savings will differ by group. In Table 8 we also show group-specific predicted interest rate
spreads that can be attributed to shopping and mortgage knowledge. The differences across groups
are most pronounced at the lower end of the group-specific shopping/knowledge distribution. For
example, at the 10th percentile, borrowers in the jumbo market pay about 12bp less than FHA
borrowers due to shopping and mortgage knowledge, whereas the jumbo-FHA difference is about
3bp at the 90th percentile.
50
In other words, the gap in knowledge and shopping is not as big
between the most savvy FHA and jumbo borrowers as the gap between the least savvy FHA and
jumbo borrowers.
The table further shows that the predicted rate spread decreases in a borrower’s FICO score
and increases in the LTV, meaning that low-FICO and high-LTV borrowers pay higher rates due
to shopping and knowledge. The same is true for borrowers with low loan amounts.
Turning to other borrower characteristics, borrowers with incomes of $175k or higher pay less due
to shopping and knowledge than borrowers with incomes of less than $35k, with a 13bp difference
at the 10th percentile. In addition, more educated borrowers on average pay less than their less
educated counterparts, and first-time homebuyers pay more than repeat homebuyers.
The magnitudes of the differences across groups in Table 8 may appear relatively small. How-
ever, it bears remembering that the right-hand-side variables of the underlying regression are coarse
50
To be clear, the percentiles are calculated within group, so across-group differences are driven only by differences
in shopping/knowledge.
24
responses to qualitative survey questions, likely leading to substantial individual-specific noise and
attenuation of the resulting coefficients.
51
With that caveat in mind, we believe the findings here lend considerable support to the mech-
anism we postulated in our earlier analysis using the rate locks and offers data. Namely, at least
some of the overpayment by many borrowers is likely due to ineffective shopping and negotiation,
reflecting a lack of financial sophistication and knowledge of the market. Such knowledge is par-
ticularly important in this setting where, as documented earlier, there is considerable dispersion in
prices across lenders, and even across branches and loan officers of the same lender. Furthermore,
comparing offers is complicated due to the multi-dimensional pricing with upfront points.
7.3 Time-series Variation in Shopping Intensity
Earlier, we saw that the locked-offer rate gap in the Optimal Blue data decreases when market
interest rates are higher, even for borrowers who do not appear constrained, and speculated that
this may partly be driven by increased shopping intensity when interest rates are higher. The
NSMO enables us to test this hypothesis directly. We estimate linear probability models of the
form:
Shopping
ijtw
= β · P MMS
it
+ ΓZ
ij
+ δ
w
+
ijtw
(2)
where Shopping
ijtw
is a binary measure of shopping intensity (discussed below) by borrower i with
loan characteristics j, loan origination month t and responding to survey wave w. P MMS
it
is our
main variable of interest, the market mortgage rate two weeks prior to loan origination. Z
ij
are
borrower and mortgage characteristics, including the measures of borrowers’ mortgage knowledge
discussed above. Finally, δ
w
are survey wave fixed effects.
As dependent variable, we use binary versions of the four main shopping variables that were
associated with lower contract interest rates in Table 7: (i) whether a borrower seriously considered
at least two lenders; (ii) whether a borrower applied to at least two lenders in search of better
terms; (iii) whether a borrower used other lenders/brokers to get information “a little” or “a lot”;
and (iv) whether a borrower used websites that provide information on getting a mortgage “a little”
or “a lot”. For each of these variables, we report regressions without other covariates (except for
survey wave fixed effects) and with the same covariates as in Table 7, except for some variables
that seem likely endogenous to the shopping effort itself.
52
Furthermore, we add the knowledge
variables used in Table 7 as well.
Panel A of Table 9 reports the results of these regressions for the full sample. We see that across
the different measures, a higher level of market mortgage rates is associated with more shopping
effort, in most cases in a statistically significant way. For instance, column (1) implies that a
1 percentage point increase in market mortgage rates increases the probability that a borrower
51
For instance, respondents likely differ in what they view as using an information source “a lot” vs. “a little”, or
being “very” vs. “somewhat” familiar with a topic.
52
These variables are whether a borrower obtained their mortgage through a broker, the term of the loan, and
whether it has an adjustable rate. We also do not include MSA fixed effects, though adding them has minimal effects.
25
considered more than one lender by 4.5 percentage points, relative to a sample average of 51
percent.
53
Column (2) shows that this coefficient is unaffected by the addition of fine borrower-
and loan-level controls, which alleviates concerns that the relationship is driven by variation in the
type of borrower that applies at different points in time (and at different levels of market rates).
The effect on the probability of applying to multiple lenders is even substantially larger, es-
pecially compared to the sample mean (which is only 19 percent). A higher PMMS rate is also
significantly associated with borrowers reporting that they obtained information from other lenders
or brokers. The association with using websites to provide information on getting a mortgage is
also positive, but not statistically significant.
Panels B to D assess the robustness of these findings in different subsamples. First, panel B
shows that the estimated coefficients remain very similar if we restrict the sample to purchase loans;
this alleviates concerns that the finding is driven by changing composition between purchase and
refinance mortgages as market rates change. Panels C and D then restrict the sample to borrowers
that are objectively or subjectively unconstrained by affordability constraints (which, if binding,
could “force” borrowers to shop more). In panel C, we only use borrowers whose debt-to-income
ratio ends up below 36 percent, suggesting that they had additional room to make larger payments.
In panel D, we restrict the sample to borrowers who responded “not at all” to the question “when
you began the process of getting this mortgage, how concerned were you about qualifying for a
mortgage?” In both subsamples, the estimated coefficients remain positive, and for the first two
shopping measures statistically significant. Thus, it does not appear that the positive relationship
between market interest rates and shopping is mainly driven by affordability constraints.
In Appendix A.6, we further complement this analysis by documenting univariate and multivari-
ate correlations between the shopping and knowledge measures, as well as between these measures
and various borrower and loan characteristics.
8 Conclusion and Policy Implications
Our empirical results provide evidence that many borrowers from the most vulnerable part of
the borrower population in the US seem to overpay for mortgages: those that are most likely
to be relatively low income, low net worth, and more likely to be first-time homebuyers. These
are the exact borrowers that various government programs effectively subsidize. If they were to
obtain mortgages from the lower end of the offer distribution, this would make their mortgage
payments more affordable and leave them with more disposable income. Alternatively, the FHA
and the GSEs could afford to raise their guarantee fees substantially without affecting final cost to
borrowers. The involved dollar amounts in this scenario are large not just at the individual level
but also in aggregate: for instance, if the average locked-offer rate gap of FHA borrowers moved
to zero (assuming nothing else changes in the market structure), this would amount to savings of
53
Over our sample period, the market mortgage rate as measured by PMMS varied from 3.31% to 4.58%.
26
about $2.75 billion/year for these borrowers.
54
Given our findings, future research should consider the effects of policies that would help bor-
rowers search and negotiate more effectively. This could take the form of required information
disclosure to borrowers of the rates available to them across different lenders in the same market
(for instance at the time they lock their rate). We recognize that this is not a straightforward
endeavor given the multi-dimensional nature of mortgage pricing in the US, but advances in tech-
nology may make this more feasible than in the past. Alternatively, future research could study
whether the problem can be alleviated if the guaranteeing agencies were to impose requirements on
the maximum locked-offer gap they allow for loans to be securitized. Of course, to understand the
effectiveness of such policies one would need to consider general equilibrium effects on the offers
that lenders make (as in Alexandrov and Koulayev 2017 and Agarwal et al. 2019).
The negative relationship between the average locked-offer rate gap and the level of market rates
that we document in Section 6 also matters for monetary policy transmission. Our findings imply
that as rates fall (e.g. in response to central bank actions), borrowers tend to do worse relative to
the rates available in the market, likely at least in part due to less shopping or negotiation. It follows
that the contract rates they end up with do not fall as much as they could, based on lenders offers,
adding another friction to the pass-through of expansive monetary policy to the mortgage market.
55
On the other hand, the pass-through of increases in policy rates to rates on new mortgages may
be dampened by more intense borrower shopping. This could be good or bad news for monetary
policy makers, depending on whether slowing the housing market through higher mortgage rates is
seen as desirable in a given situation or not.
54
This calculation is based on average FHA originations over 2015-2019 of about $230bn/year (see https://www.
hud.gov/sites/dfiles/Housing/documents/FHA_SF_MarketShare_2019Q3.pdf) multiplied by 1.2 points, which is
the upfront equivalent of the average locked-offer rate gap of +25 bp that we documented.
55
Existing work has shown that offers (as measured from investor rate sheets) respond less to increases in MBS
prices than to decreases, and less so when borrower demand is already high, which happens after falls in rates (Fuster
et al., 2017). Limited competition may also limit pass-through (Agarwal et al., 2017; Scharfstein and Sunderam,
2016). Finally, many borrowers fail to refinance when it is in their financial interest to do so (e.g., Campbell, 2006;
Andersen et al., 2020; Keys et al., 2016).
27
References
Agarwal, Sumit, Gene Amromin, Souphala Chomsisengphet, Tim Landvoigt, Tomasz Pisko-
rski, Amit Seru, and Vincent Yao, “Mortgage Refinancing, Consumer Spending, and Competition:
Evidence From the Home Affordable Refinance Program,” NBER Working Paper 21512 2017.
, John Grigsby, Ali Horta¸csu, Gregor Matvos, Amit Seru, and Vincent Yao, “Searching for
Approval,” Working Paper 2019.
Alexandrov, Alexei and Sergei Koulayev, “No Shopping in the U.S. Mortgage Market: Direct and
Strategic Effects of Providing Information,” Working Paper, CFPB 2017.
Allen, Jason, Robert Clark, and Jean-Francois Houde, “Price Dispersion in Mortgage Markets,” The
Journal of Industrial Economics, 2014, 62 (3), 377–416.
Andersen, Steffen, John Y. Campbell, Kasper Meisner Nielsen, and Tarun Ramadorai, “Sources
of Inaction in Household Finance: Evidence from the Danish Mortgage Market,” American Economic
Review, 2020, forthcoming.
Argyle, Bronson, Taylor Nadauld, and Christopher Palmer, “Real Effects of Search Frictions in
Consumer Credit Markets,” Working Paper 2017.
Bhutta, Neil and Aurel Hizmo, “Do Minorities Pay More for Mortgages?,” Review of Financial Studies,
2020, forthcoming.
Buchak, Greg, Gregor Matvos, Tomasz Piskorski, and Amit Seru, “Fintech, Regulatory Arbitrage,
and the Rise of Shadow Banks,” Journal of Financial Economics, 2018, 130, 453–483.
Campbell, John Y., “Household Finance,” Journal of Finance, 2006, 61 (4), 1553 1604.
Carlson, John A. and R. Preston McAfee, “Discrete Equilibrium Price Dispersion,” Journal of Political
Economy, 1983, 91 (3), 480–493.
Choi, James J., David Laibson, and Brigitte C. Madrian, “Why Does the Law of One Price Fail?
An Experiment on Index Mutual Funds,” Review of Financial Studies, 2010, 23 (4), 1405–1432.
Damen, Sven and Erik Buyst, “Mortgage Shoppers: How Much Do They Save?,” Real Estate Economics,
2017, 45 (4), 898–929.
DeFusco, Anthony A., Stephanie Johnson, and John Mondragon, “Regulating Household Leverage,”
Review of Economic Studies, 2020, 87 (2), 914–958.
Duncan, Doug, “Shopping Around for a Mortgage Pays Off for Consumers,” Fannie
Mae Perspectives, https://www.fanniemae.com/portal/research-insights/perspectives/
shopping-mortgage-consumers-duncan-080519.html August 2019.
Fuster, Andreas, Laurie Goodman, David Lucca, Laurel Madar, Linsey Molloy, and Paul
Willen, “The Rising Gap between Primary and Secondary Mortgage Rates,” Federal Reserve Bank of
New York Economic Policy Review, 2013, 19 (2), 17–39.
, Stephanie H. Lo, and Paul S. Willen, “The Time-Varying Price of Financial Intermediation in the
Mortgage Market,” Staff Report 805, Federal Reserve Bank of New York 2017.
28
Gathergood, John and J
¨
org Weber,“Financial literacy, present bias and alternative mortgage products,”
Journal of Banking & Finance, 2017, 78, 58–83.
Gomes, Francisco, Michael Haliassos, and Tarun Ramadorai, “Household Finance,” Journal of
Economic Literature, 2020, forthcoming.
Goodman, Laurie S, “Quantifying the tightness of mortgage credit and assessing policy actions,” BCJL
& Soc. Just., 2017, 37, 235.
Gurun, Umit G., Gregor Matvos, and Amit Seru, “Advertising Expensive Mortgages,” The Journal
of Finance, 2016, 71 (5), 2371–2416.
Hastings, Justine S., Brigitte C. Madrian, and William L. Skimmyhorn, “Financial Literacy,
Financial Education, and Economic Outcomes,” Annual Review of Economics, 2013, 5 (1), 347–373.
Horta¸csu, Ali and Chad Syverson, “Product Differentiation, Search Costs, and Competition in the
Mutual Fund Industry: A Case Study of S&P 500 Index Funds,” Quarterly Journal of Economics, 2004,
119 (2), 403–456.
Huston, Sandra J., “Financial literacy and the cost of borrowing,” International Journal of Consumer
Studies, 2012, 36 (5), 566–572.
Iscenko, Zanna, “Choices of dominated mortgage products by UK consumers,” Occasional Paper 33, Fi-
nancil Conduct Authority 2018.
Keys, Benjamin, Devin Pope, and Jaren Pope, “Failure to Refinance,”Journal of Financial Economics,
2016, 122 (3), 482 499.
Liu, Lu, “Non-Salient Fees in the Mortgage Market,” Working Paper, Imperial College London 2019.
Lusardi, Annamaria and Olivia S. Mitchell, “The Economic Importance of Financial Literacy: Theory
and Evidence,” Journal of Economic Literature, March 2014, 52 (1), 5–44.
Malliaris, Steven G., Daniel A. Rettl, and Ruchi Singh, “Is Competition a Cure for Confusion?
Evidence from the Residential Mortgage Market,” Working Paper, University of Georgia 2020.
McManus, Doug, Liyi Liu, and Mingzhe Yi, “Why Are Consumers Leaving Money On The Ta-
ble?,” Freddie Mac Economic & Housing Research Insight, http: // www. freddiemac. com/ research/
insight/ 20180417_ consumers_ leaving_ money. page , 2018.
Scharfstein, David and Adi Sunderam, “Market Power in Mortgage Lending and the Transmission of
Monetary Policy,” Working Paper, Harvard University 2016.
Stango, Victor and Jonathan Zinman, “Borrowing High versus Borrowing Higher: Price Dispersion and
Shopping Behavior in the U.S. Credit Card Market,” Review of Financial Studies, 2016, 29 (4), 979–1006.
Woodward, Susan E. and Robert E. Hall, “Diagnosing Consumer Confusion and Sub-optimal Shopping
Effort: Theory and Mortgage-Market Evidence,” American Economic Review, 2012, 102 (7), 3249–76.
Zinman, Jonathan, “Household Debt: Facts, Puzzles, Theories, and Policies, Annual Review of Eco-
nomics, 2015, 7 (1), 251–276.
29
Table 1: Summary Statistics of the Rate Lock Data
Conforming Super-Conforming Jumbo FHA
Mean St. Dev. Mean St. Dev. Mean St. Dev. Mean St. Dev.
Loan Amount (000) 255 94 544 71 720 262 222 92
Interest Rate 4.33 0.51 4.31 0.47 4.21 0.50 4.30 0.61
Discount Points Paid 0.15 0.95 0.28 0.97 0.19 0.74 0.06 1.14
FICO 742 47 750 41 763 33 669 47
LTV 81 14 80 12 77 10 93 8
DTI 35 9 36 9 31 9 42 10
First-time Homebuyer % 24 23 11 49
Refinance Share % 31 33 33 17
N. observations 2,316,400 119,894 76,941 1,092,535
Data Source: Optimal Blue
30
Table 2: Unpacking the Dispersion in Locked Interest Rates
Underwriting Grid Add Lender Controls Add Branch Controls Add LO Controls
(1) (2) (3) (4) (5) (6) (7) (8) (9) (10)
Standard Deviation 0.33 0.26 0.24 0.22 0.20 0.18 0.16 0.15 0.14 0.13
75-25 Percentile 0.36 0.28 0.26 0.22 0.20 0.17 0.15 0.14 0.13 0.12
90-10 Percentile 0.78 0.58 0.54 0.48 0.44 0.38 0.33 0.31 0.29 0.26
Underwriting Variables Grid
Lock Date x MSA F.E. Yes Yes Yes Yes Yes Yes Yes Yes Yes Yes
FICO x LTV x Program x Lock Month F.E. Yes Yes Yes Yes Yes Yes Yes Yes Yes
ZIP Code F.E. Yes Yes Yes Yes Yes Yes Yes Yes Yes
Discount Points x Program x Lock Month F.E. Yes Yes Yes Yes Yes Yes Yes Yes
Add Lender Controls
Lender F.E. Yes Yes Yes Yes Yes Yes Yes
Lender x Lock Date F.E. Yes Yes Yes Yes Yes Yes
Lender x FICO x LTV x Program x Lock Month F.E. Yes Yes Yes Yes Yes
Lender x Points x Lock Month F.E. Yes Yes Yes Yes Yes
Add Branch Controls
Branch F.E. Yes Yes Yes Yes
Branch x Lock Month F.E. Yes Yes Yes
Add Loan Officer Controls
Loan Officer F.E. Yes Yes
Loan Officer x Program F.E. Yes
Loan Officer x Lock Year F.E. Yes
Adj. R-Squared 0.59 0.75 0.78 0.81 0.83 0.85 0.88 0.89 0.89 0.90
Observations 2959539 2959539 2959539 2959539 2959539 2959539 2959539 2959539 2959539 2959539
Data Source: Optimal Blue
Notes: The dependent variable is the mortgage interest rate locked. The data covers mortgage rates locked for 277 metropolitan areas during the period between 2015-2019. We focus on 30
year, fixed rate, fully documented mortgages. “Program” refers to 12 dummy variables representing four loan programs interacted with three loan purposes. Specifications (2)-(10) also include lock
period f.e., property type f.e., cubic functions of loan amount and DTI, as well as linear functions of FICO, LTV, and (from specification (3) onward) discount points. For MSAs that span across
multiple states we include MSA x State fixed effects.
31
Table 3: Summary Statistics of the Residualized Locked Rate
Observations
90
th
10
th
Percentile Gap
Spec. (3) of Table 2 Spec. (10) of Table 2
All Mortgages 2,959,539 0.54 0.26
Program
FHA 876,640 0.71 0.31
Conforming 1,972,913 0.47 0.24
Super-Conforming 72,038 0.43 0.21
Jumbo 37,948 0.47 0.24
FICO
< 600 41,836 0.92 0.46
600, 640
219,492 0.81 0.36
640, 680
487,004 0.67 0.30
680, 740
921,992 0.55 0.27
740 1,289,215 0.44 0.22
LTV
75 500,902 0.46 0.22
75, 80
619,803 0.46 0.23
80, 95
951,705 0.50 0.25
>95 837,151 0.71 0.32
First-Time Homebuyer
No 1,981,043 0.50 0.24
Yes 978,020 0.64 0.31
Loan Purpose
Purchase 2,269,213 0.55 0.27
Cashout 344,735 0.56 0.25
Rate Refi 345,591 0.50 0.23
Lender Type
Independent Non-bank 1,878,780 0.54 0.26
Other/Unclassified 1,080,759 0.54 0.26
Data Source: Optimal Blue
Notes: This table summarizes the residualized locked mortgage rate from specifications (3) and (10) of Table 2.
32
Table 4: Summary Statistics of the Locked-Offer Rate Gap
Observations Mean St. Deviation
Percentiles
25
th
75
th
All Mortgages 64,788 0.11 0.31 -0.07 0.26
Program
FHA 14,441 0.25 0.38 0.02 0.45
Conforming 44,040 0.09 0.27 -0.06 0.22
Super-Conforming 4,478 -0.04 0.26 -0.20 0.09
Jumbo 1,829 -0.21 0.32 -0.34 -0.06
FICO
640, 660
7,406 0.23 0.40 -0.02 0.45
680, 700
9,390 0.16 0.36 -0.05 0.35
720, 740
10,207 0.12 0.30 -0.06 0.26
740+ 37,785 0.07 0.27 -0.08 0.21
LTV
75, 80
21,334 0.02 0.26 -0.12 0.16
85, 90
6,882 0.06 0.28 -0.08 0.20
90, 95
15,782 0.08 0.27 -0.08 0.22
95, 97
20,790 0.24 0.36 0.01 0.42
First-Time Homebuyer
No 32,437 0.07 0.28 -0.09 0.21
Yes 32,345 0.15 0.34 -0.05 0.32
Discount Points
-5, -0.2
14,015 0.01 0.33 -0.16 0.19
-0.2, 0.2
22,735 0.08 0.29 -0.09 0.22
0.2, 5
28,038 0.18 0.30 -0.00 0.31
Lender Type
Independent Nonbank 45,618 0.13 0.30 -0.04 0.27
Other/Unclassified 19,170 0.05 0.32 -0.13 0.23
Data Source: Optimal Blue
Notes: For each mortgage rate locked by borrowers in our data, we compute the median rate
offered by lenders in the same market on the same day for an identical mortgage. This table
summarizes the difference between each locked rate and the median offer rate (the “locked-offer rate
gap”). In the discount points category, negative values mean that the borrower receives points (also
known as a rebate or credit) while positive values mean that the borrower pays points.
33
Table 5: Regressions of the Locked-Offer Rate Gap on Borrower/Loan Characteristics, Lender-Branch Fixed Effects,
and Loan Officer Compensation
(1) (2) (3) (4) (5) (6) (7) (8)
FICO (omitted cat.: [640,660))
I
680F ICO<700
-0.056
∗∗∗
-0.044
∗∗∗
-0.043
∗∗∗
(0.007) (0.006) (0.011)
I
720F ICO<740
-0.088
∗∗∗
-0.059
∗∗∗
-0.058
∗∗∗
(0.010) (0.008) (0.013)
I
F ICO740
-0.123
∗∗∗
-0.080
∗∗∗
-0.071
∗∗∗
(0.011) (0.009) (0.013)
LTV (omitted cat.: (60,80])
I
85<LT V 90
0.017
∗∗∗
0.009 0.018
∗∗
(0.005) (0.006) (0.008)
I
90<LT V 95
0.049
∗∗∗
0.033
∗∗∗
0.031
∗∗∗
(0.006) (0.006) (0.008)
I
LT V >95
0.178
∗∗∗
0.138
∗∗∗
0.101
∗∗∗
(0.012) (0.011) (0.019)
Discount Points
I
5<P oints<0.2
-0.094
∗∗∗
0.001
(0.020) (0.006)
I
0.2<P oints5
0.108
∗∗∗
0.034
∗∗∗
(0.013) (0.012)
Loan Officer Comp (%) 0.158
∗∗∗
0.140
∗∗∗
(0.036) (0.039)
Loan amount f.e. ($10k bins) Yes Yes Yes Yes Yes Yes Yes Yes
MSA x Month f.e. Yes Yes Yes Yes Yes Yes Yes Yes
Branch f.e. Yes Yes Yes Yes Yes
Adj. R-Squared 0.116 0.480 0.442 0.154 0.505 0.456 0.157 0.476
Observations 64693 62783 14659 64693 62783 14659 64693 62783
Data Source: Optimal Blue
Notes: The dependent variable is the mortgage interest rate locked minus the median offer rate in the same market and day for an identical
mortgage. The data covers mortgage rates for 20 metropolitan areas during the period between 2016-2019. We focus on 30 year, fixed rate,
fully documented purchase mortgages. Standard errors shown in parentheses are two-way clustered at the month and lender level. Significance:
* p<0.1, ** p<0.05, *** p<0.01.
34
Table 6: The Relationship Between the Locked-Offer Gap and Treasury Yields
(1) (2) (3) (4) (5) (6) (7) (8)
Treasury Yield -0.059
∗∗∗
-0.083
∗∗∗
-0.057
∗∗∗
-0.058
∗∗∗
(0.007) (0.015) (0.014) (0.014)
Offer Spread to Prime Conforming Rate -0.176
∗∗∗
-0.177
∗∗∗
(0.026) (0.026)
Treasury Yield ×
DTI > 36 -0.066
∗∗∗
-0.090
∗∗∗
-0.064
∗∗∗
(0.008) (0.016) (0.015)
DTI 36 -0.049
∗∗∗
-0.074
∗∗∗
-0.046
∗∗∗
(0.009) (0.015) (0.012)
Borrower and Loan Controls Yes Yes Yes Yes Yes Yes Yes Yes
MSA F.E Yes Yes Yes Yes Yes Yes Yes Yes
MSA x Month F.E. Yes Yes Yes Yes Yes Yes
Branch F.E. Yes Yes Yes Yes
Adj. R-Squared 0.145 0.156 0.505 0.147 0.157 0.505 0.509 0.508
Observations 64397 64317 62398 64397 64317 62398 62783 62398
P-val. for equality of DTI coefficients 0.024 0.031 0.010
Data Source: Optimal Blue
Notes: The dependent variable is the mortgage interest rate locked minus the median offer rate in the same market and day for an identical mortgage.
The offer spread to conforming rate is defined as the average offer rate for a typical borrower in the same program in the same day minus the average offer
rate for a typical prime conforming borrower. All specifications include controls for FICO, LTV, and loan amount. The data covers mortgage rates for 20
metropolitan areas during the period between 2016-2019. We focus on 30 year, fixed rate, fully documented purchase mortgages. Standard errors shown in
parentheses are two-way clustered at the month and lender level. Significance: * p<0.1, ** p<0.05, *** p<0.01.
35
Table 7: Relationship Between Mortgage Rates and Measures of Shopping and Knowledge
Dep. var.: Interest rate spread (%) (1) (2) (3) (4) (5) (6) (7) (8) (9) (10)
Seriously considered 2 lenders -0.047*** -0.021
(0.012) (0.013)
Seriously considered 3+ lenders -0.095*** -0.049***
(0.015) (0.018)
Applied to 2+ lenders -0.052*** 0.052*
(0.013) (0.029)
Applied to 2+ lenders in search of better loan terms -0.073*** -0.092***
(0.014) (0.031)
Used other lenders/brokers to get info? A little -0.030** -0.001
(0.012) (0.013)
Used other lenders/brokers to get info? A lot -0.071*** -0.022
(0.019) (0.020)
Used web to get info? A little -0.053*** -0.042***
(0.012) (0.013)
Used web to get info? A lot -0.078*** -0.043***
(0.015) (0.015)
Used friends/relatives to get info? A little 0.001 0.004
(0.013) (0.013)
Used friends/relatives to get info? A lot 0.009 0.012
(0.018) (0.018)
Familiar with mortgage rates? Somewhat -0.099*** -0.078**
(0.033) (0.033)
Familiar with mortgage rates? Very -0.197*** -0.145***
(0.033) (0.033)
Index of mortgage knowledge (Std) -0.061*** -0.044***
(0.006) (0.006)
Most lenders offer same rate? Yes 0.033** 0.024
(0.016) (0.016)
Adj. R2 0.14 0.13 0.13 0.13 0.14 0.13 0.14 0.14 0.13 0.14
Obs. 19824 19824 19824 19824 19824 19824 19824 19824 19824 19824
Data Source: National Survey of Mortgage Originations
Dependent variable: spread between a borrower’s mortgage interest rate and the market mortgage rate prior to origination (as measured by PMMS), in percentage
points. Sample restricted to first-lien loans (without a junior lien) for single-family principal residence properties, with no more than two borrowers, and a loan
term of 10, 15, 20 or 30 years. Observations weighted by NSMO sample weights. All regressions control for origination month fixed effects, survey wave fixed effects,
MSA fixed effects, FICO score (linear term plus dummies for 11 FICO bins), LTV (linear term plus dummies for each percentage point from 79-98), indicators
for loan purpose (purchase, refinance, or cash-out refinance), 9 loan amount categories, loan program (Freddie, Fannie, FHA, VA, FSA/RHS, other), loan term,
first-time homebuyer status, single borrowers, using a mortgage broker, whether the loan has an adjustable rate, jumbo status, 6 borrower income categories, 6
borrower education categories, whether the household owns 4 different types of financial assets, race and ethnicity, metropolitan CRA low-to-moderate income
tract status, borrower age and gender, and self-assessed creditworthiness, likelihood of moving, selling, or refinancing, and risk aversion. N =19,824 instead of
19,906 because singleton observations are dropped. Robust standard errors in parentheses. * p<0.1, ** p<0.05, *** p<0.01.
36
Table 8: Summary Statistics of the Interest Rate Spread that Can Be Attributed to Shopping and
Mortgage Knowledge
Observations 10
th
Percentile Mean 90
th
Percentile
All Mortgages 19,906 -0.21 -0.35 -0.47
Program
Conforming 11,103 -0.22 -0.36 -0.47
Jumbo 679 -0.29 -0.39 -0.48
FHA 2,734 -0.17 -0.32 -0.45
FICO
600 411 -0.17 -0.31 -0.45
601-640 1,089 -0.18 -0.32 -0.45
641-680 2,195 -0.20 -0.34 -0.46
681-740 4,784 -0.20 -0.34 -0.46
> 740 11,427 -0.24 -0.36 -0.47
LTV
75 8,216 -0.23 -0.36 -0.47
76-80 3,551 -0.23 -0.36 -0.47
81-95 4,551 -0.21 -0.35 -0.47
96-97 1,805 -0.15 -0.31 -0.44
Loan Amount
<100k 3,011 -0.17 -0.32 -0.45
[100k, 200k) 7,736 -0.21 -0.34 -0.46
[200k, 300k) 4,656 -0.22 -0.36 -0.47
[300k, 400k) 2,405 -0.24 -0.37 -0.48
400k 2,098 -0.27 -0.38 -0.48
First-Time Homebuyer
No 16,717 -0.23 -0.36 -0.47
Yes 3,189 -0.15 -0.31 -0.46
Income
<35k 1,189 -0.14 -0.30 -0.43
[35k, 75k) 6,014 -0.18 -0.32 -0.45
[75k, 175k) 9,752 -0.23 -0.36 -0.47
175k 2,951 -0.27 -0.39 -0.48
Education
Less than college 3,322 -0.17 -0.31 -0.44
Some college 3,975 -0.20 -0.34 -0.46
College grad 7,017 -0.22 -0.36 -0.47
Postgrad 5,592 -0.24 -0.37 -0.48
Data Source: National Survey of Mortgage Originations
Notes: The variable we are summarizing here is the interest rate spread that is only due to shopping and knowledge
about the mortgage market (so a more negative value is better from the perspective of a borrower). We compute the
predicted value of the interest rate spread using only the displayed variables on shopping behavior and knowledge
about mortgages, in a way similar to specification (10) of Table 7 (see text for details).
37
Table 9: Relationship Between Various Binary Measures of Mortgage Shopping and Mortgage
Market Interest Rates (PMMS).
Considered 2+ lenders Applied to 2+ lenders Used other lenders Used web
for better terms to get info to get info
A. Full sample (1) (2) (3) (4) (5) (6) (7) (8)
PMMS rate 0.045** 0.045** 0.069*** 0.062*** 0.048*** 0.050*** 0.019 0.026
(0.018) (0.018) (0.014) (0.014) (0.018) (0.018) (0.018) (0.018)
Controls? No Yes No Yes No Yes No Yes
Mean of Dependent Variable 0.510 0.510 0.190 0.190 0.418 0.418 0.533 0.533
Obs. 19906 19906 19906 19906 19906 19906 19906 19906
B. Purchase loans Considered 2+ lenders Applied to 2+ lenders Used other lenders to get info
PMMS rate 0.060** 0.054* 0.077*** 0.072*** 0.049* 0.041 0.014 0.009
(0.029) (0.029) (0.024) (0.024) (0.029) (0.028) (0.029) (0.027)
Controls? No Yes No Yes No Yes No Yes
Mean of Dependent Variable 0.534 0.534 0.223 0.223 0.430 0.430 0.550 0.550
Obs. 9254 9254 9254 9254 9254 9254 9254 9254
C. DTI 36 Considered 2+ lenders Applied to 2+ lenders Used other lenders to get info
PMMS rate 0.039 0.045* 0.081*** 0.074*** 0.029 0.036 0.003 0.016
(0.025) (0.025) (0.019) (0.019) (0.025) (0.025) (0.025) (0.024)
Controls? No Yes No Yes No Yes No Yes
Mean of Dependent Variable 0.503 0.503 0.176 0.176 0.411 0.411 0.541 0.541
Obs. 10590 10590 10590 10590 10590 10590 10590 10590
D. Not concerned about qualif. Considered 2+ lenders Applied to 2+ lenders Used other lenders to get info
PMMS rate 0.041* 0.045* 0.060*** 0.052*** 0.023 0.031 -0.005 0.014
(0.024) (0.024) (0.018) (0.017) (0.024) (0.023) (0.024) (0.023)
Controls? No Yes No Yes No Yes No Yes
Mean of Dependent Variable 0.488 0.488 0.165 0.165 0.387 0.387 0.499 0.499
Obs. 11203 11203 11203 11203 11203 11203 11203 11203
Data Source: National Survey of Mortgage Originations
Notes: Sample restricted to first-lien loans (without a junior lien) for single-family principal residence properties, with
no more than two borrowers, and a loan term of 10, 15, 20 or 30 years. All four dependent variables are binary. All
regressions control for survey wave fixed effects and use NSMO analysis weights. The multivariate regressions (even
columns) further control for FICO score (linear term plus dummies for 11 FICO bins), LTV (linear term plus dummies
for each percentage point from 79-98), indicators for loan purpose (purchase, refinance, or cash-out refinance), 9 loan
amount categories, loan program (Freddie, Fannie, FHA, VA, FSA/RHS, other), first-time homebuyer status, single
borrowers, jumbo status, 6 borrower income categories, 6 borrower education categories, whether the household
owns 4 different types of financial assets, race and ethnicity, metropolitan CRA low-to-moderate income tract status,
borrower age and gender, and self-assessed creditworthiness, likelihood of moving, selling, or refinancing, and risk
aversion. Robust standard errors in parentheses. * p<0.1, ** p<0.05, *** p<0.01.
38
0
.5
1
1.5
2
Density
-.7 -.6 -.5 -.4 -.3 -.2 -.1 0 .1 .2 .3 .4 .5 .6 .7
Offered Rate Minus Median Rate
Figure 1: Offer Dispersion for Identical Mortgages
Data Source: Optimal Blue
Note: Figure shows the distribution of real-time offered interest rates, where for each offer rate we subtract the median offered
rate across lenders for an identical mortgage in the same metropolitan area. The histogram includes data between April 2016
and December 2019 from 20 metropolitan areas for 52 combinations of loan characteristics (FICO, LTV, program, loan
amount).
39
0
.05
.1
.15
Fraction
-1 -.8 -.6 -.4 -.2 0 .2 .4 .6 .8 1 1.2 1.4 1.6 1.8
Locked Rate minus Median Offer Rate
All Programs
0
.05
.1
.15
Fraction
-1 -.8 -.6 -.4 -.2 0 .2 .4 .6 .8 1 1.2 1.4 1.6 1.8
Locked Rate minus Median Offer Rate
Conventional Conforming Mortgages
0
.02
.04
.06
.08
Fraction
-1 -.8 -.6 -.4 -.2 0 .2 .4 .6 .8 1 1.2 1.4 1.6 1.8
Locked Rate minus Median Offer Rate
FHA Mortgages
0
.05
.1
.15
Fraction
-1 -.8 -.6 -.4 -.2 0 .2 .4 .6 .8 1 1.2 1.4 1.6 1.8
Locked Rate minus Median Offer Rate
Super-Conforming Mortgages
0
.05
.1
.15
.2
Fraction
-1 -.8 -.6 -.4 -.2 0 .2 .4 .6 .8 1 1.2 1.4 1.6 1.8
Locked Rate minus Median Offer Rate
Jumbo Mortgages
Figure 2: Distribution of Rate Locked Minus the Median Offer Rate for Identical Mortgages
Data Source: Optimal Blue
Note: For each mortgage rate locked by borrowers in our data, we compute the median rate offered by lenders in the
same market on the same day for an identical mortgage. This figure shows the distribution of the difference between
each locked rate and the median offer rate. The dashed line denotes the mean of the distribution.
40
0
.05
.1
.15
.2
.25
.3
.35
.4
[640,660) [680,700) [720,740) 740+
FICO
Offer Data
Locked-Offer Rate Gap
FICO coefficients and 95% CI
-.1
-.05
0
.05
.1
.15
.2
.25
.3
.35
.4
(75,80] (85,90] (90,95] (95,97]
LTV
Offer Data
Locked-Offer Rate Gap
LTV coefficients and 95% CI
Figure 3: Comparing Locked-Offer Rate Gap to Offer Differences for Different FICO and LTV
Levels
Data Source: Optimal Blue
Note: The red squares plot the coefficients on FICO bins and LTV bins from columns (1) and (4) of Table 5, where
the dependent variable is the locked-offer rate gap. The black circles are the corresponding coefficients from a
regression where the dependent variable is the offer rate (and where program fixed effects are also included).
41
Locked-Offer Rate Gap (Left)
10 Year Treasury Yield (Right)
1.5
2
2.5
3
3.5
Percent
.05
.1
.15
.2
Percentage Points
2016m1 2017m1 2018m1 2019m1 2020m1
Figure 4: The Evolution of Rate Locked Minus the Median Offer Rate and Treasury Yields
Data Source: Optimal Blue
Note: The dashed red line is the 10 year Treasury yield. The solid black line is the monthly average locked-offer gap
after controlling for borrower and loan characteristics. For the average locked-offer gap, we use the estimated month
fixed effects from a regression similar to those in Table 5 but controlling simultaneously for FICO, LTV, loan
amount, and MSA f.e..
42
Online Appendix for
Paying Too Much? Price Dispersion in the US Mortgage Market
A.1 Comparing Offer and Locked Interest Rates in Optimal Blue
to Other Data Sources
In this section we assess whether the interest rates we observe in the Optimal Blue data align with
other data sources. To begin, we compare median offer rates from Optimal Blue to offer rates
from Mortgage News Daily (MND) for various 30-year fixed-rate loan programs. MND uses several
sources of information to estimate typical offer rates, including directly obtaining rate sheets from
the largest lenders. The three panels in Figure A-2 plot median offer rates from Optimal Blue
against MND’s offer rates for conforming, FHA, and jumbo mortgages, respectively. In the top two
panels, the Optimal Blue median offer rates for conforming and FHA loans—which are the bulk of
our data—move almost in lockstep with the MND offer rates. For jumbo loans, the Optimal Blue
median offer rate exhibits a little more variation from trough to peak, but on average the level is
quite similar. Overall, these results help establish that our median offer rates from Optimal Blue
are representative of the overall market.
Next, we compare lock rates to interest rates on closed mortgages. A concern with the locks
data is that high and low lock rates may systemically be less likely to actually proceed all the way
to origination. For example, borrowers who lock in a high rate at one lender may continue to shop
around and ultimately find a better rate.
The top panel of Table A-1 compares unconditional distributions of interest rates from the
Optimal Blue locks data with interest rate distributions from other administrative data sources
on closed mortgages, by loan type. If the Optimal Blue locks are representative of closed loans,
then the rate distributions across these datasets should be very similar. The first three columns
compare distributions for FHA loans locked or closed in 2014-15. For these years, we have access
to administrative data from the Department of Housing and Urban Development (HUD) on the
universe of originated FHA loans, which serves as an ideal benchmark. In addition, we compare
the locks data to well-known and widely used Black Knight McDash servicing data, which contains
loans serviced by the largest mortgage servicers in the US. We can see in the top left portion of
Table A-1 that average and 90th percentile locked rates line up identically to both the HUD and
McDash data. Moreover, the HUD and McDash data are slightly lower at the 10th percentile,
suggesting an even wider distribution than in Optimal Blue. Table A-1 also indicates that the
distribution of FICO scores and LTVs in Optimal Blue almost mirrors the HUD data, whereas the
LPS data are skewed slightly toward less risky borrowers.
The remaining columns compare Optimal Blue locks to McDash loans in 2016-18, separately
for FHA, conforming, and jumbo loans. The most notable difference is for jumbo loans, where
we observe higher interest rates in Optimal Blue by 30-40bp, although the amount of dispersion
is similar to McDash. In Figure A-3, we plot the average, 10th and 90th percentile rates over
time from Optimal Blue locks and McDash. Rates move closely together across the distribution,
with McDash rates lagging locked rates a bit—as expected since mortgages typically do not get
originated until a few weeks after the rate mortgage rate is locked in. Again, while the levels of
rates are very similar across the two datasets for FHA and conforming mortgages, Optimal Blue
rates tend to be higher than McDash for jumbo loans, although the amount of dispersion is similar.
1
A.2 Price Dispersion in Mortgage Offers
In this appendix, we provide additional detail on our analysis of price dispersion in offer interest
rates across lenders, already briefly discussed in Section 4.1 of the main text.
There are two things to consider when thinking about the “price” of a mortgage with certain
characteristics. First, lenders do not offer a single mortgage rate to borrowers but rather a menu
with different combinations of mortgage rates and discount points to choose from. Borrowers can
pay discount points, each equal to one percent of their mortgage balance, in order to lower their
mortgage interest rate. Alternatively, they can choose negative points, known as lender credits or
rebates, in return for a higher mortgage rate. In this case, borrowers receive cash from the lender
which can be used toward closing costs. Either way, one point in upfront payments corresponds to
about 20bp in mortgage rate (so a borrower could get e.g. a 4% mortgage rate with no points, a
4.2% rate but receive one point, or a 3.8% rate by paying one point).
Second, lenders also charge origination fees. While fees are not typically considered as part of
the price of the mortgage, they are part for the total cost of securing the mortgage. We can think
of lender fees and discount points as interchangeable: from the borrower’s perspective, a lender
that charges an origination fee of one percent to originate a mortgage at 4% interest is equivalent
to a lender that charges no fees but requires the borrower to pay one discount point for a mortgage
rate of 4%.
In the Optimal Blue Pricing Insights interface, we observe how lenders compare in terms of
the sum of points and fees that they charge for a given mortgage rate, on a given day in a given
location and for certain borrower and loan characteristics. The interface allows users to specify
the key underwriting and loan characteristics, including location (MSA), FICO score, LTV, loan
amount, DTI, loan type and term (e.g. 30-year fixed), loan purpose (e.g. cashout refinance),
program (e.g. FHA or conforming), as well as details about the property (e.g. whether it is a
single-family home or a condo) and whether it will be owner-occupied or not. Furthermore, the
user specifies the desired lock period (e.g. 30 days). One could furthermore specify a given mortgage
rate for which offers should be compared (e.g. 4%), but by default the system instead shows the
comparison of points/fees for the mortgage rate at which the median lender that makes an offer
does so at (as close as possible to) zero points and fees.
An example of the resulting output is shown in Figure A-4. Lenders are sorted based on the
“price” they offer for a loan with the desired characteristics, where the price equals 100 minus the
points/fees the borrower would be charged. Thus, a price of 101 means the borrower would receive
one point, while a price of 99 means the borrower would have to pay one point to get this loan. As
can be seen in the screenshot, the range of offers in this example spans almost 4 points, which for a
typical loan of $250,000 would correspond to a difference between the cheapest and most expensive
lenders of $10,000.
As noted in the main text, we conduct searches for 100 different combinations of FICO, LTV,
program, loan amount, loan purpose, occupancy, and rate type, across 20 MSAs (at different
frequencies). For each of these searches, we then receive the underlying individual price offers for
the mortgage rate the system chooses (as explained above).
For our main analysis, we then transform these prices into the rate each lender would offer at
zero points and fees, by converting points into rates using a conversion factor that we estimate
based on the lock data. As explained in the main text, we allow for this conversion factor to be
time-varying. The estimated conversion factor averages about 21bp in rate per 1 point upfront,
which is also in line with what is typically observed in lender rate sheets. So for instance, a lender
that is shown as offering a price of 100.5 for a 4.25% mortgage rate is assigned a rate of 4.145%.
2
A.2.1 Dispersion in Offer Rates
We start by documenting the dispersion in mortgage rates available from different lenders for
identical mortgages in Los Angeles, since we have daily searches for this MSA. The first panel of
Figure A-5 shows the distribution of rates offered by different lenders for conforming mortgages
with an amount of $300k, FICO=750, LTV=80 and DTI=36. There are about 120 different lenders
offering this mortgage in Los Angeles on any given day. The histogram shows the daily offer rates
after subtracting the median (for the same day) over the period of April 2016 to December 2019.
Figure A-5 shows that the rate difference between the cheapest and the most expensive lender
is about 100bp. Moreover, even though much of the mass is in the middle of the distribution, the
tails of the distribution are rather fat. These patterns can also be seen in the other two panels of
Figure A-5, which plot the dispersion for a typical FHA mortgage and a jumbo mortgage. The
exact shape of the distribution does look different across these different mortgages, but the amount
of dispersion is similar.
Figure 1 in the main text shows the dispersion in mortgage rates available from different lenders
in all of the 20 metropolitan areas. Table A-2 shows more detailed summary statistics of the rate
dispersion in this pooled offer data, broken down by mortgage types. There are typically over 100
unique lenders on any given day making offers for each mortgage type in each location. The median
mortgage rate is higher for jumbo loans than for conforming loans reflecting in part the fact that
conforming loans are guaranteed by Fannie or Freddie in exchange for a low guarantee fee, which
is rolled into the mortgage rate. FHA mortgages have lower interest rates than other products
since borrowers also have to pay upfront (175bp) and ongoing mortgage insurance premia (85bp)
which are not part of the quoted mortgage rate. Generally, the price dispersion is a bit higher for
mortgages with low FICO scores, high LTVs and FHA mortgages. Overall, there is about a 50-55bp
difference in mortgage rates between the 10
th
percentile lender and the 90
th
percentile lender, and
a 90bp difference between the 1
st
and the 99
th
percentile lender.
Table A-3 compares the rate dispersion for a “plain vanilla” conforming mortgage with LTV
of 80 and FICO of 750 across MSAs. We see that, while there are some differences in the exact
amount of dispersion across MSAs, the qualitative points from above generalize across all of the
cities, and Los Angeles is not an outlier.
A.2.2 Dispersion in Offered Points and Fees
In this subsection we focus on the points and fees charged by lenders to originate a mortgage with
a median interest rate. The median interest rate for each mortgage type is defined exactly as in
the previous subsection: it is the interest rate at which the median lender offers a mortgage (with
given characteristics) at zero points or fees. Figure A-6 shows the distribution of points and fees
charged by different lenders to originate this median interest rate mortgage, with discount points
and fees measured as a percent of the mortgage balance. This figure shows that the range of offers
shown in the screenshot in Figure A-4 appears representative of the universe of offer distributions.
Table A-4 summarizes this dispersion for different mortgage types. The differences in the upfront
costs of a mortgage with an identical rate across lenders are very large. The difference between
the 90
th
percentile and 10
th
percentile lender is around 2.2 to 2.5% of the mortgage balance. For
a typical conforming loan of $250K that amounts to roughly a $6000 difference in upfront costs
between these lenders. Even going from the 75
th
percentile to the 25
th
percentile lender would save
about $3000 for a typical borrower with a $250k loan.
3
A.3 Matching Offers and Locks
As described in Section 3.2, we collect data on mortgage offers for 20 MSAs (some daily, others
twice or once per week) and for different loan programs (conforming, super-conforming, jumbo,
and FHA) and borrower/loan characteristics. In particular, we collect rates for FICO scores of 640,
680, 720, and 750, and LTV ratios of 70, 80, 90, 95, and 96 percent. When matching locks to these
offers, we allow for some variation in the characteristics around the values that we collect rates for,
but do so in a conservative way. What this means is that (with two small exceptions noted below)
we match locks with FICO scores slightly above the FICO value from the rate offer and with LTV
ratios slightly below the LTV value from the offer, as follows:
Offer FICO 640: Lock FICO range 640-659
Offer FICO 680: Lock FICO range 680-699
Offer FICO 720: Lock FICO range 720-739
Offer FICO 750: Lock FICO range 740-850 (maximum FICO)
Offer LTV 70: Lock LTV range 60.01-70
Offer LTV 80: Lock LTV range 75.01-80
Offer LTV 90: Lock LTV range 85.01-90
Offer LTV 95: Lock LTV range 90.01-95
Offer LTV 96: Lock LTV range 95.01-97
In choosing these ranges, we follow Fannie Mae’s loan-level pricing adjustment (LLPA) grid (https:
//www.fanniemae.com/content/pricing/llpa-matrix.pdf). This grid is also why we decided to
assign FICO scores of 740-749 the FICO 750 offer as well, and similarly for LTVs of 96.01-97 for
the LTV 96 offer. (LTV values above 95 are uncommon for GSE loans, but are very common for
FHA loans, where the modal LTV is 96.5.) We do not include some intermediate values (e.g. FICO
660-679, 700-719; LTV 80-85) since LLPAs can be different and do not always change linearly;
however, matching less conservatively in that regard does not materially affect the results.
In addition to matching on date, FICO, LTV, MSA and loan program, we also only retain
purchase mortgages with a 30 day lock period (since that is what the rate search is for). 30 days
is also the most common lock period in the data.
A.4 An Alternative to the Locked-Offer Rate Gap: Expected
Gains from Search
Our headline measure of the “locked-offer rate gap” captures how far the rate a particular borrower
locked is from what the median lender could offer them for an identical loan on the same day.
We construct this simple measure for each borrower to see how well they are doing relative to the
median lender, and to uncover which groups of borrowers do particularly badly. An alternative
approach to this is to construct a measure of expected benefits from one extra search for each
borrower by making some assumptions on how borrowers shop and what rates they obtain when
doing so.
4
We start with a simple search model similar to Carlson and McAfee (1983). Suppose that there
are n mortgage lenders who are posting mortgage rate offers on Optimal Blue for a particular
borrower type. Rates are ordered from lowest to highest:
r
1
r
2
... r
n
Borrowers only see the mortgage rates available at the lenders they meet with. Assuming each
borrower has an equal chance of meeting any one of the lenders, the probability of finding a lender
that offers the rate r is f(r) = 1/n. Suppose a borrower has already found a rate r
k
and is
considering searching one more time for a cheaper lender. The expected gain from doing so is given
by:
x
k
=
k1
X
i=1
(r
k
r
i
)f(r
i
)
=
k1
X
i=1
(r
k
r
i
)
1
n
=
"
r
k
k1
X
i=1
r
i
k 1
#
k 1
n
(A1)
Intuitively, the term in the brackets is the locked rate minus the expected rate from going to
the k 1 lenders that are offering rates lower than r
k
. Of course, the borrower does not know which
lenders are offering rates lower than r
k
, so we have to adjust the expectation by the share of these
lenders in the population, which is (k 1)/n. Therefore, this is a measure of how much money the
borrower is leaving on the table, in expectation, by not conducting one more search. Compared to
the locked-offer rate gap we use in our main analysis, where only the median available rate matters
for our assessment of “how well” a borrower did, here the width of the offer distribution also plays
a role: for a given mean of the offer distribution, x
k
will be higher when offers are more widely
dispersed (as this leads E(r|r < r
k
) to be lower).
Table A-8 summarizes the expected gains from search for different cuts of the data similar to
Table 4 in the main text. Not surprisingly, the overall level of expected gains from search measure is
larger than the locked-offer rate gap, since the expected gain is by definition non-negative. Taking
into account this difference in levels, however, all the cross-sectional patterns we are interested in
are very similar to the ones in Table 4.
Table A-9 replicates the results of Table 5 using the alternative measure of expected gains from
one more search. The results are identical in both of these tables, suggesting that the choice of
using locked-offer rate gap or the alternative measure of expected gains from one more search is
immaterial for our results. This is not very surprising, given the result in Table A-2 that the
dispersion in offer rates does not vary much with borrower/loan characteristics.
A.5 Evidence from the SCF on the Effects of Financial Literacy
and Shopping
As a complement to our analysis of the new NSMO data in Section 7, here we draw on data from
the longstanding and widely-used Survey of Consumer Finances (SCF). The SCF is a triennial,
nationally representative survey of households sponsored by the Federal Reserve Board that broadly
covers US families’ financial circumstances. It collects detailed information on families’ debts, assets,
5
income, expenses, demographics, financial institutions, credit history, and financial decision-making.
Notably, for the first time in 2016, the SCF added three questions designed by Annamaria Lusardi
and Olivia Mitchell to gauge individuals’ general financial literacy.
1
The three questions assess
understanding of basic concepts related to saving, borrowing, and investing:
1. Suppose you had $100 in a savings account and the interest rate was 2% per year. After 5
years, how much do you think you would have in the account if you left the money to grow:
more than $102, exactly $102, or less than $102?
2. Imagine that the interest rate on your savings account was 1% per year and inflation was 2%
per year. After 1 year, would you be able to buy more than today, exactly the same as today,
or less than today with the money in this account?
3. Do you think that the following statement is true or false: buying a single company’s stock
usually provides a safer return than a stock mutual fund?
For each question, interviewees have the option to respond “do not know,” or can refuse to answer.
For each respondent, we compute the fraction of questions answered correctly, including “don’t
know” and “refuse” as not having answered correctly. Across all SCF respondents in 2016, 43%
answered all three correctly, 36% answered two correctly, 16% answered one correctly, and 4%
answered none correctly.
2
For our analysis here, we focus on a subsample of SCF households that own their home and
recently took out a fixed-rate 30-year or 15-year mortgage on their home (either to refinance or
to purchase the property) between 2013 and 2016. In this subsample, 56% answered all three
financial literacy questions correctly, 31% answered two correctly, 11% answered one correctly, and
2% answered none correctly.
In Table A-11, we provide estimates of the relationship between financial literacy and the interest
rate respondents pay on their mortgage (interest rates are self-reported, and we subtract out the
average prime rate for the month when the loan was taken out). Column 1 indicates that moving
from none correct to getting all three questions correct is associated with a lower interest rate of
25 basis points. This magnitude is largely robust to adding controls. It drops a little in column 2
after controlling for credit history
3
, loan characteristics, race, income, age, and education, but then
rises back to about 25 basis points in column 3 after controlling for state fixed effects.
In addition to this measure of financial literacy, the SCF also asks respondents about how much
they shop when trying to get a loan: “When making major decisions about borrowing money or
obtaining credit, some people search for the very best terms while others don’t. On a scale from
zero to ten, where zero is no searching and ten is a great deal of searching, what number would you
(and your husband/wife/partner) be on the scale?”
4
1
A growing literature has explored the relationship between various financial outcomes and this and other metrics
of financial literacy. For a review, see Lusardi and Mitchell (2014). The only other paper examining the relationship
between financial literacy and mortgage rates is Huston (2012). More recently, Gathergood and Weber (2017) study
the relationship between financial literacy and mortgage product choice.
2
Note that these statistics and all other results reported in this section use the SCF sampling weights to adjust
for the sampling design of the SCF, which oversamples high wealth households.
3
Unlike the NSMO, we do not observe credit scores in the SCF. However, we control for any late payment in the
past year, bankruptcy in the last 4 years, and foreclosure in the last 5 years. Another caveat is that we do not observe
points or fees in the SCF, which might bias our estimates if less literate borrowers are actually paying fewer points
in return for paying higher rates.
4
Just over one-quarter of our sample of mortgage borrowers answered “10”, while less than 3% answered “0”; the
mean response was about 7.5, with a standard deviation of 2.5.
6
Table A-11 shows how shopping relates to mortgage rates in the SCF, where we have divided
the numerical responses by 10 so that the shopping variable ranges from zero to one. The results
indicate that those who report shopping the most intensely have mortgage rates that are about 25
basis points lower than those who do no shopping. And, again, this result is robust to including
a number of controls that help explain a considerable amount of the variation in reported rates.
Finally, column 6 regresses mortgages rates on financial literacy and shopping simultaneously. The
estimated coefficients on both variables are almost unchanged, indicating that both shopping and
financial literacy are independently important for the mortgage rates consumers obtain. In sum,
data from the 2016 SCF are consistent with the message from the NSMO data: borrowers with
higher financial knowledge and those who shop more tend to obtain better mortgage rates.
A.6 Correlates of Shopping Intensity and Knowledge
Section 7 strongly suggests that more intense mortgage shopping and better knowledge of the
mortgage market are associated with lower contracted rates. In this appendix, we document how
different shopping and knowledge measures are correlated with one another, and also study which
observable borrower and loan characteristics are associated with stronger reported shopping inten-
sity and higher knowledge.
In Table A-12, we report results from regressions of the four binary shopping measures already
used in Section 7.3 on the three mortgage knowledge measures introduced in Section 7.1, as well as
various other loan and borrower characteristics, most of which we turn into binary variables for ease
of interpretation. We run regressions with one covariate of interest at a time (with survey wave fixed
effects as the only additional control), or controlling for all of them jointly and further controlling
for other factors that may also affect shopping intensity (for instance, a stronger expectation of
selling the property soon). The former type of regression is called “univar.” in Table A-12 while
the latter type is called “multivar.”
In Table A-13, we report similar regressions but with the knowledge measures as dependent
variables (and only the borrower and loan characteristics as independent variables). Note that for
the first two of the three outcomes in that table, higher values correspond to more knowledge, while
for the last one, the opposite is true. We discuss the results from both tables jointly, since in some
cases they contrast in interesting ways.
The first three rows of Table A-12 indicate that borrowers that are more knowledgeable also
shop more. Of course, in this case it is difficult to rule out reverse causality, namely that the
additional shopping made them more knowledgeable (for instance, about price differences across
lenders). The fourth coefficient shows that people who say that they were “not at all concerned
about qualifying for a mortgage when they began the process of getting this mortgage” also report
shopping less.
5
This suggests that less confidence in one’s ability to qualify for a loan can have the
beneficial side effect of inducing additional shopping.
Next, we reproduce the positive relationship between PMMS and shopping measures docu-
mented in Table 9.
6
We further see that mortgage knowledge tends to be slightly lower when
PMMS is higher, although the relationship is no longer significant once other variables are con-
trolled for.
Turning to borrower and loan characteristics, we see that borrowers with higher FICO scores
are more likely to have seriously considered more than one lender, although for the other shopping
measures the evidence is more mixed. However, high-FICO borrowers tend to be substantially more
5
This self-assessed creditworthiness was also used as a control variable in Table 7.
6
The coefficients differ slightly because in this section, we use less fine control variables.
7
knowledgeable, especially when considering the univariate correlations with mortgage-rate famil-
iarity and the knowledge index. There is no significant relation between FICO and the propensity
to think that all lenders offer similar terms.
Borrowers with higher LTVs tend to shop more, but are less knowledgeable. Similarly, FHA
borrowers do not appear to shop less, but tend to be significantly less knowledgeable than other
borrowers (except that they do have a slightly higher propensity to believe in price dispersion).
Given that our earlier Optimal Blue analysis found that these groups see substantially higher
locked-offer rate gaps, these patterns suggest that knowledge may be the key differential driver
of those patterns. Similarly, we also see that borrowers with purchase loans, and especially first-
time homebuyers, report higher shopping intensity, but are substantially less knowledgeable than
refinancers (which makes sense, since the latter likely have more experience with the process).
Borrowers with larger loan amounts, and especially jumbo borrowers, both shop more and are
more knowledgeable—in line with their lower rate spreads.
Finally, in terms of borrower demographics, more educated respondents are much more likely to
shop, and have better mortgage knowledge. Income appears to have little effect on shopping once
other factors are controlled for, but still correlates significantly with knowledge. Finally, we see
that minorities appear to shop more that Non-Hispanic White borrowers (the omitted category),
but were less familiar with mortgage rates and have a lower knowledge index. However, they are
more likely to believe in price dispersion.
8
Table A-1: Comparing Mortgage Locks in Optimal Blue to Closed Mortgages
FHA Loans, 2014-15 FHA Loans, 2016-18 Conventional Conforming Conventional Jumbo
Loans, 2016-18 Loans, 2016-18
Optimal Blue HUD McDash Optimal Blue McDash Optimal Blue McDash Optimal Blue McDash
Interest Rate
10th 3.75 3.625 3.625 3.625 3.5 3.75 3.625 3.75 3.375
mean 4.1 4.1 4.1 4.4 4.3 4.4 4.3 4.3 4.0
90th 4.625 4.625 4.625 5.25 5.125 5.125 5 4.875 4.625
FICO Score
10th 628 630 641 620 629 681 686 719 726
mean 679.4 680.5 688.9 672.3 684.0 745.2 750.0 766.1 771.3
90th 744 745 754 738 751 800 802 801 803
LTV
10th 93.7 94.3 87.6 93.4 87.9 66.6 64.4 66.7 65.0
mean 95.3 95.8 93.7 95.4 93.7 83.6 82.0 77.6 82.7
90th 96.5 96.5 96.5 96.5 96.5 95.0 95.0 85.0 85.0
Loan Amount
10th 89,745 84,000 81,987 100,360 97,697 116,000 113,715 482,000 485,100
mean 187,624.3 180,450.1 173,106.5 204,065.3 203,275.5 255,892.9 255,738.2 729,963.4 850,403.6
90th 300,000 293,250 276,892 321,985 325,004 417,000 418,125 1,060,000 1,260,000
N 282,933 1,318,700 777,763 860,579 1,468,968 1,547,776 2,695,218 61,430 190,993
Data Source: Optimal Blue, HUD, Black Knight McDash
Note: All statistics are for 30-year fixed rate home purchase mortgages for owner-occupied properties. Conventional conforming include super-
conforming loans that have loan amounts under the higher loan limits in high-cost geographies. “McDash” refers to Black Knight McDash data.
9
Table A-2: The real-time interest rate dispersion for offered mortgage products with no points
and fees
Median Median Standard Percentile Differences
No. Offers Rate Deviation 75
t
h 25
th
90
th
10
th
99
th
1
st
All Offers 118 4.67 0.20 0.27 0.53 0.90
Program
FHA 117 4.08 0.22 0.32 0.59 0.93
Conforming 122 4.54 0.19 0.27 0.51 0.88
Super-Conforming 144 4.68 0.20 0.27 0.52 0.88
Jumbo 106 5.06 0.20 0.26 0.53 0.92
FICO
640 107 5.23 0.21 0.29 0.54 0.92
680 118 4.64 0.20 0.28 0.53 0.90
720 122 4.48 0.20 0.27 0.52 0.90
750 122 4.44 0.20 0.27 0.52 0.90
LTV
70 122 4.67 0.20 0.27 0.52 0.90
80 117 4.78 0.20 0.28 0.53 0.91
90 105 4.78 0.20 0.27 0.52 0.91
95 128 4.63 0.20 0.27 0.51 0.88
96 119 4.27 0.21 0.30 0.55 0.91
Data Source: Optimal Blue
Notes: This table compares real-time interest rates for identical offered mortgages (same FICO, LTV, DTI,
loan amount, location, time etc.) with no points and fees. Column 1 shows the median number of lenders offering
each mortgage product in a location on a specific day. Columns 4-6 show the difference between various percentiles
of the offer distribution.
10
Table A-3: The real-time interest rate dispersion for offered conforming mortgages with no points and fees
Median Median Standard Percentile Differences
No. Offers Rate Deviation 75
t
h 25
th
90
th
10
th
99
th
1
st
Atlanta, GA 112 4.68 0.20 0.28 0.54 0.92
Boston-Worcester-Lawrence, MA-NH-ME-CT 77 4.49 0.21 0.30 0.56 0.93
Charlotte-Gastonia-Rock Hill, NC-SC 93 4.67 0.21 0.28 0.55 0.93
Chicago-Gary-Kenosha, IL-IN-WI 103 4.57 0.20 0.28 0.53 0.90
Cleveland-Akron, OH 61 4.71 0.21 0.30 0.57 0.92
Dallas-Fort Worth, TX 136 4.67 0.21 0.29 0.55 0.93
Denver-Boulder-Greeley, CO 119 4.69 0.19 0.25 0.49 0.88
Detroit-Ann Arbor-Flint, MI 76 4.68 0.21 0.29 0.56 0.94
Las Vegas, NV 87 4.88 0.21 0.28 0.55 0.92
Los Angeles-Riverside-Orange County, CA 147 4.69 0.20 0.27 0.52 0.89
Miami-Fort Lauderdale, FL 95 4.66 0.21 0.30 0.56 0.93
Minneapolis-St. Paul, MN 73 4.65 0.19 0.26 0.51 0.89
New York-Northern New Jersey-Long Island, NY-NJ 93 4.60 0.21 0.30 0.56 0.92
Phoenix-Mesa, AZ 117 4.80 0.21 0.29 0.54 0.91
Portland-Salem, OR 88 4.77 0.20 0.27 0.52 0.88
San Diego, CA 103 4.71 0.19 0.26 0.51 0.89
San Francisco-Oakland-San Jose, CA 112 4.75 0.19 0.26 0.51 0.88
Seattle-Tacoma-Bremerton, WA 101 4.79 0.19 0.26 0.51 0.88
Tampa-St. Petersburg-Clearwater, FL 124 4.80 0.20 0.27 0.53 0.92
Washington-Baltimore, DC-MD-VA 116 4.61 0.21 0.28 0.55 0.93
Data Source: Optimal Blue
Notes: This table compares real-time interest rates for 30 year fixed rate conforming mortgages with a LTV=80, FICO=750, DTI=36, and with no
points and fees. Column 1 shows the median number of lenders offering mortgages in a location on a specific day. Columns 3-5 show the difference between
various percentiles of the offer distribution.
11
Table A-4: Dispersion in points and fees that lenders charge
to originate at the median interest rate
Percentile Differences
75
t
h 25
th
90
th
10
th
99
th
1
st
Program
FHA 1.42 2.59 3.83
Conforming 1.19 2.22 3.69
Super-Conforming 1.23 2.35 3.79
Jumbo 1.13 2.31 3.84
FICO
640 1.30 2.41 3.83
680 1.22 2.35 3.77
720 1.19 2.30 3.78
750 1.20 2.30 3.78
LTV
70 1.19 2.28 3.77
80 1.24 2.37 3.81
90 1.19 2.29 3.81
95 1.20 2.26 3.72
96 1.32 2.44 3.80
Data Source: Optimal Blue
Notes: This table compares real-time points and fees charged by
different lenders to originate identical mortgages at the median interest
rate. Points and fees are given as percent of the mortgage balance. The
median interest rate is chosen such that the median lender charges no
points and fees at this interest rate.
12
Table A-5: Summary Statistics of the Rate Locked Minus the Median Offer Rate for Identical
Mortgages by ZIP Code Demographics
Observations Mean St. Deviation
Percentiles
25
th
75
th
All Mortgages 64,788 0.11 0.31 -0.07 0.26
Median Household Income
First Tercile 21,673 0.16 0.32 -0.03 0.31
Second Tercile 21,517 0.10 0.30 -0.07 0.25
Third Tercile 21,585 0.07 0.31 -0.11 0.22
Percent College Educated
First Tercile 21,610 0.16 0.32 -0.03 0.32
Second Tercile 21,602 0.12 0.31 -0.06 0.26
Third Tercile 21,576 0.05 0.29 -0.11 0.19
Minority Share
First Tercile 21,619 0.07 0.30 -0.09 0.22
Second Tercile 21,574 0.09 0.30 -0.08 0.24
Third Tercile 21,595 0.16 0.33 -0.04 0.32
Market Share of Top 4 Lenders
First Tercile 21,711 0.11 0.28 -0.04 0.24
Second Tercile 21,513 0.09 0.31 -0.09 0.25
Third Tercile 21,564 0.12 0.34 -0.08 0.29
Data Source: Optimal Blue, American Community Survey, HMDA
Notes: For each mortgage rate locked by borrowers in our data, we compute the median rate offered by
lenders in the same market on the same day for an identical mortgage. This table summarizes the difference
between each locked rate and the median offer rate. The median household income, percent college educated,
and minority share (share of Hispanic/Latino plus non-Hispanic Black) are only observed at the ZIP code
level. The market share of the top four lenders is observed at the county level.
13
Table A-6: Regressions of the Locked-Offer Rate Gap on Observables, for FHA Loans Only
(1) (2) (3) (4) (5) (6) (7) (8)
FICO (omitted cat.: [640,660))
I
680F ICO<700
-0.033
∗∗∗
-0.039
∗∗∗
-0.061
∗∗∗
(0.009) (0.009) (0.019)
I
720F ICO<740
-0.069
∗∗∗
-0.065
∗∗∗
-0.086
∗∗∗
(0.010) (0.010) (0.019)
I
F ICO740
-0.073
∗∗∗
-0.067
∗∗∗
-0.082
∗∗∗
(0.013) (0.010) (0.016)
I
LT V >95
0.040 0.062
∗∗
0.074
(0.026) (0.024) (0.038)
Discount Points
I
5<P oints<0.2
-0.137
∗∗∗
0.021
(0.021) (0.013)
I
0.2<P oints5
0.145
∗∗∗
0.027
(0.023) (0.023)
Loan Officer Comp (%) 0.182
∗∗∗
0.180
∗∗∗
(0.057) (0.057)
Loan amount f.e. ($10k bins) Yes Yes Yes Yes Yes Yes Yes Yes
MSA x Month f.e. Yes Yes Yes Yes Yes Yes Yes Yes
Branch f.e. Yes Yes Yes Yes Yes
Adj. R-Squared 0.128 0.564 0.627 0.122 0.559 0.615 0.199 0.560
Observations 14330 12857 2965 14330 12857 2965 14330 12857
Data Source: Optimal Blue
Notes: The dependent variable is the mortgage interest rate locked minus the median offer rate in the same market and day for an
identical mortgage. Unlike in the corresponding table in the main text, here we only use two LTV bins (separated at 95) since the majority
of FHA loans have very high LTVs. The data covers mortgage rates for 20 metropolitan areas during the period between 2016-2019. We
focus on 30 year, fixed rate, fully documented purchase mortgages. Standard errors shown in parentheses are two-way clustered at the
month and lender level. Significance: * p<0.1, ** p<0.05, *** p<0.01.
14
Table A-7: Regressions of the Locked-Offer Rate Gap on Observables, for Independent Nonbank Originators Only
(1) (2) (3) (4) (5) (6) (7) (8)
FICO (omitted cat.: [640,660))
I
680F ICO<700
-0.057
∗∗∗
-0.044
∗∗∗
-0.048
∗∗∗
(0.009) (0.007) (0.012)
I
720F ICO<740
-0.093
∗∗∗
-0.066
∗∗∗
-0.057
∗∗∗
(0.011) (0.009) (0.013)
I
F ICO740
-0.126
∗∗∗
-0.087
∗∗∗
-0.069
∗∗∗
(0.012) (0.010) (0.013)
LTV (omitted cat.: (60,80])
I
85<LT V 90
0.013
∗∗
0.010 0.018
(0.006) (0.006) (0.010)
I
90<LT V 95
0.043
∗∗∗
0.030
∗∗∗
0.025
∗∗
(0.006) (0.006) (0.011)
I
LT V >95
0.178
∗∗∗
0.140
∗∗∗
0.089
∗∗∗
(0.013) (0.012) (0.016)
Discount Points
I
5<P oints<0.2
-0.078
∗∗∗
0.005
(0.025) (0.007)
I
0.2<P oints5
0.089
∗∗∗
0.017
∗∗
(0.014) (0.008)
Loan Officer Comp (%) 0.155
∗∗∗
0.139
∗∗∗
(0.036) (0.038)
Loan amount f.e. ($10k bins) Yes Yes Yes Yes Yes Yes Yes Yes
MSA x Month f.e. Yes Yes Yes Yes Yes Yes Yes Yes
Branch f.e. Yes Yes Yes Yes Yes
Adj. R-Squared 0.134 0.469 0.429 0.174 0.496 0.440 0.155 0.462
Observations 45522 44319 11720 45522 44319 11720 45522 44319
Data Source: Optimal Blue
Notes: The dependent variable is the mortgage interest rate locked minus the median offer rate in the same market and day for an
identical mortgage. The data covers mortgage rates for 20 metropolitan areas during the period between 2016-2019. We focus on 30 year,
fixed rate, fully documented purchase mortgages. Standard errors shown in parentheses are two-way clustered at the month and lender level.
Significance: * p<0.1, ** p<0.05, *** p<0.01.
15
Table A-8: Summary Statistics of the Expected Gain from Search
Observations Mean St. Deviation
Percentiles
25
th
75
th
All Mortgages 64,788 0.20 0.23 0.05 0.27
Program
FHA 14,441 0.32 0.30 0.10 0.45
Conforming 44,040 0.17 0.19 0.05 0.23
Super-Conforming 4,478 0.10 0.15 0.01 0.13
Jumbo 1,829 0.05 0.12 0.00 0.05
FICO
640, 660
7,406 0.31 0.31 0.08 0.45
680, 700
9,390 0.25 0.27 0.06 0.35
720, 740
10,207 0.20 0.22 0.05 0.27
740+ 37,785 0.16 0.18 0.04 0.22
LTV
75, 80
21,334 0.13 0.15 0.03 0.18
85, 90
6,882 0.16 0.17 0.04 0.22
90, 95
15,782 0.17 0.19 0.04 0.23
95, 97
20,790 0.30 0.29 0.09 0.42
First-Time Homebuyer
No 32,437 0.16 0.19 0.04 0.22
Yes 32,345 0.23 0.25 0.06 0.32
Discount Points
-5, -0.2
14,015 0.15 0.19 0.02 0.21
-0.2, 0.2
22,735 0.18 0.21 0.04 0.23
0.2, 5
28,038 0.24 0.25 0.07 0.32
Lender Type
Independent Non-bank 45,618 0.21 0.23 0.06 0.28
Other 19,170 0.17 0.21 0.03 0.24
Data Source: Optimal Blue
Note: For each mortgage rate locked by borrowers in our data, we compute the expected gain
from search using equation (A1).
16
Table A-9: Regressions of the Expected Gains from Search on Observables
(1) (2) (3) (4) (5) (6) (7) (8)
FICO (omitted cat.: [640,660))
I
680F ICO<700
-0.055
∗∗∗
-0.045
∗∗∗
-0.042
∗∗∗
(0.006) (0.005) (0.010)
I
720F ICO<740
-0.095
∗∗∗
-0.073
∗∗∗
-0.069
∗∗∗
(0.009) (0.007) (0.012)
I
F ICO740
-0.125
∗∗∗
-0.094
∗∗∗
-0.084
∗∗∗
(0.009) (0.007) (0.011)
LTV (omitted cat.: (60,80])
I
85<LT V 90
0.017
∗∗∗
0.013
∗∗∗
0.016
∗∗∗
(0.003) (0.003) (0.006)
I
90<LT V 95
0.034
∗∗∗
0.026
∗∗∗
0.024
∗∗∗
(0.004) (0.004) (0.006)
I
LT V >95
0.148
∗∗∗
0.121
∗∗∗
0.090
∗∗∗
(0.010) (0.009) (0.012)
Discount Points
I
5<P oints<0.2
-0.036
∗∗∗
0.004
(0.009) (0.005)
I
0.2<P oints5
0.065
∗∗∗
0.018
(0.010) (0.010)
Loan Officer Comp (%) 0.112
∗∗∗
0.099
∗∗∗
(0.026) (0.029)
Loan amount f.e. ($10k bins) Yes Yes Yes Yes Yes Yes Yes Yes
MSA x Month f.e. Yes Yes Yes Yes Yes Yes Yes Yes
Branch f.e. Yes Yes Yes Yes Yes
Adj. R-Squared 0.135 0.407 0.408 0.175 0.435 0.421 0.132 0.390
Observations 64693 62783 14659 64693 62783 14659 64693 62783
Data Source: Optimal Blue
Notes: The dependent variable is the expected gain from an additional search, given by equation (A1). The data covers mortgage rates
for 20 metropolitan areas during the period between 2016-2019. We focus on 30 year, fixed rate, fully documented purchase mortgages.
Standard errors shown in parentheses are two-way clustered at the month and lender level. Significance: * p<0.1, ** p<0.05, *** p<0.01.
17
Table A-10: The Relationship Between the Expected Gains from Search and Treasury Yields
(1) (2) (3) (4) (5) (6) (7) (8)
Treasury Yield -0.053
∗∗∗
-0.056
∗∗∗
-0.036
∗∗∗
-0.037
∗∗∗
(0.006) (0.010) (0.008) (0.008)
Offer Spread to Prime Conforming Rate -0.116
∗∗∗
-0.117
∗∗∗
(0.020) (0.020)
Treasury Yield ×
DTI > 36 -0.058
∗∗∗
-0.062
∗∗∗
-0.042
∗∗∗
(0.007) (0.011) (0.010)
DTI 36 -0.046
∗∗∗
-0.048
∗∗∗
-0.028
∗∗∗
(0.007) (0.009) (0.005)
Borrower and Loan Controls Yes Yes Yes Yes Yes Yes Yes Yes
MSA F.E Yes Yes Yes Yes Yes Yes Yes Yes
MSA x Month F.E. Yes Yes Yes Yes Yes Yes
Branch F.E. Yes Yes Yes Yes
Adj. R-Squared 0.165 0.181 0.438 0.166 0.182 0.438 0.441 0.441
Observations 64396 64316 62397 64396 64316 62397 62782 62397
P-val. for equality of DTI coefficients 0.037 0.028 0.012
Data Source: Optimal Blue
Notes: The dependent variable is the expected gain from an additional search, given by equation (A1). The offer spread to conforming rate is defined
as the average offer rate for a typical borrower in the same program in the same day minus the average offer rate for a typical prime conforming borrower.
All specifications include controls for FICO, LTV, and loan amount. The data covers mortgage rates for 20 metropolitan areas during the period between
2016-2019. We focus on 30 year, fixed rate, fully documented purchase mortgages. Standard errors shown in parentheses are two-way clustered at the month
and lender level. Significance: * p<0.1, ** p<0.05, *** p<0.01.
18
Table A-11: Relationship between Interest Rate Spreads and Measures of Financial Literacy and Shopping in the Survey
of Consumer Finances
(1) (2) (3) (4) (5) (6) (7)
Financial Literacy (Fraction Correct) -0.247** -0.202** -0.246** -0.245**
(0.110) (0.097) (0.097) (0.100)
Shops Around for Credit -0.262*** -0.259*** -0.230*** -0.222**
(0.090) (0.084) (0.085) (0.087)
Loan Characteristics Yes Yes Yes Yes Yes
Borrower Characteristics Yes Yes Yes Yes Yes
State Fixed Effects Yes Yes Yes
Observations 820 816 816 821 817 817 816
R-squared 0.011 0.15 0.225 0.009 0.151 0.222 0.229
Data source: 2016 Survey of Consumer Finances (SCF)
Notes: Sample comprised of households that took out a 15 year or 30 year fixed-rate home purchase or refinance mortgage in 2013-2016
for their principal residence. Outcome variable is the interest rate (self-reported) on the first lien mortgage relative to the average Freddie
Mac PMMS prime rate for a loan of the same term in the month the mortgage was taken out. The Financial Literacy variable refers to
the fraction correct on three questions designed by Lusardi and Mitchell and asked in the 2016 SCF. The Shopping Around variable is a
self-reported value between 0 and 10 gauging the degree to which respondents shop for credit; we divide responses by 10 so that the range is
0 to 1. The loan characteristics we control for in specifications (2), (3) and (5)-(7) include loan program, loan term, property type, and loan
purpose (purchase, refinance or cash out). Borrower controls include indicators of whether they were late on any payment in the past year,
had a bankruptcy in the last 4 years, had a foreclosure in the last 5 years, as well as controls for income, education, age and race/ethnicity.
* p<0.1, ** p<0.05, *** p<0.01.
19
Table A-12: Relationship Between Various Binary Measures of Mortgage Shopping and Character-
istics of Borrower and Loan.
Considered 2+ lenders Applied to 2+ lenders Used other lenders Used web
for better terms to get info to get info
Univar. Multivar. Univar. Multivar. Univar. Multivar. Univar. Multivar.
(1) (2) (3) (4) (5) (6) (7) (8)
Very familiar with mortgage rates 0.057*** 0.045*** -0.007 0.009 0.022*** 0.010 0.003 0.008
(0.008) (0.009) (0.007) (0.007) (0.008) (0.009) (0.008) (0.009)
Index of mortgage knowledge (Std) 0.047*** 0.033*** 0.005* 0.006 0.030*** 0.022*** 0.038*** 0.041***
(0.004) (0.004) (0.003) (0.004) (0.004) (0.004) (0.004) (0.004)
Most lenders offer same rate? Yes -0.085*** -0.076*** -0.052*** -0.051*** -0.071*** -0.063*** -0.017 -0.014
(0.012) (0.012) (0.010) (0.010) (0.012) (0.012) (0.012) (0.011)
Not concerned about qualifying for mtg. -0.047*** -0.076*** -0.052*** -0.044*** -0.068*** -0.095*** -0.073*** -0.092***
(0.008) (0.009) (0.006) (0.007) (0.008) (0.009) (0.008) (0.009)
Market mortgage rate (PMMS) 0.045** 0.046** 0.069*** 0.063*** 0.048*** 0.050*** 0.019 0.027
(0.018) (0.018) (0.014) (0.014) (0.018) (0.018) (0.018) (0.018)
FICO/100 0.015** 0.017** -0.015*** -0.002 0.008 0.013* -0.005 0.017**
(0.006) (0.007) (0.005) (0.006) (0.006) (0.007) (0.006) (0.007)
LTV/100 0.051** 0.007 0.130*** 0.052*** 0.049** 0.045* 0.187*** 0.088***
(0.020) (0.025) (0.015) (0.019) (0.020) (0.025) (0.020) (0.025)
Loan amount > 200k 0.081*** 0.034*** 0.029*** 0.018** 0.083*** 0.049*** 0.061*** 0.009
(0.008) (0.009) (0.006) (0.008) (0.008) (0.009) (0.008) (0.009)
Jumbo 0.116*** 0.042** 0.017 0.000 0.116*** 0.047** -0.018 -0.073***
(0.020) (0.020) (0.016) (0.017) (0.020) (0.021) (0.020) (0.020)
FHA -0.004 -0.000 0.031*** -0.007 -0.010 -0.005 0.031*** 0.005
(0.011) (0.013) (0.010) (0.011) (0.011) (0.013) (0.011) (0.013)
VA/FSA -0.005 0.002 0.005 -0.013 0.009 0.014 0.003 0.019
(0.012) (0.014) (0.010) (0.011) (0.012) (0.014) (0.012) (0.014)
Purpose = home purchase 0.045*** 0.037*** 0.058*** 0.041*** 0.023*** 0.013 0.030*** -0.064***
(0.008) (0.010) (0.006) (0.008) (0.008) (0.010) (0.008) (0.010)
First-time homebuyer 0.048*** 0.023* 0.067*** 0.016 0.019* 0.003 0.148*** 0.110***
(0.011) (0.013) (0.009) (0.012) (0.011) (0.013) (0.010) (0.013)
At least college degree 0.087*** 0.053*** 0.028*** 0.018** 0.076*** 0.053*** 0.133*** 0.090***
(0.008) (0.009) (0.006) (0.007) (0.008) (0.009) (0.008) (0.009)
Household income > 100k 0.060*** 0.004 0.004 -0.010 0.050*** -0.000 0.057*** 0.015
(0.008) (0.010) (0.006) (0.008) (0.008) (0.010) (0.008) (0.010)
White Hispanic 0.033** 0.032** 0.057*** 0.042*** 0.014 0.010 0.052*** 0.043***
(0.016) (0.016) (0.013) (0.014) (0.016) (0.016) (0.016) (0.016)
Black 0.061*** 0.067*** 0.060*** 0.052*** -0.000 -0.010 0.055*** 0.052***
(0.017) (0.017) (0.014) (0.015) (0.016) (0.017) (0.017) (0.016)
Asian 0.115*** 0.061*** 0.030** 0.003 0.119*** 0.071*** 0.149*** 0.088***
(0.017) (0.017) (0.014) (0.015) (0.017) (0.017) (0.016) (0.016)
Other race 0.063*** 0.055** 0.046** 0.034* 0.038 0.028 0.057** 0.041*
(0.024) (0.024) (0.020) (0.020) (0.024) (0.024) (0.024) (0.022)
Mean of Dependent Variable 0.510 0.190 0.418 0.533
Adj. R2 0.04 0.03 0.03 0.07
Obs. 19906 19906 19906 19906
Data Source: National Survey of Mortgage Originations
Note: Sample restricted to first-lien loans (without a junior lien) for single-family principal residence properties,
with no more than two borrowers, and a loan term of 10, 15, 20 or 30 years. All four dependent variables are
binary. Observations weighted by NSMO sample weights. The univariate regressions (odd columns) only feature
one of the covariates in the table, along with survey wave fixed effects. The multivariate regressions (even columns)
simultaneously control for all the variables listed in the table, survey wave fixed effects, and the following additonal
variables: indicators for single borrowers, cash-out refinances, whether the household owns 4 different types of financial
assets, metropolitan CRA low-to-moderate income tract status, borrower age and gender, and self-assessed likelihood
of moving, selling, or refinancing, as well as risk aversion. Robust standard errors in parentheses. * p<0.1, ** p<0.05,
*** p<0.01.
20
Table A-13: Relationship Between Various Measures of Mortgage Knowledge and Characteristics
of Borrower and Loan.
Very familiar with Knowledge Index Thinks all lenders
mortgage rates (std) offer same terms
Univar. Multivar. Univar. Multivar. Univar. Multivar.
(1) (2) (3) (4) (5) (6)
Market mortgage rate (PMMS) -0.061*** -0.022 -0.074** -0.003 -0.017 -0.024
(0.018) (0.017) (0.037) (0.034) (0.040) (0.039)
FICO/100 0.113*** 0.046*** 0.179*** 0.018 0.002 0.001
(0.006) (0.007) (0.013) (0.014) (0.008) (0.009)
LTV/100 -0.398*** -0.049** -0.658*** -0.096** 0.117*** 0.081**
(0.019) (0.023) (0.041) (0.049) (0.027) (0.035)
Loan amount > 200k 0.117*** 0.023*** 0.331*** 0.079*** -0.020** -0.015
(0.008) (0.009) (0.016) (0.018) (0.010) (0.012)
Jumbo 0.173*** 0.023 0.501*** 0.103*** -0.124*** -0.121***
(0.017) (0.017) (0.037) (0.037) (0.027) (0.028)
FHA -0.189*** -0.031** -0.344*** -0.063** -0.022 -0.040**
(0.011) (0.013) (0.023) (0.025) (0.015) (0.017)
VA/FSA -0.055*** 0.001 -0.116*** -0.047* 0.021 0.009
(0.012) (0.013) (0.025) (0.026) (0.015) (0.017)
Purpose = home purchase -0.168*** -0.051*** -0.181*** 0.009 0.044*** 0.043***
(0.008) (0.009) (0.016) (0.019) (0.010) (0.014)
First-time homebuyer -0.322*** -0.206*** -0.413*** -0.156*** 0.012 -0.043**
(0.010) (0.013) (0.021) (0.025) (0.014) (0.017)
At least college degree 0.067*** 0.014* 0.285*** 0.147*** 0.006 -0.000
(0.008) (0.008) (0.016) (0.017) (0.011) (0.012)
Household income > 100k 0.180*** 0.067*** 0.457*** 0.174*** -0.010 0.001
(0.008) (0.009) (0.015) (0.018) (0.010) (0.013)
White Hispanic -0.104*** -0.021 -0.224*** -0.061** -0.075*** -0.066***
(0.016) (0.015) (0.032) (0.030) (0.021) (0.021)
Black -0.102*** -0.027 -0.074** 0.059* -0.131*** -0.116***
(0.017) (0.017) (0.032) (0.032) (0.022) (0.023)
Asian -0.042** -0.070*** -0.086** -0.230*** -0.102*** -0.079***
(0.017) (0.016) (0.035) (0.034) (0.022) (0.023)
Other race -0.076*** -0.029 -0.070 -0.004 -0.115*** -0.110***
(0.024) (0.023) (0.051) (0.048) (0.033) (0.032)
Mean of Dependent Variable 0.617 -0.025 0.682
Adj. R2 0.14 0.16 0.02
Obs. 19906 19906 10275
Data Source: National Survey of Mortgage Originations
Note: Sample restricted to first-lien loans (without a junior lien) for single-family principal residence properties,
with no more than two borrowers, and a loan term of 10, 15, 20 or 30 years. The dependent variables are binary
except in columns (3)-(4), where the knowledge index is standardized to have mean 0 and standard deviation 1 (in
unweighted sample). Observations weighted by NSMO sample weights. The univariate regressions (odd columns)
only feature one of the covariates in the table, along with survey wave fixed effects. The multivariate regressions (even
columns) simultaneously control for all the variables listed in the table, survey wave fixed effects, and the following
additonal variables: indicators for single borrowers, cash-out refinances, whether the household owns 4 different types
of financial assets, metropolitan CRA low-to-moderate income tract status, borrower age and gender, and self-assessed
likelihood of moving, selling, or refinancing, as well as risk aversion. Robust standard errors in parentheses. * p<0.1,
** p<0.05, *** p<0.01.
21
0
.2
.4
.6
.8
1
Empirical CDF
-2 -1 0 1 2
Discount Points Paid
Conforming
FHA
Jumbo
Figure A-1: The Empirical Cumulative Distribution of Discount Points Paid, by Program
Data Source: Optimal Blue
Note: Figure shows cumulative share of borrowers that paid up to a certain amount of discount points; negative values
represent credits/rebates. Data includes purchase and refinance rate locks in 2015-2019.
22
2.75
3
3.25
3.5
3.75
4
4.25
4.5
4.75
5
5.25
5.5
Jan 2016 Jan 2017 Jan 2018 Jan 2019 Jan 2020
Mortgage News Daily
Freddie Mac PMMS
Optimal Blue Insight
Conforming Mortgages
2.75
3
3.25
3.5
3.75
4
4.25
4.5
4.75
5
5.25
5.5
Jan 2016 Jan 2017 Jan 2018 Jan 2019 Jan 2020
Mortgage News Daily
Optimal Blue Insight
FHA Mortgages
2.75
3
3.25
3.5
3.75
4
4.25
4.5
4.75
5
5.25
5.5
Jan 2016 Jan 2017 Jan 2018 Jan 2019 Jan 2020
Mortgage News Daily
Zillow
Optimal Blue Insight
Jumbo Mortgages
Figure A-2: Comparison of Average Offer Rates from Optimal Blue with Mortgage News Daily
Data
Data Source: Optimal Blue, Mortgage News Daily, Freddie Mac, Zillow
Note: The Optimal Blue Data are for borrowers with FICO=750, DTI=36, with no points/fees, and LTV=80 for conforming
and jumbo, and LTV=96.5 for FHA. The Mortgage News Daily (MND) data reflect rates for “top-tier borrowers, and we
adjust the MND rates assuming they include 0.5% points and fees.
23
Figure A-3: Comparison of Locked Interest Rates from Optimal Blue with Interest Rates on Closed
Originations in McDash
Data Source: Optimal Blue, Black Knight McDash
Note: The Optimal Blue series lead the McDash series because for Optimal Blue we observe the date when the loan terms are
locked, while in McDash we observe when a loan is originated.
24
Figure A-4: Screenshot of Sample Offer Distribution from Optimal Blue Pricing Insights
Data Source: Optimal Blue
Note: Figure shows an example of the real-time distribution of offers across lenders in the same metropolitan area for a loan
with given characteristics and at a note rate of 5.125%. Lenders are sorted by “price”, which equals 100 + the points
(rebate/credit) the lender pays to the borrower (so “102” means the borrower receives two points at closing, while “98” means
they would have to pay two points). The mortgage note rate for which offers are shown is chosen such that the median lender
offers a price as close as possible to 100. For actual lenders using the interface, an orange dot would show their position in the
distribution.
25
0
.01
.02
.03
.04
Fraction
-.6 -.5 -.4 -.3 -.2 -.1 0 .1 .2 .3 .4 .5 .6
Offered Rate Minus Median Rate
Conforming, $300k, FICO=750, LTV=80, DTI=36, No Points/Fees
0
.01
.02
.03
.04
Fraction
-.6 -.5 -.4 -.3 -.2 -.1 0 .1 .2 .3 .4 .5 .6
Offered Rate Minus Median Rate
FHA, $300k, FICO=680, LTV=96, DTI=36, No Points/Fees
0
.01
.02
.03
.04
.05
Fraction
-.6 -.5 -.4 -.3 -.2 -.1 0 .1 .2 .3 .4 .5 .6
Offered Rate Minus Median Rate
Jumbo, $700k, FICO=750, LTV=80, DTI=36, No Points/Fees
Figure A-5: Interest Rate Offer Dispersion for Identical Mortgages in Los Angeles
Data Source: Optimal Blue
Note: The spread is defined as the difference between real-time mortgage rate offers and the median offer rate for identical
mortgage products. The histogram includes daily data between April 2016 and December 2019.
26
0
.01
.02
.03
.04
Fraction
-3 -2.5 -2 -1.5 -1 -.5 0 .5 1 1.5 2 2.5 3
Points and Fees (% of mortgage balance)
Distribution in Points/Fees Charged for Mortgages at the Median Interest Rate
Figure A-6: Dispersion in Points and Fees Lenders Charge for Identical Mortgages at the Median
Interest Rate
Data Source: Optimal Blue
Note: Points and fees are given as percent of the mortgage balance. The median interest rate is calculated as the
rate at which the median lender charges no points and fees.
27