Fair Lending and Machine Learning
(Part 1)
This image from 1939 shows us the practice of “Redlining” in Detroit. The term stems from the policies developed by the Home Owners Loan Corporation (HOLC), in which neighborhoods would be ranked according to their perceived investment risk. The “worst” neighborhoods, marked in red, had something in common: they were the home of racial and ethnical minorities.
Until 1968, the racist pretext to write off entire neighborhoods as “hazardous” was not only the norm but completely legal.
The consequences are still felt today, as the denial of access to homeownership propagated the patterns of racial and economic subordination that we see today.
In order to avoid both the intentional and unintentional denial of credit to segments of the population, the concept of Fair Lending was born.
What is Fair Lending?
Fair Lending concerns itself with investigating the decisions, and the systems that guide the access to credit, in the effort to avoid unjustified differentiation.
What is an unjustified basis for differentiation?
These are attributes we should not consider when deciding someone’s creditworthiness. We designate these “protected attributes”. They may have:
Practical irrelevance
Attributes that do not have material basis on the outcome.
Example: Race or gender alone do not affect someone’s creditworthiness, and as such, they should not be considered.
and/or
Moral irrelevance
Attributes that we morally object to using, regardless of influence on the outcome.
Example: Disability status could affect a person’s ability to pay back a loan, but we morally object to using it, because we do not want people with disabilities to be subjected to a subordinate status in our society.
They always have:
Social salience
Attributes that are socially salient, and that have served as basis for systematic, unjustified differentiation in the past.
Example: Race has historically been used as basis to disfavour individuals and communities.
The two doctrines of discrimination law
These guiding principles of US discrimination law inform the ethics that guide fair lending. They are called disparate treatment and disparate impact, and they aim to address different causes of unfair discrimination.
Disparate Treatment
This principle describes the more “obvious” instances of discrimination, such as considering race directly to make a decision.
Whether this has the goal to favor or disfavor the protected group is irrelevant; Considering a protected attribute (or a proxy) in a decision is always a violation of this principle.
It describes instances of discrimination that are:
Formal
When a protected attribute is considered directly for a decision.
Example: An underwriter that makes decisions based on gender of the loan applicant.
and / or:
Intentional
Using a proxy for a protected attribute, as a purposeful attempt to discriminate.
Example: In order to deny minorities access to credit, neighbourhoods populated by minorities are considered “Hazardous”. There is an attempt to conceal the racist pretext.
Disparate Impact
This principle describes the less “obvious”, but no less significant instances of discrimination.
For example, a mortgage lender that has the policy to deny loan applications for less than $60,000. At first glance, this would seem like a neutral policy: but closer examination would reveal that this policy disproportionately excludes minority applicants from the houses they can typically afford.
It describes instances of discrimination that are at the same time:
Facially neutral
They seem non-discriminatory at first glance, but disproportionately affect members of a protected class.
Example: A mortgage lender has a policy to deny loan applications for less than $60,000. However, minority applicants are disproportionately excluded of the houses they can typically afford.
and
Unintentional
The negative impact is unintentional (otherwise it would be disparate treatment).
Example: The same mortgage lender was not aware of that policy’s disproportionate exclusion of minority applicants.
and also
Avoidable
A different rule or policy that is less disparate in its impact exists, and there is no business necessity that would impede using it.
Example: The same mortgage lender could have removed that rule to deny loan applications under $60,000, as it would not negatively impact their ability to do business. Since disparate impact existed, was unintentional, and was avoidable, the mortgage lender is liable for disparate impact violation.
Where can these Fair Lending violations happen today?
Manual Underwriting
The traditional form of underwriting, performed by human experts.
Humans have a notorious tendency to be prejudiced. Often, the biases of the human underwriter will lead them to see a borrower as less creditworthy because of their race or gender, undermining their chance to be judged on their merits, and denying them opportunities in life.
Manual underwriting has risks in terms of Disparate Treatment (due to conscious and unconscious bias) and of Disparate Impact (due to unjustified policies that exclude protected classes).
Automated Underwriting
Increasingly, the process of underwriting is automated by machine learning algorithms.
Leveraging the attributes and the default outcomes of past borrowers, these algorithms learn the patterns of what makes a good borrower better than any human could ever hope to.
However, these models are trained on past data: that data can be tainted by the prejudiced decisions of former human underwriters.
Automated Underwriting systems have risks in terms of Disparate Impact, as biased data can lead to the exclusion of protected classes.
What does machine learning mean for Fair Lending?
As underwriting increasingly becomes the domain of machine learning, complex machine learning algorithms judge the financial future of millions of people.
We should recognize its potential to create a fairer lending process: done correctly, we can automate underwriting in a formalized, evidence-based way, and avoid human prejudice and bias, as studies have shown. (Gates, Perry, Zorn (2002))
We must also ensure that these models, using data tainted by the prejudices of the former human decision makers, do not simply perpetuate the same patterns of exclusion and disenfranchisement. We must avoid “digital redlining”.
Fair Lending should guide a new era of machine learning: The era where the goal of these systems is not to simply to maximize predictive power, but also to create systems that support human values.
