"Targeted Principal Forgiveness Is Effective: Mortgage Modification and Financial Crisis,"
with Joelle Scally (2022).
Research into the Global Financial Crisis finds forgiving mortgage principal ineffective at stemming defaults, and argues that borrowers default because of illiquidity, not strategically. We argue the opposite: targeted forgiveness is effective, and default is better explained by quantifying how illiquidity interacts with borrowers’ strategy. We embed these interactions in a computational heterogeneous structural model, introducing idiosyncratic default penalties. Differing penalties explain borrowers' differing deviations from pure-financial optimality. We run the model on heterogeneous microdata, estimating penalties from credit scores and payment histories. Forgiving low-score, deep-underwater borrowers would have eliminated nearly all their defaults, with net gain for lenders.
"Mortgage Default: A Heterogeneous-Agent Model,"
with Joelle Scally (2022).
We introduce a loan-level model of mortgage default with heterogeneity in
borrower characteristics and mortgage terms, including idiosyncratic
penalties for default. Borrowers’ penalties determine how closely their
behavior hews to the predictions of the double-trigger or strategic models.
The state space varies loan-to-loan based on all of the loan’s, borrower's,
property's, and neighborhood's idiosyncratic characteristics. We test the
model on a high-performance computing cluster against real data drawn from
linked databases with billions of observations of hundreds of simultaneous
attributes. The model predicts defaults out-of-sample, fits cross-sectional
characteristics of the distribution of mortgage performance, and classifies
likelihood of default with high accuracy and better than all known
benchmarks.
"Machine Lending: Hidden Bias, Hidden Risk" (2022).
I estimate a machine-learning model of mortgage default in tandem with a
heterogeneous structural model. While ML models typically achieve excellent
in-sample fit when trained against massive datasets, such datasets may also lead
to overfitting the low default rates that prevailed throughout the recent
prolonged period of economic expansion, in which virtually no households
experienced negative equity. ML models' extrapolation of low expected defaults
may thus depend on unrealistic assumptions about continued economic expansion.
Present-day ML models are often opaque and provide forecasts that are not
interpretable to users, making it correspondingly more difficult for lenders to
isolate the degree to which their forecasts hinge upon such unrealistically
optimistic assumptions. ML model opacity can also lead such models to repeat
and amplify historic discriminatory biases in lending. I illustrate how
structural modeling can be leveraged to obviate both perils by augmenting ML
model forecasts.
"Climate Change and Financial Assets: The Effects of Stormwater on Mortgage Default",
with Alex Gelber, Tom Corringham, and Hai Long Duong (2021).
We establish that climate change poses risks to financial assets. We exploit a
unique dataset with high-resolution data on adverse weather. Using a
difference-in-difference approach, we identify the effect of stormwater runoff
on mortgage default. Residential property owners inadequately self-insure
against flood risk from stormwater runoff, leading mortgages backed by
properties in areas exposed to increased runoff to go delinquent and default at
higher rates. Moderate storms, not just headline disasters, exacerbate
mortgage default. Losses are subsidized by borrowers in other geographies, but
discrimination and redlining considerations prevent the GSEs from correcting
the mispricing in loan-level price adjustments. Losses are projected to grow
substantially in the next several decades.
"Credit Availability Did Expand Before the Global Financial Crisis" (2016).
Scholars remain divided on whether the Global Financial Crisis was fueled by significantly looser
credit underwriting standards in the early 2000s, disagreeing not only about whether loose
standards caused the crisis, but even on whether standards were loose in that era. I examine three
different loan-level mortgage origination datasets in the US to try to disentangle why scholars
disagree on this question. I show that linking loans at the property-level, which is only possible for
some of those datasets, is necessary to see the higher origination loan-to-value ratios that obtained
before the crisis. Moreover, I show that leverage rose especially for less creditworthy borrowers,
many of whom would have been excluded entirely from getting a mortgage in times of tighter
credit. The simultaneous expansion in low-LTV financing diluted average leverage, obscuring
the reality that unprecedented cheap credit was available pre-Crisis.
"Endogenous Leverage and Credit in an Agent-Based Model of the Housing Market,"
with John Geanakoplos, Ravi Jagadeesan, Emily Dodwell, and Jesse Wang (2015).
Did house prices rise before the Global Financial Crisis because lenders
expanded credit? Or did lenders expand credit because of the economic forces
that led prices to rise? We model endogenous credit extension in a computational
agent-based model of the housing market. Profit-maximizing, rational,
forward-looking lenders charge interest rates to borrowers with different
loan-to-value ratios and different credit scores based on the lenders' estimates
of those borrowers' propensities to default. When lenders extrapolate future
default rates from prior history, then expanding house prices shield borrowers
from negative equity and fuel an artificially low level of defaults. Lenders
respond by loosening credit standards, which feeds back into borrowers' demand
for housing, fueling a fragile bubble. The bubble bursts when relatively few
initial defaults cascade, leading lenders to choke off further credit and
subsequent demand for housing to dry up. The model captures the time series and
distributional features of lending and house prices observed throughout the
Global Financial Crisis and provides a micro-foundation for studying how credit
availability fuels bubbles.
"An Agent-Based Model of the Housing Market Bubble in Metropolitan Washington DC,"
with Doyne Farmer, John Geanakoplos, Peter Howitt, et al. (2014).
With Robert Axtell, Benjamin Conlee, Ernesto Carella, Doyne Farmer, John Geanakoplos,
Jon Goldstein, Matthew Hendrey, Peter Howitt, David Masad, and Nathan Palmer.
Published in Housing
markets and the macroeconomy: challenges for monetary
policy and financial stability---a conference by Deutsche Bundesbank, the German
Research Foundation (DFG) and the International Monetary
Fund. We develop a computational model of a regional housing market. Over a million
distinct agents buy, sell, and rent houses according to different behavior
rules, which depend on demographic, financial, and housing stock characteristics
we estimate using data in the Washington, D.C. metropolitan area
from 1997 -- 2009.
We use both individual record-level matching and statistical inference on
several dozen disparate datasets to simulate a single joint distribution of
household characteristics. Households' transactions endogenously generate a
housing market bubble and crash that resembles the observed history not only in
the timing and magnitude of the boom and bust in home prices, but also in other
aggregate dynamics such as time-on-market, homeownership rate, and vacancy rate
and in distributional characteristics such as house prices across tiers of
building quality and loan performance across bands of credit quality. We use the
model to study the drivers of the bubble. We show that low risk-free interest
rates do not generate a house price bubble when credit availability is
restricted, whereas loose credit contributes to a bubble even without low
# V2
details {
display: inline !important;
}
details > summary {
display: inline;
text-align: justify;
cursor:pointer;
}
details > summary > p{
text-align: justify;
cursor:pointer;
}
span.research {
display: inline-block;
text-align: justify;
text-justify: inter-word;
transition-delay: 0s;
transition-duration: 0.2s;
transition-property: all;
transition-timing-function: ease-in-out;
user-select: text
}
details.description summary::after {
content: attr(data-open);
opacity: 0;
font-size: 5px;
}
details.description[open] summary::after {
content: attr(data-open);
opacity: 1;
font-size: 14px;
}
details.description summary::before {
content: attr(data-close);
opacity: 0;
font-size: 5px;
}
details.description:not([open]) summary::before {
content: attr(data-close);
opacity: 1;
font-size: 14px;
}
details.description summary::after,
details.description summary::before {
display: inline-block;
transition: all .4s ease-in-out;
transform-origin: center bottom;
}