The FICO problem
The FICO score was a genuine reform. Before credit scoring standardized borrower evaluation in the mid-twentieth century, lending decisions were made by loan officers exercising personal judgment. Discretionary systems create predictable problems: worthy borrowers get rejected for reasons unrelated to their ability to repay, and the people making decisions can use their power unevenly, whether intentionally or not. A Federal Reserve Bank of Philadelphia study analyzing the near-universe of U.S. mortgage applications from 1994 to 2018 found that variation in loan officer subjectivity alone produced measurable, persistent gaps in approval rates across otherwise comparable applicants. FICO replaced that discretion with something objective, legible, and scalable: a single number derived from a borrower’s credit accounts, repayment history, utilization rate, and the age and mix of their existing debt. By making creditworthiness measurable across a large and diverse borrower population, the scoring system helped expand consumer credit considerably and remains the dominant metric lenders use to determine who gets a loan and on what terms.
FICO scores what it can see, however, and what it can see is limited to a specific slice of financial behavior. Borrowers who have not used much credit, because they are young, recently arrived in the country, or simply have not had reason to borrow, produce little or nothing for the model to evaluate. The Federal Reserve estimates that roughly 32 million American adults are effectively unscoreable: about 7 million have no credit history at all, and 25 million have files too thin to generate a reliable score. According to the CFPB, about 15 percent of Black and Hispanic consumers are credit invisible, compared to 9 percent of white consumers; an additional 13 percent of Black consumers and 12 percent of Hispanic consumers have records too sparse to produce a reliable score under widely-used models.
The exclusion problem takes two distinct forms, and they affect borrowers on opposite ends of the loan market. The first is the medical resident earning $60,000 a year with a high debt-to-income ratio and a thin credit file, who needs a large personal loan. Her income will likely triple within a few years as she completes her residency and enters practice, making her a low default risk over the life of a loan. A conventional underwriting model, evaluating her current financials and limited credit history rather than her earnings trajectory, would characterize her as risky and either decline her application or approve it at unfavorable rates. The second is the hourly worker who needs $300 to cover an emergency car repair before his next paycheck. He has a bank account, steady income, and no history of financial mismanagement, but no credit history, which means FICO cannot evaluate him at all. The Federal Reserve’s Survey of Household Economics and Decisionmaking found that 37 percent of American adults could not cover a small emergency expense using cash, savings, or a credit card, and 6 percent reported using a payday loan, pawn loan, or auto title loan in the prior year, products the CFPB estimates carry annualized rates of nearly 400 percent. The payday lending industry serves approximately 12 million American borrowers annually. It exists largely to fill the gap left by a mainstream credit system that treats the absence of a credit file as a disqualification.
Why the fix took so long
Cash flow analysis, which involves evaluating a borrower’s income, expenses, and account behavior over time rather than their bureau-reported credit history, is not a new concept. Banks have always analyzed income and expenses when underwriting mortgages and commercial loans, and the evaluation of a borrower’s repayment capacity is a well-established part of the underwriting process. The problem with applying that logic to consumer credit, particularly small-dollar consumer credit, was economic rather than analytical.
A $500 installment loan at a 36 percent annual rate generates roughly $30 in revenue. Manual review of a borrower’s bank statements, however brief, can easily consume that margin once staff time, compliance overhead, and processing costs are accounted for. The economics of large loans absorb that cost comfortably. The economics of small loans never did, which meant that alternative data analysis remained a good idea that mainstream lenders had no practical incentive to implement for consumer credit at scale.
Improvements in data aggregation, open banking infrastructure, and machine learning over the past decade changed the calculation: a borrower’s deposit account history can now be analyzed in seconds at a cost that no longer overwhelms the revenue on a small loan. Regulatory guidance followed the technology. In December 2019, the Federal Reserve, FDIC, OCC, NCUA, and CFPB issued a joint interagency statement endorsing the use of alternative data, specifically cash-flow data derived from bank accounts, in credit underwriting, noting that improving the measurement of income and expenses through cash flow evaluation “may be particularly beneficial for consumers who demonstrate reliable income patterns over time from a variety of sources rather than a single job.” The following May, the same agencies issued principles for small-dollar lending that explicitly approved deposit account activity as a basis for underwriting decisions. An October 2025 Federal Reserve brief called automated cash-flow underwriting a promising area for innovation and noted that it had finally made small-dollar alternative data lending economically viable, a threshold that manual analysis never cleared.
How lenders are using alternative data
Upstart: Education and employment as credit signals
Upstart, founded in 2012, was among the first platforms to systematically incorporate non-bureau data into personal loan underwriting. Its model evaluates education, field of study, employment type, and work history alongside traditional credit signals, using machine learning to weight more than 2,500 variables when assessing a borrower’s repayment likelihood. The underlying premise is that a borrower’s human capital, meaning their earnings trajectory and employment stability, predicts repayment capacity more accurately than the length of their credit history, particularly for borrowers who are early in their financial lives and have not yet had time to accumulate a meaningful credit record.
Upstart’s own 2024 Access to Credit Report, which compared its model against a hypothetical conventional benchmark using the same applicant pool, found that the Upstart model approved 43 percent more applicants and produced APRs 33 percent lower. The equity gains were substantial: the model approved 52 percent more Black applicants and 57 percent more Hispanic applicants than the conventional benchmark, at lower rates. A separate Harvard Business School study, using a CFPB-developed counterfactual model, reached similar conclusions, finding that the gains were concentrated among borrowers with higher incomes and educational credentials but thin credit files, and were strongest in areas with higher concentrations of minority and foreign-born residents, the populations for whom the gap between FICO-implied risk and actual default behavior was largest.
Petal: Underwriting from bank transaction data
Petal built its underwriting model entirely from raw bank account transaction data rather than bureau-reported credit history. When a borrower applies for a Petal credit card, the model pulls their transaction history and categorizes it into income, fixed expenses such as rent and utilities, and discretionary spending. It then evaluates net cash flow, income stability, and bill-payment behavior over time to generate what Petal calls a CashScore, a measure of creditworthiness derived from how a borrower actually manages their money rather than how long they have been managing debt. Borrowers with no credit file, evaluated using CashScore, performed at levels comparable to borrowers with prime bureau scores.
Petal subsequently spun the underwriting technology out as Prism Data, licensing it to other lenders as an API-based service, making cash-flow underwriting something any institution can add to its existing stack without building the capability independently.
SoFi: Cash flow and future earnings for thin-file borrowers
SoFi targets a specific segment of the credit-constrained population: borrowers who are financially healthy and earning well but have not accumulated enough credit history for bureau scores to represent them accurately. Its model layers free cash flow analysis and member-level behavioral data on top of bureau-reported scores, allowing it to assess a borrower’s actual financial position rather than the age and mix of their credit accounts.
For borrowers with strong earning potential but moderate current income, SoFi also incorporates future earnings projections into its underwriting assessment. A medical resident with a high debt-to-income ratio would appear risky to a conventional model evaluating current financials. SoFi’s model accounts for the probability that the same borrower will be earning significantly more within a few years, adjusting the default risk estimate accordingly. The platform’s advantage compounds over time: by the time a SoFi member applies for a loan, the platform has typically accumulated months of their deposit, investment, and spending behavior across its financial services products, giving it a richer view of their finances than a credit bureau can provide from account-level reporting alone. SoFi originated a record $10.5 billion in loans in Q4 2025 across personal, student, and home lending, with net charge-off rates on personal loans tracking well below its stated tolerance.
Cash App Borrow: Closed-loop behavioral underwriting for subprime borrowers
Cash App Borrow operates in the segment of the market that conventional lenders have most consistently declined to serve: small-dollar loans to borrowers with subprime credit profiles. The product offers loans between $20 and $500, with an average loan size of $153, and over 70 percent of its users carry FICO scores below 580, a population that most traditional lenders would decline at the application stage.
The underwriting model runs entirely on behavioral data generated inside the Cash App ecosystem. When a user requests a loan, the model evaluates their aggregate deposits, peer-to-peer transfer patterns, Cash App Card spending, Afterpay buy-now-pay-later payment history, retail investing activity, and prior Borrow repayment history. Bureau scores are not the primary input. Block, Cash App’s parent company, found in internal testing that removing bureau scores from the model entirely increased approvals by 38 percent without any corresponding increase in losses. The product’s repayment rate sits at 97 percent, across a borrower population FICO would have characterized as high-risk.
FICO Score XD: The incumbent responds
In response to the exclusion problem its original model created, Fair Isaac developed FICO Score XD, a parallel scoring system that uses utility payments, mobile phone bills, and cable subscription payments to generate scores for borrowers who fall outside conventional scoring criteria. Because nearly every American adult pays at least some of these bills, the model can score a substantially larger share of the population than the standard model. FICO Score XD now covers approximately 96 percent of U.S. adults, compared to 81 percent under the standard model. Fair Isaac building a product to compete with its own standard is, at minimum, evidence that the standard has limits.
Demonstrated but not yet adopted
The regulatory framework endorsing alternative data has been in place since 2019. Fintechs have demonstrated the model across different borrower segments, loan sizes, and underwriting approaches. Despite this, adoption across the lending industry remains limited. A 2024 survey of 125 banks, credit unions, fintechs, and other lending institutions found that only 43 percent of lenders supplement credit scores with alternative data in their risk assessments, even as 90 percent said that access to more alternative data would help them approve more worthy borrowers.
The friction is structural. Incorporating transaction-level data into underwriting requires either direct relationships with data aggregators or open banking infrastructure that many institutions have not yet built. Model risk management requirements impose compliance and validation costs that scale differently for large and small institutions. And the FICO score benefits from the self-reinforcing logic of any dominant standard: it is familiar, widely accepted by secondary market investors, and deeply embedded in origination and servicing systems that were designed around it. Replacing or supplementing it requires organizational commitment that an interagency statement can encourage but not compel.
Alternative data is past the proof-of-concept stage. The borrower populations that conventional models cannot evaluate, the 7 million credit invisible, the 25 million with thin files, and the larger number whose scores understate their actual creditworthiness, have been shown across multiple lenders and loan types to be more financeable than their bureau records suggest. The question now is a narrower one: whether the institutions that serve most of the consumer lending market will update their underwriting to reflect that, or whether the gap between what the data shows and what the industry practices persists for the same reason it always has, because the cost of changing has so far outweighed the incentive to do so.