Cash-flow underwriting got a blessing and no plumbing
Five regulators’ December 2019 statement favored cash flow data because consumers permission it, it comes from bank records and lenders can explain it. It expressly left out how that data moves, so lenders had to build permission, provenance, explanation and evidence themselves.
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
Five regulators’ December 2019 statement favored cash flow data because consumers permission it, it comes from bank records and lenders can explain it. It expressly left out how that data moves, so lenders had to build permission, provenance, explanation and evidence themselves.
- The week On December 3, 2019, five federal financial regulators issued a joint statement on the use of alternative data in credit underwriting.
- What the statement endorsed, and what it left out The document was an interagency statement, not a regulation, and it offered no safe harbor.
- Permission: consent the lender has to prove The statement said consumers can expressly permission access to their cash flow data, which enhances transparency and control.
- Provenance: the path from the bank to the model The statement called bank account records reliable sources.
- Explanation: specific reasons from transaction data The statement said creditors’ use of cash flow data can generally be explained and disclosed to the borrower, as ECOA and the Fair Credit Reporting Act may require.
- Evidence of lower risk: testing the claim The statement said that some alternative data, including cash flow data, may present no greater risks than traditional data, and that data directly related to consumers’ finances may present lower risks than other data.
- Where the argument is weakest One objection is that the agencies were clear.
- What to build first The four parts depend on each other.
- With hindsight Each of the four conditions the statement left to lenders has since been the subject of federal action, and much of that action has been withdrawn, reopened or reversed.
The week
On December 3, 2019, five federal financial regulators issued a joint statement on the use of alternative data in credit underwriting. The Federal Reserve Board, the Consumer Financial Protection Bureau, the Federal Deposit Insurance Corporation, the National Credit Union Administration and the Office of the Comptroller of the Currency signed it together. The agencies’ press release said it covered underwriting by banks, credit unions and non-bank financial firms. The statement itself addressed “bank and non-bank financial firms,” which it called “firms.”
The statement defined alternative data as information not typically found in the credit files of the nationwide consumer reporting agencies, or customarily provided by consumers on credit applications. It cited the CFPB’s February 2017 request for information on alternative data and modeling techniques, and the bureau’s May 2015 report on consumers with no credit record. A footnote limited the scope. The statement applied to the use of consumer data in the credit process, not to the furnishing, compilation or transfer of that data.
The agencies said alternative data may improve the speed and accuracy of credit decisions. They said it may help firms evaluate consumers who cannot obtain credit in the mainstream system, and may enable better pricing or terms. They noted that firms were also using such data in fraud detection, marketing, pricing, servicing and account management, though the statement focused on underwriting.
One kind of data got particular attention. The agencies said some firms were automating the use of cash flow data, derived from bank account records, to evaluate ability to repay. They gave four reasons it may present no greater risk than traditional data. Evaluating income and expenses is a well-established part of underwriting. Cash flow data are specific to the borrower and generally come from reliable sources such as bank account records. Consumers can expressly permission access to them. And creditors’ use of cash flow data can generally be explained and disclosed to the borrower, as the Equal Credit Opportunity Act and the Fair Credit Reporting Act may require.
The statement also described “Second Look” programs, in which firms use alternative data only for applicants who would otherwise be denied. It said such programs may improve credit opportunities and must still comply with consumer protection law. It reminded banks that alternative data use should be consistent with safe and sound operations, and it pointed to the agencies’ model risk management guidance from 2011 and 2017.
The compliance message was proportionate. A well-designed compliance management program should analyze the relevant consumer protection laws before a firm uses alternative data. Data that present greater consumer protection risks warrant more extensive compliance management, including testing, monitoring and controls. The agencies said they may offer further information as they learned more, invited feedback by email to each agency, and said firms may consult their regulators when planning such uses.
Two footnotes reached beyond consumer underwriting. One said appropriate data controls for banks include rigorous assessment of the quality and suitability of data to support prudent banking operations. Another said alternative data may also improve the speed, cost and accuracy of small business underwriting. For technology and data leaders, in our view, the footnote on scope mattered most, because it left the data supply chain for lenders to govern themselves.
The December 2019 statement read as encouragement, and it was. Five regulators said in one voice that alternative data, including cash flow data from bank accounts, may make credit decisions faster and more accurate. Less prominent was a footnote. The statement applied to how consumer data is used in the credit process. It did not apply to “the furnishing, compilation, or transfer of such data.”
That exclusion matters because each reason the agencies gave for comfort depended on the excluded part. Cash flow data were said to come from reliable sources, but reliability depends on how the data travels from the bank to the lender. Consumers were said to permission access, but permission is a record someone has to create, store and honor. Use was said to be explainable, but explanation depends on how raw transactions are turned into model inputs. And the lower-risk claim depended on evidence that the agencies did not supply.
Our argument is that the statement endorsed cash flow underwriting on four conditions it did not provide, and that a lender adopting it had to build each one. The four parts follow the statement’s own reasons, in its order. The first is permission: the consumer’s consent, its scope and its revocation. The second is provenance: how account data reaches the lender and how reliable it remains in transit. The third is explanation: turning cash flow features into the specific reasons a declined applicant must receive. The fourth is evidence of lower risk: the testing that shows the data treats applicants fairly and predicts repayment.
The parts do not overlap. Permission concerns the consumer’s choice, provenance the data’s path, explanation the individual decision and evidence the program’s overall outcomes. Together they cover what a lender controls once it decides to use cash flow data. A first section sets out what the statement endorsed and left out. A sixth section tests the argument against objections and against different kinds of lender. The last sets out what to build first. Lenders should confirm their obligations with counsel and their regulators.
What the statement endorsed, and what it left out
A statement, not a rule
The document was an interagency statement, not a regulation, and it offered no safe harbor. It described how five agencies viewed existing law as it applied to a new practice. That form has advantages: it can be issued quickly and signals supervisory attitude. Its limit is that it lacks the force of a rule, and it can be reinterpreted as personnel and priorities change.
The statement named the laws in play. Fair lending laws, prohibitions on unfair, deceptive or abusive acts or practices, and the Fair Credit Reporting Act applied as appropriate. Its footnotes pointed to adverse action notice requirements under Regulation B and the Fair Credit Reporting Act, and to model risk guidance for banks. None of those were new. The novelty was the signal that regulators saw cash flow data as a promising and relatively low-risk input.
Who the statement reached, and who would examine it
The statement addressed bank and non-bank firms in one text, and the agencies’ press release named credit unions among them. Each kind of lender answers to its own supervisor, and the five signatories included the three federal banking agencies, the credit union regulator and the CFPB.
That breadth meant the same practice would be judged by different examiners with different priorities. A bank could expect model risk reviewers to ask about validation. A non-bank lender could expect consumer protection examiners to focus on disclosures and complaints. The statement gave all of them the same starting text, and each would read it through its own exam procedures.
Proportionality as the organizing idea
The statement tied compliance effort to risk. It cited the Uniform Interagency Consumer Compliance Rating System, which says expectations for compliance management vary with risk and other factors. Data with greater consumer protection risk warranted more extensive testing, monitoring and controls.
Proportionality is sensible and shifts the burden of proof. A lender that claims its data is low risk must be able to show why. The statement offered reasons why cash flow data might be low risk, but each lender’s program still had to earn that rating with its own evidence.
Why cash flow data was singled out
The agencies said data directly related to consumers’ finances, such as cash flow data, may present lower risks than other data. Cash flow data describes income and expenses directly. The statement noted it may be especially useful for consumers who earn reliable income from several sources rather than a single job.
A footnote described what cash flow analysis can measure. It listed categories of income and expenses, including fixed costs such as housing and the size of variable spending. It also listed how an account has been managed over time, such as residual balances. Those are features a lender computes from raw transactions. The statement treated them as close to traditional underwriting, and in substance they are.
The excluded half
Raw transactions do not arrive at a lender by themselves. They can reach it in several ways. A data aggregator may connect to the consumer’s bank through an application programming interface. It may instead log in with the consumer’s own credentials and copy what the screen shows, a practice known as screen scraping. Or the applicant may upload statements directly.
Each path raises its own questions about consent, security, accuracy and liability. The statement set all of them aside. That was a reasonable choice for a short statement about underwriting. It also meant the agencies endorsed an output while saying nothing about the supply chain that produced it.
What the agencies asked for in return
The statement closed with an invitation. The agencies welcomed feedback by email and said firms may consult their regulators when planning alternative data uses. A lender planning a cash flow program could describe its design to its regulator in advance, and record the discussion. Such a record does not bind the regulator, but it shows the lender raised the questions early.
The agencies also said they may offer further information as their understanding grew. A lender reading that line in 2019 should have expected the guidance to move, and should have built controls that would survive a stricter reading.
Permission: consent the lender has to prove
What “expressly permission” requires in operation
The statement said consumers can expressly permission access to their cash flow data, which enhances transparency and control. In operation, that sentence describes a set of records. The lender needs to show what the consumer agreed to share, for what purpose, for how long and through which intermediary. It needs to show when the consumer agreed and how the request was presented.
Consent captured in an aggregator’s interface is not automatically the lender’s record. If the lender cannot retrieve the consent text and timestamp for a given applicant, it cannot show what was permitted. A complaint or an examination would then turn on documents held by a vendor.
Scope and duration
Underwriting needs a snapshot, which may be several months of transactions at the time of application. Some products ask for more. Servicing teams may want continuing access to monitor accounts. Fraud teams may want to re-check identity. Marketing teams may want to use the same connection for offers.
Each of those is a different purpose. A consumer who agreed to a one-time underwriting pull did not necessarily agree to monitoring. The statement noted that firms were using alternative data in servicing and marketing as well, and focused only on underwriting. A lender that reused an underwriting connection for other purposes was outside the comfort the statement offered.
What the consumer actually sees
Consent is only as good as the screen that obtains it. A consumer applying for a loan may move through a lender’s form, then an aggregator’s window, then the bank’s login page. Each step may carry its own terms. The consumer may not know that a third company is involved, or that it may keep the data.
A lender can improve this with design choices. Name the intermediary on the lender’s own page before the handoff. State the purpose, the accounts and the number of months requested. Say whether access ends after the decision. Test the flow with real applicants and record where they drop out. A high drop-out rate at the connection step is a fairness signal as well as a conversion problem.
Revocation
Permission that cannot be withdrawn is weak permission. A consumer may revoke access through the bank, the aggregator or the lender. The lender needs to know when access ends and must stop pulling data. It also needs a rule for data already collected: how long it is kept, for what purposes and when it is deleted.
Joint accounts and other people’s data
Bank account data can describe more than one person. A joint account shows a spouse’s pay and spending. A household account shows payments to and from relatives. An applicant can consent to share an account, but the other holder has not consented to the lender’s analysis of their transactions.
The statement did not address this. A lender can limit the problem by using derived features about the account as a whole rather than storing named counterparties, and by stating in its disclosure that shared accounts include others’ information. Where co-applicants apply together, it can obtain consent from each.
Controls a lender would put in place
- A consent record per applicant, stored by the lender, with the disclosure text, scope, duration and timestamp.
- A purpose tag on every data pull, so underwriting, servicing, fraud and marketing uses can be separated and audited.
- A revocation feed from the aggregator, tested regularly, that stops further pulls within a defined time.
- A retention schedule for raw transactions and derived features, set separately from loan file retention.
Provenance: the path from the bank to the model
Three ways the data arrived
The statement called bank account records reliable sources. The records are reliable at the bank. What the lender receives may differ. Screen scraping depends on page layouts that can change without notice, so fields can be missed or misread. Interfaces are more stable but vary in what they expose. Uploaded statements depend on document parsing and can be altered before upload.
A lender using cash flow data therefore inherits the error rate of its intermediary. A misclassified deposit can turn steady income into irregular income. A missed account can hide a loan payment. These errors flow straight into the features the model uses.
The aggregator as a critical third party
For a bank, an aggregator that supplies underwriting data is a third party supporting a critical activity. The agencies’ third-party risk expectations already applied. Due diligence would cover the aggregator’s security, its data quality controls, its subcontractors and its financial condition. Contract terms would cover data accuracy, incident notice, audit rights and exit.
Security deserves particular attention. Credential-based screen scraping, as the CFPB later described it, uses the consumer’s own login credentials. An aggregator that scrapes then holds credentials for many consumers at many banks. A breach at that point exposes accounts far beyond the lender’s own customers. The lender did not create that risk, but its product encouraged it.
Contract terms for an aggregator
- Accuracy commitments for transaction categorization, with a defined sampling method and remedies when error rates exceed agreed limits.
- Notice of changes to connection methods, data fields or source bank coverage before they take effect.
- Prompt notice of any security incident affecting consumer credentials or data, with cooperation in investigation.
- Limits on the aggregator’s own use of data collected for the lender’s applicants.
- Delivery of consent and revocation records to the lender in a usable form.
- Exit terms, including data return and deletion, and support during a move to another provider.
Concentration on a few intermediaries
A lender that connects through one aggregator shares it with every other lender that aggregator serves. A failure at the aggregator can stop cash flow underwriting at all of them at once. For a lender that has moved much of its thin-file volume onto cash flow data, that is an operational resilience question. A second provider, or a manual fallback, limits the damage.
When the source changes
A lender depends on banks it has no contract with. A source bank can change its website, alter its interface or restrict third-party access. Any of those can break connections for that bank’s customers overnight. The lender needs monitoring that detects a sudden drop in successful connections by source bank, and a fallback, such as statement upload, that keeps applicants moving.
Uploaded statements and altered documents
Statement upload avoids credential sharing and adds a different risk. Documents can be edited before upload, and parsing software can misread layouts. A lender that accepts uploads needs checks for tampering, such as comparing totals and balances across pages, and a review path when parsing confidence is low. It should also record which applicants used uploads, because the path an applicant used may affect outcomes and should be part of fairness testing.
Measuring data quality
A lender can measure provenance quality directly. Useful metrics include the share of applications where the connection fails, the share where fewer accounts or months arrive than requested, and the rate of transaction categorization errors found in sampled reviews. Each should be tracked by aggregator and by source bank, because problems can cluster at particular institutions.
The connection to the explanation
Provenance errors also affect fairness. If connection failures are more common at banks that serve particular communities, applicants from those communities may be pushed back to traditional underwriting or declined for incomplete data. That pattern would not show in model performance metrics. It would show only if the lender tracked connection outcomes by applicant group.
Explanation: specific reasons from transaction data
The legal requirement
The statement said creditors’ use of cash flow data can generally be explained and disclosed to the borrower, as ECOA and the Fair Credit Reporting Act may require. Regulation B requires a creditor that takes adverse action to give a statement of the specific reasons. The CFPB later stressed that those reasons must be specific and accurate whatever technology produced the decision.
Cash flow features fit that requirement better than many alternatives. Reasons about income that is too low for the credit requested, or obligations that are too high relative to income, follow directly from what a cash flow model measures. The difficulty lies in the mapping from hundreds of computed features to a few accurate reasons.
From features to reasons
A model may use dozens of features derived from transactions: average balance, number of overdrafts, volatility of deposits, share of income from each source, ratio of fixed to variable costs. To produce reasons, the lender needs a method that ranks which features drove each decision and a table that maps those features to reason statements a consumer can understand.
Both pieces need governance. The ranking method must be stable and tested. The mapping table must be reviewed by compliance and kept in step with model changes. If the model is retrained and the table is not updated, notices may cite reasons that no longer drive decisions.
Testing notices before they are sent
Reason accuracy can be tested like any other output. Each month, sample adverse action notices and compare the stated reasons with the model’s own attribution for those applications. Where they differ, find out why. Also check that reasons are understandable to a consumer with no technical knowledge. A reason that names a computed feature, such as deposit volatility, may be accurate and still unhelpful.
Disputes and corrections
Explanation is also the start of correction. An applicant told that income was insufficient may know that a deposit was missed or misread. The lender needs a path for the applicant to point out the error and for staff to re-pull or review the data. Without it, provenance errors become permanent decisions. We see tracking how often such reviews change outcomes as one of the better measures of data quality a lender has.
Who owns the reason table
The mapping from features to reasons can fall between teams. Data scientists build the model and its attribution method. Compliance owns the notice. Operations sends it. When the model changes, no one may be told to update the table. Naming one owner, which we would place in compliance with a technical counterpart, and making table review a required step in model release, closes that gap.
Pricing and counteroffers
Adverse action is broader than denial. As the CFPB’s 2022 circular summarized the rule, adverse actions include denying an application, terminating an existing account, making unfavorable changes to the terms of an existing account and refusing to increase a credit limit. A cash flow model used to set price or limit can trigger notice duties too. A lender that uses cash flow data only for pricing should not assume it has escaped the explanation problem.
Second Look programs
The statement singled out Second Look programs, where alternative data is used only for applicants who would otherwise be declined. These programs raise their own explanation question. If an applicant is declined on the traditional score and then declined again on cash flow review, the notice must reflect the reasons actually used. Lenders need to decide, and document, how the two stages combine into one set of reasons.
The Fair Credit Reporting Act angle
If the cash flow data comes from a party that is a consumer reporting agency under the Fair Credit Reporting Act, separate obligations follow. They include adverse action disclosures naming the source and the consumer’s right to dispute. Whether a given aggregator is a consumer reporting agency depends on facts the statement did not address. Lenders should get a written legal view for each data source.
Evidence of lower risk: testing the claim
What the agencies did and did not say
The statement said that some alternative data, including cash flow data, may present no greater risks than traditional data, and that data directly related to consumers’ finances may present lower risks than other data. It said data presenting greater consumer protection risks warrant more extensive compliance management. It did not say cash flow data was safe. It left the lender to show where its own program fell.
Fair lending testing
Cash flow features can correlate with protected characteristics. Income patterns differ by age. Deposit sources may include public assistance, and Regulation B lists income from a public assistance program as a prohibited basis. Spending categories may reflect family status or religion. We would test whether features, and the model as a whole, produce different outcomes for different groups, and whether less discriminatory alternatives exist. In 2019 Regulation B’s commentary said neutral factors in credit scoring systems could still be challenged under the “effects test.” The CFPB’s April 2026 amendments deleted that language and added that the Act does not recognize the effects test, so under federal ECOA this testing became a choice rather than an answer to an effects-test claim. Counsel can say which other laws still apply.
That testing needs demographic information the lender may not hold. Where it does not, testing has to rely on estimates of group membership. Estimates carry error, so the method, its limits and the way results were read all need to be documented.
Searching for less discriminatory alternatives
Testing for disparities is half the work. The other half is looking for a model that serves the lender’s legitimate business needs with less disparity. With cash flow data, the search has many levers. A feature can be dropped, capped or replaced with a less correlated measure. The look-back window can change. Thresholds can be adjusted for second-look applicants.
The search should be documented: which alternatives were tried, how they performed on repayment and on disparity, and why the chosen model was preferred. Even where a lender keeps its original model, that record shows the choice was deliberate. It is also the record a supervisor or plaintiff could ask to see.
Model risk management
For banks, the statement pointed to the agencies’ model risk management guidance, which it said contains principles for managing risk related to models that may use alternative data. A cash flow model adds a data dependency that traditional scores do not have: the model is only as stable as the aggregator feed. Validation should cover data inputs as well as model logic.
Thin files and the missing comparison
The applicants for whom cash flow data matters most are those with little or no credit history. They are also the hardest to evaluate, because there is no traditional score to compare against. A lender cannot show that cash flow data performs better than a score that does not exist. It can show how approved thin-file applicants repay over time, against applicants with similar cash flow profiles who also had scores. That comparison takes years to mature, so it should start with the first loans, not after the program has scaled.
Monitoring after launch
A model validated at launch can drift. The applicant population changes with marketing and the economy. The aggregator changes its categorization. Source banks change what they expose. Each can shift feature values without any change to the model itself.
Ongoing monitoring should therefore cover inputs as well as outputs. Track feature distributions by month and by source bank. Track approval and loss rates by estimated group. Set thresholds that trigger review, and record what was done when they fired.
Reporting to committees and boards
Senior oversight needs a short, stable set of measures. They include the volume of cash flow decisions, approval and loss rates against traditional underwriting, fair lending results, data quality metrics and complaints. Reported quarterly, they let a risk committee see whether the program still deserves its low-risk label.
Predictive value
The case for cash flow data rests on the claim that it predicts repayment, especially for thin-file applicants. A lender should test that claim on its own population before relying on it. A Second Look program offers a natural test design: compare repayment for applicants approved on the second look with those approved on the first, over enough time to observe losses.
Where the argument is weakest
The statement was explicit about its scope
One objection is that the agencies were clear. The footnote said what the statement did not cover, and no careful lender would read it as blessing the data path. On this view, our argument criticizes a document for staying on topic.
The objection is fair to the drafting. It is weaker as a guide to risk. The reasons the statement gave for comfort all depended on the excluded half. A lender that relied on those reasons, as the statement invited, inherited the risks the footnote set aside.
The rules for the data path were coming
A second objection is that the data path was always going to be regulated separately, under section 1033 of the Dodd-Frank Act, which gives consumers rights to their financial data. Lenders could wait for that rule.
Waiting would have meant running cash flow underwriting on unregulated connections for years. The underwriting decisions made in that period still needed consent records, data quality controls and accurate reasons. A future rule on data access would not cure past gaps.
Cash flow data carries its own privacy cost
A third objection runs the other way. Cash flow data may be riskier than the statement suggested. Transaction histories reveal where people shop, which clinics they visit, which causes they support and which places of worship they give to. A lender that pulls this data for underwriting holds a detailed record of private life.
The objection strengthens the argument. It belongs mainly under permission and provenance: what the consumer agreed to share, and how the data is kept and limited. A lender that minimizes what it collects and stores, keeping derived features rather than raw transactions where it can, reduces both privacy and fairness risk.
A non-binding statement did not justify the effort
A fourth objection is that the statement bound no one. Lenders could treat it as a friendly signal and build only what existing law clearly required.
Existing law already required most of what is described here. Adverse action notices, fair lending compliance, third-party oversight for banks and the protection of customer data all predated the statement. The statement changed which data lenders were encouraged to use, not the duties attached to using it. What a non-binding statement could not do was protect a lender when supervisors or courts later read those duties strictly.
Differences by kind of lender
A large bank could have model risk teams, fair lending analysts and vendor management functions able to absorb this work. Its challenge was coordination across them.
A community bank or credit union that bought cash flow underwriting as part of a vendor platform faced a different problem. Its exposure lay in the vendor contract and in its own ability to understand the model. It needed the vendor to provide consent records, reason code mappings and fair lending test results, and should have treated the absence of any of them as a reason to walk away.
A non-bank lender supervised by the CFPB faced the consumer protection expectations without the bank model risk guidance. It still had to meet ECOA and the Fair Credit Reporting Act. Without the model risk discipline banks were expected to apply, its validation and data quality controls had to be built by choice rather than by supervisory habit.
A credit union serving members with irregular income had the strongest case for cash flow data and could have the fewest resources to govern it. For that lender, buying a well-documented service was safer than building one.
Where the argument does not hold
For lenders that obtained cash flow data only from their own customers’ deposit accounts, the provenance part largely falls away. The data path is internal and permission can come through existing account terms, subject to privacy law. The explanation and evidence parts still apply. So does the privacy point: using a customer’s own deposit history to decide on credit is a new purpose, and the customer should be told about it in plain words at the time of application.
What to build first
The four parts depend on each other. Explanation is unreliable if provenance is poor. Evidence of lower risk is meaningless if the data path excludes some applicants. Permission records are the foundation, because every later control refers to what the consumer agreed to.
- Own the consent record. Store the disclosure, scope, duration and timestamp for every applicant, in the lender’s systems, retrievable by application number.
- Treat the aggregator as critical. Run full third-party due diligence, set data quality metrics by source bank, and write accuracy, incident and exit terms into the contract.
- Govern the reason mapping. Document how features become adverse action reasons, review the table with compliance, and re-test it at every model change.
- Test fairness at both stages. Track connection failures and outcomes by estimated applicant group, and test each feature and the full model for disparities and alternatives.
- Measure predictive value. Follow approved applicants long enough to confirm that cash flow data predicts repayment in the lender’s own population.
Three measures show progress. The first is the share of cash flow decisions with a retrievable consent record. The second is the connection failure rate by source bank. The third is the agreement rate between sampled adverse action reasons and model attribution. Each should be reported to the committee that owns credit risk.
We think the staffing is modest relative to the lending volume involved. One product owner can hold permission and provenance. A compliance lead can own reasons and notices. A model risk or analytics lead can own fairness testing and monitoring. For a smaller lender, a vendor can supply much of the evidence, as long as the contract requires it and someone inside reads it.
An examiner reviewing a cash flow program would ask for each of these artifacts. A lender that can produce them can show its use of alternative data is responsible, which is the standard the statement set.
With hindsight
Each of the four conditions the statement left to lenders has since been the subject of federal action, and much of that action has been withdrawn, reopened or reversed.
On explanation, the CFPB issued Circular 2022-03, published in the Federal Register on June 14, 2022. It said ECOA and Regulation B require specific and accurate reasons for adverse action, whatever technology is used. Creditors may not use complex algorithms when that prevents them from giving such reasons. A further circular, 2023-03, addressed the proper use of Regulation B’s sample notice forms. On May 12, 2025, the CFPB withdrew both circulars, among many guidance documents, saying much of that guidance had imposed compliance burdens outside notice-and-comment rulemaking. The withdrawal did not amend the adverse action requirement in Regulation B itself. The 2019 interagency statement was not on that withdrawal list.
On provenance and permission, the CFPB issued its personal financial data rights rule under section 1033 on November 18, 2024, effective January 17, 2025. It required banks, credit unions and other providers to make consumer data available to consumers and authorized third parties. A data provider could not meet its duty to third parties by letting them screen scrape, and the rule set limits on how third parties collect, use and keep data. Authorized third parties had to apply an information security program meeting the Gramm-Leach-Bliley Act safeguards rules. Compliance dates ran by institution size from April 1, 2026 through April 1, 2030. An earlier piece, finalized in June 2024, set the attributes a standard-setting body needed for CFPB recognition, and was folded into the November rule.
That rule did not settle the matter. A bank and two trade associations sued in the Eastern District of Kentucky, in Forcht Bank, N.A. v. Consumer Financial Protection Bureau. After the change of administration, the court paused summary-judgment briefing for 90 days while the CFPB reviewed its position, and extended the compliance deadlines by the same period, moving the first from April 1 to June 30, 2026. On July 29, 2025, the CFPB asked the court to stay the case because it had decided to reconsider the rule, and the court granted the stay. On August 22, 2025, the CFPB sought comment on who may act as a consumer’s representative, on data security, on whether the costs of data access should be shared, and on the compliance dates, which it said it planned to propose extending. On October 29, 2025, the court enjoined the CFPB from enforcing the rule until it has completed its reconsideration. A Federal Register search on November 14, 2026 found no proposed or final revision of the rule. Lenders that depend on aggregated account data were left with a rule they could not plan around and no date for its replacement.
On evidence of lower risk, the ground shifted too. On April 22, 2026, the CFPB finalized amendments to Regulation B, effective July 21, 2026, providing that ECOA does not authorize disparate-impact liability. The same rule further defined discouragement of prospective applicants and added conditions for special purpose credit programs. Commenters opposing it had argued that disparate-impact liability was particularly important in some kinds of lending, including automated, AI-driven credit models. For lenders, the federal test of a model’s fairness narrowed, while other laws and supervisors may still apply.
The statement’s central judgment, that cash flow data sits closer to traditional underwriting than other alternative data, was not among the guidance the CFPB withdrew in May 2025. Its assumption that the surrounding legal framework was stable did not hold. The bank model risk guidance the statement cited moved as well: on April 17, 2026, SR 26-2, issued by the Federal Reserve, the OCC and the FDIC, superseded and replaced SR 11-7. The data path, the guidance on explanations, the fairness test and the model risk guidance all moved between 2022 and 2026. We would expect lenders that had built their own controls for each part to have been the least exposed to those swings.
Spotted an error? Tell us and we will check it and correct the article where needed.