Housing Finance and Algorithmic Discrimination: What Different Sides Are Protecting

In 2021, The Markup and Associated Press published an analysis of 2.4 million conventional mortgage applications filed across the United States. Controlling for factors like loan amount, neighborhood, and applicant income, they found that major U.S. lenders denied applications from Black borrowers at roughly twice the rate of white borrowers with comparable financial profiles. The finding landed in the middle of a debate that had been running since at least the 1970s — about whether American mortgage lending discriminates, whether the discrimination is intentional or structural, and whether the algorithmic underwriting systems that now process the vast majority of mortgage applications make things better or worse.

The case for algorithms is not trivial: a system that applies the same criteria to every application can't give a worse deal to a Black borrower because a loan officer happened to be having a bad day, or made assumptions about a neighborhood based on who lives there. Robert Bartlett, Adair Morse, Richard Stanton, and Nancy Wallace, in a widely cited 2019 NBER working paper, found that algorithmic fintech lenders discriminate against Black and Latino borrowers less than face-to-face lenders — but still charge them 5.3 basis points more for purchase mortgages and 2.0 basis points more for refinances, even after controlling for credit risk. The algorithm isn't neutral; it's just differently biased. The question everyone is actually fighting about is: why, and what to do about it.

That question sits at the intersection of several genuinely difficult problems. How do you measure discrimination in systems that are formally race-blind? What counts as a "legitimate" credit factor when the factors we use were shaped by decades of racially segregated lending and wealth accumulation? What does fair lending law actually require — disparate treatment (intent) or disparate impact (effect)? And is the right frame for this debate discrimination in the current system, or the encoded legacy of historical discrimination, or both at once? Different answers generate different prescriptions that cannot easily be reconciled because they flow from different diagnoses.

What fair lending enforcement advocates are protecting

The argument that the existing legal framework — the Fair Housing Act, the Equal Credit Opportunity Act, the Community Reinvestment Act — establishes the tools needed to address lending discrimination, and that what's required is rigorous enforcement rather than system redesign. The Home Mortgage Disclosure Act (1975) requires lenders to report data on applications, approvals, denials, and loan pricing by race, gender, and geography — one of the most granular public datasets on discrimination in any industry. The 1977 Community Reinvestment Act requires federally regulated banks to demonstrate that they are meeting the credit needs of the communities in which they operate, including low- and moderate-income neighborhoods. The CFPB, created by the Dodd-Frank Act (2010), has enforcement authority over the largest lenders. The Supreme Court's 2015 ruling in Texas Department of Housing and Community Affairs v. Inclusive Communities Project confirmed that the Fair Housing Act prohibits practices with discriminatory effects — not only those with discriminatory intent — closing the loophole that would allow a lender to claim compliance simply by not explicitly considering race. Fair lending enforcement advocates are protecting the recognition that this architecture exists, that it has produced documented results when enforced, and that dismantling or circumventing it — even in the name of innovation — removes the accountability structures that make discrimination visible.

The HMDA data argument: the disparity data is real, it is public, and it obligates a regulatory response — the question is whether regulators are actually using it. The Boston Federal Reserve's 1996 study by Alicia Munnell and colleagues, which analyzed over 3,000 mortgage applications in the Boston area controlling for forty-five creditworthiness factors, found that Black and Hispanic applicants were 60 percent more likely to be denied than white applicants with equivalent financial profiles. The study was contested vigorously — critics argued the controls were imperfect, that the sample was small — but subsequent research using larger datasets has consistently replicated the finding. The Markup's 2021 analysis, using HMDA data across 17 million applications, found disparities in denial rates that persisted even after controlling for every factor available in the dataset. Fair lending advocates argue that the persistence of this finding across decades, datasets, and methodologies is itself significant: the debate about whether HMDA data can definitively prove discrimination has become a way to avoid responding to what it plainly shows.

The Community Reinvestment Act argument: a law that requires banks to document investment in the communities they take deposits from is the most direct available lever for addressing geographic credit deserts, and it has never been fully implemented. The CRA was passed in response to the practice of "redlining" — the systematic refusal to lend in Black and minority neighborhoods, named for the red lines that federal agencies and lenders drew on maps to designate neighborhoods as investment risks. The law requires banks to demonstrate affirmative service to low- and moderate-income areas; CRA ratings affect approval for mergers and acquisitions. Critics argue the CRA is a blunt tool with significant compliance theater — that banks can satisfy CRA requirements through community development loans in wealthy areas while maintaining credit deserts in others. Fair lending advocates agree the CRA needs modernization — the 2023 OCC/FDIC/Fed joint rulemaking represented the first major update since 1995 — but they argue that expanding and enforcing it, rather than trading it for market mechanisms, is the right direction.

What algorithm transparency and audit advocates are protecting

The argument that the algorithmic systems now processing the majority of mortgage applications — including Fannie Mae's Desktop Underwriter and Freddie Mac's Loan Product Advisor — produce racially disparate outputs that cannot be detected, challenged, or corrected by existing enforcement frameworks because the systems are opaque. Automated underwriting systems (AUS) make or heavily shape lending decisions for the vast majority of conforming mortgages in the United States. Lenders are not required to disclose what factors these systems weigh, how they weight them, or why a given application received a particular recommendation. A lender can truthfully tell a rejected applicant that the denial reflects the AUS finding, without being able to explain what produced it. Under these conditions, disparate impact claims face a nearly insurmountable evidentiary problem: the plaintiff must demonstrate that a facially neutral practice caused a disparate effect and that an alternative practice would serve the lender's legitimate purpose with less discriminatory effect. Without access to the algorithm's logic, constructing that argument is nearly impossible. Algorithm transparency advocates are protecting the principle that any system making high-stakes decisions about people's lives must be auditable by the people it affects — that opacity is not a neutral design choice but a transfer of power from applicants and regulators to lenders.

The proxy variable problem: algorithms that formally exclude race can still encode race through correlated variables — zip code, school district, employer, internet usage patterns — and the system cannot be evaluated for discrimination without knowing what variables are in the model. The Equal Credit Opportunity Act prohibits lenders from using race, color, national origin, sex, religion, marital status, or age as credit factors. It does not prohibit using zip code, which correlates strongly with race in a country where residential segregation has produced racially homogeneous neighborhoods. It does not prohibit using property value, which correlates with neighborhood racial composition because decades of discriminatory appraisal practices produced lower appraised values in Black neighborhoods. It does not prohibit using employment type, educational background, or dozens of other variables that predict creditworthiness partly because they predict race. The interagency 2024 final rule on automated valuation models — which requires quality control policies intended to ensure AVMs produce credible and accurate estimates free of discriminatory factors — begins to address this problem. Audit advocates argue it is insufficient without disclosure requirements that would allow independent researchers and regulators to test whether the models are in fact producing race-neutral outputs.

The fintech amplification argument: digital lenders, which can reach borrowers outside traditional branch networks and process applications at scale, have the potential to expand credit access significantly — but realizing that potential requires accountability structures that currently don't exist. The Bartlett et al. study found that fintech lenders, because they process applications through algorithms rather than face-to-face interactions, do reduce discrimination relative to traditional lenders — the disparity in approval rates is smaller. But the disparity in pricing persists: Black and Latino borrowers who are approved by fintech lenders still pay more, and the study's authors argue this pricing disparity is consistent with algorithmic systems detecting race through correlated features and adjusting pricing accordingly. Algorithm transparency advocates are protecting the recognition that the potential benefits of automated lending — scale, consistency, reduced individual bias — cannot be realized while the systems themselves remain unauditable.

What traditional credit standards advocates are protecting

The argument that FICO scores and standard underwriting criteria — debt-to-income ratio, loan-to-value ratio, payment history, credit utilization — are empirically validated predictors of default risk, that using them is required by sound lending practice, and that racial disparities in mortgage approvals reflect genuine differences in creditworthiness rather than discrimination in the current system. This position is often caricatured as defending redlining by other means, but its strongest version is worth taking seriously. FICO scores predict default with genuine accuracy across racial groups. A lender that approved loans based on factors less predictive of repayment would face higher default rates and threaten the financial soundness of the lending system — including harming borrowers who received loans they couldn't repay. The Equal Credit Opportunity Act explicitly requires that credit decisions be based on creditworthiness factors; using race directly would violate the law and introduce an element that is both legally prohibited and analytically irrelevant to the prediction of repayment. Traditional credit standard advocates argue that the claim that current underwriting criteria are discriminatory proves too much: if any metric that correlates with race is a proxy for race, then virtually no underwriting criteria survive scrutiny, and the result is a credit system that cannot price risk.

The individual loan officer bias argument: algorithmic systems reduce discrimination by removing the discretionary judgment of individual loan officers, who evidence suggests apply different standards to applicants of different races — and that the way to address remaining disparities is to improve algorithmic models, not to add human discretion back. The pre-algorithmic era of mortgage lending was not a golden age for Black borrowers. The HOLC redlining maps, FHA underwriting guidelines that explicitly downgraded neighborhood ratings for racial composition, and individual bank loan officer practices that denied applications from Black professionals with excellent credit histories are all well-documented. Traditional credit standards advocates argue that the shift toward algorithmic underwriting represents genuine progress — a system that applies the same rules to every applicant is more equitable than one where the outcome depends on who's across the desk. The Bartlett et al. finding that algorithmic lenders discriminate less than face-to-face lenders is, in this reading, evidence that the direction is right, and that the residual disparities in algorithmic lending are better addressed by improving the models than by reintroducing human judgment.

The wealth gap upstream argument: racial disparities in mortgage approvals are the downstream effect of upstream disparities in wealth, income, and credit history — themselves products of historical discrimination — and the credit system cannot be the site of the remedy for harms that originated elsewhere. This is the position that most complicates the debate. Credit scores and income are racially disparate because decades of discriminatory employment, education, and wealth-building practices produced racially disparate outcomes. A lender that approved loans based on race-adjusted standards would be attempting to correct for upstream injustice through downstream credit decisions — which critics argue is neither the correct tool nor the lender's appropriate role, and which creates real risk of harm if borrowers receive loans they genuinely cannot repay. Traditional credit standards advocates are protecting the recognition that not every institution that produces racially disparate outcomes is the site of the discrimination, and that misidentifying the level of the problem generates solutions aimed at the wrong target.

What alternative data and structural reform advocates are protecting

The argument that the FICO scoring architecture is itself the problem — not because FICO discriminates intentionally, but because it is structurally thin-file for communities historically excluded from formal credit markets — and that incorporating alternative data sources would simultaneously improve predictive accuracy and reduce racial disparities. FICO scores require a credit history: typically, at least one account open for six months and at least one account reported to a credit bureau in the past six months. Approximately 26 million Americans are "credit invisible" — they have no credit file — and another 19 million are "unscorable" with existing models. These populations are disproportionately Black, Latino, and low-income. The inability to score them is not a neutral technical limitation; it reflects a definition of creditworthiness built around formal credit markets that these communities were historically excluded from. Alternative data advocates argue that rent payments, utility payments, and telecommunications payments are as predictive of mortgage repayment behavior as credit card payment history — and that incorporating them would extend credit to borrowers who are reliable payers but lack the formal credit history the current system requires. Fannie Mae's 2022 announcement that it would factor positive rent payment history into Desktop Underwriter recommendations was described by the agency as potentially benefiting over 17 percent of previously unscorable applicants. VantageScore 4.0, which incorporates trended data including utility and telecom payments, similarly scores millions of borrowers that FICO 8 cannot.

Richard Rothstein's The Color of Law argument: the racial wealth gap that algorithmic lenders are measuring was not produced by private discrimination or individual bad choices but by explicit government policy — and this history obligates a policy response, not merely a private sector adjustment. Rothstein's 2017 book documents, at granular local level, the federal, state, and municipal policies that created racially segregated neighborhoods: FHA guidelines that explicitly refused to insure mortgages in integrated neighborhoods; Veterans Administration policies that denied GI Bill benefits to Black veterans seeking homes in the suburbs being built for returning white soldiers; explicitly racially restrictive covenants that were written into deeds; public housing sited in ways designed to maintain segregation. The argument is not that individual lenders today are acting in bad faith, but that the wealth and homeownership gaps those policies created compound across generations — and that the credit metrics lenders use today are measuring, in part, the distance of a family from a government-subsidized wealth-building opportunity that was explicitly denied to Black families in the postwar decades. Structural reform advocates are protecting the recognition that this history changes what counts as "race-neutral" credit scoring: a system measuring creditworthiness that was shaped by racially discriminatory wealth policy is not neutral simply because it doesn't mention race.

The Community Reinvestment Act expansion argument: the CRA only covers federally insured depository banks, excluding mortgage companies, credit unions, and fintech lenders that now originate a majority of mortgages — and this coverage gap means that the institution that most expanded credit access to communities of color faces no affirmative obligation to serve them. Non-bank mortgage lenders — companies like Quicken Loans (now Rocket Mortgage), United Wholesale Mortgage, and loanDepot — now originate more than 60 percent of mortgage loans in the United States, including a substantial share of FHA loans in communities of color. None of them are subject to CRA requirements. Ta-Nehisi Coates's 2014 essay "The Case for Reparations" documented in detail how contract buying — a practice in which Black buyers could not obtain conventional mortgages and instead entered predatory installment contracts that gave them no equity and resulted in eviction if they missed a payment — stripped wealth from Black communities in postwar Chicago at the same moment white families were building equity through FHA-subsidized mortgages. The structural reform argument is that the CRA, properly extended to cover all mortgage originators, would create an affirmative obligation that the current market does not impose.