Mortgage Lending and Denial Disparities in Chicago Through HMDA

Abstract

The Home Mortgage Disclosure Act file records who applies for a mortgage in Chicago, who is approved, who is turned away, and at roughly what price. It leaves out the one number lenders weigh most, namely the applicant's credit score. This paper joins the published lending literature to a fresh descriptive reading of the real public HMDA registers for Cook County from 2018 through 2023, a pooled body of 1,466,670 credit decisions ^[1]. The approach is an anatomy of a single application, traced from the lender's decision to the geography it lands in. The disparities are large and they hold. Black applicants were denied at 34.1 percent against 14.4 percent for non-Hispanic white applicants, a ratio of 2.37 to 1, with Latino applicants at 26.5 percent and Asian applicants at 17.6 percent ^[1]. The ordering does not break in any of the six years. Income, the only serious control the public file allows, barely moves the result. Inside the highest income band the Black-white gap is still 17.7 percentage points ^[1]. Among borrowers a lender already approved, 18.6 percent of Black borrowers received a higher-priced loan against 2.6 percent of white borrowers, a gap wider than the one at the door ^[1]. Lending concentrates in a short list of institutions tilted heavily toward white borrowers. One planned test came back empty. The 1938 redlining grade of a Chicago tract does not predict its modern denial rate ^[1]. Following recent Federal Reserve evidence, we read all of this as the careful measurement of a raw disparity, not as proof of present-day discrimination, because the missing credit score puts that claim out of reach ^[4].

What the file records and what it leaves out

A mortgage application is a small bundle of facts that a lender converts into one decision. Most of those facts never reach the public. The Home Mortgage Disclosure Act is the law that pries part of the bundle open. Congress passed it in 1975, at the end of a decade of organizing by community groups on the South and West Sides of Chicago and in cities like it, who argued that banks were taking deposits from Black neighborhoods and lending the money back out somewhere whiter. The charge was disinvestment, and it was hard to prove because the lending data did not exist in any usable public form. HMDA created that record. It requires most lenders to report, loan by loan, who applied, what they asked for, where the property sits, and what the lender decided. A year later the Community Reinvestment Act of 1977 gave regulators a tool to act on what the data showed. The two statutes were built as a pair, one to make lending visible and one to put a thumb on the scale toward the neighborhoods that had been starved. Fair-lending research has run on the HMDA file ever since.

For most of its life the file was coarse. It told you the race of the applicant, the action the lender took, and the census tract, and not much that would let an analyst ask why a given loan was denied. That changed with the 2018 reporting year. After the financial crisis, the Dodd-Frank Act moved HMDA rulemaking to the Consumer Financial Protection Bureau and ordered the file widened. The expanded register added the fields that turn a head count into something closer to an anatomy. An analyst can now read an applicant's reported income, the loan amount, a debt-to-income measure, a derived loan-to-value ratio, and a rate-spread field that flags whether a loan that closed carried an unusually high price. Strung together, those fields let you follow one application from the request, through the decision, to the price the borrower paid if the loan funded. That is the spine of this paper.

One field is still missing, and it is the one that matters most. The public HMDA file does not contain the applicant's credit score. In the underwriting that ends in an approval or a denial, the score is the single number lenders rely on more than any other, a compact summary of repayment history that sits at the center of nearly every automated and manual decision. The Federal Reserve economists who ran HMDA analysis for years wrote the standard methodological guide to the file, and they are blunt about the limit. The data can describe the pattern of lending and were never built to separate creditworthiness from the treatment of equally creditworthy applicants, because the central creditworthiness variable is not in the file ^[12]. Every serious reader of HMDA has had to work inside that wall. We work inside it too, and we put the limit at the front of the paper rather than burying it in a footnote, because it governs every number that follows.

That constraint fixes what kind of paper this is, and the honest description is narrow. This is not an experiment. We did not audit lenders, send matched testers, interview borrowers, or build a model that predicts who should have been approved. What we did is smaller and, we think, still worth doing. We took the real public HMDA loan and application registers for Cook County, the home of Chicago, and pooled the six reporting years from 2018 through 2023 into one body of 1,466,670 credit-decision records, pulled directly from the FFIEC and CFPB Data Browser that publishes them ^[1]. From that file we computed descriptive breakdowns. Denial rates by race and ethnicity. The same rates inside income bands. The share of approved loans that came back higher-priced. The reasons lenders gave for the denials. The concentration of lending across institutions. And a geographic overlay against the 1930s redlining map. These are counts and ratios, not coefficients from a fitted model. They describe what the file holds. They do not reach behind it.

The descriptive work is paired with the published literature for a specific reason. Four decades of mortgage-lending research, much of it built on HMDA, has already charted the ground a raw gap sits in. It has measured how large unconditional disparities tend to be, how much they shrink when fuller underwriting data is added, and how much survives. Reading our Chicago numbers against that body of work is what keeps the interpretation honest. A raw gap of a given size means one thing if the literature finds such gaps mostly dissolve under proper controls, and something quite different if they mostly persist. We lean on that comparison throughout, naming the sources as we go, so a reader can always tell our figures from the certified analysis apart from claims borrowed from researchers who held data we do not.

The order of the paper follows one application through the system. We open on the decision itself, the unconditional denial rate by race and ethnicity, because that is the front door and the largest number. We apply the strongest control the file allows, reported income, and watch how little the gap gives. We open the denied files and read the reasons lenders wrote down, which is where the missing credit score shows up as an absence rather than a presence. We turn to the borrowers who cleared the bar and look at what their loans cost. We ask who the lenders are and how concentrated the market is. And we lay the whole thing over the 1938 grade lines to test whether the old boundary still predicts where credit is denied today. That last test failed, and we report it as a null instead of dressing it up as a confirmation.

Two disciplines hold across all of it. Every empirical number is either ours, computed from the certified analysis and marked as such, or it carries the name of the source it came from. We invent nothing. Where we are unsure a figure belongs to us, we leave it out. And we make no causal claim. The numbers describe patterns in who was approved, who was denied, and who was charged more. They do not establish why. Wherever a gap survives a control, the missing credit score is named in the same breath, because that absence is the reason a surviving gap cannot be read as a measure of discrimination. What follows is an anatomy, not a verdict.

The neighborhoods the numbers sit in

Chicago is an unusually sharp place to read a lending file, because the lines the file might trace were drawn here early and drawn hard. The terms that organize the national redlining story were minted on this ground. When the Home Owners' Loan Corporation graded the city in the late 1930s, it marked the Black Belt on the South Side in red and labeled it hazardous, which told lenders not to write mortgages there. The grades were not the cause of segregation so much as its accounting, but they hardened a pattern that decades of private and public action had already begun, and they handed lenders a map. The historical record on this is not ours to add to, but it is the context every modern Chicago lending number sits inside, and the literature we cite establishes its weight.

The mechanisms that filled the red zones are well documented. Restrictive covenants kept Black families out of white blocks by private contract until the Supreme Court declined to enforce them in 1948. The Federal Housing Administration would underwrite new subdivisions only on terms that excluded Black buyers, which is the central thread of Richard Rothstein's account of segregation as the product of explicit policy rather than private accident. And when mortgages were unavailable, contract selling moved in to fill the gap, an instrument Beryl Satter reconstructed from her own father's law practice on the West Side. A speculator bought a building cheap from a fleeing white owner, sold it on contract to a Black family at a steep markup, and held the deed until the final payment, so a single missed month could erase years of payments and let the seller start over with the next family. The family paid like owners and were protected like nobody. None of that is in the HMDA file. But it is the recent past out of which the present distribution of credit, income, and home equity in Chicago was produced, and it is why a raw lending gap measured in 2020 cannot be read as if it appeared from nowhere.

That history runs directly into the data through one channel the literature has measured carefully. Aaronson, Hartley, and Mazumder use the boundaries of the HOLC maps themselves, comparing tracts that fell just inside a downgraded line against tracts just outside it, and find that the 1930s grades caused lasting reductions in homeownership and house values and lasting increases in segregation, working through constrained access to credit over many decades ^[7]. The map did not merely record where Black Chicagoans lived. It changed what their property was worth and who could borrow against it, and the effect compounded across two and three generations. The closer of the two crises also matters here. Reid and colleagues document the dual mortgage market of the years before 2008, in which Black and Latino borrowers were steered disproportionately into subprime and high-cost loans and then foreclosed at more than twice the white rate ^[10]. The foreclosure wave fell hardest on exactly the neighborhoods the 1930s map had marked, stripping equity that had taken decades to build. By the time our window opens in 2018, the population applying for mortgages in Chicago carries that history in its balance sheets, whether or not any 2018 lender ever saw a 1938 map.

We rehearse this not to argue it again but to set the frame. The disparities the HMDA file shows are unconditional, and an unconditional gap can reflect many things at once, including the unequal distribution of the very risk factors the file cannot measure. That unequal distribution is not natural. It has a documented history, much of it written in Chicago. Holding both facts together, namely that the modern gap is large and that the history behind it is real, is the only honest way to read what comes next. Neither fact lets us skip to a causal claim about any individual lender, and we do not try.

What the lending literature already established

Our descriptive numbers do not arrive in an empty room. Researchers have been pulling on the HMDA file and on matched-pair audit studies for the better part of fifty years, and the shape of their findings tells you what to expect from a raw Chicago gap and how seriously to take it. Setting our work against theirs is the only way to say which of our results is genuinely new and which is one more instance of a pattern that has been visible for decades. The short version runs in three parts. Large raw disparities are well documented. Fuller data shrinks them substantially. Something usually survives. The argument has always been about how much, and why.

The study that set the terms is the Boston Fed work of the mid-1990s by Munnell and her colleagues, still the reference point for almost everything written since. They started from HMDA records and then did the expensive thing. They went to the lenders and collected the loan-file variables the public data lacks, the debt ratios and credit histories and loan-to-value detail that underwriters actually use. In the raw data, minority applicants in Boston were roughly 2.7 times as likely to be denied as white applicants ^[2]. That number alone would have meant little, because the raw file did not show what the underwriters saw. The contribution was what happened when they added thirty-eight omitted measures of creditworthiness. The disparity fell to roughly 1.6 times ^[2]. Honest readers take two lessons from that single result, and the order matters. First, the raw HMDA gap overstates how unequal lending is once you can see what lenders see, because the gap shrank by a wide margin under proper controls. Second, it did not vanish. Race still carried an independent association with denial after the fullest controls then available, and pinning down that residual is the project three decades of subsequent work inherited.

Pull back from any single city and the long arc comes into view in the meta-analysis by Quillian, Lee, and Honoré, who pooled nineteen observational mortgage-lending studies and sixteen audit experiments running from the 1970s into the 2010s ^[6]. The two kinds of evidence answer different questions, and it is worth keeping them apart. An observational study like ours, or like Munnell's, reads the records of real applications and measures the gap that remains after whatever controls the data allow, so it always carries the worry that an unmeasured variable is doing the work. An audit study sends matched testers, identical on paper but differing by race, to the same lenders, which closes the unobserved-variable gap by construction but tests only the earliest stage of the process, the inquiry and preapproval, rather than the full underwriting of a funded loan. Quillian and colleagues find that both literatures point the same way. Discrimination declined measurably across the four decades, with a stubborn residual that did not close, and the sharpest improvement came in the outright denial of credit availability rather than in its terms ^[6]. Decline and persistence together is the frame our cross-year Chicago numbers have to be read inside. A disparity that holds steady across recent years is consistent with their picture of a problem that has receded from its mid-century extremes without disappearing, not with a story of either steady worsening or resolution. It also locates our work squarely on the observational side, with the unobserved-variable worry fully in force, which is the worry the missing credit score keeps alive on every page.

The single most important recent result for our purposes, and the one we treat as the honesty anchor for the whole paper, comes from the Federal Reserve work of Bhutta, Hizmo, and Ringo ^[4]. They use data that includes the credit scores and detailed underwriting variables the public HMDA file lacks, and they exploit both human and algorithmic credit decisions to separate the part of a denial gap that observable risk explains from the part that looks like differential treatment of similar applicants. Observable risk factors explain most of the racial disparity in denials. After accounting for what lenders can actually see, the residual differential-treatment gap, the part consistent with applicants of equal risk being treated differently, is on the order of one to two percentage points ^[4]. That finding does two things to our analysis at once. It tells us a large raw denial gap is exactly what you would expect even in a market where the treatment-based component is small, because observable risk is unequally distributed across groups for reasons that trace back through the history sketched above. And it tells us, sharply, that our own unconditional Chicago numbers cannot be read as a discrimination estimate, because we lack precisely the variables that did most of the explanatory work in their study. We return to this at every point where a gap survives a control.

Pricing carries its own debate, and the two poles of it matter for the section where we look at what approved loans cost. Bhutta and Hizmo show that apparent interest-rate gaps by race and ethnicity are largely offset once discount points are taken into account ^[3]. A discount point is cash paid at closing to buy down the monthly rate, so two borrowers with the same underlying offer can end up at different rates simply by paying different amounts upfront. On their reading, minority and white borrowers face the same menu of rate-and-point combinations and sort to different places on it, trading an upfront fee for a lower monthly rate or the reverse, so a raw rate gap can overstate unequal treatment ^[3]. Bartlett, Morse, Stanton, and Wallace cut the other way. They find that risk-equivalent Black and Latino borrowers still pay more on both GSE and FHA loans, a residual they estimate costs minority borrowers in the neighborhood of 450 million dollars a year ^[5]. Their FinTech result is the part most relevant to a market increasingly run by algorithms. Automated lenders, stripped of any face-to-face encounter, narrow the pricing gap relative to loan officers but do not close it, which suggests part of the gap survives even when the most obvious channel for in-person bias is removed ^[5]. The two papers are less contradictory than they look. They use different data and different definitions, and read together they bracket the truth, that some of a raw pricing gap reflects borrowers sorting across a shared schedule and some of it does not. Our rate-spread numbers sit inside that bracket, and we will not pretend to resolve which share is which.

Closest to our own setup is the work of Lewis-Faupel and Tenev, who use the post-2018 expanded HMDA fields, the same enlarged file we draw on, to study approval and pricing disparities ^[9]. Conditioning on the observable creditworthiness measures the new file added shrinks both the approval and the pricing gaps, they find, yet leaves disparities that stay economically meaningful, and they report an asymmetry worth carrying forward. Omitted factors do a better job explaining the pricing gap than the approval gap ^[9]. That is a direct signal about the data we hold. It suggests that adding the variables we cannot see would compress our numbers without erasing them, and that the approval gap in particular is the more resistant of the two to observable controls.

There is also a body of national HMDA work tying the modern file back to the 1930s map, which bears directly on the geographic test we run later. Mitchell and colleagues at the National Community Reinvestment Coalition read HMDA back to 1981 and find that the average HOLC-redlined neighborhood received about 3,000 fewer mortgages than a Best-graded area over four decades, net of housing-availability controls ^[8]. That is a finding about cumulative lending volume measured over a long horizon, and it is the precedent against which we have to read our own much shorter and narrower slice. Holding it in view ahead of time is what keeps us from overclaiming when our slice comes back flat.

Laid end to end, this literature defines a corridor. Raw HMDA gaps are large, real, and well attested. Fuller data shrinks them, in the Boston case by roughly the distance from 2.7 times to 1.6 times ^[2], and the Federal Reserve work suggests the treatment residual in denials is only a point or two once everything observable is held constant ^[4]. Something nonetheless survives, more durably for approval than for price ^[9]. Our Chicago numbers live inside that corridor. They are unconditional by necessity, and the job of the sections that follow is to report their size and shape precisely while never pretending they sit anywhere but at the raw, uncontrolled end of the range these researchers have mapped.

How we built the file

The source is the public loan and application register that the Federal Financial Institutions Examination Council and the Consumer Financial Protection Bureau publish through the HMDA Data Browser ^[1]. The register is a U.S. government work, free to redistribute, and it is the same file every fair-lending researcher cited above either used or descends from. We took the records for Cook County, identified by the county FIPS code 17031, and pooled the six reporting years from 2018 through 2023, the full run of the expanded post-Dodd-Frank file as it stood when we pulled it. That gives 1,466,670 records once the year files are stacked ^[1]. The geography is deliberate. Cook County is not the whole Chicago metropolitan area, which spreads across five collar counties as well, but it holds the city of Chicago and the great majority of the population whose lending history the redlining literature is about, and confining the analysis to one county keeps the tract-level overlay against the 1938 map tractable. A reader who wants the six-county metro can rerun the same code against the additional FIPS codes, and the posted files make that straightforward ^[1].

Before any rate is computed, two coding choices have to be pinned down, because both move the headline numbers and both have a defensible convention behind them. The first is how a credit decision is counted. HMDA's action-taken field distinguishes a loan that was originated from one that was approved but not accepted, denied, withdrawn by the applicant, closed for incompleteness, or purchased from another institution, and it flags preapproval requests separately. A denial rate is only meaningful against the applications a lender actually decided on the merits, so our denominator is the sum of originated, approved-not-accepted, and denied files. Withdrawn and incomplete files carry no decision and are excluded. Purchased loans are someone else's origination and are excluded. Preapproval-only outcomes are not folded in as denials. That is the standard fair-lending denominator, and it runs a little leaner than the denied-over-all-records cut that a casual query against the file would return, which is why our 2022 numbers come in slightly below the raw aggregation the Data Browser hands back for the same county and year ^[1].

The second choice is how race and ethnicity are assigned. HMDA reports ethnicity and race in separate fields, and a person can be both Hispanic and, say, white or Black by those fields. To produce mutually exclusive groups that do not double-count, we follow the convention the Federal Reserve uses in its own HMDA work. Anyone recorded as Hispanic by derived ethnicity is counted as Latino or Hispanic, and everyone else is assigned by derived race. The four groups we track, white non-Hispanic, Black, Latino or Hispanic, and Asian, therefore do not overlap, and they line up with the categories the literature we compare against reports ^[4]. This is a defensible convention rather than the only possible one, and it has consequences worth naming. It places Black Hispanic applicants in the Latino group rather than the Black one, and it collapses a great deal of heterogeneity inside the Asian and Latino categories, both of which span national origins with very different lending histories. We do not break those subgroups out, because the certified analysis pools them, and we flag the limitation rather than papering over it.

A large share of the file does not carry a clean race label at all. For 143,856 of the pooled credit decisions, race is recorded as not available ^[1]. We neither drop these records nor distribute them across the four groups. We hold them as their own category, report their denial rate alongside the others, and exclude them from the four-group comparisons, because their composition is unknown and any assignment would distort the groups they were folded into. Treating the unlabeled block as a visible category rather than a silent deletion is part of being straight about how much of the file is fully classified.

The analysis itself is a set of aggregations, not a model. For each cut we report counts and the rates and ratios built directly from them. Denial rates are denials over credit-decision applications. Income bands are built from HMDA reported income, in thousands, and we hold the band fixed while comparing groups, with records missing income kept in an explicit unknown band and never imputed. The higher-priced share is the count of originated first-lien loans whose reported rate spread is 1.5 or more, over the count of originated first-lien loans carrying any reported spread. Lender concentration ranks reporters by originated dollars using the HMDA legal entity identifier, the unique code that names each institution in the file. The geographic overlay assigns each Cook County census tract to a 1938 HOLC grade by a tract-centroid point-in-polygon test against the digitized 1938 zones already held in our repository, then groups tracts by grade. Every one of these is a descriptive operation. None of them estimates a coefficient, controls for a covariate in a regression sense, or predicts a counterfactual decision. The whole pipeline is posted so any reader can reproduce a number, change a denominator, or extend the geography ^[1].

We checked the build against the live source rather than trusting our own stacking. Pulling a Cook County 2022 aggregation directly from the Data Browser API on a broad denominator returned roughly 18.8 percent denial for white applicants and 36.1 percent for Black applicants ^[1]. Our pooled-method 2022 figures, on the leaner credit-decision denominator, are 16.3 and 34.8 percent ^[1]. The two are close and differ in exactly the direction the denominator difference predicts, which is the confirmation we wanted that the curated file reproduces the public one rather than drifting from it. With the file built and the conventions fixed, the rest of the paper is reading.

The decision, in raw numbers

The first thing to fix is the denominator, because a denial rate is only as trustworthy as the set of applications it divides by. We restrict the count to credit-decision applications, the files a lender actually adjudicated on the merits. In HMDA's action-taken coding that means loans originated, applications approved but not accepted by the borrower, and applications denied. We drop files that the applicant withdrew or that closed for incompleteness, since those carry no lending decision, and we do not count preapproval-only actions as denials. The denial rate we report is therefore denials over adjudicated applications, the figure fair-lending analysts use, and it runs slightly leaner than a naive count of denials over every record in the file. Race and ethnicity follow the mutually exclusive coding the Federal Reserve uses. Anyone Hispanic by derived ethnicity is counted as Latino or Hispanic, and everyone else is assigned by derived race, so the four main groups do not overlap.

Pooled across 2018 through 2023 in Cook County, the picture is stark and easy to state. White non-Hispanic applicants were denied at 14.4 percent across 477,229 adjudicated applications. Black applicants were denied at 34.1 percent across 126,447 applications, which is 2.37 times the white rate. Latino or Hispanic applicants were denied at 26.5 percent across 156,819 applications, 1.84 times white. Asian applicants were denied at 17.6 percent across 78,906 applications, 1.22 times white ^[1].

Set the four rates side by side and a few features deserve a careful read before anyone overinterprets them. The Asian rate sits closest to the white rate, a gap of about three percentage points that is real but of a different order than the others. The Black and Latino bars are the ones that pull away, the Black bar more than double the white and the Latino bar nearly so. That contrast, near-parity at one end and a doubling at the other, is the central descriptive fact of this dataset, and the ordering carries no cause on its face. The bars give the size of the gap at the front door, the rate at which different groups were turned away before a single adjusting variable is brought in. They do not say why. The spread is consistent with several stories at once, including the unequal distribution of the very risk factors the file cannot measure.

A fifth group sits outside the four bars by design. The file records race as not available for 143,856 applications, a large block of the pooled credit decisions, and those applications were denied at 22.5 percent, 1.56 times the white rate ^[1]. We hold this category out of the four main groups and report it on its own line rather than folding it in or dropping it quietly, because its composition is unknown and assigning it would distort whatever group it was mixed into. Naming the size of the unlabeled block is part of being honest about how much of the file carries a clean race or ethnicity tag and how much does not. The mutually exclusive coding that governs the four main groups, with Hispanic ethnicity taking precedence and race assigned thereafter, mirrors the convention the Federal Reserve uses in its own HMDA reporting, which is why our group definitions line up with the literature we compare against.

It helps to place our raw Chicago gap next to the most familiar number in this field. The Boston Fed study found minority applicants roughly 2.7 times as likely to be denied as white applicants in the unadjusted HMDA data ^[2]. Our pooled Black-white ratio of 2.37 to 1 sits a little below that, and our Latino-white ratio of 1.84 to 1 lower still, which tells us the order of magnitude is the familiar one, not something peculiar to Chicago or to these years. The comparison cuts two ways, and both belong here. It means our numbers are not freakish. They are the kind of raw disparity HMDA has shown for decades. And it is a reminder of what came next in Boston, where adding thirty-eight creditworthiness variables pulled that 2.7 figure down toward 1.6 ^[2]. We have no such variables to add. The 2.37 we report is the starting point of that kind of exercise, not its conclusion, and the sections that follow apply the only adjustment the public file permits, income, and find it does far less than the Boston controls did.

For now the claim is deliberately small. At the front door of the Cook County mortgage market, pooled over six years, Black applicants were turned away at more than twice the white rate and Latino applicants at nearly twice, while Asian applicants faced a much smaller gap. That is the size of the disparity before any control. What it means is the work of the rest of the paper, and the first question is whether the gap is a feature of one unusual market year or a constant across the cycle.

A gap that holds across the cycle

The years 2018 through 2023 were not a calm stretch in the mortgage market, which makes them a useful test. Rates fell to historic lows through 2020 and 2021 as refinancing surged, then climbed sharply as the Federal Reserve tightened, and the mix of who was applying and for what shifted underneath all of it. If the racial denial gap were an artifact of one year's conditions, a single refinance wave or a single rate shock, it would wobble as the market moved. Following the disparity year by year is how you tell a structural pattern from a transient one.

The lines move together, and that is the first thing to see. The white non-Hispanic denial rate starts at 20.3 percent in 2018, falls through the easy-credit years to 11.2 percent in 2020 and 10.8 percent in 2021, then climbs back to 16.3 percent in 2022 and 19.8 percent in 2023 ^[1]. The Black line traces the same shape one tier up, from 41.2 percent in 2018 down to 29.0 percent in 2020 and 29.1 percent in 2021, then back to 34.8 and 38.9 percent ^[1]. Latino and Asian rates swing in the same rhythm, the Latino line running from 34.8 percent down to 20.6 percent and back to 30.4 percent, the Asian line from 24.9 percent down to 12.9 percent and back to 22.2 percent ^[1]. Every group got approved more easily when credit was loose and turned away more often when it tightened, which is what a market responding to the same macro conditions looks like.

What does not move is the ordering. In all six years the Black and Latino lines sit above the white line, and the Asian line sits between, closer to white. The cycle changes the level of denial for everyone at once and never reshuffles the rank. The absolute Black-white gap does breathe a little with the cycle, but less than the swing in the level would suggest. It runs near 20.9 points in 2018, narrows to about 17.8 points in 2020 when credit was loosest, and opens back toward 19.1 points by 2023 ^[1]. The ratio tells almost the opposite story from the gap, which is exactly why both deserve to be on the page. Measured as a multiple of the white rate rather than a difference from it, the Black-white disparity is widest in the easy-credit years, reaching about 2.69 times the white rate in 2021, and narrowest at the ends of the window ^[1]. The two measures diverge because the white baseline itself swings so much. When the white rate falls to near 11 percent, even a Black rate that has also fallen looks like a much larger multiple, while the raw point gap is at its smallest. Reporting only the ratio would make the loose years look the most unequal, and reporting only the gap would make the tight years look that way, so we keep both and let the reader see the whole shape rather than the half that flatters a story. The direction stays fixed through all of it. No year in the window shows a Black or Latino denial rate at or below the white rate. A disparity that survives a full rate cycle, a refinance boom, and a sharp tightening is not a property of one market year. It is a property of the market.

The 2022 figures make a clean single snapshot, useful because it is the year we can most directly check against the raw source. In 2022 the Black denial rate was 34.8 percent against a white rate of 16.3 percent, a ratio of 2.14 to 1, with Latino at 27.5 percent and Asian at 19.2 percent ^[1]. Those are our pooled-method numbers, computed on the credit-decision denominator described earlier. When we pulled a live Cook County aggregation straight from the CFPB Data Browser for 2022 on a slightly different and broader denominator, it returned roughly 18.8 percent for white applicants and 36.1 percent for Black applicants ^[1]. The two do not match to the decimal, and they are not supposed to, since one counts only adjudicated credit decisions and the other is a coarser denied-over-records cut. They sit close enough to confirm that our numbers reproduce against the underlying public file rather than drifting from it, and the small difference is the denominator discipline we described, working as intended.

The persistence itself echoes the meta-analytic finding of Quillian, Lee, and Honoré, whose four-decade synthesis describes denial-side disparities as the more durable feature of mortgage lending even as discrimination on other margins receded ^[6]. A gap that holds its shape across every year of a volatile cycle fits that durability. We stop short of reading anything causal into the cyclical swings themselves. The fall and rise of denial rates tracks the rate environment and the changing composition of applicants, refinancers crowding in when rates drop and thinning out when they climb, and separating those compositional shifts from anything else would take variables we do not have. The defensible claim is the structural one. Across six years and a full turn of the cycle, the racial ordering of denial in Cook County never broke.

Holding income roughly constant, and watching the gap stay

If a denial gap shrinks to nothing once you account for who could actually afford the loan, it tells a very different story than one that holds firm. Income is the strongest control the public HMDA file lets us apply, and the honest report is that it barely moves the result. This is the most careful section of the paper, and it has to be, because it is the one most easily misread in both directions, by those who would call the surviving gap proof of bias and by those who would call its shrinkage proof of innocence. Neither reading is available from these numbers.

The method is deliberately coarse. We bind the comparison inside HMDA income bands built from reported income and compare each group's denial rate against the white rate of applicants in the same band. We do not smooth, model, or impute. Reported income is missing for roughly a third of records, and rather than guess at those values we keep them in an explicit unknown band and report it alongside the rest, so the reader can see exactly how much of the file carries an income figure and how much does not. We did not use debt-to-income as a primary control, even though the expanded file nominally carries it, for two concrete reasons. It is missing for about 35 percent of records, and it arrives as an inconsistent mix of bucketed strings and raw integers that cannot be pooled cleanly without imposing assumptions we would rather not make. Income, for all its coarseness, is the cleaner instrument, and it is the one we lean on.

Walking up the income ladder, the gap does not behave the way the affordability story predicts. The Black-minus-white denial gap is 12.5 percentage points in the under-50k band, then it widens rather than narrows as income rises, reaching 15.3 points in the 50k-to-75k band, 16.6 points from 75k to 100k, 18.9 points from 100k to 150k, and 17.7 points at 150k and above ^[1]. The Latino-minus-white gap climbs the same way, from a near-negligible 1.9 points at the bottom to 6.9, then 9.0, then 11.9, and 12.7 points at the top ^[1]. These are gaps against white applicants reporting the same income, and they grow as you move toward the high end of the distribution rather than collapsing there.

It is worth looking at the two halves of each gap separately, because the white side and the Black side move in opposite directions. The white denial rate behaves exactly as an affordability story predicts. It falls steadily from 34.4 percent in the lowest band to 17.2, 13.9, 11.7, and 9.7 percent at the top, cut by roughly two-thirds as income climbs ^[1]. The Black rate barely descends. It starts at 47.0 percent in the lowest band and lands at 27.5 percent in the highest, and across the three middle bands it is nearly flat, sitting at 32.5, 30.5, and 30.6 percent ^[1]. Income buys a white applicant a much better chance of approval and buys a Black applicant comparatively little. That divergence is what opens the gap as you climb, and it is the kind of pattern income alone cannot account for.

The headline is best stated as a comparison between the raw gap and the within-band gap. Pooled across all incomes, the raw Black-white denial gap is 19.7 percentage points. Confining the comparison to the highest income band, 150k and above, brings it to 17.7 points, a reduction of only about 10 percent ^[1]. The Latino-white gap does not shrink at all under the same exercise, moving from a raw 12.1 points to 12.7 points within the top band ^[1]. Put concretely, the highest-earning Black applicants in the file, those reporting 150k or more, were denied at 27.5 percent, well above the 14.4 percent pooled white rate and only a little below the 34.4 percent rate that the poorest white applicants faced ^[1]. A Black household clearing 150,000 dollars a year was turned away at a rate close to that of a white household earning under 50,000. Income, the best lever the public file offers, moves the Black-white gap by a tenth and moves the Latino-white gap not at all.

Here is where the discipline has to be absolute. This exercise is incomplete by construction, and the reason is the one named at the top of the paper. The public file has no credit score. Income and credit score are correlated but they are not the same thing, and the variable that does the heaviest lifting in any underwriting decision is the one we cannot put in the comparison. A gap that survives an income control is therefore not a discrimination estimate. It is a gap that income alone does not explain, which is a much weaker and much more defensible claim. The Federal Reserve work makes exactly this point with the data we lack, finding that observable risk factors, credit score foremost among them, account for most of the racial denial disparity, with the treatment-based residual on the order of one to two percentage points once everything underwriters see is held constant ^[4]. The 17.7-point gap we measure inside the top income band is not that residual. It is what is left after one of several relevant variables is held roughly constant, with the most important one still missing.

The pattern we observe lines up with the closest study to our own. Lewis-Faupel and Tenev, working the same post-2018 expanded file, find that conditioning on the observable creditworthiness measures shrinks the approval gap while leaving a disparity that stays economically meaningful, and that the approval gap is more resistant to observable controls than the pricing gap is ^[9]. Our income result is a coarser version of that finding, produced with a single control, and it points the same way. The gap compresses a little and refuses to disappear. Reading our 17.7 points through their lens and the Fed's together, the most we can say is that part of what we see would likely be absorbed by the credit and underwriting variables we do not hold, in the manner the Fed documents ^[4], and part of it might not be, in the manner Lewis-Faupel and Tenev find for approval specifically ^[9]. Which part is which is not a question this file can answer.

So the careful conclusion is narrow on purpose. Within income bands the Black-white denial gap barely moves and the Latino-white gap holds steady, which establishes that income is not the explanation for the raw disparity. It does not establish what is. The single most important candidate explanation, applicant credit risk as summarized by a score, sits outside the data entirely, and until it is brought in no version of this comparison can isolate present-day treatment. We have measured how much income explains, namely very little, and we have stopped there.

The reasons lenders write down

When a lender denies an application it records why, choosing from a fixed menu of reason codes, and the public file carries up to four of them per denial. Reading that distribution by race is tempting, and it has to be handled with care. These are lender-reported characterizations, not audited findings. A single denial can carry more than one code. The reasons describe how the institution chose to label its own decision rather than an independently verified cause, and the menu is coarse. What follows is a descriptive breakdown of how denials were characterized across groups, and nothing more.

Among denied applications pooled across 2018 through 2023, the primary reasons sort differently by group in a way that points straight at the hole in the data. Black denials were attributed to credit history far more often than white denials, 31.8 percent against 20.0 percent ^[1]. White and Asian denials leaned harder on debt-to-income, cited as the primary reason in 31.5 percent of white denials and 36.4 percent of Asian denials, against 23.5 percent for Black denials ^[1]. The other categories run closer together. Collateral was the primary reason in 19.5 percent of white denials, 17.6 percent of Black, 16.6 percent of Latino, and 17.1 percent of Asian ^[1]. Incomplete application landed at 12.1 percent for white, 10.0 percent for Black, 10.8 percent for Latino, and 10.5 percent for Asian denials ^[1]. Latino denials sit between the poles, with debt-to-income at 30.2 percent and credit history at 26.0 percent ^[1]. The smaller reason codes round out the picture without disturbing it. Unverifiable information, insufficient cash, and employment history each account for only a few percent of denials in every group, and the differences among groups on those codes are small. The racial split concentrates almost entirely in the two big reasons, credit and capacity, while the property-side and paperwork reasons look much the same across the board.

The credit-history skew in Black denials is the single most consequential line in this section, and it has to be read precisely. It is exactly the channel the public file cannot see directly. The file records that a lender cited the applicant's credit history as the reason for the denial. It does not record the credit score behind that citation. We can see that credit history was invoked roughly half again as often in Black denials as in white ones, and we cannot see whether the underlying scores justify that difference, overstate it, or understate it. This is the precise point at which the missing variable becomes visible as an absence. The reason codes gesture at the very factor that would let an analyst adjudicate the gap, and then the file goes dark exactly where the score would be.

That is why the breakdown supports no causal reading in either direction. A higher credit-history share in Black denials is consistent with the Federal Reserve finding that observable risk factors, with the credit score at their center, account for most of the racial difference in denials ^[4]. If Black applicants in this file carried lower scores on average, for reasons that trace back through the dual market and the longer history of unequal credit access, then credit history would show up more often as the stated reason without any difference in how a given score was treated. The reason codes are equally consistent with that story and with others, and they cannot distinguish among them. The Boston Fed lesson applies directly. Creditworthiness variables absorb much of a raw denial gap but not all of it ^[2]. The reason codes are a shadow of those variables, informative about which factors lenders invoked and silent on the values underneath.

Read at face value and no further, the distribution says something modest and real. Across groups, denials hung on different stated grounds. Credit history did more of the work in Black denials, debt-to-income did more in white and Asian ones, and collateral and incompleteness were cited at similar rates throughout. That is a fact about how denials were labeled. It is not a fact about whether equally situated applicants were treated alike, because the field that would settle the question, the credit score sitting behind every credit-history code, is the one the public file does not contain.

What approval costs

Clearing the underwriting bar is not the end of the story, because price varies by who you are even among those a lender said yes to. The next slice of the file looks only at originated first-lien loans and asks how often the loan came back higher-priced. We define higher-priced the way HMDA does, a reported rate spread of 1.5 or more over the comparable benchmark, the threshold the rules draw from the Home Ownership and Equity Protection Act convention. A loan exempt from rate-spread reporting cannot be classified this way, and those exempt originations are about 11 percent of the total, so they sit outside this denominator. What remains is a clean count of priced first liens that did or did not cross the higher-priced line.

The pooled 2018 to 2023 result is the widest ratio anywhere in our analysis. Among 341,979 priced white first liens, 2.6 percent were higher-priced. Among 68,402 Black first liens, the share was 18.6 percent. Latino borrowers came in at 12.6 percent across 96,450 loans, and Asian borrowers at 2.7 percent across 54,024 ^[1]. That puts the Black higher-priced rate at 7.1 times the white rate and the Latino rate at 4.8 times, and the ordering does not break in any single year of the window ^[1]. Asian and white borrowers, who sat closest on the denial measure, sit almost on top of each other here as well, within a tenth of a percentage point, while the Black and Latino shares pull far away.

It helps to remember what the higher-priced flag stands for in a household's budget. The rate spread that trips the flag is the gap between a loan's price and the going rate for low-risk borrowers, and even a modest spread compounds over a thirty-year mortgage into a large sum, a difference of tens of thousands of dollars across the life of a typical loan and a heavier monthly payment the whole way through. A borrower carrying a higher-priced loan also has less cushion when income falls, which is one of the threads Reid and colleagues trace from high-cost lending into the foreclosure wave ^[10]. So the flag is not a bookkeeping nicety. It marks loans that cost their holders real money and left them more exposed, and it falls on Black borrowers in Cook County at seven times the white rate.

What makes this section sharper than the denial section is the selection working against the gap. These are not applicants. They are borrowers, people the lender already approved and funded. If approval screens out weaker files, the pool that survives should look more alike on the things that drive price, not less, and the gap among them should compress relative to the gap at the door. The opposite shows up. The denial ratio at the front was 2.37 to 1. The price ratio among the funded is 7.1 to 1, nearly triple it ^[1]. The disparity does not shrink once the riskiest applications are removed from view. On this measure it grows. That widening under a tighter screen is the detail that makes the pricing result harder to wave off than the denial result, and it is worth stating plainly even though the file cannot tell us why it happens.

The published work gives us two readings of a pricing gap like this one, and honesty requires holding both. Bhutta and Hizmo find that headline interest-rate differences by race are largely offset once discount points are counted, because Black, Latino, and white borrowers face the same point-to-rate schedule and choose different spots on it, trading cash at closing for a lower rate or the reverse ^[3]. A rate-spread incidence number does not see the points, the fees, or the cash a borrower brought to the table, so part of what we are counting could be borrowers landing at different places on a shared menu rather than being offered different menus. Against that, Bartlett and colleagues find that risk-equivalent Black and Latino borrowers still pay more on both GSE and FHA loans, a residual they put at roughly 450 million dollars a year nationally, and they find that algorithmic pricing narrows the gap without closing it ^[5]. Lewis-Faupel and Tenev, working the same expanded post-2018 fields we use, report that omitted observable factors do more to explain the pricing gap than the approval gap, which cuts the other way and warns against reading too much into a raw price differential ^[9]. The deeper backdrop is the dual market that Reid and colleagues document, in which Black and Latino borrowers were routed into high-cost and subprime products before 2008 and then foreclosed at more than twice the white rate, a sorting whose aftermath still shapes who carries a thin or damaged credit file today ^[10].

We sit inside that disagreement rather than resolving it. The rate spread is an incidence flag, not a decomposition. We can report that a Black borrower in Cook County was about seven times more likely than a white borrower to walk away with a higher-priced first lien, and that this held through a full rate cycle. We cannot, from the public file, attribute that to points versus offers versus unobserved risk, because the score and the closing detail that would separate those channels are not in front of us. The number sizes the pricing gap among the approved. It does not explain it. What it does establish, against the intuition that screening should equalize the survivors, is that the disparity among funded borrowers is the largest one in the file, not the smallest.

Who lends, and the geography that did not hold

Two structural questions close the anatomy, namely who is doing the lending and where the money comes to rest. Both are answerable from the file. One returns the answer we planned for. The other returns a flat no.

Start with concentration. Ranking Cook County lenders by originated dollars from 2018 to 2023, the market is narrow at the top. Among the top 25 lenders by origination dollars, the top 10 hold 68.2 percent of those dollars ^[1]. Inside that set the racial tilt is steep. Across the top 15 lenders the median ratio of white-applicant origination dollars to Black-applicant origination dollars is 10.9 to 1, and four of those lenders run above 17 to 1, with the most lopsided at 29.5 to 1 ^[1]. The largest lender in the set placed 14.4 billion dollars with white applicants against 1.3 billion with Black applicants over the period ^[1]. This is a descriptive ratio, not an allocation rule, and it reflects who applied to which institution as much as how any institution decided. As a shape, though, it lines up with what WBEZ and City Bureau found when they read Chicago HMDA directly, that lenders put 8.4 times more home-purchase money into white neighborhoods than Black ones, with a single white neighborhood drawing more lending than all the majority-Black neighborhoods combined ^[11]. Our reproducible top-lender ratio is a different cut of the same data and lands in the same register, which is the point of computing it. The two numbers are not the same statistic, ours by lender and theirs by neighborhood, but they describe one market, and they agree on its direction and rough magnitude.

The spread across lenders is as telling as the median. The ratios are not bunched. They run from below 3 to 1 at a couple of institutions up to nearly 30 to 1 at one, which means the white tilt is not a uniform property of the market so much as a feature that some lenders carry far more heavily than others. A median of 10.9 with that kind of range says the typical large Cook County lender placed roughly eleven dollars with white applicants for every dollar with Black applicants, and that a handful placed far more skewed books than that. We are careful about what the ratio can and cannot mean. It is built from originated dollars, so it mixes together how many loans an institution made, how large those loans were, and who walked in the door to apply, and the file does not let us separate a lender that turned Black applicants away from one that simply drew few Black applicants because of where it markets and operates. A high ratio is a description of an institution's book, not a finding about its conduct. What the concentration figures establish, taken together, is structural rather than individual. A large share of Chicago's mortgage credit flows through a short list of institutions whose lending, in dollar terms, is heavily weighted toward white borrowers, which is the same market shape the local reporting found from the neighborhood side ^[11].

Now the geography, delivered as the null it is. The plan was to overlay modern lending on the 1938 redlining map and show the money concentrating where the appraisers once drew the favored grades, the way the national NCRC work would predict ^[8]. We assigned each Cook County tract to a 1938 grade by taking the tract centroid and testing which digitized HOLC zone, if any, it falls inside. Of 683 graded zones and 1,332 tract centroids, 766 centroids landed inside a graded zone, which is 57.5 percent of Cook tracts, with the rest outside any 1938 boundary ^[1]. Within the tracts we could grade, modern denial rates come out essentially flat. Formerly A and B tracts denied at 18.8 percent, formerly C at 20.9 percent, and formerly D, the redlined grade, at 20.8 percent ^[1]. The D-to-A/B ratio is 1.11, a hair above parity and nothing like the doubling we see across racial groups. Dollars per application tell the same non-story, landing at 308,075 dollars in the A/B tracts and 308,882 in the D tracts, a ratio of 1.00 ^[1]. The 1938 grade does not predict the modern denial rate, so we do not chart it as if it did.

One signal does survive the overlay, and it is demographic rather than financial. Tracts the appraisers marked D in 1938 are 26.3 percent Black-applicant today, against 17.6 percent in the formerly A/B tracts ^[1]. The old line still sorts who lives where and who applies from where. It no longer sorts the denial rate at the tract level once you are looking at applications. The redlining map, in other words, still shapes the demography of the modern application pool while having gone quiet on the lending decision itself, and the disparity we measure rides with the applicant's race rather than with the grade the area carried eighty years ago.

This has to be set carefully against the redlining literature, because a null here is not a refutation there. Aaronson, Hartley, and Mazumder use the boundaries of the HOLC maps to show that the 1930s grades caused lasting reductions in homeownership and house values and lasting increases in segregation through constrained credit ^[7]. Mitchell and colleagues, reading HMDA since 1981, find the average redlined neighborhood received about 3,000 fewer mortgages than a Best-graded one over four decades, net of housing-availability controls ^[8]. Those are volume-and-homeownership findings measured over long horizons. Ours is a tract-level denial-rate and dollars-per-application measure pooled over six recent years. A flat denial rate across grades does not overturn a deficit in cumulative lending volume or a gap in homeownership, because they are not the same quantity. A neighborhood can be denied at the county-average rate today and still sit on two generations of foregone mortgages and suppressed equity, which is exactly the cumulative quantity the NCRC measured and we did not. What our slice says, and only what it says, is that in Cook County from 2018 to 2023 the modern denial disparity rides with the applicant's race, not with the grade the area carried in 1938. We report that as a null because it is one, and because reporting it any other way would be the kind of overclaim this paper is built to avoid.

How much this can and cannot say

Every number in this paper sits behind one binding limit, and it is worth stating once, cleanly. Public HMDA does not contain the applicant's credit score, the single variable underwriters weight most heavily. With that field absent, no result here isolates present-day discrimination, and every result is descriptive. This is the reading the recent Federal Reserve work asks for. Bhutta, Hizmo, and Ringo find that observable applicant risk factors explain most of the racial denial disparity and that the residual differential-treatment gap is on the order of one to two percentage points once those factors are in the model ^[4]. A raw gap is therefore an upper bound on what differential treatment could account for, not a measurement of it, and the closer one reads, the more of the raw gap a fuller file tends to absorb.

The secondary limits shape specific numbers above, and each one is worth naming. Race and ethnicity here are mutually exclusive, with anyone Hispanic by derived ethnicity counted as Latino and everyone else placed by derived race, mirroring Federal Reserve practice. A large block of applications, 143,856 of the pooled credit decisions, carries no available race and is held out of the four main groups and reported on its own line, denied at 22.5 percent ^[1]. Our denominator is credit-decision applications, originated, approved-not-accepted, and denied, which means withdrawn files, incomplete files, and purchased loans are excluded and preapproval-only actions are not counted as denials, so our rates run slightly below a naive denied-over-everything figure. Reported income is missing for roughly a third of records, shown as an explicit unknown band and never imputed. Debt-to-income arrives as a mix of bucket strings and raw integers and is about 35 percent missing, which is why we declined to use it as a primary control even though it would otherwise be the natural second axis. The higher-priced share is restricted to originated first-lien loans carrying a reported rate spread, leaving out the roughly 11 percent of originations exempt from that reporting. And the HOLC overlay reaches only 57.5 percent of Cook tracts by the centroid test, which returns a null we report as a null. The centroid method is itself coarse. A tract whose center falls just outside a 1938 boundary is dropped even if much of its area lay inside a graded zone, and a tract is assigned a single grade even where the historical map split it, so the overlay is best read as a first pass rather than the last word on the geography. A finer areal-overlap method might recover some of the dropped tracts, though nothing about the flat denial rates across the grades we could assign suggests the larger picture would change.

One point about reproducibility is worth keeping in mind, since it is the main thing a descriptive paper like this one offers in place of a causal claim. The curated Cook County file and the code that builds every table are posted, so a reader is not asked to take our aggregations on trust ^[1]. The numbers can be rerun, the denominator can be redrawn, the income bands can be recut, and the geography can be widened from Cook County to the six-county metro by adding the collar-county FIPS codes. What a reader cannot do with the posted file is add the credit score, because the public release does not contain it, and that is the one change that would convert the descriptive gaps here into the conditional estimates the Federal Reserve work produces ^[4]. Reproducibility extends exactly as far as the public file does and stops at the same wall we stop at.

A word about what those limits do and do not threaten. They are limits of scope, not signs that the headline numbers are fragile. The denial gap, the cross-year persistence, the income result, and the pricing gap all rest on hundreds of thousands of records each and all point the same way, and none of them depends on the missing-data categories we set aside. What the limits foreclose is the next question, not this one. They stop us from converting a measured gap into a discrimination estimate, and they stop us from claiming the 1938 map drives the modern decision. They do not put the size and shape of the raw disparity in doubt. That distinction is the whole posture of the paper. We are confident about what we measured and silent about what we could not.

Two older guides keep this honest. Munnell and colleagues showed three decades ago that adding thirty-eight omitted creditworthiness variables to a raw Boston HMDA model cut the minority denial gap from roughly 2.7 times to about 1.6 times, with race still mattering but the raw figure plainly overstating the conditioned one ^[2]. That is the mechanism we are missing, running in the direction it always runs. Avery, Brevoort, and Canner, writing as Fed economists who built much of the HMDA analytic practice, are blunt about reading the file past its design. It documents the pattern of lending and was never meant to settle the question of cause ^[12]. We take that instruction at face value. The disciplined posture is to claim less, which here means we have measured the size and the shape of a raw disparity. That is a real and useful thing to have measured, and we stop there.

What is left standing

Read in one pass, the file traces a single application through a pipeline. It enters at a raw denial rate roughly double the white rate. It meets the only serious control the public data allows, coarse income bands, and the gap barely gives, holding at 17.7 percentage points for Black applicants even in the top income band ^[1]. If it is denied, the stated reason points more often at credit history for Black applicants, the one channel the file flags and cannot show, since the score behind the code is absent. If it is approved, it faces a pricing gap wider than the denial gap, with Black borrowers about seven times as likely to land a higher-priced first lien ^[1]. All of this happens inside a lending market concentrated in a handful of institutions and tilted heavily toward white borrowers ^[1], mapped onto a geography where the 1938 grade line has gone quiet on denial and the applicant's race has not.

Two endpoints deserve to be held without softening either. The 1938 redlining boundary is no longer the fault line in Chicago denial data, flat across grades in our six-year pool, even as it still sorts who applies from where ^[1]. And the missing credit score means the race gap that survives crude controls is not a discrimination estimate, only the residual we can see with the field that matters most removed. Both of those are uncomfortable for a tidy narrative, and both are what the data say. The temptation in this kind of work is to make the old map the villain and the modern gap its direct inheritance, a clean line from 1938 to 2023. The file will not support that line at the level of the lending decision, and saying so is part of the job.

What stands across the literature and our numbers together is narrower than the raw figures and more durable than any single year. A disparity of consistent direction and large size runs through Chicago mortgage lending, and crude observable controls do not dissolve it. Lewis-Faupel and Tenev and the Federal Reserve work locate that surviving difference partly in unobserved risk and partly in a stubborn residual that conditioning does not erase ^[9][4]. We cannot split those two with the public file, and we have not pretended to. The slice is reproducible, and the files are posted so any reader can rerun it for a different year, a different denominator, or a different geography ^[1]. The analysis that would answer the question this one raises, the one that includes the credit score, is the one the public file forecloses, and naming that gap is itself part of the accounting. Visible numbers are where accountability starts. They are not where it ends.