Note: The following article addresses a CFPB Supervisory Highlights document released on January 17, 2025 under former Director Chopra. At the time of this writing, the CFPB's future is very uncertain with the new Acting Director halting most Bureau activities - including supervisory examinations. Nevertheless, due to the significant issues raised in this Supervisory Highlights document - and their continued relevance to consumer lenders regardless of the CFPB's future - I encourage fair lending compliance risk managers and credit risk professionals to consider the opinions expressed in this article.
One Last Thing Before We Go ....
Well, it took four years for the Biden Administration's CFPB to disclose how the technologists' credit model disparate impact framework has influenced its supervisory examinations of credit scoring models. Unfortunately, rather than issuing a formal set of fair lending examination procedures and standards, we get a last-minute drop of a "Special Edition" Supervisory Highlights document that is frustratingly short on substance and rationale - yet profoundly disruptive to long-standing regulatory precedent and industry fair lending compliance programs.
My reaction after an initial read-through:
So they expect us to use race/ethnicity data to build credit scoring models now??
Isn't that illegal?
Don't they know that?
Hmmm....
Now, before I dive into the major problems of this final coda to the CFPB's Chopra chapter, I want to acknowledge some reasonably helpful information it provides.
First, while not new, there are some basic fair lending compliance risk management expectations for credit scoring models.[1] Specifically, supervised lenders should have:
Formal fair lending controls that effectively detect and remediate disparate impact discrimination caused by certain credit model attributes that are "not directly related to consumers’ finances and how consumers manage their financial commitments" (i.e., "alternative data"). This reflects what I call the traditional credit model disparate impact framework - i.e., a long-standing compliance process designed to identify model attributes that: (1) lack a clear and causal relationship to an applicant's creditworthiness, (2) represent "artificial, arbitrary, and unnecessary" barriers to credit access, and (3) disproportionately disadvantage prohibited basis groups. As the CFPB notes, if such attributes are identified, the lender's controls should include a requirement to adequately document the associated business needs for such attributes and document the lender's consideration of "comparably accurate inputs with less discriminatory effects."
Importantly, this fair lending control operates at the individual model attribute level. That is, each model attribute is reviewed both qualitatively (i.e., Does it represent an "artificial, arbitrary, and unnecessary" barrier to credit, and does it have a questionable causal relationship to an applicant's creditworthiness?) and quantitatively (i.e., Does it disproportionately disfavor one or more prohibited basis groups?).
Formal fair lending controls that effectively detect and remediate disparate treatment discrimination caused by certain credit model attributes that - individually or in combinations - serve as potential demographic proxies. Like the previous section, this expectation reflects a long-standing fair lending compliance testing activity in which lenders evaluate the credit model's individual predictive attributes for excessive correlations with an applicant's demographic profile (e.g., applicant zip codes or job categories). Such controls should include policies on:
What specific criteria identify a potential "demographic proxy"?
How are the contributions of such potential proxies measured relative to the lender's documented legitimate business needs?
What threshold contribution level is necessary for a potential proxy to be "business justified"?
How does one search for less discriminatory alternative attributes even when such thresholds are attained?
Both of these fair lending control expectations align with long-standing regulatory requirements for credit scoring models established after the 1974 Equal Credit Opportunity Act ("ECOA"). They are described formally in federal bank regulatory guidance documents such as the OCC's 1997-24 Bulletin on Credit Scoring Model Examination Guidance, the FFIEC's 2009 Interagency Fair Lending Examination Procedures, and the OCC's 2023 Fair Lending Handbook (and its prior 2010 version).
Next, there is some helpful new information related to more complex, modern-day credit scoring models - specifically:
For credit scoring models containing hundreds, if not thousands, of predictive attributes, the CFPB: (1) notes lenders' "difficulties in being able to effectively monitor" the demographic proxy risks of all attributes, and (2) "found that institutions did not have a process for ensuring adequate review of input variables for fair lending risks before those variables were selected as model inputs".[2] Industry proponents of such models who argue that variable-level disparate impact analyses are infeasible in the age of AI/ML should pay particular heed here.
Finally, the CFPB reiterates its concerns about advanced technologies used to explain individual credit decisions produced by complex credit scoring models. Specifically, supervised lenders are expected to have clear policies and procedures to validate the specificity and accuracy of "explainability methodologies" used to identify Adverse Action Notification ("AAN") reasons associated with an applicant's insufficient credit score. This requirement is consistent with the CFPB's prior 2022-03 Circular on the topic - about which I have written in detail.[3]
Now, let's turn to the new content that I find particularly problematic.
The CFPB Prescriptively Directs Certain Lenders To Use Advanced Technologies Relying On Borrower Race Data To Train "Debiased" Credit Scoring Models
According to the Supervisory Highlights document, CFPB examiners - in partnership with the Bureau's recently-hired technologists - have engaged in the following supervisory activities:
"The exam teams conducted statistical analyses of the institutions’ underwriting and pricing practices and found disproportionately negative outcomes for Black or African American and Hispanic applicants when compared to white applicants. Certain credit scoring models contributed to disparities in multiple card products, ... and the exam team’s analysis suggested that the way the institutions developed or implemented their credit scoring models contributed to some of those disparities." (emphasis mine)
"In response to these findings, the exam teams identified potential alternative models to the institutions’ credit scoring models using open-source automated debiasing methodologies. These alternative models used the same credit strategy, machine learning algorithm, and external configuration variables as the institutions’ models. For most of the credit scoring models reviewed, the exam team identified potential alternative models that appeared able to meaningfully reduce disparities while maintaining comparable predictive performance as the institutions’ original models." [4] (emphasis mine)
"As this work suggested there may be appropriate less discriminatory alternative models that would meet the institutions’ legitimate business needs, the exam teams directed the institutions to search for less discriminatory alternatives to the credit scoring models that produced prohibited-basis disparities. Examiners directed that searches include using methods capable of meaningfully assessing and adjusting their models to identify and evaluate potential less discriminatory alternatives to meet the documented legitimate business needs associated with use of the specific model." (emphasis mine)
Before describing the specific issues with this model-level disparate impact testing and remediation directive, I first note that model-level testing has been a key part of most lenders' fair lending compliance programs since the CFPB adopted demographic proxies in 2014. However, for most lenders, such testing evaluates whether the credit model has equal predictive accuracy across demographic groups - a disparate impact metric carefully designed: (1) to focus exclusively on the credit scoring model's specific purpose (i.e., credit risk measurement), (2) to exclude separate downstream credit policies - such as credit score thresholds - impacting actual credit decisions (covered by separate fair lending testing), and (3) to consider both the model's predicted default rates AND each group's actual default rates to ensure that actual creditworthiness considerations are included in the disparate impact test. If potential disparate impact is found, remediation generally involves refining the credit model's specification (i.e., the attributes included in the model) to reduce the relative predictive error disadvantaging prohibited basis applicants.
However, in adopting the technologists' credit model disparate impact framework as described in the quoted sections above, the CFPB inexplicably discards the industry's traditional model-level testing and remediation framework in favor of a complex, black-box algorithmic approach. In this method, borrower race/ethnicity data is used to create latent demographic proxies that adjust estimated credit scores to drive lending outcome equity across these demographic groups. Ironically, by adopting this advanced technology-based approach, the CFPB has created its own technological exception to the long-standing disparate impact foundations that underpin industry fair lending compliance risk management programs.
In what follows, I lay out my specific issues with this approach.
What's Wrong With the CFPB's Directive To Use Automated Debiasing Methodologies?
The CFPB embraces an equity-based measure of disparate impact that is, in my opinion, inconsistent with ECOA and the Supreme Court's ("SCOTUS's") disparate impact interpretive opinions.
In this Supervisory Highlights document, the CFPB notes that its credit scoring model examinations identified "disparities" in underwriting and pricing outcomes at certain institutions. For example,
"The exam teams found disparities in underwriting and pricing outcomes ... ",
"The exam teams conducted statistical analyses of the institutions’ underwriting and pricing practices and found disproportionately negative outcomes for Black or African American and Hispanic applicants when compared to white applicants.", and
"[E]xaminers directed the institutions to test such models for prohibited basis disparities ...".
While not explicitly stated, my discussions with industry participants indicate that these "disparities" appear to be simple comparisons of "raw" (i.e., unconditional) average credit outcomes across demographic groups (e.g., Adverse Impact Ratios ("AIRs")). Accordingly, the CFPB appears to be formally embracing an equity-based definition of disparate impact in which unequal lending outcomes - regardless of applicant creditworthiness - are the sole prima facie evidence.
In my opinion, this is clearly inconsistent with ECOA, which explicitly states:
"The purpose of this regulation is to promote the availability of credit to all creditworthy applicants without regard to race, color, religion, national origin, sex, marital status, or age ..." 12 CFR Sec. 202(1)(b) (emphasis mine)
And, as I discuss in Algorithmic Justice: What's Wrong With the Technologists' Credit Model Disparate Impact Framework, SCOTUS has consistently stated in its disparate impact interpretive opinions that statistical disparities alone do not represent prima facie evidence of disparate impact discrimination. For example,
"A disparate-impact claim relying on a statistical disparity must fail if the plaintiff cannot point to a defendant's policy or policies causing that disparity. A robust causality requirement ensures that “[r]acial imbalance . . . does not, without more, establish a prima facie case of disparate impact” and thus protects defendants from being held liable for racial disparities they did not create. Wards Cove Packing Co. v. Antonio, 490 U. S. 642, 653 (1989), superseded by statute on other grounds, 42 U. S. C. § 2000e–2(k)." (emphasis mine)
Accordingly, to be consistent with ECOA and SCOTUS's interpretive opinions, traditional disparate impact analyses have focused on specific credit model attributes that may be causing a disparate impact on otherwise creditworthy applicants. Such analyses necessarily focus on the marginal effects these attributes have on lending outcomes across otherwise creditworthy groups.
Curiously, while the CFPB appears to ignore these precedents in adopting the technologists' equity-based definition of disparate impact, it is careful to avoid explicitly stating this in the Supervisory Highlights document - opting instead for general references to "disparities". This is puzzling, for if the CFPB now intends to hold lenders accountable for a new, progressive interpretation of disparate impact (as it appears to be doing in examinations and possibly in enforcement actions), why not communicate this clearly to allow supervised lenders to understand fully these expectations?
The open-source automated debiasing methodologies embraced by the CFPB explicitly use borrower race data during model training to produce more equitable lending outcomes.
As outlined in my earlier research on these methodologies, automated debiasing tools such as fairness regularization and adversarial debiasing reduce prohibited basis lending disparities by altering how the credit model assigns weights to each predictive attribute when computing a credit score (i.e., increasing some attribute weights while decreasing others). This attribute re-weighting is specifically designed to leverage correlations between applicant credit profiles and their associated demographics - effectively creating "reverse proxy effects" that improve average credit scores for prohibited basis applicants - typically at the expense of lower average credit scores for control group applicants.
Based on my research, there appear to be many troubling issues with these algorithmic debiasing methodologies that, unfortunately, have gone unacknowledged and unaddressed by most proponents - such as:
As automated algorithms, they are unaware of whether any attributes whose weights they change constitute "artificial, arbitrary, and unnecessary barriers" to credit per SCOTUS's disparate impact court opinions.[5] Instead, they ruthlessly alter the weights on whatever attributes are necessary to fulfill their dual objectives of maximizing fairness and predictive accuracy - including attributes that reflect fundamental causal underwriting criteria like debt-to-income ratios or recent bankruptcies.[6]
By altering the credit model's estimated predictive relationships in the pursuit of equitable lending outcomes, the algorithms may undermine the model's statistical validity, conceptual soundness, and ability to satisfy (if required) Regulation B's "empirically-derived, demonstrably and statistically sound" criterion. (See OCC 2011-12 Supervisory Guidance on Model Risk Management and 12 CFR § 1002.2(p))
The algorithms create a trade-off between inter-group and intra-group fairness, disadvantaging some prohibited basis applicants to achieve more equitable overall approval rates between groups. This occurs because the LDA credit model adjusts applicants' credit scores based on the statistical relationships of their credit attributes with their demographic profiles. Since these statistical relationships are imperfect, some prohibited basis applicants will have their credit scores adjusted downward even though the average credit score of their demographic group is improved. Accordingly, while a particular demographic group, as a whole, may experience an improved approval rate, some members of that group - who would have been approved under the original credit model - may now be denied.[7]
Employing dual training objectives (i.e., fairness and accuracy) impacts the stability and robustness of the LDA credit model solutions produced by the algorithmic debiasing methodologies. According to my research,[8] very small changes in model training samples can lead to significantly different sets of LDA credit model weights - meaning that the credit attributes primarily responsible for the original model's "disparate impact" are inherently unstable - a feature that is conceptually inconsistent with traditional disparate impact theory.
The addition of a second model training objective, namely fairness, complicates the lender's ability to generate specific and accurate Adverse Action Notifications ("AAN") compliant with Regulation B. Specifically, to boost the lender's equity-based fairness performance, the automated debiasing algorithm will lower the credit scores of many applicants to levels that change their credit decision (vs. the original model) from approval to denial. In such cases, what AAN reason should lenders provide to these applicants, given that the denial is really tied to fulfilling the fairness objective rather than the accuracy objective?[9]
Notwithstanding the above issues, I believe the most grievous issue associated with these automated debiasing methodologies is that lenders must include borrower demographic data (i.e., race/ethnicity, gender, and/or age) into the LDA model development process[10] to enable the debiasing algorithm to:
Learn the statistical relationships between applicant credit attributes and their associated demographic profiles.
Learn how to use these relationships to target its alteration of specific credit attribute weights to achieve greater credit score distributional alignment across the demographic groups.
This is a grievous issue for the simple reason that the use of demographic data in credit evaluation systems has always been considered a violation of ECOA - which explicitly states:[11]
Except as provided in the Act and this regulation, a creditor shall not take a prohibited basis into account in any system of evaluating the creditworthiness of applicants. - 12 CFR Sec. 202(b)(1) (emphasis mine)
Except as otherwise permitted or required by law, a creditor shall not consider race, color, religion, national origin, or sex (or an applicant’s or other person’s decision not to provide the information) in any aspect of a credit transaction. - 12 CFR Sec. 202(b)(9) (emphasis mine)
In fact, this prohibition against the use of borrower demographic information - either directly or via proxies - forms the basis of much of the federal bank regulatory agencies' fair lending examination procedures associated with credit decisions and credit scoring models.[12] As just one of many examples, the FFIEC's Interagency Fair Lending Examination Procedures states:
"Including variables in a credit scoring system that constitute a basis or factor prohibited by Regulation B is an overt indicator of discrimination."
While it is true that these LDA credit models do not explicitly consider applicant race (or other prohibited bases) when computing credit scores at run time, it is also true that such explicit data is unnecessary given how algorithmic debiasing uses applicant race data during LDA credit model training. In particular, as discussed above, algorithmic debiasing creates a "reverse demographic proxy effect" during LDA credit model training to improve lending equity. Specifically, by learning the correlations between the applicants' demographic characteristics and their associated credit attributes during model training, LDA credit models mask their use of demographic data at run time.
This means that instead of needing an applicant's demographic information to compute a "debiased" credit score, the LDA credit model implicitly infers the applicant's demographics from their credit attributes. In traditional versions of these debiased credit models, these proxy effects are implicitly encoded into the altered credit model weights during LDA model training. However, in more recent evolutions of these methodologies, the reverse proxy effects are isolated into a separate "fairness correction" model - similar to Meta's Variance Reduction System - where credit score fairness adjustments are estimated based on the applicant's credit profile.[13] This allows some proponents to claim (falsely, in my opinion, but I am not a lawyer) that the resulting LDA credit models comply with ECOA since they do not explicitly use applicant demographic data at run time.
But if the use of reverse proxy effects is not a violation of ECOA, why are federal bank regulatory agencies and fair lending compliance professionals simultaneously seeking to identify and remediate other demographic proxies in credit scoring models that may harm prohibited basis groups?
In fact, in the same Supervisory Highlights document, the CFPB states:
"Examiners identified risks ... including difficulties in being able to effectively monitor whether any variables, individually or in combination, acted as a proxy for prohibited bases under ECOA." (emphasis mine)
"The institutions also were directed to review the input variables before using those variables in models, including assessing the variable’s relationship to the creditworthiness of applicants and whether the variable may operate as a proxy for prohibited bases under ECOA." (emphasis mine)
It is contradictory (and, frankly, troubling) for the CFPB to instruct lenders to identify and remediate demographic proxies affecting lending outcomes across, say, racial groups, while simultaneously directing the use of advanced technologies based on demographic proxies to intentionally affect lending outcomes across racial groups. According to ECOA, NO demographic proxies should be used whatsoever.
Lastly, another argument I often hear to counter the claim that algorithmic debiasing - by using demographic data - violates ECOA is that the lender can prevent such violations simply by employing two separate teams during model development. One team creates the credit model without access to demographic data, and the second team (perhaps sitting in Compliance or Legal) applies the algorithmic debiasing process using the demographic data. If the second team identifies an LDA, then it instructs the first team on how to change the original credit model to align with the approved LDA version. Voila! Since the first team didn't explicitly use demographic data to create the credit model, there is no violation!
In my non-lawyerly opinion, this is a distinction without a difference. Ultimately, the credit scores assigned to the company's applicants have been altered by the demographic data used to estimate the LDA credit model, and that's really what should matter.
The CFPB is directing lenders to use these automated debiasing methodologies without evident regard for their reliability or their potential to create unintended risks for lenders and their consumers.
Interestingly, although the CFPB has publicly expressed concerns multiple times about the accuracy and reliability of post-hoc explainability methodologies used to produce AANs, it does not express similar concerns here about the even more complex algorithms they want lenders to employ to debias credit scoring models.
For example, in the final section of the Supervisory Highlights document, the CFPB revisits the concerns first raised in its 2022-03 Circular: Adverse action notification requirements in connection with credit decisions based on complex algorithms. In particular:
"While some creditors may rely on post-hoc explanation methods to identify reasons for an adverse action, the creditors still must be able to validate the accuracy of those methods." (emphasis mine)
"In recent examinations ... [e]xaminers also found that the institutions had not validated that their processes for selecting reasons produced accurate results. To address these findings, examiners directed the institutions to, among other things, test and validate the methodologies used to identify principal reasons in adverse action notices." (emphasis mine)
Ironically, although they require validation testing to ensure the accuracy and reliability of the post-hoc explainability methodologies used to identify principal reasons in AANs, the CFPB does not require such validation testing for the open-source automated debiasing methodologies they instruct certain institutions to implement. Considering the significant safety-and-soundness risks involved in altering credit model weights using black-box fairness algorithms, along with the significant impacts on individual consumer credit outcomes, it is surprising that the CFPB: (1) has not shared its own validation testing of these algorithmic fairness tools to justify its confidence in directing their adoption by certain supervised lenders, and (2) has not acknowledged the inconsistency in its stance on validating post-hoc explainability tools but not algorithmic fairness tools.
What Are We To Make of This?
Although the Trump Administration's rollback of equity-focused policies and practices may alter the CFPB's (and the DOJ Civil Rights Division's[14]) stance on employing these "automated debiasing methodologies" and the technologists' credit model disparate impact framework on which they are based, there's also a chance that the subtle embedding of reverse demographic proxies into underwriting (and marketing) models goes unnoticed - at least temporarily.
Nevertheless, even if the CFPB's new leadership reverses course on these policies and practices (or if the CFPB is "reimagined"), many state-level lawmakers, regulators, enforcement agencies, and private plaintiffs may still engage in this area with views similar to those expressed in this Supervisory Highlights Special Edition. And, if a Democratic administration takes over in 2029, they may revive these views retroactively. Clearly, lenders will need to navigate these treacherous waters for some time.
In the interim, you can read my further thoughts on this thorny area - including an alternative to the technologists' credit model disparate impact framework, in Algorithmic Justice: What's Wrong With The Technologists' Credit Model Disparate Impact Framework and The Road to Fairer AI Credit Models: Are We Heading in the Right Direction?
* * *
ENDNOTES:
[1] I note that lenders should define "credit scoring models" broadly to include related risk / decision models, credit policy rules, and credit score threshold logic used to render final credit decisions, pricing, and loan amounts. For example, fraud models, pricing models, credit strategies and overlays, hard-policy declines, credit score cut-offs, and similar automated underwriting artifacts should typically be included in a lender's assessment of disparate impact risk.
[2] I have expressed similar concerns with the compliance and safety-and-soundness risks of "high-dimensional" credit scoring models. See Four Potential Pitfalls of AI Credit Scoring Models and Six Unanswered Fair Lending Questions Hindering AI Credit Model Adoption.
[3] For my analysis of the CFPB's 2022-03 Circular "Adverse action notification requirements in connection with credit decisions based on complex algorithms", see "Dark Skies Ahead: The CFPB's Brewing Algorithmic Storm". For my discussion of the specific challenges associated with these validation activities, see "Using Explainable AI To Produce ECOA Adverse Action Reasons: What Are The Risks?".
[4] Unfortunately, the CFPB fails to describe: (a) the specific "open-source" automated debiasing methodologies they employed, (b) what they mean by "same credit strategy" and same "configuration variables", and (c) how they specifically measured "predictive performance" (to determine that their LDA credit models maintained "comparable predictive performance" to the original models). These are all critical inputs for configuring - and assessing the results of - the automated debiasing methodologies they expect certain lenders to adopt. And the Bureau's silence on these key criteria does nothing to alleviate broader industry concerns about these methodologies.
[5] See SCOTUS's Griggs vs. Duke Power Co. disparate impact opinion ("Policies, whether governmental or private, are not contrary to the disparate-impact requirement unless they are “artificial, arbitrary, and unnecessary barriers.”)
[6] Some industry practitioners address this criticism by blocking the algorithm from adjusting the weights on a certain list of "untouchable" credit attributes. However, if this approach is taken, then - for logical consistency - it is also necessary to adjust the Adverse Impact Ratio for the effects of these benign attributes. But, this adjustment would lead to an LDA model whose lending outcomes are no longer equitable, as the portion of the overall disparity caused by these untouchable credit attributes would persist, thereby challenging the basis for the equitable disparate impact interpretation. Unfortunately, this is just one of several logical inconsistencies in the technologists' disparate impact framework.
[7] See the following section of Fool's Gold 4: The Brittleness of Algorithmic LDA Credit Models for further details.
[8] See Fool's Gold 4: The Brittleness of Algorithmic LDA Credit Models for further details.
[9] Of course, if a lender is truly removing a discriminatory "artificial, arbitrary, and unnecessary" attribute, I would expect that this attribute disadvantages prohibited basis applicants by overestimating their expected default rates and, correspondingly, advantages control group applicants by underestimating their expected default rates. In such cases, this attribute truly reduces credit access for otherwise creditworthy prohibited basis applicants and improperly expands credit access to otherwise non-creditworthy control group applicants. In such contexts, this AAN issue is moot since - after the attribute removal - the new control group denials are not denied because of fairness considerations; they are denied because their excessive credit risk is now accurately measured.
This is not necessarily the case, however, with automated debiasing methodologies as the algorithm is indifferent to whether expected default rates are overestimated or underestimated at the demographic group level when adjusting the model's weights. In fact, the original credit model may significantly underestimate the prohibited basis group's expected default rates, yet the algorithm could still identify LDA model versions that further boost prohibited basis approvals (often at the expense of further overestimating the control group's expected default rates). Accordingly, LDA models created by these automated debiasing methodologies might deny creditworthy applicants whose expected default rates are deliberately overestimated solely to improve the lender's equity-based fairness performance - with these credit decisions being unrelated to the applicants' actual credit risk.
[10] Ironically, laws prevent lenders from collecting data on the race/ethnicity and gender of applicants for all non-mortgage credit products to prevent these factors from influencing credit decisions. And, to make matters even worse, since lenders have no access to this data by law, the CFPB expects them to use statistical proxies of these demographic attributes for LDA credit model development - despite their known (and growing) inaccuracies. See, for example, BISG Proxy Bias Redux: The Impact of the 2020 U.S. Census Data.
[11] Ironically, the CFPB explicitly refers to ECOA's non-discrimination purpose before it then proceeds to describe how it used borrower race data to find LDA credit models for an institution's credit card credit scoring model.
[12] One common exception here is the use of an applicant's age in an empirically-derived, demonstrably and statistically sound credit scoring system. See 12 CFR Sec. 202(6)(b)(2).
[13] Some industry participants have adopted a "two-stage" approach to these LDA credit models - similar to the structure of Meta's Variance Reduction System - that does not alter the weights of the original credit model. For example, on February 8, 2024 Zest AI filed for a patent titled "Machine Learning Model Correction," Publication No. US 20240046349A1 - detailing a methodology that trains a second "fairness correction" model using adversarial debiasing techniques and borrower race data. Similarly, Upstart announced recently that it is also using a secondary fairness adjustment model based on adversarial debiasing as part of its "Universal LDA" framework. In both cases, the predictions of these stand-alone fairness correction models are then combined with those of the original unaltered credit model to produce debiased credit scores that improve lending outcome equity. My point here is that whether the borrower race proxy effects are integrated into the original credit model's weights or through a separate standalone "fairness correction" model, borrower race information is still being used to generate credit scores.
[14] See the DOJ's Meta Settlement in which Meta deployed a "Variance Reduction System" to address bias in specific machine-learning models that were purportedly causing discrepancies between the demographics of users who actually saw certain HEC-related ads and the demographics of the intended target audience.
© Pace Analytics Consulting LLC, 2025.