ArticlePDF Available

Automated underwriting in mortgage lending: Good news for the underserved?

Authors:
  • Freddie Mac

Abstract and Figures

Automated underwriting (AU) systems have become the tool of choice in mortgage lending decisions. While these systems provide significant benefits to mortgage originators and investors, questions have been raised about their impact on underserved populations. The questions focus on the relative accuracy of AU compared with manual underwriting and whether AU has increased the flow of mortgage credit to underserved consumers.Using information from Freddie Mac's Loan Prospector AU service, we provide statistics useful in examining these issues. The data strongly support our view that AU provides substantial benefits to consumers, particularly those at the margin of the underwriting decision. We find evidence that AU systems more accurately predict default than manual underwriters do. We also find evidence that this increased accuracy results in higher borrower approval rates, especially for underserved applicants.
Content may be subject to copyright.
Automated Underwriting in Mortgage Lending:
Good News for the Underserved?
Susan Wharton Gates
Freddie Mac
Vanessa Gail Perry
George Washington University
Peter M. Zorn
Freddie Mac
Abstract
Automated underwriting (AU) systems have become the tool of choice in mortgage
lending decisions. While these systems provide significant benefits to mortgage origina-
tors and investors, questions have been raised about their impact on underserved popu-
lations. The questions focus on the relative accuracy of AU compared with manual
underwriting and whether AU has increased the flow of mortgage credit to underserved
consumers.
Using information from Freddie Mac’s Loan Prospector AU service, we provide
statistics useful in examining these issues. The data strongly support our view that
AU provides substantial benefits to consumers, particularly those at the margin of
the underwriting decision. We find evidence that AU systems more accurately predict
default than manual underwriters do. We also find evidence that this increased accu-
racy results in higher borrower approval rates, especially for underserved applicants.
Keywords: Affordability; Homeownership; Mortgages
Introduction
Over the past six years, automated underwriting (AU) systems have
become the tool of choice in mortgage lending decisions. From govern-
ment-sponsored enterprises (GSEs) to large banking institutions to
mortgage insurers, all major mortgage market participants use some
form of AU to evaluate the relative riskiness of mortgage applicants.
The ubiquitous nature of such systems is not in dispute. What has been
questioned is whether the widespread adoption of statistically based
mortgage underwriting is good for underserved populations.
This article sheds some light on that debate, particularly as it refers to
the GSEs. Following a brief discussion of the GSEs and their credit risk
management, we summarize the history of AU and outline the major
Housing Policy Debate ·Volume 13, Issue 2 369
© Fannie Mae Foundation 2002. All Rights Reserved. 369
zorn.qxd 09/17/02 9:15 AM Page 369
370 Susan Wharton Gates, Vanessa Gail Perry, and Peter M. Zorn
questions and concerns it has raised. We then use mortgage application
and performance data from Freddie Mac’s Loan Prospector AU service
to provide a set of statistics useful in examining AU’s impact on
consumers.
We draw two basic conclusions from this analysis: (1) Compared with
traditional manual underwriting, AU more accurately predicts default,
and (2) AU’s greater accuracy results in higher borrower approval
rates, especially for underserved applicants. Because we lack a compre-
hensive model of the mortgage delivery system, our conclusions cannot
be definitive. Nonetheless, the analysis we provide strongly supports
our view that AU provides substantial benefits to consumers, particu-
larly those at the margin of the underwriting decision.
The mission of the GSEs
The GSEs, Freddie Mac and Fannie Mae, are shareholder-owned corpo-
rations chartered by Congress to create a stable flow of funds to mort-
gage lenders in support of homeownership and rental housing. The
GSEs do not originate mortgages themselves but rather fulfill their
mission by purchasing mortgages originated by lenders. To a large
extent, these purchases are funded by packaging mortgages into securi-
ties and selling them to investors, with the GSEs guaranteeing the
timely payment of principal and interest. As a result of their secondary
market activity, the GSEs give lenders broad access to capital markets,
making mortgage funds readily available to consumers. These second-
ary market activities are conducted on a large scale. Since its inception
in 1970, for example, Freddie Mac has purchased more than $2 trillion
in residential mortgages.
A presumption implicit in the granting of the GSEs’ charters is that
secondary market activity will positively affect the primary mortgage
market, which will, in turn, yield significant benefits to the consumer.
A more explicit presumption is that even though the GSEs do not origi-
nate mortgages themselves, their targeted purchases will significantly
expand the homeownership opportunities of underserved populations.
Toward this end, the U.S. Department of Housing and Urban Develop-
ment has since 1994 set goals for the GSEs’ purchase of mortgages for
underserved borrowers.
Many studies have assessed the effect of GSE activity on the housing
market in general and on consumers in particular. Although universal
agreement is lacking, most observers agree that the GSEs’ actions have
brought desirable stability to mortgage markets and have significantly
Fannie Mae Foundation
zorn.qxd 09/17/02 9:15 AM Page 370
Automated Underwriting: Good News for the Underserved? 371
broadened the investor base for housing (see Miller and Pearce 2001).
Moreover, Bostic and Surette (2001), for example, find evidence sup-
porting the view that the broad activities promulgated by the GSEs
result in improved opportunities for homeownership, particularly for
minority and lower-income families.
This article has a more specific focus. We ask whether the AU systems
introduced by the GSEs have, on balance, benefited underserved popu-
lations. We attempt to answer this question with simple statistics on
the population of Loan Prospector applications and subsequent Freddie
Mac purchases. Such an approach necessarily yields conclusions that
are more suggestive than definitive. However, given the scant research
on the impact of AU, this approach seems warranted.
Risk management by the GSEs
Because the GSEs themselves do not originate mortgages, effective
partnership with lenders is key to their successful risk management.
The GSEs generally purchase mortgages from lenders under the terms
of separate, annually renegotiated master contracts. Each contract
specifies the fees lenders must pay to have the GSEs guarantee the
credit risk of delivered loans (expressed as basis points of the loan
amount). Guarantee fees are fixed for all loans delivered under the
contract, and although fees may differ from lender to lender, they are
single-valued for any given lender (i.e., guarantee fees are better char-
acterized as an example of average cost rather than marginal/risk-based
cost pricing).
Lenders’ behavior is central to managing the GSEs’ credit risk. Loans
delivered under master contracts arrive at the GSEs in bulk. The GSEs
do not assess the credit quality of individual loans before purchasing
them. Instead, master contracts require lenders to warrant that all
delivered loans conform to the GSEs’ underwriting standards. In this
manner, lenders act as the GSEs’ agents.
To mitigate the principal-agent risk that lenders fail to perform their
contractual obligations, the GSEs assess the credit quality of loans
after purchase. Under the terms of the contract, loans that do not con-
form to underwriting guidelines can be returned to lenders for repur-
chase. However, because repurchase is costly, terms are governed by
extensively detailed requirements. Consequently, forced repurchase is
relatively rare, and some of the GSEs’ purchases that do not meet their
underwriting standards remain in their portfolios.
Housing Policy Debate
zorn.qxd 09/17/02 9:15 AM Page 371
372 Susan Wharton Gates, Vanessa Gail Perry, and Peter M. Zorn
Given that GSEs insure credit risk on mortgages they do not originate,
accurate risk assessment is essential. Credit risk management is also
enhanced by a better ability to clearly and precisely communicate this
assessment to lenders.
AU
In 1995, Freddie Mac introduced Loan Prospector, its statistically based
AU system. AU itself, however, is not a new phenomenon. Since the
1970s, such systems have been used in automobile and credit card lend-
ing, where borrower credit is the major empirical driver of default.
As noted in the summary by Straka (2000), the increased complexity of
mortgage underwriting, particularly the strong role of borrower equity,
delayed attempts to introduce AU to the nation’s mortgage markets. By
the mid-1990s, however, attempts to marry consumer credit scoring
models with mortgage models began to bear fruit. Designed along the
lines of the three “Cs” of manual underwriting—credit, capacity, and
collateral—custom mortgage scoring models incorporated credit score
data summarizing applicants’ credit histories with mortgage loan appli-
cation data. These data include, for example, loan-to-value (LTV) ratios,
property type, borrower debt ratios, and, in some instances, proprietary
market forecasts.
Model developers initially experimented with different means of evaluat-
ing credit and mortgage data. Methods such as artificial intelligence,
however, quickly gave way to statistically based models that predict mort-
gage default based on the actual performance of millions of mortgages.
By focusing solely on those mortgage or credit characteristics relevant to
repayment, statistically based AU systems can offset variables indicating
higher risk with other factors in the application that have been proven
statistically to reduce risk. This ability to account for the multitudinous
ways in which risk factors influence each other is one of the major advan-
tages automated approaches have over manual underwriting.
In 1996, Federal Reserve economists examined the use of credit scores
in the mortgage underwriting decision and concluded that they are sta-
tistically valid predictors of default. These researchers further con-
cluded that credit scoring could result in more loan approvals:
In principle, a well-constructed credit-history scoring system holds
the promise of increasing the speed, accuracy, and consistency of the
credit evaluation process while reducing costs. Thus, credit scoring
can reduce risk by helping lenders weed out applicants posing exces-
sive risk and can also increase the volume of loans by better identi-
fying creditworthy applicants. (Avery et al. 1996, 627)
Fannie Mae Foundation
zorn.qxd 09/17/02 9:15 AM Page 372
Automated Underwriting: Good News for the Underserved? 373
From the perspective of credit guarantors such as the GSEs, AU pro-
vides significant advantages over manual underwriting. A key benefit is
AU’s superior ability to clearly and precisely communicate GSEs’ risk
standards to lenders. Loans that meet Freddie Mac’s underwriting
guidelines, for example, receive a simple “accept” or “accept-plus” des-
ignation from Loan Prospector; the remaining loans receive a “caution”
designation, as well as specific feedback about why the loan is consid-
ered higher risk.1This crisp delineation of risk contrasts with the nec-
essarily extensive Freddie Mac Seller/Servicer Guide on which Loan
Prospector was initially based and which still is used for manually
underwritten loans. As a result, lenders using Loan Prospector receive
clearer, more precise feedback on whether loans meet Freddie Mac’s
underwriting standards and are therefore better able to deliver loans
that meet these standards.
AU further reduces principal-agent problems by decreasing the extent
of asymmetric information inherent in the GSE-lender relationship (see
Cutts, Van Order, and Zorn 2001). Sophisticated statistical modeling
combined with large numbers of observations offer the potential for AU
to be far more accurate than manual underwriting, even though man-
ual underwriters sometimes have more detailed borrower information
at their disposal. In addition, AU benefits the lender agent by virtually
eliminating the GSEs’ ability to force loan repurchase except in cases of
fraud or failure to deliver the specified loan.
AU systems have also brought significant benefits to borrowers. The
process of obtaining a mortgage under manual underwriting is slow,
cumbersome, and costly. Mortgage applicants typically wait weeks to
learn whether they have been approved for a loan. Rejected borrowers
receive little information about why their applications were rejected or
how they can improve their overall mortgage credit profile. Moreover,
the subjectivity of the approval and feedback process under manual
underwriting makes mortgage lending more vulnerable to fair lending
violations, intended or otherwise.
The growing use of AU is not surprising, given the potentially signifi-
cant benefits accruing to all parties in the mortgage transaction. Today
it is estimated that AU is used to originate approximately 60 to 70 per-
cent of residential mortgages, and the trend shows no sign of abating.
Roughly 65 to 70 percent of Freddie Mac’s current loan purchases come
through Loan Prospector. Freddie Mac alone has processed well over
10 million conventional, conforming mortgage applications using this
Housing Policy Debate
1Loans receiving an “accept-plus” classification require less documentation than those
getting an “accept.”
zorn.qxd 09/17/02 9:15 AM Page 373
374 Susan Wharton Gates, Vanessa Gail Perry, and Peter M. Zorn
system. Other widely used AU systems include Fannie Mae’s Desktop
Underwriter, Wells Fargo’s ECS service, and Countrywide’s CLUES.
AU is not only used in the conventional, conforming market, but else-
where as well. In 1996, for example, a custom version of Loan Prospec-
tor was developed and approved for loans insured by the Federal
Housing Administration. The following year, the Department of Veter-
ans Affairs approved the use of Loan Prospector. Automated systems
are also used to underwrite jumbo mortgages. Using Loan Prospector,
for example, Freddie Mac can assess the risk of loans that exceed the
conforming loan limit and thus are ineligible for purchase by the GSE.
AU, however, has yet to penetrate the subprime segment of the mort-
gage market to any significant extent.
Concerns over AU
To be sure, the industry’s rapid adoption of AU has not occurred with-
out some controversy. Many of the questions center on the implications
of AU for underserved populations, as described in the following
sections.
Does AU simply codify social inequities?
AU relies heavily on objective, electronically verifiable data such as
credit scores, financial reserves, and LTV ratios. One concern is that
the reliance on historical performance data simply codifies and perpetu-
ates social inequities (Barefoot 1997). As noted by Bunce, Reeder, and
Scheessele (1999), the codification of underwriting guidelines that are
not representative of the mortgage behavior of all groups could result
in unintentional discrimination. To avoid this outcome, the specifica-
tion of the mortgage scoring models must “assign risk based on objec-
tive criteria” (Bunce, Reeder, and Scheessele 1999, 41).
Another concern is that minority and lower-income households may be
systematically underrepresented in the baseline populations used to
develop the models. This is particularly problematic because some con-
sumer advocates argue that minority consumers differ in their use of
credit and financial services. As a result, models with these data may
not be appropriate for all subgroups (see Avery et al. 2000).
Finally, to the extent that wealth and credit variables, as well as the
electronic verification thereof, are unevenly distributed, underserved
populations may remain underserved under AU (Barefoot 1997; Glass-
man and Wilkins 1997).
Fannie Mae Foundation
zorn.qxd 09/17/02 9:15 AM Page 374
Automated Underwriting: Good News for the Underserved? 375
Will AU reverse progress in expanding homeownership among
underserved populations?
Some advocates argue that AU’s objective, quantitative nature has cre-
ated a mortgage “meritocracy” that disadvantages underserved popula-
tions (Steinbach 1998). That is, just at the time when gains have been
made in expanding access to mortgage markets, AU introduces a new
set of immutable rules. As noted by Bunce, Reeder, and Scheessele
(1999), marginal borrowers who have been helped most in recent years
by the rise of affordable lending programs are disproportionately repre-
sented among borrowers rejected under strict statistically based under-
writing systems.
Does the lack of transparency around AU serve as a foil for
discrimination?
This issue relates to the “black box” nature of AU systems. In virtually
all cases, the details of underlying models are considered proprietary by
their developers and, consequently, are not disclosed to the public. The
failure to reveal information about AU systems and credit scoring in
general has fueled allegations of intended and unintended discrimina-
tion in mortgage lending (Quinn 2000). Although AU systems have
recently become more transparent (see, for example, the corporate Web
sites of Freddie Mac, Fannie Mae, and Fair, Isaac & Company), issues
nonetheless remain.
The concerns raised about AU are not trivial and merit significant
attention from mortgage researchers and policy makers. Perhaps the
most fundamental question is whether AU really is better than manual
underwriting. It has been argued that manual underwriting, because it
relies on personal contact and subjective information, is more accurate
than AU. Although statistically based models can efficiently deal with
mainstream loan characteristics, human underwriters are thought by
some to be better at making appropriate exceptions that benefit under-
served populations (Glassman and Wilkins 1997).
In many ways, therefore, the question of adverse outcomes under AU
turns on the issue of accuracy. If AU is less accurate, it is reasonable to
expect that underserved populations, broadly speaking, will fare worse
than under manual underwriting. However, if AU is more accurate,
then the question becomes one of measuring the impact of this
increased underwriting accuracy on mortgage consumers.
Housing Policy Debate
zorn.qxd 09/17/02 9:15 AM Page 375
376 Susan Wharton Gates, Vanessa Gail Perry, and Peter M. Zorn
Data and analytical approach
We use Freddie Mac proprietary data to explore AU’s accuracy and
impact on borrowers, particularly underserved populations. We first
examine the accuracy of AU across different borrower groups and as
compared with manual underwriting. To assess the impact of AU on
underserved mortgage applicants, we next analyze the percentage of
lower-income borrowers (whose household income is equal to or less
than 100 percent of the area median) and minority borrowers (other
than white, non-Hispanic) among Freddie Mac’s purchases of loans
processed through Loan Prospector.
Our analysis is based on three different data sources. To assess loan-
level performance, we use data from Freddie Mac purchases of 1994
and 1995 originations of conventional, conforming mortgages secured
by one-unit owner-occupied properties. Although these loans were not
initially scored using Loan Prospector (which in fact did not exist at
the time they were originated), these loans were later graded using an
emulated version of the model.2Because these loans are part of Freddie
Mac’s portfolio, we can also track their delinquency performance. The
data we provide extend through the second quarter of 2001.
To compare the relative accuracy of AU with manual underwriting, we
also assess the performance of nearly 1,000 loans originated in 1993 and
1994 and purchased by Freddie Mac in 1995 as part of an affordable
housing initiative with a major lender. As part of the purchase require-
ments, Freddie Mac had these loans manually underwritten, with
underwriters grading them as investment quality (“accept”) and nonin-
vestment quality (“caution”). For the purposes of this study, we subse-
quently re-underwrote these loans using emulated versions of Loan
Prospector’s 1995 and 2000 models. Finally, we attach delinquency per-
formance data for these loans through the second quarter of 2001.
While not randomly drawn from Freddie Mac purchases, this sample
provides a unique test of the effects of AU. By and large, these loans
are at the margin of the highest acceptable risk to Freddie Mac—pre-
cisely the loans for which critics have expressed the most concern.
Finally, we analyze all applications processed through Loan Prospector
from 1995 to 2000 and ultimately purchased by Freddie Mac. Over this
period, Freddie Mac purchased roughly one-third of the applications
processed by Loan Prospector, while two-thirds of the applications were
never delivered.
Fannie Mae Foundation
2The emulated versions of Loan Prospector very closely approximate the classifications
of the production version, but by necessity must deal with occasional missing variables.
zorn.qxd 09/17/02 9:15 AM Page 376
Automated Underwriting: Good News for the Underserved? 377
These data provide an important first look into how AU affects the
origination process of lenders and the delivery of loans to the GSEs.
However, it is worth emphasizing that our study takes a macro
approach to the issues surrounding AU. By virtue of the data used, this
article paints in broad strokes. In particular, we focus on net impacts
(whether more applicants are helped than hurt). Consumer advocates,
with their street-level view of such issues and their focus on individual
outcomes, may find our broad assertions and findings interesting but
ultimately unsatisfying.
This caveat notwithstanding, we believe that our bird’s eye view pro-
vides a useful starting point for addressing consumer issues related to
AU. As will be seen, manual underwriters would have accepted some
mortgage applications that are rejected by AU systems. However, the
data strongly suggest that, on average, more underserved applicants
are accepted under AU than under manual underwriting.
Statistical findings
We use our data to provide simple statistics that address the questions
of accuracy and approval rates.
Accuracy
An AU system can be described as accurate if the loans it predicts to be
low risk actually perform substantially better than loans predicted to be
high risk. Loan Prospector has precisely these characteristics, as illus-
trated in figure 1. According to Freddie Mac purchases of loans origi-
nated in 1994 and 1995, loans rated caution by Loan Prospector’s 2000
model experienced default at four times the average rate of all loans. By
contrast, loans rated accept-plus experienced default at just one-fifth the
average default rate. (Similar results hold for the 1995 version.)
Moreover, Loan Prospector predicts default for both low-income and
minority borrowers. Low-income borrowers rated caution default at
four times the average, while minority borrowers rated caution default
at roughly five times the average. Clearly, there is substantial evidence
that Loan Prospector accurately predicts default.3
Housing Policy Debate
3To add statistical rigor to this assertion, we also use the three Loan Prospector risk
classes to estimate a logit model of 90-day delinquency separately for all borrowers,
lower-income borrowers, and minority borrowers. In all three estimations, we find that
the estimated coefficients for the risk classes are each statistically significantly differ-
ent from zero, as well as statistically significantly different from each other.
zorn.qxd 09/17/02 9:15 AM Page 377
378 Susan Wharton Gates, Vanessa Gail Perry, and Peter M. Zorn
Figure 1. Relative 90-Day Delinquency Rates of 1994 and 1995 Originations
Note: Delinquency rates are expressed relative to the overall average.
A more challenging question is whether Loan Prospector is more accu-
rate than manual underwriting. We address this question by using the
subsample of roughly 1,000 loans purchased by Freddie Mac as part of
an affordable housing initiative. Figure 2 compares the actual perform-
ance of the affordable loans as assessed by manual underwriters and
Loan Prospector. Panel A illustrates the comparison between loans rated
by manual underwriters and those rated by the 1995 version of Loan
Prospector. Panel B shows the same comparison for the 2000 version.
As can be seen, some loans were rated accept by Loan Prospector but
caution by manual underwriters, while others were rated caution by
Loan Prospector but accept by manual underwriters. These two groups
of loans—the “swap-ins” and “swap-outs”—are of most interest.
Panel A shows that the 1995 version of Loan Prospector does a better
job of distinguishing between high- and low-risk loans than manual
underwriters do. In particular, 90-day delinquency rates for swap-ins
are one-fifth the average, and these loans perform like those rated
accept by both Loan Prospector and manual underwriters. Similarly,
90-day delinquency rates for swap-outs are 1.75 times the average, and
these loans perform like those rated caution by both Loan Prospector
and manual underwriters. Loan Prospector, therefore, provides more
accurate predictions in these cases than manual underwriters do.4
0.2 0.3 0.2
1.1 1.4
1.9
4.0 4.0
4.9
0
3
6
All Borrowers Lower-Income
Borrowers Minority
Borrowers
Accept Plus Accept Caution
Percent
Relative Delinquency Rate
Fannie Mae Foundation
4The statistical validity of this claim is supported by the results of a logit estimation
explaining 90-day delinquency as a function of the four alternative risk classifications
shown in panel A (Loan Prospector accept/caution interacted with manual underwrit-
ing accept/caution). Each estimated coefficient in the model is statistically significant
zorn.qxd 09/17/02 9:15 AM Page 378
Automated Underwriting: Good News for the Underserved? 379
Figure 2. Relative 90-Day Delinquency Rates for Affordable Housing Loans
Note: Delinquency rates are expressed relative to the overall average.
A. 1995 Version of Loan Prospector versus Manual Underwriting
0.21 0.21
1.52 1.75
0
2
4
Loan Prospector Risk Class Accept Accept Caution Caution
Manual Risk Class Accept Caution Caution Accept
Loans (%) 24.0 20.8 27.7 27.5
B. 2000 Version of Loan Prospector versus Manual Underwriting
0.66 0.75
2.55
3.52
0
2
4
Relative Delinquency Rate
Relative Delinquency Rate
Loan Prospector Risk Class Accept Accept Caution Caution
Manual Risk Class Accept Caution Caution Accept
Loans (%) 44.7 42.7 5.8 6.9
Housing Policy Debate
from zero. More important, the coefficient on swap-ins is statistically different from
the coefficients on both swap-outs and loans assessed caution by both systems, but
insignificantly different from loans rated accept by both systems. Similarly, the coeffi-
cient on swap-outs is statistically different from the coefficients on swap-ins and loans
assessed accept by both systems, but insignificantly different from loans rated caution
by both systems. Clearly, Loan Prospector is the more accurate system—it identifies
swap-ins that perform like accepts and swap-outs that perform like cautions.
zorn.qxd 09/17/02 9:15 AM Page 379
380 Susan Wharton Gates, Vanessa Gail Perry, and Peter M. Zorn
Subsequent refinements have only enhanced the predictive power of
the system. Panel B compares the 2000 version of Loan Prospector with
manual underwriting. Again, the performance of swap-ins and swap-
outs is most relevant, and again swap-ins perform like loans rated
accept by both systems while swap-outs perform like loans rated cau-
tion by both systems.5Where the underwriting systems disagree, both
versions of Loan Prospector clearly predict the actual outcomes for
these lower-income borrower data better than manual underwriters do.
The two panels of figure 3 repeat these statistics for the minority bor-
rower subsample among these affordable housing loans. Again, in both
panels, swap-ins perform like loans rated accept by both systems, while
swap-outs perform like loans rated caution by both systems. Thus, both
versions of Loan Prospector clearly predict the actual outcomes of
minority borrowers in these data better than manual underwriters do.6
Approval rates
AU systems can be said to benefit potential borrowers if they result in
increased approvals for applicants. The data presented in figures 2 and
3 allow us to make this assessment. Comparisons between the 1995
version of Loan Prospector and manual underwriting show greater
accept rates with manual underwriting (there are fewer swap-ins than
swap-outs). This outcome is not surprising; it reflects Freddie Mac’s
initially conservative setting of risk standards when first introducing
Loan Prospector. Over time, however, Freddie Mac rapidly expanded
accept rates as the tool became more accurate and the company gained
experience with and confidence in the new technology. The B panels of
figures 2 and 3 show the dramatic results of these changes. Not only is
the 2000 version of Loan Prospector far better at predicting risk than
manual underwriting, it has a net swap-in of 36 percent for all afford-
able housing loans and a net swap-in of 29 percent for the minority
borrowers subgroup. Clearly, there is a strong case for the argument
that AU expands mortgage approvals relative to manual underwriting.
Fannie Mae Foundation
5The logit estimation associated with panel B similarly shows that swap-ins identified
by Loan Prospector perform like accepts and that swap-outs perform like cautions. In
this estimation, however, the coefficient on swap-outs is insignificantly different from
zero, where a zero coefficient implies that relative 90-day delinquency rates for swap-
outs are predicted to be roughly 4.8 times the average.
6We repeat the logit estimations associated with figure 2 and obtain similar results—
accepts perform worse than cautions, swap-ins identified by Loan Prospector perform
like accepts, and swap-outs perform like cautions. For the estimation associated with
panel B, we find that the coefficient on swap-outs is insignificantly different from zero,
as is the coefficient on loans rated caution by both Loan Prospector and manual under-
writers. In this instance, a zero coefficient implies predicted relative 90-day delin-
quency rates of roughly 2.3 times the average.
zorn.qxd 09/17/02 9:15 AM Page 380
Automated Underwriting: Good News for the Underserved? 381
Figure 3. Relative 90-Day Delinquency Rates for Affordable Housing Loans to
Minority Borrowers
Note: Delinquency rates are expressed relative to the overall average.
Freddie Mac continues to broaden the definition of an accept loan as it
enhances and fine-tunes the statistical models in Loan Prospector. Fig-
ure 4 compares the accept rates on all applications processed by Loan
Prospector in 2000 (both the 1995 and 2000 versions). The results show
that the newer version has substantially higher accept rates. For exam-
ple, the 1995 version approved 23 percent of black borrowers, compared
with 54 percent under the 2000 version—an increase of 31 percentage
A. 1995 Version of Loan Prospector versus Manual Underwriting
0.37 0.37
1.14 1.38
0
2
4
Accept Accept Caution Caution
Accept Caution Caution Accept
16.4 12.6 36.6 34.5
B. 2000 Version of Loan Prospector versus Manual Underwriting
0.71 0.71
2.14 2.18
0
2
4
Loan Prospector Risk Class
Manual Risk Class
Loans (%)
Reduction in
Delinquency Rate
Accept Accept Caution Caution
Accept Caution Caution Accept
38.9 41.2 7.9 12.0
Loan Prospector Risk Class
Manual Risk Class
Loans (%)
Reduction in
Delinquency Rate
Housing Policy Debate
zorn.qxd 09/17/02 9:15 AM Page 381
382 Susan Wharton Gates, Vanessa Gail Perry, and Peter M. Zorn
Figure 4. Accept Rates for Loan Prospector Applications in 2000
points. The accept rate for Hispanic and nonminority applicants grew
by 31 and 27 percentage points, respectively.
While it is difficult to quantify this effect precisely, it is reasonable to
assume that steady increases in the share of loans rated accept by Loan
Prospector exert a positive influence on mortgage originations and, in
turn, homeownership. One thing we can show precisely, however, is the
share of all loan applications rated accept or better by Loan Prospector
and ultimately purchased by Freddie Mac. This trend is upward-sloping
for both low- and moderate-income and minority borrowers. Figure 5
shows that the share of minority purchases rose from 8.5 percent in
1995 to 14.9 percent in 2000. Similarly, the low- and moderate-income
share of Freddie Mac purchases rose from 32 percent to 41 percent dur-
ing the same period. By using Loan Prospector, therefore, Freddie Mac
has purchased an increasing share of loans to underserved populations.
Discussion
Freddie Mac’s experience suggests that AU is more accurate and ap-
proves more underserved applicants than manual underwriting does. It
is tempting to conclude, based on these findings, that AU’s increased
accuracy results in increased homeownership opportunities, particularly
23.2
31.5
49.8
53.5
63.3
76.7
0
25
50
75
100
Black
Borrowers Hispanic
Borrowers Nonminority
Borrowers
1995 Loan Prospector Scorecard
2000 Loan Prospector Scorecard
Percent
Fannie Mae Foundation
zorn.qxd 09/17/02 9:15 AM Page 382
Automated Underwriting: Good News for the Underserved? 383
Figure 5. Freddie Mac Purchases of Loan Prospector Applications
for underserved families. Because of the attenuation of the secondary
market from the actual origination of mortgages, it is difficult to estab-
lish a definitive correlation between GSE accept rates, mortgage origina-
tions, and, consequently, homeownership. This caveat notwithstanding,
we contend that AU’s improved accuracy is good news for underserved
households. Not only is this view consistent with the statistical findings,
it is intuitively appealing.
31.9 34.7 36.9 35.6
41.0 41.0
0
25
50
PercentPercent
1995 1996 1997 1998 1999 2000
A. Percentage of Low- and Moderate-Income Borrowers
8.5
10.1 10.8 10.2
13.2
14.9
0
10
20
1995 1996 1997 1998 1999 2000
B. Percentage of Minority Borrowers
Housing Policy Debate
zorn.qxd 09/17/02 9:15 AM Page 383
384 Susan Wharton Gates, Vanessa Gail Perry, and Peter M. Zorn
Underwriting is ultimately a prediction game. Because the mortgage-
default function is extremely asymmetric (defaults are rare but very
costly), the challenge is to approve as many low-risk loans as possible
while rejecting the few that impose the highest risk/cost. Idiosyncratic
events such as unemployment and divorce make it impossible to know
with certainty ex ante how an individual mortgage applicant will per-
form over time. The best developers of mortgage scoring models can do
is to accurately predict applicants’ probability of default.
Given this high degree of uncertainty in identifying the loans that will
ultimately default, mortgage lenders traditionally price risk on average
(e.g., there is relatively little variation in the pricing of prime loans).
Under a system of average pricing, lower-risk loans are priced above
cost, effectively helping to cover the losses on the higher-risk loans that
are more likely to default. This cross-subsidization works well for all
market participants as long as everyone has similar information about
the loan in question. It can break down, however, when some partici-
pants have better information on loan quality and are therefore able to
select against others, leaving them with lower-quality loans. As noted
by Cutts, Van Order, and Zorn (2001), average cost pricing and the
resulting concern over adverse selection causes mortgage guarantors
such as Freddie Mac and Fannie Mae to set a maximum risk they are
willing to take (set an underwriting/risk standard cutoff). The higher
the GSEs’ confidence in their ability to assess risks, the higher the
aggregate level of risk they will tolerate.
The potential impacts of increased underwriting accuracy are illus-
trated in figure 6. The dashed line represents the distribution of
predicted default probabilities for mortgage applicants under an under-
writing model that does a better job of separating mortgage applicant
risk. More accurate models reduce the share of applicants to the right
of the risk cutoff by moving more applicants to the tails of the default
distribution and disproportionately more applicants to the low-risk
tail. This shift represents the ability of more accurate models to more
closely predict the ex post outcome that the vast majority of applicants
will not default on their mortgages.
By itself, this shift in distribution ensures that a greater percentage
of mortgage applications meet the maximum risk cutoff. At the same
time, however, increased accuracy (i.e., tighter confidence intervals
around model predictions) enables the risk cutoff to shift outward to a
higher predicted default probability while maintaining a constant over-
all risk standard. This also results in a greater number of applicants
meeting the maximum risk cutoff. It is not surprising, therefore, that
the combination of these effects yields the observed empirical outcome;
the increased accuracy of AU results in higher accept rates.
Fannie Mae Foundation
zorn.qxd 09/17/02 9:15 AM Page 384
Automated Underwriting: Good News for the Underserved? 385
Figure 6. Effect of Introducing More Accurate Underwriting Models
It is also not surprising that the increased accuracy of AU benefits to a
larger extent underserved populations. This group tends disproportion-
ately to have higher-risk values for the attributes commonly used when
underwriting mortgages. As a result, underserved populations stand to
gain the most from AU’s enhanced ability to better distinguish between
low- and high-risk applicants at the margin of acceptable risk.
Figure 7 uses the distribution of Fair, Isaac & Company (FICO) scores
for minority and nonminority applicants processed through Loan
Prospector in 2000 to illustrate this point. For example, 43 percent of
minority applications have FICO scores falling in the 580 to 679 range,
arguably the area of close calls in underwriting. By contrast, 32 percent
of nonminority applications fall within this range.
Conclusions and challenges
An examination of Freddie Mac data suggests that AU systems
are more accurate than manual underwriting. Furthermore, AU’s
improved accuracy appears to be good for consumers: On average, more
applicants are swapped in than are swapped out. Underserved popula-
tions, particularly, appear to benefit from the system’s greater accuracy.
Improved risk assessment reduces the uncertainty of predicting mort-
gage defaults, thus allowing risk managers to set a more lenient risk
standard while maintaining a constant overall level of risk. The result-
ing broadening in underwriting standards has the effect of expanding
Predicted Default Probability
Less accurate underwriting model
Risk cutoff with a less accurate model
Risk cutoff with a more accurate model
Share of Population
More accurate underwriting model
Housing Policy Debate
zorn.qxd 09/17/02 9:15 AM Page 385
386 Susan Wharton Gates, Vanessa Gail Perry, and Peter M. Zorn
Figure 7. FICO Score Distributions for Loan Prospector Applications in 2000
homeownership opportunities, particularly for underserved popula-
tions. As noted by Bunce, Reeder, and Scheessele (1999), Fannie Mae
and Freddie Mac, for example, seem to be more confident and willing
to purchase low-income loans if they are processed through their
mortgage-scoring models. Both government-sponsored enterprises have
recently introduced new 3-percent down payment programs, based on
their confidence in their automated underwriting systems.
The improved understanding of credit risk has also led to the develop-
ment of an array of more flexible mortgage products that better serve
the needs of borrowers who could not meet traditional underwriting
standards. For example, conventional mortgage loans traditionally
require at least a 10 percent down payment and enough cash to cover
closing costs with two to three months of reserves. Potential borrowers
have to provide evidence of stable employment and detailed income
documentation; these requirements place seasonal and self-employed
workers at a disadvantage. In addition, potential borrowers have to
provide evidence of a good credit history, based on payment records for
consumer loans such as credit cards and car loans.
AU, however, encourages increased flexibility in mortgage products—
low or no down payment requirements, higher debt-to-income ratios,
reduced cash reserve requirements, flexible employment standards,
reduced points and fees, and reduced mortgage insurance. Additional
mortgage products focus on “A-minus” or subprime loans for borrowers
with prior credit issues (Peterson 2001). According to Harvard’s Joint
11
7
43
32 32 35
14
27
0
25
50
< 580 580–679 680–749 > 750
Minority Nonminority
Percent
Fannie Mae Foundation
zorn.qxd 09/17/02 9:15 AM Page 386
Automated Underwriting: Good News for the Underserved? 387
Center for Housing Studies, the advent of new loan products such as
these has enabled “more income-constrained and cash-strapped bor-
rowers at the margin to qualify for mortgage loans” (1998, 1).
Increased underwriting flexibility offers significant advantages for
underserved populations. For example, a recent study published by the
Research Institute for Housing America finds that borrowing con-
straints, such as are embodied in underwriting guidelines, dispropor-
tionately restrict access to homeownership across racial and ethnic
groups. In the case of white and Hispanic applicants, borrowing con-
straints serve to delay home purchase, while primarily excluding black
applicants. The study estimates that removing borrowing constraints
would increase the homeownership rate by 4 percentage points, ceteris
paribus (Rosenthal 2001).
Increased accuracy, as we readily acknowledge, results in the rejection
of some borrowers who might have fared better with manual under-
writing. However, the data indicate that this swap-out is far less than
the swap-in resulting from AU’s ability to identify and approve borrow-
ers who otherwise would have been rejected by manual underwriters.
The compelling evidence for this claim is that aggregate approval rates
are much higher under AU than under manual underwriting. From a
broad public policy perspective, therefore, AU appears to provide a net
expansion of homeownership opportunities, even though not every
applicant will benefit.
Increased accuracy will also likely drive the mortgage market toward
risk-based pricing because increased accuracy and strong competition
will tend to eliminate the cross-subsidization inherent in average cost
pricing. While it is beyond the scope of this article to quantify the net
impact of risk-based pricing on consumers, we expect it to be largely
positive. Small changes will likely occur in the prime market; con-
sumers with the strongest credit profiles will pay slightly less for
mortgage credit, while those with weak profiles will pay slightly more.
More significantly, however, risk-based pricing will probably lower the
cost of mortgage credit for consumers who mistakenly find themselves
in the higher-cost subprime segment of the market because this seg-
ment is far less competitive than the prime market, and mortgage
prices vary widely. On average, therefore, we expect that risk-based
pricing will result in a broader allocation of mortgage credit at com-
petitive prices that accurately reflect a borrower’s risk.
Finally, despite these positive developments in mortgage lending,
underserved borrowers are still less likely to be approved for a loan.
This differential mirrors broader societal inequities in financial capac-
ity and credit, which are key variables in both automated and manual
Housing Policy Debate
zorn.qxd 09/17/02 9:15 AM Page 387
388 Susan Wharton Gates, Vanessa Gail Perry, and Peter M. Zorn
underwriting. The challenge for the mortgage industry is to expand
homeownership opportunities by reducing the differential credit effect
for low-income and minority borrowers. We believe that to achieve this
end, four key issues must be addressed:
“Unbanked” consumers
According to the Federal Reserve’s Survey of Consumer Finances, in
1998 approximately 9.5 percent of U.S. households did not have any
type of bank transaction account, such as a checking or savings
account, and over 13 percent did not have a checking account (Ken-
nickell, Starr-McCluer, and Surette 2000). A critical issue is how to
extend homeownership opportunities to unbanked consumers, who dis-
proportionately include low-income and minority households. Because
AU models rely on records established through traditional banking
relationships, potential home buyers who lack these relationships are
at a distinct disadvantage. Policy makers, regulators, GSEs, and lenders
must continue to explore ways to increase accessibility to financial
services.
Financial literacy
To achieve financial stability in our increasingly cashless society, con-
sumers need to become familiar with information about credit and
accumulation of wealth. Not all consumers obtain this information
from reliable sources, potentially producing differences in its use.
Historically, black households have had fewer educational opportunities
relative to the majority population. In addition to differential access to
education, differences—including the effects of discrimination—in
income, employment, access to financial and credit markets, and accu-
mulation of wealth have undoubtedly contributed to a lack of cumula-
tive experience with the financial system. Due to a lack of trust or
familiarity with formal sources, members of this group are also more
likely to rely on informal sources (Ards and Myers 2000).
In the face of rising consumer debt burdens and concerns about preda-
tory lending, government regulators, policy makers, and consumer
advocates worry that many consumers lack the knowledge to success-
fully navigate the complex financial marketplace. According to Alan
Greenspan, the mortgage industry “should work to educate consumers
on evaluating the broad array of products offered by financial service
providers and to empower them to make the choices that contribute to
their overall economic well-being” (2001).
Fannie Mae Foundation
zorn.qxd 09/17/02 9:15 AM Page 388
Automated Underwriting: Good News for the Underserved? 389
Home buyer education
Increasingly, lenders are offering affordable mortgage products that
require borrowers to undergo home buyer education (McCarthy and
Quercia 2000). The requirements vary, but all are based on the premise
that borrowers who better understand how to obtain and maintain a
mortgage are less likely to become delinquent or default. Hirad and
Zorn (2002) find some support for this premise in an empirical exami-
nation of the effect of home buyer education on credit risk. These
results have profound implications for developing credit-scoring models
and for expanding homeownership opportunities. Although further
research is necessary, these findings suggest, for example, that home
buyer counseling can be used to increase the success of affordable lend-
ing programs for both lenders and borrowers.
Digital divide
Home buyers increasingly use the Internet to gather information about
homes for sale, to obtain credit and homeownership counseling, to shop
for as well as to apply for mortgages, and to communicate with agents
and lenders. There are substantial differences in Internet access and
usage across income, racial, and ethnic groups. This digital divide could
potentially result in informational disadvantages for lower-income and
less educated consumers that could deter them from becoming home-
owners. A Vanderbilt University study reports, for example, that white
households use the Internet more than black households do and are
significantly more likely to do so because they have a computer at home
(Hecht 2001). Clearly, the Internet cannot solely be relied on as a distri-
bution channel for mortgage credit.
In summary, over the past few years AU systems such as Freddie Mac’s
Loan Prospector have become increasingly accurate in assessing mort-
gage credit risk. By reducing the uncertainty of the mortgage allocation
process, AU systems have enabled the mortgage industry, in aggregate,
to assume more credit risk. Underserved borrower populations have
been the chief beneficiaries of this increased accuracy. To expand mar-
kets further, refinements in mortgage risk assessment must be accom-
panied by a greater focus by the mortgage industry on technological
and informational disparities.
Housing Policy Debate
zorn.qxd 09/17/02 9:15 AM Page 389
390 Susan Wharton Gates, Vanessa Gail Perry, and Peter M. Zorn
Authors
Susan Wharton Gates is Director of Public Policy at Freddie Mac. Vanessa Gail Perry is
Assistant Professor of Marketing at George Washington University. Peter M. Zorn is
Vice President, Housing Economics, at Freddie Mac.
We gratefully acknowledge the substantial insights and suggestions of Amy Cutts,
James Follain, Jeff Markowitz, and two anonymous referees, as well as the extraordi-
nary research assistance of Abdighani Hirad, Cynthia Waldron, and Suzy Ptaszynski.
Any errors are our responsibility alone. The opinions expressed in this article are those
of the authors and do not necessarily reflect the views of Freddie Mac or its board of
directors.
References
Ards, Sheila, and Samuel Myers Jr. 2000. Credit and Knowledge: The African-American
Experience. Presented at the annual meeting of the Southern Economic Association,
Crystal City, VA, November 10.
Avery, Robert B., Raphael Bostic, Paul S. Calem, and Glenn B. Canner. 1996. Credit
Risk, Credit Scoring, and the Performance of Home Mortgages. Federal Reserve Bul-
letin 82(7):621–48.
Avery, Robert B., Raphael Bostic, Paul S. Calem, and Glenn B. Canner. 2000. Credit
Scoring: Statistical Issues and Evidence from Credit Bureau Files. Real Estate Econom-
ics 28(3):523–47.
Barefoot, JoAnn S. 1997. The Next Fair-Lending Risk: Disparate Impact. ABA Banking
Journal 89(5):28–32.
Bostic, Raphael W., and Brian J. Surette. 2001. Have the Doors Opened Wider? Trends
in Homeownership Rates by Race and Income. Journal of Real Estate Finance and Eco-
nomics 23(3):411–34.
Bunce, Harold L., William J. Reeder, and Randall M. Scheessele. 1999. Understanding
Consumer Credit and Mortgage Scoring: A Work in Progress at HUD. Presented at the
Fannie Mae Foundation Research Roundtable: Making Fair Lending a Reality in the
New Millennium, Washington, DC, June 30.
Cutts, Amy Crews, Robert A. Van Order, and Peter M. Zorn. 2001. Adverse Selection,
Licensing, and the Role of Securitization in Financial Market Evolution and Pricing.
Unpublished working paper. Freddie Mac.
Glassman, Cynthia A., and Howard M. Wilkins. 1997. Credit Scoring: Probabilities and
Pitfalls. Journal of Retail Banking Services 19(2):53–56.
Greenspan, Alan. 2001. The Importance of Education in Today’s Economy. Keynote
address to A Federal Reserve System Community Affairs Research Conference on
Changing Financial Markets and Community Development, Washington, DC, April 5–6.
World Wide Web page <http://chicagofed.org/cedric/2001/conferencebook.pdf> (accessed
May 2, 2002).
Fannie Mae Foundation
zorn.qxd 09/17/02 9:15 AM Page 390
Automated Underwriting: Good News for the Underserved? 391
Hecht, Ben. 2001. Bridging the Digital Divide. Journal of Housing and Community
Development 58(2):14–17, 45.
Hirad, Abdighani, and Peter M. Zorn. 2002. A Little Knowledge Is a Good Thing:
Empirical Evidence on the Effectiveness of Prepurchase Homeownership Counseling.
In Low-Income Homeownership: Examining the Unexamined Goal. Washington, DC:
Brookings Institution. Forthcoming.
Joint Center for Housing Studies. 1998. State of the Nation’s Housing. Cambridge, MA:
Harvard University.
Kennickell, Arthur, Martha Starr-McCluer, and Brian J. Surette. 2000. Recent Changes
in U.S. Family Finances: Results from the 1998 Survey of Consumer Finances. Federal
Reserve Bulletin 86:1–29.
McCarthy, George W., and Roberto G. Quercia. 2000. Bridging the Gap between Supply
and Demand: The Evolution of the Homeownership Education and Counseling Indus-
try. Institute Report No. 00–01. Research Institute for Housing America.
Miller, James C., and James Pearce. 2001. Freddie Mac and Fannie Mae: Their Funding
Advantage and Benefits to Consumers. Unpublished manuscript. Freddie Mac.
Peterson, Paul. 2001. Keeping Promises. Mortgage Banking 61(6):70–77.
Quinn, Lawrence Richter. 2000. Credit Score Scrutiny. Mortgage Banking 60(12):50–55.
Rosenthal, Stuart S. 2001. Eliminating Credit Barriers to Increase Homeownership:
How Far Can We Go? Working Paper No. 01–01. Research Institute for Housing
America.
Steinbach, Gordon H. 1998. Making Risk-Based Pricing Work. Mortgage Banking
58(12):10–20.
Straka, John. 2000. A Shift in the Mortgage Landscape: The 1990s Move to Automated
Credit Evaluations. Journal of Housing Research 11(2):207–32.
Housing Policy Debate
zorn.qxd 09/17/02 9:15 AM Page 391
zorn.qxd 8/30/02 10:30 AM Page 392
... Accountability is necessary to ensure that articial intelligence (AI) is deployed responsibly, especially given the wide applicability of AI algorithms to several automated decision making contexts with 'high stakes' [40,57,58,81]. While automated decision systems (ADS) [106] have the potential to make more ecient and fairer decisions than their human counterparts [45,52], they could also produce harmful outcomes, worsening inequality in society [20,23,50,54,92,94,98]. Through accountability relationships, the actors responsible for harms caused by the ADS can be obligated to provide 'accounts' to the individuals who are harmed; the individuals or their representatives may then judge the accounts, and seek to impose consequences if necessary [121]. ...
Preprint
Full-text available
Accountability, a requisite for responsible AI, can be facilitated through transparency mechanisms such as audits and explainability. However, prior work suggests that the success of these mechanisms may be limited to Global North contexts; understanding the limitations of current interventions in varied socio-political conditions is crucial to help policymakers facilitate wider accountability. To do so, we examined the mediation of accountability in the existing interactions between vulnerable users and a 'high-risk' AI system in a Global South setting. We report on a qualitative study with 29 financially-stressed users of instant loan platforms in India. We found that users experienced intense feelings of indebtedness for the 'boon' of instant loans, and perceived huge obligations towards loan platforms. Users fulfilled obligations by accepting harsh terms and conditions, over-sharing sensitive data, and paying high fees to unknown and unverified lenders. Users demonstrated a dependence on loan platforms by persisting with such behaviors despite risks of harms such as abuse, recurring debts, discrimination, privacy harms, and self-harm to them. Instead of being enraged with loan platforms, users assumed responsibility for their negative experiences, thus releasing the high-powered loan platforms from accountability obligations. We argue that accountability is shaped by platform-user power relations, and urge caution to policymakers in adopting a purely technical approach to fostering algorithmic accountability. Instead, we call for situated interventions that enhance agency of users, enable meaningful transparency, reconfigure designer-user relations, and prompt a critical reflection in practitioners towards wider accountability. We conclude with implications for responsibly deploying AI in FinTech applications in India and beyond.
... However, there are also good examples of AI systems helping to tackle that bias, for example reducing racial inequality in the criminal justice system [15] or automated financial underrating systems can be helpful for applicants with an undervalued credit history [8]. ...
Article
The paper discovers an essence and importance of introduction of ethical dimension in all phases of artificial intelligence (AI): development of concept and source code, implementation in real-life applications and support and improvement of existing solutions. Modern society largely depends on cybertechnologies most of which are using elements of AI and ethical aspects of it is of paramount importance.
... E.g.,Gates et al. (2002) showed how the adoption of automated underwriting in mortgage lending contributed to the increase of approval rates for minority and low-income applicants by 30% while improving the overall accuracy of default predictions. ...
Preprint
Algorithms learn rules and associations based on the training data that they are exposed to. Yet, the very same data that teaches machines to understand and predict the world, contains societal and historic biases, resulting in biased algorithms with the risk of further amplifying these once put into use for decision support. Synthetic data, on the other hand, emerges with the promise to provide an unlimited amount of representative, realistic training samples, that can be shared further without disclosing the privacy of individual subjects. We present a framework to incorporate fairness constraints into the self-supervised learning process, that allows to then simulate an unlimited amount of representative as well as fair synthetic data. This framework provides a handle to govern and control for privacy as well as for bias within AI at its very source: the training data. We demonstrate the proposed approach by amending an existing generative model architecture and generating a representative as well as fair version of the UCI Adult census data set. While the relationships between attributes are faithfully retained, the gender and racial biases inherent in the original data are controlled for. This is further validated by comparing propensity scores of downstream predictive models that are trained on the original data versus the fair synthetic data. We consider representative & fair synthetic data a promising future building block to teach algorithms not on historic worlds, but rather on the worlds that we strive to live in.
... The results of Study 6 suggest that algorithmic decision-making may make people less likely to take actions against decisions that might, in fact, be discriminatory. We believe that algorithms have tremendous potential to unveil and eventually rid decision-making from human biases (Kleinberg et al., 2020), leading to more accurate and equitable results (Gates et al., 2002). Yet, to the extent that algorithmic bias exists and is hard to detect and eradicate (Hao, 2019), the use of algorithms might have societal implications that should not be ignored. ...
Article
Full-text available
Algorithms have been the subject of a heated debate regarding their potential to yield biased decisions. Prior research has focused on documenting algorithmic bias and discussing its origins from a technical standpoint. We look at algorithmic bias from a psychological perspective, raising a fundamental question that has received little attention: are people more or less likely to perceive decisions that yield disparities as biased, when such decisions stem from algorithms as opposed to humans? We find that algorithmic decisions that yield gender or racial disparities are less likely to be perceived as biased than human decisions. This occurs because people believe that algorithms, unlike humans, decontextualize decision-making by neglecting individual characteristics and blindly applying rules and procedures irrespective of whom they are judging. In situations that entail the potential for discrimination, this belief leads people to think that algorithms are more likely than humans to treat everyone equally, thus less likely to yield biased decisions. This asymmetrical perception of bias, which occurs both in the general population and among members of stigmatized groups, leads people to endorse stereotypical beliefs that fuel discrimination and reduces their willingness to act against potentially discriminatory outcomes. (PsycInfo Database Record (c) 2021 APA, all rights reserved).
... AI with the combined efforts of big data and ML (BD/ML) enables speedy, flexible, tailored, cost effective by charging reduced fee by a fraction compared to human advisors, convenient, creating wider range of services, making all services available to a broader range of users, as well as offering better services, heightening efficiency, enhancing quality, creating happy customers with raised customer satisfaction levels [8,9,21,60,61]. Robo-advisory systems simplifies customers' user experience [12,37], translate user interface into the language of the customer's choice by welcome customers in customer's preferred language while walking through products and services in the preferred language of choice [12]. ...
Chapter
Full-text available
Banks soundness plays a crucial role in determining economic prosperity. As such, banks are under intense scrutiny to make wise decisions that enhances bank stability. Artificial Intelligence (AI) plays a significant role in changing the way banks operate and service their customers. Banks are becoming more modern and relevant in people’s life as a result. The most significant contribution of AI is it provides a lifeline for bank’s survival. The chapter provides a taxonomy of bank soundness in the face of AI through the lens of CAMELS where C (Capital), A(Asset), M(Management), E(Earnings), L(Liquidity), S(Sensitivity). The taxonomy partitions opportunities from the main strand of CAMELS into distinct categories of 1 (C), 6(A), 17(M), 16 (E), 3(L), 6(S). It is highly evident that banks will soon extinct if they do not embed AI into their operations. As such, AI is a done deal for banks. Yet will AI contribute to bank soundness remains to be seen.
Article
Today, humans interact with automation frequently and in a variety of settings ranging from private to professional. Their behavior in these interactions has attracted considerable research interest across several fields, with sometimes little exchange among them and seemingly inconsistent findings. In this article, we review 138 experimental studies on how people interact with automated agents, that can assume different roles. We synthesize the evidence, suggest ways to reconcile inconsistencies between studies and disciplines, and discuss organizational and societal implications. The reviewed studies show that people react to automated agents differently than they do to humans: In general, they behave more rationally, and seem less prone to emotional and social responses, though this may be mediated by the agents’ design. Task context, performance expectations and the distribution of decision authority between humans and automated agents are all factors that systematically impact the willingness to accept automated agents in decision-making - that is, humans seem willing to (over-)rely on algorithmic support, yet averse to fully ceding their decision authority. The impact of these behavioral regularities for the deliberation of the benefits and risks of automation in organizations and society is discussed.
Article
Racial discrimination in housing has long fueled disparities in homeownership and wealth in the United States. Now, automated algorithms play a dominant role in rental and lending decisions. Advocates of these technologies argue that mortgage lending algorithms reduce discrimination. However, “errors in background check reports persist and remain pervasive,” and algorithms are at risk for inheriting prejudices from society and reflect pre-existing patterns of inequality. Additionally, algorithmic discrimination is often challenging to identify and difficult to explain or prosecute in court. While the Federal Trade Commission (FTC) is responsible for prosecuting this type of discrimination under the Fair Credit Reporting Act (FCRA), their enforcement regime “has inadequately regulated industry at the federal and state level and failed to provide consumers access to justice at an individual level,” as evidenced by its mere eighty-seven enforcement actions in the past forty years. In comparison, 4,531 lawsuits have been brought under the FCRA by other groups in 2018 alone. Therefore, the FTC must update its policies to ensure it can identify, prosecute, and facilitate third-party lawsuits against a primary driver of housing discrimination in the 21st century: discrimination within algorithmic decision making. We recommend that the FTC issue a rule requiring companies to publish a data plan with all consumer reporting products. Currently, the FTC recommends that companies make an internal assessment of the components of the proposed data plan to ensure that they are not in violation of the FCRA. Therefore, requiring that these plans be published publicly does not place undue burden on companies and empowers consumers to advocate for themselves and report unfair practices to the FTC. Coupled together, these will reduce the costs of investigation and enforcement by the FTC and decrease the discriminatory impact of automated decision systems on marginalized communities.
Article
Survey respondents disagree strongly about the dispersion of future returns and, increasingly, macroeconomic uncertainty. Such disagreement about risk may raise asset prices when collateralized debt products allow investors to realize perceived gains from trade. Investors who expect low volatility in collateral cash-flow appreciate senior debt as riskless. Those who expect high volatility, in contrast, value the upside potential in junior debt or equity claims. We show how such self-selection may have had a sizeable effect on the prices of RMBS and CDOs before the crisis, as investors disagreed about the volatility of aggregate economic conditions and their importance for default rates in collateral pools.
Article
Full-text available
Homeownership among U.S. families increased notably in recent years, from 63 percent in 1989 to 66.2 percent in 1998. This article examines this trend and the factors contributing to it. We find that (1) homeownership rose for all racial, ethnic, and income groups; (2) the differences in homeownership between minority and nonminority families and between middle-income and lower-income families declined significantly; and (3) changes in family-related characteristics explain homeownership trends among only the top two income quintiles. Among the lower two income quintiles, family-related characteristics explain almost none of the increase in homeownership. This pattern suggests that the favorable economic climate of the last decade, changes in mortgage and housing markets, and changes in the regulations governing those markets account for the increase in homeownership among lower-income families. Copyright 2001 by Kluwer Academic Publishers
Article
Full-text available
Using data from the Federal Reserve Board's two most recent Surveys of Consumer Finances, this article provides a detailed picture of changes in the financial condition of U.S. families between 1995 and 1998. The financial situation of families changed notably in the three-year period. While income continued a moderate upward trend, net worth grew strongly, and the increase in net worth was broadly shared by different demographic groups. A booming stock market accounts for a substantial part of the rise in net worth, but the data also suggest that improvements in financial circumstances extended to many families that did not own stocks. The indebtedness of families grew, but less rapidly than their assets. Nonetheless, compared with 1995, debt repayments in 1998 accounted for a larger share of the income of the typical family with debt, and the proportion of debtors who were late with their payments by sixty days or more in the year preceding the survey was also higher.
Article
This article reviews the shift in the 1990s from traditional manual mortgage underwriting to mortgage scoring models and statistical automated underwriting (AU). AU has become the predominant mortgage underwriting method, helping to catalyze a broad wave of technological advancement. Still, further significant changes are likely. Examining statistical AU's history, growth, and underlying factors and concerns, this article especially considers the question of how far AU should be relied upon to evaluate more marginal and higher-risk mortgage applicants (traditionally harder cases). Supporting a cautious, continued expansion of statis- tical AU into higher-risk loans, the article reveals some provocative evidence on statistical scorecards versus traditional underwriters and explores fair lending strengths of AU. Future overall prospects for likely AU and related research areas and developments are summarized.
Article
This paper models the structure of a financial market that is composed of two types of institutions, banks and securities or secondary markets. The model analyzes the development of the securities market as a way of trading off its lower cost of securitization with adverse selection due to asymmetric information possessed by banks, using a simple adverse selection model in the tradition of Akerlof (1970). Successive modifications, including licensing contracts similar to those of Leland (1979) and Chan and Leland (1982) solve some of the adverse selection problems by "licensing"loan sellers in an effort to provide minimum quality standards.
Article
Although credit scoring offers benefits to lenders and borrowers, its use raises important statistical issues that may affect the ability of scoring systems to accurately quantify an individual's credit risk. The evidence from a national sample of credit-bureau records suggests that concerns about omitted-variable bias may be justified, as local economic factors show significant correlations with credit scores. Copyright American Real Estate and Urban Economics Association.