by investors directly; and b) the markups and fees that the originator charges. The resulting rates
and fees are the best offer each lender could make to a consumer who would request a mortgage
from them.
We conduct daily searches in one local market (Los Angeles), twice-weekly searches in four
markets, and weekly searches for 15 additional markets.
4
We collect offer rate distributions for
100 different loan types, differing across the following dimensions: FICO score, loan-to-value ratio,
loan type (conforming, FHA, jumbo), loan purpose (purchase or cash-out refinance), occupancy
(owner-occupied or investor), rate typ e (30-year fixed or 5/1 adjustable), and loan amount. The
mortgages require full income, asset and employment documentation, and are used to finance single
unit homes.
Two limitations of the offe rs data are: (i) we are not able to track institutions over time or
match them directly to the lenders in the lock data, si nce there is no fixed lender identifier; (ii) the
time series so far is relatively short: we started systematically tracking offers in April 2016.
In the main analysis, we primarily use the offered rates as a benchmark for the rates that
borrowers lock. However, in the online appendix we present a separate analysis documenting price
dispersion in offers only, which is also substantial and of independent interest.
3 Dispersion in Locked Mortgage Rates
In this section we document the magnitude of price dispersion in mortgage rates that borrowers
lock. Dispersion in mortgage rates can arise for multiple reasons, some of which are differences
in borrower characteristics, mortgage characteristics and lender characteristics. We are interested
in investigating whether identical borrowers who choose the same mortgage product, in the same
market, at the same time pay different prices. To investigate this, we regress locked mortgage 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 b e tween 75th-25th or 90th-10th percentiles.
Table
2 shows the results from various specifications, estimated on the same set of 1.94 million
locked loans over the 2015-2018 period.
5
Across all columns, we control for a basic set of variables,
which consist of fully interacted bins of values for FICO, LTV, DTI, loan amount, as well as loan
program (conforming, super-conforming, jumbo, FHA). The resulting “grid” takes 7,680 unique
values. To allow for variation within t he grids, we furthermore linearly control of each of the four
continuous variables. In addition, we add a fixed effect for whether a locked loan is a refinance,
and for the length of the lock period.
6
4
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.
5
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 (7) in all specifications.
6
The lock period typically varies from 15 to 90 days, with 30 and 45 days being the most common choices. A
7