Prism Data

Without the models, all that is left is raw bank statements. Models do every step. They categorize disorganized transaction data into clean account activity and derive thousands of trended attributes across income, assets, expenses and behavior. They project future income and produce consortium trained scores that predict twelve month default, first party fraud and first payment default. The company states it has seen more financial transaction data over a longer period than any other provider in the category, which is a modeling advantage rather than a distribution one.

The output is a score and a set of attributes that the lender applies within its own credit policy, so approval, pricing and portfolio decisions stay with the accountable institution, and the design is explicitly for dropping into existing loan applications rather than replacing the decision process. Adverse action reason codes accompany the score, which means a declining lender can state why rather than pointing at an opaque number. What is absent is any description of confidence handling on an individual score, or guidance on how a lender should treat a cash flow assessment that conflicts sharply with a bureau score.

Model governance is visible in a shape a model risk function would recognize. The score is versioned publicly across successive releases in 2022 and 2024. Predictive power is expressed using a standard separation statistic rather than a marketing figure, and the training base is described as millions of records across lenders and products. The score is characterized as additive to conventional assessment rather than a replacement, which is a testable claim.

Independent research validated the efficacy of the approach. The limit is that the headline improvement is qualified as average lift based on the company's own analysis across client portfolios. No absolute figure is published, and there is no external performance audit.

A major credit bureau distributes the score as part of its own open banking suite, so an incumbent resells an alternative data product rather than competing with it. A named decision platform is supported. A lender serving independent contractors is named, with its chief risk officer quoted on using the score, income and attribute products together. A large consumer finance company took a strategic investment position, and its chief strategy officer commented publicly on the platform.

Independent research from a nonprofit innovation center studied the score empirically. Published operational figures show sub second response times and better than 99.9 percent availability, on nearly a decade of experience serving regulated institutions.

Prism Data answers the pooling question by making pooling the disclosed product. The flagship is described as the first consortium based cash flow underwriting model, built from millions of anonymized, consumer permissioned records drawn from many different banks, credit products and customer segments. The stated reasoning is that the best models rest on many observations and diverse data. Every participant therefore benefits from what the whole consortium contributes. The contribution is anonymized and consumer permissioned. It is a candid answer, since the alternative would be to obscure that a client's data improves a competitor's score.

Each step of the privacy architecture is described. The client pulls deposit account data through its own systems or any aggregator. That data is de identified before it is shared with the company, and processing returns outputs without the underlying records ever arriving attached to an identity. The consortium models behind the score were built from anonymized, consumer permissioned records, so both the technique and the consent basis are named.

Delivery is also offered through consumer reporting agency and non agency channels, so an institution can choose the legal framework that fits its use. That is a privacy posture designed around the statute rather than asserted alongside it.

Security assurance is the one conspicuous omission in an otherwise exceptionally well documented compliance posture. There is no public attestation, certification, trust center or enumerated framework. A major credit bureau has taken the score into its own product line, and banks and fintechs run it in live credit decisions, so it has passed security review repeatedly. Publishing the control set would complete an otherwise thorough disclosure.

Both governing United States statutes are named, and the obligations they create are addressed operationally rather than asserted. The company states compliance with the Fair Credit Reporting Act (FCRA) and the Equal Credit Opportunity Act (ECOA). It confirms the score may be used for approve, decline and pricing decisions, and supplies adverse action reason codes so lenders can meet their notification duty.

Delivery runs through both consumer reporting agency and non agency channels to suit different use cases. The company publishes an explanation of how disclosure duties differ between the two, including the consumer's right to request within sixty days the nature of non agency information used against them. That is a vendor engineering around the statute and teaching its customers the mechanics.

The decisive fairness evidence is independent rather than self reported. A nonprofit innovation center studied cash flow underwriting empirically across six non bank providers, including users of this score. It found the models appeared to predict creditworthiness within protected populations at least as well as traditional scores and attributes, and better in select cases.

The company states its models undergo rigorous fair lending review and have been shown not to create disparate impact. It supplies adverse action reason codes so declined applicants learn the reasons, and it explicitly targets consumers with no, thin or stale files. It also publishes the uncomfortable half: as many as 20 percent of prime and super prime consumers score poorly on cash flow and are higher risk. That shows the method reallocates credit rather than simply loosening it.

Recourse reaches the affected individual by law rather than by goodwill. Adverse action reason codes are supplied with the score, so a declined applicant is told which factors drove the outcome rather than that an algorithm said no. Delivery through a consumer reporting agency channel brings the full statutory dispute and correction apparatus with it.

The company also publishes the mechanics of the alternative path, including that a consumer declined on non agency information may request within sixty days the nature of the information used. The lender is served too, since documented reason codes and fair lending review are what it produces when an examiner asks. What is missing is anything covering the vendor's own error.

The data path is described end to end. Deposit account data originates with the client and is pulled through any aggregator or directly from client systems. It is de identified, then reaches the platform through aggregators, decision engines or a single endpoint connection. The consortium training base is characterized by breadth across banks, products and segments. The models are the company's own.

No individual aggregator is named, which matters because coverage and parsing quality vary between them and determine what the score sees. There is no published subprocessor list or hosting arrangement.

Every practical route into a lender's stack is covered and two are named. A major credit bureau carries the score inside its own open banking product set, which reaches institutions that would never contract with a startup directly, and compatibility with a named decision platform lets lenders deploy and iterate their cash flow strategies without engineering work.

Beyond those, data can arrive through any aggregator, through decision engines, or directly from client systems via a single endpoint interface, and outputs return in under a second with published availability. The company's stated ambition is to become common infrastructure the whole industry uses, and the integration surface is built for it.

Where the service is hosted is not published, and there is no region selection or residency commitment. Exposure is reduced by design rather than by disclosure, since data arrives de identified and the service is domestic. Lenders operating under supervisory expectations for third party processing would still expect the arrangement documented.

Pricing is not public, and neither is packaging or the basis of charge. The product line spans several scores and attribute sets that would ordinarily price separately, and distribution through a credit bureau introduces a second commercial route that is equally undescribed. Nothing indicates whether charge falls per score pulled, per applicant, by volume tier or as a license.

Product coverage spans essentially the whole consumer credit market, with the score usable for credit cards, personal loans, mortgages, auto loans, buy now pay later, cash advance, earned wage access and lease to own, plus small business credit, and specialised variants exist for short term and high risk lending where default behaviour differs. Buyers span banks, lenders and fintechs, reached directly and through a credit bureau's distribution. Applicant coverage is the deliberate extension: the score works for consumers with no credit file, thin files or stale files who conventional scoring cannot assess at all.