Cash-Flow Underwriting: Adoption, Performance & Risk

Cash-flow underwriting is already in use, but its reach and results are uneven

Cash-flow underwriting uses account activity, balances, receipts and payment timing to inform a credit decision. The public evidence now extends beyond vendor announcements: banks use their own deposit histories, mission-based lenders analyze linked accounts or statements, merchant platforms use sales records, and mortgage systems can incorporate account cash flow. These are different implementations rather than one universal replacement for credit-bureau information. [21, 26, 32, 36]

The attraction is understandable. Transactions can reveal current resources, irregular income, a cash cushion or a payment-date mismatch that past credit performance does not fully describe. The same information can reduce paperwork, support a second look, change a credit limit or help size a loan. Better prediction, more affordable terms, less customer effort and greater lender profitability remain separate outcomes; improvement in one does not establish the others.

As of the October 4, 2026 research check, there is credible evidence of live use and incremental predictive information in studied populations. The sources reviewed do not establish a current market-wide adoption percentage, a universal realized loss reduction or a causal improvement in borrower wellbeing. This article distinguishes documented workflows, historical surveys, retrospective studies, vendor claims and original hypothetical examples. A recheck date does not make an old launch or old loan cohort a new event.

Income verification, risk prediction and affordability answer different questions

These functions can use overlapping transactions but require different evidence. Income verification reconstructs the source and likely continuity of resources. A risk model estimates an outcome over a defined period. An affordability assessment compares resources, obligations, buffers and payment dates for the proposed product. Treating their outputs as interchangeable obscures what a lender has actually established.

An account-connection service, transaction classifier, consumer report, score and lending policy perform different jobs. The meaning of a decision depends on who owns each transformation, what information is missing and how a result can be explained or corrected. The lender still chooses the amount, price, term and decision threshold.

Scroll horizontally to see all columns.

Why lenders adopt it: six different economic and customer problems

The business case varies by product. Own-bank deposit history can support a quick small-dollar offer for an established customer. External account data can reconsider a decline or reveal capacity that a thin credit file misses. Statement extraction can reduce repetitive staff work even when the underlying lending policy is unchanged. Merchant receipts can support revenue-linked offers. Mortgage infrastructure can add information inside an existing eligibility process. None of these benefits necessarily requires a machine-learning model. [21, 23, 26, 30, 14, 36]

For a borrower, the gain may be a decision that reflects recent circumstances, less document collection, a better-fitting amount or a payment date aligned with income. For a lender, it may be lower acquisition friction, more informed line assignment, reduced processing expense or additional profitable customers. If repeated login attempts, missing account coverage and manual corrections simply transfer work to the applicant, an automated connection can fail to improve the overall experience.

Adoption therefore has several stages: a supplier is selected, a pilot begins, data reaches an actual decision, a measurable share of applications uses it, and funded loans mature. A source establishing one stage does not establish the next. The adopter tables below are an evidence map, not a census or a ranking.

Documented bank and community-lender workflows

These cases use deposit histories, connected accounts or bank statements for identifiable credit decisions. The community-lender cases are described in FinRegLab’s June 2025 implementation research, which includes interviews and pilots rather than a representative adoption survey.

Scroll horizontally to see all columns.

Second looks, business cards and merchant finance

The European examples have specific geographic boundaries. Merchant-platform data describes sales activity more directly than a complete household or business budget; external bank links can add another view. Documentation of a live capability does not measure how many decisions use it.

Scroll horizontally to see all columns.

Mortgage infrastructure and a documented statement workflow

A mortgage-system feature can be operational without being used for every lender or casefile. An extraction workflow can be implemented without producing an independently measured credit-loss benefit.

Scroll horizontally to see all columns.

Named selections and integration intentions

These vendor announcements remain evidence of a relationship and its stated scope. The research did not establish a comparable funded-volume or matured-loss series for these particular integrations.

Scroll horizontally to see all columns.

SoFi illustrates why an underwriting approach and a vendor rollout are different

SoFi’s Q2 2025 filing describes personal- and student-loan underwriting that combines credit-bureau reports, industry scores, proprietary models and debt-capacity analysis indicated by borrower free cash flow. That lender disclosure supports a hybrid underwriting approach. It does not isolate a Cash Atlas treatment group or attribute changes in company-wide charge-offs to that product. The dated filing is evidence of the disclosed approach, not proof of the latest operational coverage. [47]

This distinction matters throughout the market. A bank can already use internally held transactions before selecting an external provider. A provider can also supply verification without changing the approve/decline model. Counting every relationship as a new cash-flow credit deployment overstates what the public record establishes.

How much adoption? The denominator changes the answer

No defensible current national adoption rate was established in this research. A percentage of surveyed professionals, a percentage of institutions, a percentage of applications and a percentage of funded loans answer different questions. Institutions that use cash flow for a narrow second-look segment may still have most loans decided by other routes.

Historical surveys are useful directional evidence only when their bases remain visible. The Nova Credit/Researchscape survey’s 57% applies to 109 respondents who already used alternative data, within a 185-person US sample surveyed in October 2022. It is not 57% of all US lenders, and its rounded percentage does not disclose an exact cash-flow-user count. [41]

Connectivity and sentiment are still further removed from actual credit use. The often-cited 83% openness figure summarized by Plaid describes willingness to consider new data. The 71% open-banking implementation figure cited in a later UK report is an infrastructure statistic whose underlying denominator was not established in this review. Neither is used here as a cash-flow adoption estimate. [42, 43, 44]

Scroll horizontally to see all columns.

The provider market spans different functions and score definitions

The reviewed products occupy different layers. A familiar score scale does not make two models equivalent, and higher values do not always mean lower risk. A comparison needs the exact version, outcome definition, performance horizon, minimum data history and population on which the product was validated.

The table preserves the provider specifications checked September 29, 2026, rather than implying that every product detail was newly reverified on October 4. FICO’s May 20, 2026 general-availability announcement is distinct from the reviewed Plaid documentation’s beta label for LendScore. Product availability alone does not establish lender penetration. [5, 7]

Scroll horizontally to see all columns.

What better prediction means, and what it leaves unanswered

AUC measures ranking: roughly, how often a model places an observed adverse-outcome case above a non-adverse case in risk. It does not directly measure a dollar loss rate or establish that a predicted 3% risk actually occurs 3% of the time. Calibration concerns that second question. KS measures the maximum separation between score distributions for outcome groups. An improvement in KS, AUC, approvals and net losses uses different units and cannot be exchanged one-for-one.

A complete comparison separates a change in information from a change in algorithm. The four relevant configurations are a conventional model with bureau inputs, the same model with added cash flow, a more flexible model with bureau inputs and that flexible model with both. A combined improvement cannot all be attributed to cash flow when most of the measured change comes from the algorithm. Product, outcome horizon, risk cutoff and sample selection also affect the comparison.

Consumer evidence: useful incremental information, mostly pre-pandemic loans

FinRegLab’s July 2025 study analyzed credit accounts originated April 2018–March 2019 and a twelve-month composite distress outcome: 90-plus-day delinquency, charge-off, repossession, foreclosure or bankruptcy. The research supports incremental predictive information in the population studied. It does not describe a randomized production rollout. Support from JPMorgan Chase and Capital One is disclosed in the publication. [2]

The strongest tested combination used bureau plus cash-flow information with XGBoost. Figure 1 reports bureau-only logistic-regression AUC of 0.8653; its increments imply approximately 0.8831 for bureau-only XGBoost and 0.8854 for combined-data XGBoost. Cash-only XGBoost was 0.7987. Those derived combined figures are shown as such, rather than presented as independent new calculations from raw loan data. [2]

At the study’s 3% risk cutoff, baseline simulated approval was 65.18%. Adding cash-flow data with the conventional model contributed about 0.5 percentage points; changing the model contributed about 2.5 points; doing both about 3.0 points. Most of that comparison’s gain came from the modeling change. These are simulated approval decisions for a retrospective sample, not observed incremental loans, dollars saved or a promise for credit-invisible applicants. [2]

Scroll horizontally to see all columns.

Who entered the consumer study matters as much as the headline

The technical appendix begins with 750,266 linked observations and retains 424,546, or 56.59%. Exclusions include 2,588 without credit scores, 90,342 with low-quality bank data and 232,790 with insufficient primary-checking information. Those are research filters; they are not a production account-connection failure rate. Matching and account-history requirements leave a relatively scored, prime and higher-income sample. People rejected on every application are not represented by an observed newly funded loan in this design. [3]

The held-out sample contains 107,789 observations and 2,468 adverse outcomes, or 2.29%, from July and November 2018 and March 2019 originations. Other months supplied development and validation data. It is meaningful unseen-data testing, but largely contemporaneous with development rather than an untouched later-cycle test. Performance after inflation, changed payment behavior, different bank coverage or a later recession remains a separate question. [3]

Small-business evidence has a different outcome and validation design

FinRegLab’s June 2025 research with Howell and Matsumoto studies 38,021 originated loans from two anonymous online nonbank lenders, dated February 13, 2015–January 19, 2024, after excluding loans too recent to have an observed status. Nonperformance combines charge-off, more-than-60-day delinquency, or forbearance plus modification. Its 17% mean nonperformance is therefore not a fixed-horizon annual default rate. [4]

The random-forest comparison improves AUC from 0.652 to 0.663 overall and from 0.599 to 0.622 for firms younger than five years whose owners have FICO below 700. This is incremental ranking information in these originated loans. Testing uses an outcome-stratified random 80/20 split, with a further validation split inside training, rather than a forward-vintage holdout. Selection by the two lenders and unequal seasoning constrain transfer to rejected applicants or another product. [4]

The reported 16% mean APR, available for 36,252 observations, describes those loans. It is not evidence that adding cash flow reduced borrowing costs. More flexible access, lower prices and improved business survival would each require their own comparison. [4]

Scroll horizontally to see all columns.

Earlier studies add support without supplying a universal effect size

FinRegLab’s 2019 research examined six nonbank providers: Accion, Brigit, Kabbage, LendUp, Oportun and Petal, with independent empirical analysis by Charles River Associates. Cash-flow metrics were predictive where loan-level outcomes were available and often added information beyond traditional scores. Products, populations and available comparisons differed, while demographic and inclusion analysis were not available uniformly. The work supports useful signals across studied cases rather than a single pooled loss benefit or blanket fair-lending clearance. [46]

The CFPB’s July 2023 research linked March 2019 Making Ends Meet survey responses with subsequent credit records. It studied self-reported savings, regular saving/no overdrafts and bill-payment proxies, not a new API-connected transaction model. Effective regression samples were below 1,000, with scores below 720 and a two-year 90-plus-day delinquency outcome. Only accumulated savings was statistically significant in its primary weighted analysis. The approximate 70% adjusted lower delinquency association for that proxy is not a 70% loss reduction caused by deploying cash-flow underwriting. The report cautions that estimates are imprecise and cannot support demographic breakdowns. [45]

Vendor performance claims use incompatible measures

Vendor evidence can help identify a hypothesis or intended use, but the claims below are not an independent head-to-head contest. The public materials do not provide a common matched dataset, common loss horizon or complete reproducible methodology. Product launches and commercially selected customer cases may emphasize favorable results.

Scroll horizontally to see all columns.

A production forecasting model need not be a default model

Parafin’s May 19, 2026 technical report says ParaFormer runs daily to generate monthly sales forecasts used to size merchant offers. Its retrospective comparison covered approximately 15,000 funded businesses at January, April, July and October 2025 snapshots, comparing forecasts with realized sales, the previous production forecaster and repayment outcomes; repayment stress was observed at 65% of advance duration, approximately 175 days. This is company research, not independent prospective validation. [52]

The distinction is consequential: Parafin says the model forecasts seasonal sales conditional on the business remaining healthy, while separate models address deterioration and churn. It reports an individual-level correlation of −0.06 between its forecast-lift measure relative to the production forecaster and terminal loss. A revenue-forecast improvement therefore is not itself a demonstrated reduction in portfolio losses. The disclosed backtest does not establish an industry-wide adoption rate or a causal borrower-welfare benefit. [52]

An adverse adoption example: A4CB later tightened startup policy

FinRegLab’s June 2025 implementation report describes Allies for Community Business observing increased loan losses by 2023 following a broader package of underwriting changes, particularly among startups. A4CB subsequently halved the maximum startup loan size and broadened debt-to-income and nonsufficient-funds criteria; reported losses slowed afterward. This account belongs alongside the favorable adoption examples. [27]

The evidence is an interview-based implementation case, not an audited causal evaluation. It supplies neither a comparable standardized loss-rate denominator nor an isolated cash-flow treatment group. Cash-flow data cannot be identified as the cause of the earlier deterioration or the subsequent improvement. Bundled policy changes, borrower mix, loan size and economic conditions can all matter. The narrower observation is that a data-enabled process still encountered losses and changed sizing and policy after outcomes developed. [27]

Resource classification determines what a ratio means

Hypothetical example: an account receives $6,000 during a month: $3,600 net pay, $1,200 transferred from savings, $800 of new borrowing and a $400 merchant refund. Calling all $6,000 recurring earnings overstates the identified pay by $2,400, or 66.7%. Savings may be a usable buffer, borrowing creates an obligation and a refund reverses earlier spending. These facts have different meanings even though every item is an inflow.

An auditable interpretation connects source account, transaction, classification, derived attribute and decision. Opening and closing balance reconciliation can reveal missing events; pending/posted duplicates, reversals and internal transfers can otherwise inflate apparent activity. Gross business receipts differ from owner income after costs and taxes. A successfully connected account need not represent every household account or every obligation.

Missing rent or debt payments may mean another account is used. A recurring deposit may be wages, public assistance, personal support or a transfer. Authentic account records can also contain staged deposits, circular transfers or temporary borrowed balances. Data authenticity and sustainable economic resources are separate questions.

Regulation B §1002.6(b)(5) permits assessment of amount and probable continuance while restricting automatic discounting or exclusion on specified protected grounds, including covered income sources and part-time status. Its commentary addresses individualized evaluation. A transaction label such as benefits or gig income does not itself resolve probable continuance. [15]

Worked example: a positive monthly surplus can hide a payment shortfall

Assume a hypothetical applicant starts with $1,700, receives two $1,800 paychecks, and has $2,900 of existing monthly expenses and debt payments. A proposed $500 installment leaves a positive $200 monthly surplus. The payment is nevertheless scheduled before the first paycheck, creating a $300 cash shortfall. The table is a projected cash path, not an assumption that the bank will permit an overdraft; a negative figure indicates an unmet payment need.

Moving the proposed installment from day 3 to day 6 eliminates that shortfall under these exact assumptions, while leaving the same $1,900 month-end balance. If the first paycheck slips to day 10, the problem returns. The appropriate analysis therefore includes timing, a buffer and plausible delays. Rescheduling can address a mismatch; it cannot create income or cure a persistent deficit.

This simplified example excludes unexpected expenses, fees and payment-return effects. A production assessment needs the actual bill calendar, available funds and uncertainty around both amounts and dates. The useful output is the lowest projected available balance across the period, together with the assumptions driving it.

Scroll horizontally to see all columns.

Variable income changes the buffer and the product decision

A second hypothetical borrower earns alternating monthly amounts of $2,000 and $6,000. Average income is $4,000. With $2,500 of existing outflows and a $500 proposed payment, the average surplus is $1,000, but every low-income month consumes $1,000 of savings. Two consecutive low months require $2,000 just to cover the assumed deficits, before a safety margin. A verified $3,000 reserve and a $200 reserve support different conclusions even with the same average income.

Longer histories can reveal seasonality and distinguish recurring from exceptional income, but short histories or changed employment introduce uncertainty. A stress scenario or conservative percentile is an analytical assumption rather than an automatic universal haircut. Self-employment taxes and protected income require separate treatment; the applicable income-evaluation rule concerns individual circumstances. [15]

For credit cards, line size and future utilization matter. Regulation Z §1026.51 requires consideration of minimum-payment capacity based on income or assets and current obligations at opening and line increases. Its safe-harbor estimation method assumes full use of the proposed line from the first billing-cycle day; the broader rule permits a reasonable estimation method. A cash-flow score alone does not document this assessment. [17]

For a long installment loan, a low recent default forecast cannot establish affordability over a multi-year term. Payment size, interruptions to income, variable rates, emergency expenses and dependence on refinancing can change the cash path. For short installments, the next pay cycle may be informative, yet overlapping obligations and failed-payment timing can dominate. These are product-specific analytical issues; the card rule is not a universal statutory test for every installment product.

A business cash cycle and a mortgage file need different interpretation

Hypothetical small-business example: a merchant receives $60,000 in monthly deposits, but $10,000 is a transfer from another owned account and $5,000 is a new loan. Operating receipts are therefore $45,000 before assessing expenses. With $35,000 of operating outflows, the apparent operating cash surplus is $10,000, not $25,000. Taxes, owner withdrawals, capital expenditure, seasonality and existing debt payments may further change the resources available. This is a classification example, not a business valuation or a lending limit.

Analysis: receipt timing also matters. A business that buys inventory before collecting sales can have healthy annual margins and still need working capital. A schedule aligned with its cash cycle may be more useful than a larger limit with inflexible payment dates. Testing such a design requires observing costs and customer outcomes; transaction history alone does not establish that the proposed terms are suitable.

Mortgage use has a separate boundary. For covered transactions, Regulation Z § 1026.43 includes repayment-ability and verification requirements, including reasonably reliable third-party records for income or assets relied upon. An account-data feed can contribute evidence; a cash-flow score does not by itself establish satisfaction of those requirements or an investor’s purchase criteria. Match the evidence to the applicable mortgage standard and the facts of the file. [20]

From a connection to a funded loan: the whole applicant funnel

A hypothetical process in which 70% of invited applicants consent, 80% of those connect and 90% of connected accounts meet data standards reaches usable data for only 50.4% of invitees. A strong result in that subset leaves 49.6% outside the same evidence path. Unsupported institutions, refusal, technical failures, stale permission and genuinely insufficient history are different reasons for missing data.

Applicants willing and able to link accounts can differ from those who cannot or decline. The meaningful sequence is eligibility, invitation, consent, connection, usable history, decision, approval, funding and seasoned performance, with counts and elapsed time at each stage. A second-look conversion rate based only on successful links omits the people lost before a model produces a score.

An optional second-look invitation, randomized under appropriate legal and ethical controls, can help estimate the overall effect of offering that process on eligible applicants. It does not reveal outcomes for never-funded counterfactual loans. Retrospective analysis of originated loans and reject-inference methods cannot manufacture those missing outcomes. Conditional results among successful connectors remain different from results across everyone invited.

Hypothetically, ten adverse outcomes among 500 loans equal 2%, with an approximate 95% Wilson interval of 1.1%–3.6% under independent binary observations and complete comparable follow-up. That interval omits correlated economic shocks, selection, incomplete seasoning and dollar severity. A small favorable pilot and a durable production result remain different evidentiary stages.

Risks begin before the model and continue after approval

The following mechanisms are analytical failure modes, not claims about their prevalence at the named institutions. They explain how a technically valid feed or a better ranking model can still lead to unreliable decisions. The final column describes evidence that helps distinguish a controlled process from an untested assumption.

Plaid’s servicing documentation illustrates a concrete boundary. Its recurring Consumer Report service became generally available August 20, 2026. Reports can still be generated from successful connections when other linked accounts fail, so report availability does not establish complete account coverage. Retrieval is restricted to the documented account-review/collection or consumer-written-instruction purposes; origination decisioning purposes are not accepted for these reports. A credit-decision report requires a separate creation flow with an appropriate decisioning purpose. These are vendor-documented product constraints, not evidence of lender adoption or a legal conclusion about every use. [53]

Scroll horizontally to see all columns.

A useful validation comparison includes calibration, seasoning and economic value

The relevant outcome is product-specific: a 180-day payment-risk estimate is not proof of three-year installment affordability. Model version, observation window, geographic coverage, target population, missing-data treatment and outcome horizon define what a result means. Performance in all applicants does not automatically transfer to a deliberately selected second-look decline band.

The four-way data/model comparison becomes more informative with an untouched later-vintage test and tests across providers, income patterns and channels. Common risk or approval constraints make tradeoffs interpretable. Calibration, adverse-outcome counts and uncertainty matter alongside AUC and KS. Repeated borrowers, common sectors and macroeconomic shocks can make nominal sample counts overstate statistical independence.

Count-based delinquency, gross charge-offs, recoveries and net losses per funded dollar can move differently. A lower incidence of problems can coexist with higher loss dollars when new loans are larger or recoveries weaker. Consistent horizons, seasoning and censoring treatment are essential to separating a maturing loan book from deterioration. A production change can also alter prices, amounts and take-up, so a fixed historical sample does not reproduce every consequence.

Operational reversibility is another distinct issue. A previously validated fallback, a reproducible decision record and known escalation responsibility allow a degraded process to be narrowed or restored. A single universal numeric tolerance would ignore differences in risk appetite, base rates, outcome delay and statistical power. Retraining alone cannot repair missing accounts or an unsuitable repayment date.

The economics can work even with modest model lift, or fail despite it

Hypothetical economics: 10,000 evaluation attempts at an assumed $2 data cost each cost $20,000. Another 1,000 manual reviews at $8 add $8,000. If 400 genuinely incremental funded loans contribute $100 each after funding, servicing and expected credit losses, gross incremental contribution is $40,000 and the remainder is $12,000 before fixed implementation and governance costs. These assumptions are not vendor price quotes.

Variable-cost break-even is 280 incremental loans at that contribution. At 250, the process loses $3,000 before fixed costs. Borrowers who would have funded through the baseline are not incremental growth merely because they moved channels. Lower prices, larger balances and changed risk mix can also change the contribution per loan.

Failed attempts, re-pulls, minimum contractual commitments, manual fallback, disputes, validation, fraud, capital, monitoring and eventual vendor replacement affect the full economics. Forecast contribution becomes more credible when reconciled to actual vintages as losses emerge. Staff time saved has economic value when work is removed or capacity is productively redeployed; extraction speed alone does not measure that result.

Customer effort has value too: elapsed time, repeated document requests, correction rates and completion without a working connection influence both access and cost. A modest predictive gain can accompany a valuable process improvement. Conversely, a stronger score can produce a poor commercial outcome when application friction and unresolved errors overwhelm the incremental lending benefit.

Explanations and corrections concern the final lending decision

Regulation B §1002.9(b)(2) requires specific principal reasons when adverse-action reasons are provided; failure to meet internal standards or a qualifying score is insufficient. Commentary connects reasons to factors actually considered or scored and addresses mixed judgmental/scoring systems. A FCRA credit-score-factor disclosure does not by itself satisfy ECOA’s reason requirements. A purchased score can be one input without explaining the lender’s final action. [16]

A decision can depend on eligibility, fraud checks, a score cutoff, an affordability overlay or a manual exception. Replaying the entire decision from retained application-time records tests whether the result and its explanation agree. Payroll mislabeled as a transfer calls for a different correction from genuinely insufficient income; missing information is also distinct from an adverse decision supported by information already available. [16]

Consumer-reporting roles, permissible purposes, service-provider obligations, data security and correction responsibilities depend on the actual arrangement. A vendor compliance label cannot establish the lender’s compliance. This article does not resolve every FCRA/GLBA classification, federal or state fair-lending issue, or changing legal theory. Predictive parity is not legal clearance.

Current governance and access status, dated October 4, 2026

The Federal Reserve’s April 17, 2026 SR 26-2 supersedes SR 11-7 and SR 21-8. It emphasizes risk-based practices tailored to the institution’s size, complexity and model-risk profile; the Fed letter says it is expected to be most relevant to Fed-regulated organizations exceeding $30 billion in assets. It is supervisory guidance, not a universal underwriting statute. Responsibilities across data transformations, model versions, overrides and third parties remain relevant to a cash-flow system’s actual use. [18]

The CFPB compliance page, last modified January 6, 2026 and checked in this research, reports that Section 1033 rule compliance dates were stayed on October 29, 2025 and identifies the August 22, 2025 reconsideration notice. This is the agency’s posted status, not a new independent review of the complete litigation docket. Original deadlines therefore cannot establish universal access, while a stay of compliance dates does not erase every contract or other data obligation. [19]

The agencies’ December 2019 alternative-data statement acknowledged potential gains in speed, accuracy and access alongside consumer-protection concerns. It did not endorse an individual provider or exempt a lender from applicable law. Clear permissions, purpose limits, retention choices and practical fallback remain separate from a model’s predictive performance. [1]

FinRegLab’s April 9, 2026 machine-learning framework adds practitioner discussion of reason-code aggregation, explanation limitations, monitoring and third-party responsibilities. It is a practice framework, not a new outcome experiment. Its references predate the April 17 SR 26-2 revision, so its historical regulatory citations cannot silently substitute for the later guidance. [49]

What remains unknown, and what future evidence could change

The strongest supported conclusion is that cash-flow information can add useful context and predictive information, with verified operational examples across several credit markets. The size and durability of benefits remain specific to the population, decision, product and data path. The reviewed sources do not establish a generalized realized net-loss benefit, causal reductions in APRs or fees, or a causal improvement in disposable income or wellbeing. This describes the limits of this research set, not a claim that such evidence can never exist.

FinRegLab’s April 2026 small-business financial-health initiative broadens the questions to stability, resilience and owner wellbeing using CDFI borrower cash-flow data. Capital One supports the initiative and Plaid supplies connectivity. Its measurement framework and future testing plans are a research development; they are not results demonstrating that lending already caused those benefits. [50]

An informative adoption update would move a named institution from an announcement to an identifiable production workflow, then disclose application coverage and funded volumes. An informative performance update would supply cohorts, denominators, vintages, consistent horizons, realized losses and borrower terms. Later-originated data, more substantial no-file or irregular-income samples and prospective evidence would address limitations in the existing studies.

Further revisions can preserve the original event and publication dates while adding newly verified evidence. A new press release without those details does not close an old performance gap, and the absence of a new public disclosure does not prove that no operational change occurred. Selection, actual use, reach, risk, affordability and economics remain distinct parts of the story.

Sources

1. Federal financial regulators — joint alternative-data statement, December 3, 2019; reviewed September 29, 2026Official releaseBack to text: ↑
2. FinRegLab — Advancing the Credit Ecosystem, July 2025 main empirical paper; reviewed September 29, 2026Source · PDFBack to text: ↑1↑2↑3↑4
3. FinRegLab — July 2025 technical appendix and sample limitations; reviewed September 29, 2026Source · PDFBack to text: ↑1↑2
4. FinRegLab — Sharpening the Focus: Using Cash-Flow Data to Underwrite Financially Constrained Businesses, June 2025; reviewed September 29, 2026Source · PDFBack to text: ↑1↑2↑3↑4
5. FICO — next-generation UltraFICO general availability and attributed performance claims, May 20, 2026SourceBack to text: ↑1↑2↑3
6. Experian — Cashflow Score launch, scale and KS-based claim, March 25, 2025SourceBack to text: ↑1↑2
7. Plaid — current Consumer Report / Plaid Check documentation and LendScore beta specification; undated, accessed September 29, 2026SourceBack to text: ↑1↑2
8. Mastercard — Payment Risk Insights score and 180-day horizon, April 7, 2026SourceBack to text: ↑
9. Prism Data — CashScore product description and attributed lift claim; undated, accessed September 29, 2026SourceBack to text: ↑1↑2
10. Nova Credit — SoFi relationship expansion and implementation intention, October 15, 2024SourceBack to text: ↑1↑2
11. Nova Credit — Imprint / Alloy integration announcement, June 25, 2025SourceBack to text: ↑1↑2
12. Nova Credit — Chase selection of Cash Atlas and Credit Passport, September 3, 2025SourceBack to text: ↑1↑2
13. Nova Credit — PayPal U.S. cash-flow underwriting selection, September 4, 2025SourceBack to text: ↑1↑2
14. Ocrolus — attributed TAB Bank statement-analysis customer story; undated, accessed September 29, 2026SourceBack to text: ↑1↑2↑3↑4
15. CFPB — current Regulation B §1002.6(b)(5) and commentary on income evaluation; accessed September 29, 2026Official textBack to text: ↑1↑2
16. CFPB — current Regulation B §1002.9 and commentary on specific reasons and incompleteness; accessed September 29, 2026Official textBack to text: ↑1↑2
17. CFPB — current Regulation Z §1026.51, card ability to pay and payment-estimation safe harbor; accessed September 29, 2026Official textBack to text: ↑
18. Federal Reserve — SR 26-2 revised model-risk guidance, April 17, 2026; supersedes SR 11-7 and SR 21-8Official sourceBack to text: ↑
19. CFPB — Personal Financial Data Rights compliance page, modified January 6, 2026; reports October 29, 2025 court stay of compliance dates; accessed September 29, 2026Official sourceBack to text: ↑
20. CFPB, Regulation Z § 1026.43, repayment-ability and verification provisionsOfficial textBack to text: ↑
21. Huntington current Standby Cash eligibility explanation; undated page checked October 4, 2026SourceBack to text: ↑1↑2↑3
22. Huntington Standby Cash launch; 2021-06-01SourceBack to text: ↑
23. Bank of America Balance Assist no-action-letter request, underwriting description; 2020-11Official source · PDFBack to text: ↑1↑2
24. Bank of America community banking solutions; undated page checked October 4, 2026SourceBack to text: ↑
25. Bank of America Q4 2021 presentation; 2022-01-19SourceBack to text: ↑1↑2
26. Allies for Community Business loans: published credit criteria; undated page checked October 4, 2026SourceBack to text: ↑1↑2↑3
27. FinRegLab Transforming Small Business Credit; 2025-06-03Source · PDFBack to text: ↑1↑2↑3↑4↑5↑6
28. FinRegLab implementation research summary; 2025-06Source
29. Ascendus Business Line of Credit; undated page checked October 4, 2026SourceBack to text: ↑
30. N26 overdraft support: open-banking second look; undated page checked October 4, 2026SourceBack to text: ↑1↑2
31. Capital on Tap explains its use of Open Banking; 2026-01-26SourceBack to text: ↑
32. Square loan eligibility and optional bank connection; undated page checked October 4, 2026SourceBack to text: ↑1↑2
33. PayPal Working Capital official US product page; undated page checked October 4, 2026SourceBack to text: ↑
34. PayPal global small-business lending scale; 2025-05-05SourceBack to text: ↑
35. YouLend official platform FAQ; undated page checked October 4, 2026SourceBack to text: ↑
36. Fannie Mae risk factors evaluated by DU; 2025-02-05SourceBack to text: ↑1↑2↑3
37. Fannie Mae DU cash-flow FAQs; 2026-07SourceBack to text: ↑
38. Fannie Mae rent and cashflow use cases; undated page checked October 4, 2026Source
39. Freddie Mac cash flow capability in LPA; 2022-11-07SourceBack to text: ↑
40. Freddie Mac 2025 technology update; 2025-07-14SourceBack to text: ↑
41. Nova Credit / Researchscape 2022 survey; 2022-10Source · PDFBack to text: ↑1↑2↑3
42. Plaid cash-flow data industry-standard article; 2023SourceBack to text: ↑
43. Plaid Consumer Report launch; 2024-06-04SourceBack to text: ↑
44. Datos / Plaid Cash Flow Underwriting: Reshaping the UK Lending Landscape; 2025-06Source · PDFBack to text: ↑1↑2
45. CFPB research and endnotes; research checked October 4, 2026Official sourceBack to text: ↑
46. 2019 empirical report, executive summary; research checked October 4, 2026Source · PDFBack to text: ↑
47. SoFi Q2 2025 filing, printed page 60; research checked October 4, 2026Source · PDFBack to text: ↑
48. Joint statement; research checked October 4, 2026Official release · PDF
49. 2026 framework; research checked October 4, 2026Source · PDFBack to text: ↑
50. Project report page; research checked October 4, 2026SourceBack to text: ↑
51. Parafin and Tekion, Pay Over Time launch; September 10, 2026; rechecked October 4, 2026SourceBack to text: ↑
52. Parafin Research, ParaFormer production forecasting and company backtest; May 19, 2026; rechecked October 4, 2026SourceBack to text: ↑1↑2
53. Plaid Check servicing documentation; August 20, 2026 general availability and current purpose/partial-connection rules; checked October 4, 2026SourceBack to text: ↑