The 2015 HMDA rule sent credit scores and underwriting results to regulators and left the public view open

The CFPB’s October 2015 Regulation C rule added credit scores, automated underwriting results and loan identifiers to lenders’ reports and left public release to a later balancing test. Under 2018 guidance those fields stay out of public data, and the public schema omits them.

Looking Back revisits past weeks one at a time, up to the most recent, and asks what it meant for the banks, health systems and public agencies we serve. The series was written in 2026, with the benefit of hindsight, and each article says which week it looks back on.

Two-minute summaryThe dek and the first sentence of each section, quoted from the article

The CFPB’s October 2015 Regulation C rule added credit scores, automated underwriting results and loan identifiers to lenders’ reports and left public release to a later balancing test. Under 2018 guidance those fields stay out of public data, and the public schema omits them.

  1. The week The rule text carries its own dates.
  2. The record: fields that came from models and systems The rule’s list of reportable data runs to 38 numbered items in section 1003.4(a). [2] Several were familiar, such as loan amount, action taken and property location.
  3. The identifiers: keys that made records linkable Commenters had raised the risk of typing errors in a long identifier, and one noted that manual input might be needed to keep the information accurate. [2] The check digit was the Bureau’s answer.
  4. The audiences: what regulators receive and what the public sees Before the rule, institutions had to make their loan/application registers available to the public, with certain fields redacted to preserve privacy. [2] The rule removed that requirement.
  5. The perimeter: who reports and how often The release put the effect in one number.
  6. How the rule reached different institutions A large bank above the quarterly threshold faced every part of the rule at once: new fields, new identifiers and a reporting cadence measured in quarters. [2] Where its loans come through several origination channels, the rule’s requirement to report the result of the automated underwriting system used meant capturing those results across each channel, in the same structure.
  7. Objections, in order The objection is that the rule told lenders what to report and nothing about how to run their data.
  8. Checks a data or compliance leader could run These checks are our suggestions, not requirements of Regulation C. They apply to any dataset that includes model outputs sent to a regulator.
  9. With hindsight The privacy decision the 2015 rule deferred came in December 2018.

The week

On October 15, 2015, the Consumer Financial Protection Bureau announced that it had finalized a rule updating the reporting requirements of the Home Mortgage Disclosure Act regulation, Regulation C. [1] The release said the rule would shed more light on consumers’ access to mortgage credit, and that the Bureau was working with other federal agencies to streamline how lenders submit the data. [1]

The rule text carries its own dates. It was signed on October 13, 2015, and published in the Federal Register on October 28, 2015, at 80 FR 66128. [2] Its stated effective date was January 1, 2018, with some amendments effective on January 1, 2017, others on January 1, 2019, and one set on January 1, 2020. [2]

The release set out the background. It said HMDA, originally enacted in 1975, requires many lenders to report information about the home loans for which they receive applications or that they originate or purchase. [1] It said that in 2014, 7,062 financial institutions reported information about approximately 11.9 million mortgage applications, preapprovals and loans. [1] It said the Dodd-Frank Act directed the Bureau to expand the dataset, and that the existing data did not give adequate information about loan features such as adjustable-rate and non-amortizing loans. [1]

The rule described what it changed in three groups. [2] The first was coverage. It adopted uniform loan-volume thresholds for banks, credit unions and nondepository lenders. An institution had to report if it originated at least 25 closed-end mortgage loans or at least 100 open-end lines of credit in each of the two preceding calendar years. It also had to meet the rule’s other coverage criteria. [2] It adopted a dwelling-secured standard for consumer-purpose loans and lines of credit, so that closed-end home-equity loans, home-equity lines of credit and reverse mortgages were covered. [2]

The second group was the data itself. The rule grouped the required data into four categories: information about applicants, borrowers and the underwriting process; about the property; about the features of the loan; and certain unique identifiers. [2] Its examples for the first category were age, credit score, debt-to-income ratio and automated underwriting system results. [2] Its examples of identifiers were a universal loan identifier, the property address, a loan originator identifier and a legal entity identifier for the institution. [2]

The rule also said what it left out. It did not adopt the proposed requirements to report qualified mortgage status or the initial draw on an open-end line of credit. [2] The release added that the final rule dropped the proposed “risk-adjusted, pre-discounted interest rate” and the proposal to require reporting of all dwelling-secured commercial-purpose transactions. [1]

The third group was reporting and disclosure. Institutions would still submit their data by March 1 for the preceding calendar year. [2] Institutions that report large volumes would also submit data quarterly for the first three quarters of the following year. [2] The rule set that threshold at 60,000 transactions in the preceding calendar year, excluding purchased loans, and said that on 2013 data, 29 institutions would have met it. [2]

The rule removed the requirement that each institution provide its own disclosure statement and modified loan/application register to the public. [2] Instead, institutions would give a written notice that the information was available on the Bureau’s website. [2] On what the public would see, the Bureau adopted the balancing test it had proposed in 2014. The test would decide whether and how HMDA data should be modified before public release to protect privacy. [2]

The release described the burden side. It said a new standardized threshold would reduce the number of banks and credit unions required to report by an estimated 22 percent. [1] It said many amended requirements aligned with industry data standards, and that the Bureau had completed a pilot of a new web-based tool for collecting the data. [1] It said most provisions would take effect on January 1, 2018, that lenders would collect the new information in 2018, and that they would report it by March 1, 2019. [1]

The release closed on privacy. It said the Bureau was looking at ways to improve public access to HMDA data modified to protect applicant and borrower privacy. It said the Bureau was exploring strategies and techniques to address both privacy and market transparency. [1] The rule had settled what lenders would collect and send. It had not yet settled what of that the public would see.

The October 2015 HMDA rule is remembered for its length and its new data points. For a technology or data leader at a lender, its more useful feature is structural. It separated three things that the old regime had tied together: what a lender collects, what it reports to its regulator, and what the public sees. [2]

Our argument is that the rule moved algorithmic outputs into the regulator’s dataset while leaving the public dataset to a later decision. Credit scores, the name and version of the scoring model, the automated underwriting system used and its result all became reportable fields. [2] Whether any of them would be published was handed to a balancing test the Bureau would apply afterwards. [2] In our reading, that split turned HMDA from a disclosure exercise into a data governance problem with two audiences.

That framing matters because the work it created sat with data owners. A field such as the automated underwriting result does not live in a compliance spreadsheet. It comes out of a loan origination system, an interface to a government-sponsored enterprise’s or federal insurer’s tool, and the records of the person who ran it. [2] The rule made those systems part of the reporting chain.

We take the rule in four parts, each a different layer of the data it created:

  1. The record: which fields a covered loan file now had to produce, and where they came from.
  2. The identifiers: the keys that made each record linkable to others.
  3. The audiences: the split between what regulators receive and what the public sees.
  4. The perimeter: who has to report, and how often.

The parts do not overlap. The first is content, the second linkage, the third disclosure and the fourth scope. After them we look at how the rule reached different institutions, take the objections in order and end with checks a data or compliance leader could run.

The record: fields that came from models and systems

What was added

The rule’s list of reportable data runs to 38 numbered items in section 1003.4(a). [2] Several were familiar, such as loan amount, action taken and property location. Others were new and came straight from underwriting. Item 15 required, except for purchased loans, the credit score or scores relied on in making the credit decision and the name and version of the scoring model used to generate each score. [2]

Item 35 required, again except for purchased loans, the name of the automated underwriting system the institution used to evaluate the application and the result that system generated. [2] The rule defined an automated underwriting system narrowly. It meant an electronic tool developed by a securitizer, a federal government insurer or a federal government guarantor. The tool provides a result on the applicant’s credit risk and on whether the loan is eligible to be originated, purchased, insured or guaranteed by that entity. [2]

Item 16 required the principal reason or reasons the institution denied the application, if applicable. [2] Other added fields described pricing and terms: total loan costs or points and fees, interest rate, loan term, introductory rate period and non-amortizing features. [2] The release described the purpose as screening for possible fair lending problems and helping institutions and regulators focus on the riskiest areas. [1]

Why the source of a field matters

Each of these fields has an origin, and the origins differ. A debt-to-income ratio is a calculation, and its inputs live in the underwriting file. A credit score is an output of a model the lender licenses and does not build. An automated underwriting result is the output of a model another institution built, returned through an interface. [2]

For a data team, that difference sets the work. A field calculated in-house needs a documented formula and a record of which inputs were “relied on”, the rule’s own phrase for credit scores. [2] A field returned by another party’s system needs a captured response, a timestamp and a way to show which run of the system produced the result reported. A field that names a model version needs a version history.

In our reading, the rule did not require any of those artifacts by name. It required the reported values to be accurate. Accuracy, for a field produced by a model, depends on knowing which model ran, on which inputs, at which point in the file. The record the rule created was a record of machine decisions, and the lender had to be able to stand behind it.

The fields it declined

The rule’s omissions show where the Bureau drew a line. It did not require qualified mortgage status, and it did not require the initial draw on an open-end line of credit. [2] The release added the proposed risk-adjusted, pre-discounted interest rate to the list of dropped fields. [1]

Each dropped field would have been another derived value with its own lineage. We do not infer from the record opened here why each was dropped beyond what the release says: that changes were made after considering public comments. [1] The practical effect was a shorter list of derived values to trace, which still included the two model outputs.

Definitions borrowed from other law

The rule did not invent its own meaning of credit score. It said the term has the meaning set forth in 15 U.S.C. 1681g(f)(2)(A), a provision of the Fair Credit Reporting Act. [2] That choice tied a HMDA field to a definition that already governed disclosures lenders make under credit reporting law.

For a data team, a borrowed definition is useful and constraining. It is useful because the field means the same thing in two regimes, so one captured value can serve both. It is constraining because a lender cannot redefine the field to match what its systems happen to store. If a system holds an internal risk grade rather than a score within that definition, the grade is not the reportable value.

The automated underwriting definition worked the same way. By limiting it to tools developed by a securitizer or a federal government insurer or guarantor, the rule excluded a lender’s own internal decision engines from that field. [2] In our reading, a lender running its own model alongside an agency tool reported the agency tool’s result, and its own model’s output did not appear in the field at all. The regulator’s copy captured the external model, not the internal one.

What an examiner could ask to see

The rule did not list the evidence an examiner would request. It is reasonable to expect, in our reading, that a reported field invites a question about its support. For the model-output fields, the support would be the captured score report or underwriting response, the version identifier and the record of who acted on it.

A practical test follows. Pick a sample of reported applications and try to reproduce each credit score and underwriting result from stored evidence alone, without asking the loan officer. Where the evidence is a screenshot or a note, the field is only as reliable as the person who made it. Where it is a stored system response tied to the loan identifier, the field can be checked by anyone. This test is our suggestion, not a requirement of the rule.

Demographic data and its collection method

The rule also changed how ethnicity, race and sex were captured. When an application was taken in person and the applicant did not provide the information, institutions had to report whether it was collected on the basis of visual observation or surname. [2] When applicants provided it themselves, institutions had to let them self-identify using disaggregated ethnic and racial categories. [2]

That is a data provenance requirement in plain form. The value is not enough; the rule wanted to know how it was obtained. In our reading, it set a pattern that the model-output fields followed less explicitly: a reported value carries an implied claim about its source.

The identifiers: keys that made records linkable

A loan identifier with a check digit

The rule required a universal loan identifier for each covered loan or application. [2] The Bureau added a check digit to it, after a commenter suggested one. [2] The rule explained that a check digit is generated by applying a mathematical function to a sequence of characters, so the sequence can be validated at a later point in the process. [2]

Commenters had raised the risk of typing errors in a long identifier, and one noted that manual input might be needed to keep the information accurate. [2] The check digit was the Bureau’s answer. For a systems team, it meant the identifier had to be generated once, validated, carried through every system that touched the loan and never retyped.

Identifiers for the institution and the originator

The rule also required a legal entity identifier for the institution and a loan originator identifier. [2] The Bureau’s later documents describe that field as the unique identifier assigned by the Nationwide Mortgage Licensing System and Registry for the mortgage loan originator. [3] [4] Together with the property address, these fields linked each record to a specific loan, a specific lender and a specific person who originated it. [2]

Linkage is what makes a dataset useful to an examiner. It is also what makes it sensitive. A record keyed to a property address and an originator can be joined to other records about the same property and the same person. The rule’s identifiers raised the value of the dataset to supervisors and, in the same move, raised the stakes of releasing it.

What linkage asks of a lender

The engineering consequences were concrete. A lender needed one place that minted the loan identifier and a rule for when it was assigned. It needed the institution identifier stored as reference data. It needed the originator identifier validated against licensing records, and a policy for which originator to report where several handled a file.

The rule spoke to that last point. It said an institution that establishes and follows a reasonable, written policy for determining which individual originator has primary responsibility for the reported transaction complies with the requirement. [2] In our reading, the Bureau accepted a documented decision rule in place of a single right answer. That is a governance pattern, and it recurs: write the rule down, apply it consistently, and the field is defensible.

The audiences: what regulators receive and what the public sees

Moving disclosure to the agencies

Before the rule, institutions had to make their loan/application registers available to the public, with certain fields redacted to preserve privacy. [2] The rule removed that requirement. Institutions would instead post a notice that their register, as modified by the Bureau to protect privacy, could be obtained on the Bureau’s website. [2]

The rule described the change directly: disclosure of HMDA data shifted entirely to the agencies. [2] Institutions would no longer provide their HMDA data directly to the public, only a notice saying where to get it. [2] For a lender, that removed a publication task. It also removed control over the public form of its own data.

The balancing test

The rule explained the legal basis. The Dodd-Frank amendments added a provision directing the Bureau to develop regulations that modify or require modification of itemized information for the purpose of protecting applicant and borrower privacy. [2] The Bureau interpreted the statute to call for a balancing test, weighing privacy risks against the benefits of disclosure in light of HMDA’s purposes. [2]

The rule recounted the proposal’s view on which the Bureau had sought comment, and the final rule stated the same view. Privacy interests arise under the test only where disclosure may both substantially facilitate identifying an applicant or borrower in the data and disclose information that is not otherwise public and may be harmful or sensitive. [2] The rule said that, based on its analysis to date, the Bureau believed some of the proposed new data points may create privacy concerns sufficient to warrant some degree of modification, including redaction, before public disclosure. [2]

What the rule did not do was decide which fields would be released, in what form. That decision was left for later. [2] In our reading, this is the center of the rule for data leaders. The Bureau collected fields whose public release it had not yet decided, and said so.

Two audiences, two risk profiles

The split created two copies of the same data with different exposure. The regulator’s copy held everything reported, including credit scores and model results. [2] The public copy would hold whatever survived the balancing test. [2]

In our reading, a lender’s risk lay mostly in the first copy. The release said the new underwriting and pricing data would help both institutions and regulators focus attention on the riskiest areas for fair lending problems. [1] A wrong credit score or a mislabeled underwriting result would show up there, whatever the public saw. In our reading, the split made accuracy of model-derived fields a supervisory matter first and a public one second.

Why the timing mattered

The rule took effect for most provisions on January 1, 2018. [2] The privacy decision had no date in the rule. That left lenders building pipelines for fields whose public fate was open. A lender deciding how much to invest in, say, validating automated underwriting results had to assume the regulator would use them, and could not assume the public would not see them.

That uncertainty is a design input. When the audience for a dataset is not yet fixed, the safer engineering assumption is the widest plausible audience. In our reading, the 2015 rule asked lenders to build to that assumption until the Bureau decided.

A dataset designed for comparison

The release made clear what the regulator’s copy was for. It said the new underwriting and pricing information would enhance the ability to screen for possible fair lending problems. It said the data would also help the Bureau and others monitor multifamily, affordable and manufactured housing. [1] Screening and monitoring are comparative activities. They look for differences between lenders, places and groups of borrowers.

A comparative dataset punishes inconsistency more than a descriptive one. If one lender records the debt-to-income ratio relied on and another records a recalculated value, the difference between them is an artifact that a screening method may read as a pattern. In our reading, the rule’s precision about “relied on” values and its borrowed definitions were attempts to make lenders’ fields comparable. The burden of that comparability fell on each lender’s data definitions.

Free text as a risk

Some values could not be reduced to codes. The later guidance referred to free-form text fields used to report several values. They included race, ethnicity, the scoring model’s name and version, the principal reasons for denial and the underwriting system name. [4] Free text is where identifying details can leak into a dataset by accident, which is why, in our reading, the guidance named those fields for exclusion from public release wholesale.

For a lender, free text in a regulatory file deserves the same review as free text in any other shared dataset. A denial reason typed by hand can contain a name, an account number or a description of a health condition. The regulator’s copy would hold it. A data team that scans its HMDA file for unexpected personal data before submission treats the file as what it became in 2015: a dataset with two audiences.

The perimeter: who reports and how often

Thresholds by volume

The rule replaced a patchwork of coverage tests with uniform loan-volume thresholds: 25 closed-end loans or 100 open-end lines of credit in each of the two preceding calendar years. [2] It kept the existing depository criteria on assets, location in a metropolitan statistical area, the federally related test and at least one first-lien home purchase loan or refinancing. [2] For nondepository institutions, it replaced the old tests with the loan-volume thresholds and kept the location test. [2]

The release put the effect in one number. It said the new threshold would reduce the number of banks and credit unions required to report by an estimated 22 percent. [1] It said those lenders receive a low volume of applications and originate a low volume of loans, so the change would not compromise the usefulness of the dataset. [1]

Two tests that do not borrow from each other

The rule kept closed-end and open-end coverage separate. An institution meeting only the 25 closed-end threshold did not have to report its open-end lending. One meeting only the 100 open-end threshold did not have to report its closed-end lending. [2] For a lender, that meant two counts, two answers and the possibility of reporting one product line and not the other.

Transactional coverage changed alongside. The rule covered most consumer-purpose transactions secured by a dwelling, including closed-end home-equity loans, home-equity lines of credit and reverse mortgages. [2] It covered commercial-purpose loans only if they were for home purchase, home improvement or refinancing, and it excluded home improvement loans not secured by a dwelling and all agricultural-purpose loans and lines of credit. [2]

Those definitions are business rules a system has to apply. A home-equity line booked in a consumer lending platform separate from the mortgage platform still had to be counted and, if covered, reported. In our reading, the rule’s coverage changes reached systems that had never fed HMDA reporting before, and that is where data gaps were likely to appear.

Quarterly reporting for the largest

The rule added quarterly submission for institutions with at least 60,000 covered transactions, excluding purchased loans, in the preceding calendar year. [2] The Bureau lowered the threshold from the 75,000 it had proposed and said that, on 2013 data, 29 institutions would have reported quarterly in 2014, and that those institutions reported approximately 49 percent of all transactions reported under HMDA in 2013. [2]

Quarterly reporting changes the engineering problem. Annual reporting can tolerate a long cleanup at year end. Quarterly reporting needs data that is correct close to the time the loan is acted on. The quarterly requirement, effective January 1, 2020, required submission “within 60 calendar days after the end of each calendar quarter except the fourth quarter.” [2]

The perimeter as a variable

The thresholds were stated as numbers in the rule, and numbers can change. The rule fixed them; it did not make them permanent. As the hindsight section shows, the thresholds were a part of the rule that moved after 2015. [5] [6] [7] For a lender near a threshold, that meant coverage was a recurring question, answered from its own origination counts each year.

A new submission channel

The release described the Bureau working with the other members of the Federal Financial Institutions Examination Council and the Department of Housing and Urban Development to modernize data submission. [1] It said the Bureau had completed a pilot of a web-based collection tool and that implementing the technology would reduce manual and paper-based systems used by regulators and reporting institutions. [1]

A web-based submission channel moves validation earlier. Errors that once surfaced in a regulator’s processing can be returned at the point of filing. For a lender, that is welcome and demanding. It shortens the cycle between submitting and correcting, and it makes the quality of the file visible to the filer before anyone else sees it. In our reading, the channel and the check digit worked together: one validated the file, the other validated each loan’s key.

How the rule reached different institutions

Large banks

A large bank above the quarterly threshold faced every part of the rule at once: new fields, new identifiers and a reporting cadence measured in quarters. [2] Where its loans come through several origination channels, the rule’s requirement to report the result of the automated underwriting system used meant capturing those results across each channel, in the same structure.

For such a bank, the rule also raised the value of internal consistency. In our reading, its regulator could compare its fields with those of peers. A scoring model version recorded one way in one channel and another way elsewhere would read as noise at best and as a pattern at worst.

Community banks and credit unions

A community bank or credit union near the 25-loan threshold faced a different calculation. [2] The release said small depository institutions with low loan volume would no longer have to report. [1] One just above the line would report the full set, with fewer staff and less automation.

For that institution, the burden was less in the systems than in the people. A small lender may take applications on paper or through a vendor’s portal. Each reported credit score or underwriting result had to be traced to a file. In our reading, the rule’s alignment with industry data standards, which the release emphasized, helped where a vendor already captured those fields. [1]

Nondepository lenders

Nondepository mortgage lenders lost the old asset-size and volume tests and were covered by the same loan-volume thresholds as banks. [2] The rule’s background noted that depository institutions report HMDA data to their supervisory agencies, while nondepository institutions report to the Department of Housing and Urban Development. [2]

For a nonbank lender, the record fields came from the same tools a bank uses: credit scores from bureaus and results from the same government-sponsored or federal systems. [2] The difference was in who read the data and how its examination program used it. The rule did not change that; it changed what was in the file.

Institutions that buy loans

Purchased loans were carved out of several new fields, including credit score and automated underwriting result. [2] They were also excluded from the quarterly threshold count. [2] An aggregator that buys loans reported a narrower record for those loans. In our reading, the carve-outs recognized that the buyer did not run the underwriting and could not attest to its outputs.

Vendors that sit in the chain

Every model-output field passes through at least one party other than the lender. The credit score comes from a consumer reporting agency’s product. The underwriting result comes from a tool built by a securitizer, insurer or guarantor. [2] A lender may also run origination on a vendor’s platform, which captures and stores both.

The rule placed the reporting duty on the financial institution. [2] It did not move any of it to the vendors whose systems produced or held the values. In our reading, that made vendor contracts part of HMDA compliance. A lender whose platform vendor could not export the scoring model version, or could not retain the underwriting response tied to the loan, had a reporting gap it could not close alone.

The contractual fixes are ordinary. They include a data dictionary that maps platform fields to the regulation’s items and a commitment to retain source responses as long as the lender needs them. They also include notice before the platform changes how a field is captured. The rule mentions them nowhere. Each follows from the fact that the rule’s fields are produced outside the compliance function.

Objections, in order

“HMDA is a disclosure law, not a data governance law”

The objection is that the rule told lenders what to report and nothing about how to run their data. That is accurate as to the text. [2] It misses what reporting a model output requires. A lender that reports a credit score “relied on in making the credit decision” must know which score that was. [2] The rule did not mandate lineage. It made lineage the only way to be confident the field was right.

“The public would never see the sensitive fields”

In October 2015, nobody could say that from the rule. The Bureau said some proposed new data points may create privacy concerns and adopted a balancing test without deciding outcomes. [2] A lender had to build assuming either result. As it turned out, the later guidance said several of those fields would be excluded from public data. [4] The objection is right in hindsight and was not available as a planning assumption.

“The thresholds made this a large-lender problem”

The release estimated the thresholds would reduce the number of banks and credit unions required to report by 22 percent. [1] They did not remove lenders making 25 closed-end loans a year. [2] For a credit union just above the line, the full record applied. In our reading, the rule spread new data duties well below the largest institutions, even as it trimmed the bottom.

“Aligning with industry standards made it easy”

The release said alignment with established industry data standards would mitigate burden for many lenders. [1] Alignment helps when the standard field is already captured and reliable. It does not help when the field exists in the standard but is filled inconsistently, or when a lender’s systems hold it in free text. The later guidance referred to free-form text fields used to report some values, including the scoring model’s name and the underwriting system’s name, which suggests in our reading that not everything mapped cleanly. [4]

“Public release was the Bureau’s problem, not the lender’s”

Formally, yes. The rule shifted disclosure entirely to the agencies and left each lender only a notice to post. [2] A lender no longer prepared a redacted register for the public. [2]

The objection overlooks reputation and accuracy. A lender’s public HMDA data would still carry its name and legal entity identifier, whatever modifications the Bureau applied. Errors in the reported fields would flow into the public copy in whatever form survived the test. The Bureau owned the redaction. The lender still owned the truth of what was redacted.

“Collecting more data creates its own privacy risk”

This is the strongest objection, and the rule partly concedes it. The Bureau said some proposed new data points may create privacy concerns. [2] Its answer was to manage the risk at the point of public release, through the balancing test, rather than at collection. [2]

That answer leaves the regulator’s copy, and the lender’s own copy, holding more sensitive data than before. The objection holds as to those copies. Where the argument still stands is on the governance point: once the data must be collected, the lender’s choice is how well it protects and traces it, not whether to hold it. The statute and the rule fixed the collection; the lender owned the controls.

Checks a data or compliance leader could run

These checks are our suggestions, not requirements of Regulation C. They apply to any dataset that includes model outputs sent to a regulator. Organizations should confirm their own obligations with counsel.

  • Lineage for model outputs: for each reported credit score and underwriting result, can the institution show which model or system produced it, which version, and when?
  • One source of identifiers: is the loan identifier minted once, validated by its check digit and never retyped downstream?
  • Written decision rules: where a field needs a choice, such as which originator had primary responsibility, is there a written policy, and is it applied the same way across channels?
  • Provenance of demographic data: does each record show how ethnicity, race and sex were obtained?
  • Threshold monitoring: does someone own the annual count that decides coverage, for both closed-end loans and open-end lines?
  • Vendor terms: do platform and data contracts commit the vendor to export each reportable field, retain the source responses and give notice before changing how a field is captured?
  • Free-text review: is every free-text field in the file scanned for personal details that do not belong there before submission?
  • Reproducibility sample: can a reviewer who did not work the file reproduce a sample of reported scores and underwriting results from stored evidence alone?
  • Two audiences: has the institution reviewed what it sends to its regulator separately from what the public dataset shows, and does its data risk assessment cover both?

The last check carries the rule’s structural lesson. The 2015 rule made the regulator’s copy and the public copy different objects. [2] A lender that governs only the copy it sees published has governed the smaller of the two.

The same structure now appears wherever a regulator collects model outputs it may not publish. A lender that built HMDA lineage well built a pattern it can reuse. Capture the output with its version and inputs, key it to a stable identifier, and assume the regulator will read it line by line.

With hindsight

The privacy decision the 2015 rule deferred came in December 2018. The Bureau released final policy guidance on its website on December 21, 2018, and published it in the Federal Register on January 31, 2019. [4] It applied to data compiled in or after 2018 and disclosed from 2019. [4]

The guidance said the Bureau intended to exclude several fields from the public loan-level data. They were the universal loan identifier, the application and action-taken dates, the property address, the credit score or scores relied on, the originator identifier and the underwriting system result. [4] It also named free-form text fields for exclusion, including the name and version of the credit scoring model, the principal reasons for denial and the automated underwriting system name. [4] It said the Bureau would reduce the precision of other fields, disclosing loan amount and property value as the midpoint of a $10,000 interval, binning age, and binning and top-coding the debt-to-income ratio. [4]

The guidance gave a reason for one exclusion that bears on data leaders. It said low credit scores and high debt-to-income ratios may suggest a borrower’s vulnerability to scams relating to debt relief or credit repair. [4] It called the guidance non-binding and said the Bureau would commence a notice-and-comment rulemaking on public disclosure in May 2019. [4]

As checked on November 16, 2026, the FFIEC HMDA documentation site publishes the public loan/application register schema. It lists fields for the applicant’s and co-applicant’s credit score type, fields named aus_1 through aus_5 and denial reason fields. [8] It lists no field for the credit score value, the automated underwriting result, the universal loan identifier or the property address. [8] The models’ outputs went to regulators. In public, by the schema’s field list, they remain out.

Congress narrowed who reports the full record. The Economic Growth, Regulatory Relief, and Consumer Protection Act was signed on May 24, 2018. It added partial exemptions for insured depository institutions and insured credit unions that originated fewer than 500 closed-end loans, or 500 open-end lines, in each of the two preceding calendar years. [3] The exemptions do not apply to an institution with certain negative Community Reinvestment Act examination ratings. [3]

The Bureau’s interpretive and procedural rule, effective September 7, 2018, said the partial exemptions cover 26 data points and that partially exempt transactions still require the remaining 22. [3] Its table of covered data points included the universal loan identifier, property address, credit score, reasons for denial and debt-to-income ratio. [3] It designated a non-universal loan identifier for institutions that choose not to report a universal one. [3]

The coverage thresholds moved more than once. In 2017, before they took effect, the Bureau temporarily raised the open-end threshold to 500 for 2018 and 2019, and in October 2019 extended that to January 1, 2022. [5] A rule published May 12, 2020 raised the closed-end threshold from 25 to 100, effective July 1, 2020, and set the permanent open-end threshold at 200 from January 1, 2022. [5]

Litigation reversed part of that. On September 23, 2022, the United States District Court for the District of Columbia vacated the 2020 rule as to the higher closed-end threshold. [6] The Bureau’s technical amendment, effective December 21, 2022, restored the 2015 threshold of 25 closed-end loans to the Code of Federal Regulations. [6] As checked on November 16, 2026, the eCFR text of 12 CFR 1003.2, up to date as of November 12, 2026, states thresholds of 25 closed-end mortgage loans and 200 open-end lines of credit. [7]

What held up from October 2015 is the split between audiences. The model-output and identifier fields the rule added to the regulator’s copy are, by the 2018 guidance and the current public schema, absent from the public copy. [4] [8] What did not hold up is the perimeter as written. The open-end threshold rose from 100 to 200, a statute exempted part of the record for smaller insured lenders, and the closed-end threshold went up and came back. [3] [5] [6] [7]

In our view, a durable lesson for a lender’s data leaders is the one the 2015 rule built into its structure. Model outputs that support a credit decision are records a regulator may read line by line. The public view of them can change by guidance; the supervisory view did not depend on it.

Spotted an error? Tell us and we will check it and correct the article where needed.

  • HMDA
  • Regulation C
  • CFPB
  • data governance
  • mortgage lending
  • privacy