AiCurio

The removal test leaves a loan database. A deep learning artificial neural network is the entire product, described as the first commercially available machine learning model for residential mortgage loan performance and cash flows, and everything the company sells is an application of it: life of loan forecasting, portfolio surveillance, loss mitigation prioritisation and refinance targeting are all readouts from the same model. There is no rules based alternative that produces loan level monthly cash flow projections eight years forward.

The system recommends and a servicer decides, which is stated consistently in the company's own language of model recommended actions, prescribed loss mitigation and default strategies, and guidance for month to month life of loan decisioning. That framing keeps the accountable institution in the decision, which matters because the downstream actions include whether to pursue loss mitigation with a struggling borrower.

What is not described is any threshold, confidence indication or review point, and nothing states how a servicer should treat a prediction that conflicts with what it knows about a particular borrower's circumstances.

The accuracy claim is unusually well specified rather than merely large. It names the unit, the individual loan; the quantity, all monthly cash flows including principal, interest and servicing costs; the horizon, up to 96 months forward; and the logical entailment, that predicting every cash flow necessarily means predicting every default and prepayment at the same accuracy. Model validation is stated as completed.

The training window is the strongest element and it is rarely available: 22 years of loan and payment history spanning the housing collapse and its recovery, so the model has observed a complete credit cycle rather than a benign one, which is the single most important property for a mortgage default model. One caveat belongs on the record: the published figure rises from 96 to 97 to 98 percent across sources over time with no explanation of what changed.

Three commercial relationships are named with executives quoted. An outsourced mortgage services provider selected the platform for predictive analytics, its chief executive noting the depth of data and expecting it to serve his own banking and lender clients. A mortgage solutions firm launched a product built on it, describing the engine by name as the base of its own offering. A third arrangement pairs the model with a title decisioning product through an affiliated platform.

The training corpus is the other evidence, at more than 100 million loan records and several billion monthly payment records across 22 years. What is missing is commercial scale: no funding, revenue, portfolio volume or customer count is disclosed, and most published coverage clusters around 2020 with one later confirmation.

No data boundary statement was located. The model is trained on an industry wide historical corpus rather than on any one customer's book, which limits the usual concern, and the platform additionally runs inside other vendors' products as an embedded engine, so portfolio data flows through arrangements with intermediaries as well as directly. Nothing states whether a customer's own portfolio data contributes to model improvement, what happens to it when a contract ends, or how the embedding partners' obligations are constituted.

No data protection agreement, retention schedule, subprocessor list or sourcing disclosure was located, and the corpus is among the largest personal financial datasets described anywhere in this index: monthly payment histories for more than 100 million American mortgages spanning 22 years, with up to 300 data elements attached to each monthly record. That is a detailed longitudinal account of how tens of millions of households met or missed their largest obligation. Sources are described only as public and private, and nothing states what permissions underpin the private portion or how borrower level records are held.

No attestation, certification, trust centre or enumerated framework was located. Multiple partners have embedded or resold this engine inside their own products, which implies vendor assessment has been passed and creates a further obligation, since those partners inherit the security posture of a component they did not build. For a company holding payment histories on over a hundred million mortgages, a published control set is the obvious missing artifact.

The regulatory grounding is in the leadership rather than in a citation. The chief executive is a consumer finance and securitisation attorney recognised nationally in mortgage and banking law, who has represented financial institutions in securitisation transactions exceeding a trillion dollars over three decades, so the company is run by someone whose profession is the rules this product operates under.

Fair lending testing is named as a completed exercise, which engages the relevant regime directly. What holds this below the top grade is that no statute, regulation or supervisor is identified anywhere, so a buyer cannot see which obligations were tested against.

This vendor states something almost nobody else in the credit lane does: that it completed model validation and fair lending testing, and that the model passed without any statistically significant bias or economic impact. Naming the test and reporting the result is materially more than describing an intention, and it is the disclosure Stratyfy omits while selling bias mitigation as a product. Three things hold it below the top grade.

The claim is self reported, with no methodology, validator or date published. There is no demographic outcome data of the kind Upstart files publicly. And one use case carries its own exposure, since the model identifies which borrowers should be targeted for refinance marketing, and deciding who is offered better terms is a fair lending question in the same family as deciding who is approved.

No guarantee, indemnity or falsifiable commitment was located beyond the accuracy claim itself, which is asserted rather than warranted. The institutional customer can test predictions against realised performance over time, which is genuine recourse in a forecasting product.

The borrower has none and is materially affected: the model prescribes loss mitigation and default strategy for individual loans and determines who receives refinance offers, so a household's treatment in difficulty and its access to better terms both turn on a prediction it will never see, cannot contest and is not told about.

The input side is described by category with useful specificity, covering loan, property and market data from public and private sources, supplemented by economic and housing data, and quantified at more than 100 million loan records and several billion payment records with up to 300 elements per monthly record. That tells a buyer the shape and scale of the evidentiary base even without individual suppliers named. The model is the company's own.

What is not disclosed is any provider behind the private data, which matters because licensing terms on loan level performance data determine what the resulting model may lawfully be used for, and no hosting or subprocessor arrangement appears.

The most substantive integration is embedding rather than connection: a mortgage solutions provider built and launched its own analytics product with this engine named as its base, which means another vendor staked its product on the model and disclosed the dependency.

A second arrangement combines the predictions with a title clearance decisioning product so that originators see credit, collateral and title readiness together, and a third supplies the analytics to an outsourcing provider serving its own bank and lender clients. Three routes to market through partners is a real distribution position. What is not published is any direct integration into servicing or loan management systems, and no developer documentation was located.

No hosting provider, region selection, residency commitment or private deployment option was located. Exposure is domestic and therefore simpler than for international vendors, and the material involved is borrower level payment history at very large scale, held on behalf of regulated servicers and investors whose own examiners would expect the processing arrangement to be documented.

No pricing, packaging or basis of charge was located. One value figure is published, that model recommended actions can improve portfolio cash flows by more than 100 basis points annually and more on distressed pools, which frames the return without indicating the cost. Nothing states whether charge falls per loan analysed, per portfolio, by subscription or through the partners who embed the engine.

The buyer set spans the full chain of parties holding mortgage cash flows: loan owners, servicers, investors, originators, banks and credit unions, with the company describing its users as all owners and servicers of the mortgage cash waterfall. Instrument coverage reaches beyond whole loans into servicing rights and non performing and reperforming pools, which is the secondary market side of mortgage and the corner this index has had nobody in.

Functional breadth runs from origination targeting through portfolio surveillance to loss mitigation strategy. The limit is scope by design: this is United States residential mortgage only, and the company says so plainly.