RiskSpan

Only part of this business is AI. Without it, a very substantial business remains: a loan level data platform holding performance records on tens of millions of mortgages, and a normalization and validation pipeline. Scenario and stress testing run across more than seventy asset classes. The flagship prepayment and credit models are quantitative rather than learned.

The credit model is built on a delinquency transition matrix and the prepayment model on a two component turnover and refinance framework, both classical econometric constructions. The AI line is real but narrower and more recent. CascAIde, introduced in 2024, applies model driven extraction and a rules engine to portfolio risk management.

Machine learning is also applied to model validation to surface blind spots conventional techniques miss, and the company states it operates production systems processing billions of performance records. The shape is an established analytics platform with an embedded intelligence line.

The division of labor is clear, and the controls on it are only partly specified. The platform forecasts and values; a person trades, provisions or hedges on the output. Nothing in the material describes the system taking an action on its own. Analytics run automatically overnight at loan level so results are ready each morning. Users get interactive model diagnostics for back testing, a scenario library and the ability to save custom rate scenarios reflecting their own economic view.

That is a real mechanism for a user to interrogate the output rather than accept it. No threshold, tolerance or escalation trigger is described anywhere. Nothing states what happens when an overnight run produces results outside expected bounds, what the extraction component does when it cannot read a document confidently, or what the rules engine is permitted to decide without a person.

The models are open to inspection in a way that lets an institution validate them itself, without depending on someone else's testing. The quantitative team holds public monthly calls walking through how the prepayment and credit models tracked against actual behavior, naming what moved and why. Performance is discussed in the open on a fixed cadence rather than on request. Versioned update notes describe methodology changes in specific terms.

They include the split of the non qualified mortgage prepayment model into a unified turnover component and a refinance component segmented by documentation type, and the rebuild of the credit model onto a delinquency transition matrix framework. Clients receive interactive diagnostics for back testing rather than static reports. The company separately sells model validation and model risk governance as a service and publishes research on applying machine learning to model validation. One limit: all of this covers the quantitative models, and nothing comparable is published about the behavior of the AI extraction component.

The scale figures are substantial and the customer attribution is not. Published case studies carry quantified results, including an implementation that reduced client resource support needs by 46 percent and an asset manager replacing an inflexible dealer supplied risk system. Every one of them identifies the client only by type and location. Corporate scale is stated as clients managing over $30 trillion in assets and more than 45 million active mortgage loans.

External recognition is real but is not an outcome: a technology award from a housing finance trade publication and a listing on a major cloud marketplace. No financial institution is named anywhere in the published material, so no quantified result has a named customer beside it.

No safety practice, output screening, red teaming or data boundary is described. The pooled data question is pointed here because the business model depends on aggregation: the platform combines public agency performance data, purchased third party data and client contributed portfolio data into one analytical estate, and its models improve as that estate grows.

Nothing states whether performance data contributed by one institution informs models serving a competing institution, whether a client can exclude its holdings from model development, or what becomes of contributed data when a relationship ends. For clients who are trading counterparties to each other in the same collateral, that silence is a commercially material gap rather than a formality.

There is no public privacy policy detail, processing agreement, subprocessor list or retention schedule. The subject matter is loan level records describing individual borrowers, including credit score, loan terms, delinquency status and payment history. Those records cover tens of millions of active loans and are combined from public agency releases, purchased data and client portfolios.

Nothing states how long client contributed loan files are retained, what happens to them when a client leaves, or how borrower level attributes are segregated between the public and private sources they were assembled from.

Security is undocumented in public, with no certification, attestation, trust center or security page. The data position, rather than the company's size, makes that gap conspicuous. The platform ingests and stores loan level performance records covering tens of millions of active American mortgages, together with client portfolio holdings. Its buyers are banks, asset managers and insurers whose own third party risk processes will require evidence before onboarding.

A listing on a public cloud marketplace involves a commercial vetting step and is not a security assurance. A firm operating at this scale for these buyers is likely to have an attestation that is simply not published.

As an analytics and services supplier to regulated institutions, the company holds no license of its own. Its work sits close to supervised processes without being supervised itself. Outputs feed allowance calculations under CECL. It provides model validation and model risk governance services that institutions use to satisfy supervisory expectations on model risk, and its material notes regulators among the audience for loan performance analytics. No regulator engagement, examination, registration or independent assurance over its own validation practice is public.

No fairness testing, differential outcome analysis or bias disclosure appears, and the exposure is real even though no borrower ever meets this vendor. Credit models projecting delinquency migration at loan level feed which pools get bought, at what price, and what allowance a lender books against them, and the prepayment model is segmented explicitly by documentation type, which correlates with borrower category.

Nothing published examines whether forecast accuracy differs systematically across borrower cohorts, geographies or product types, which matters because a model that is well calibrated on average and poorly calibrated on a subpopulation prices that subpopulation's loans wrongly. The company publishes work on using machine learning to find model validation blind spots, which is the right instinct pointed at other people's models rather than at this question in its own.

The borrower has no relationship with this vendor and no knowledge of it, yet its forecasts influence whether their loan is bought, how their servicing is valued and what loss is provisioned against them. No liability position, accuracy warranty, error rate or remediation commitment is published. Between vendor and client the position is also unstated.

That gap matters because the output feeds regulatory capital and accounting judgments, which leaves open what an institution may rely on and what remains its own responsibility. Extensive documentation and back testing evidence support the institution's defense of a model it has adopted. That is a different thing from a stated allocation of responsibility.

Disclosure is generous at the data layer and silent at the model layer. Data sources are named, including a major property data provider for non qualified mortgage loan level performance, the government sponsored enterprises for agency performance data, and Snowflake for client integration. Nothing names a model provider, family, version or hosting arrangement for the AI extraction component, and nothing states whether any third party model sits in the pipeline. Knowing where the training data came from without knowing what was trained on it leaves an institution unable to enumerate its own downstream dependencies.

Several systems are named, and they sit at the data layer rather than the system of record. Snowflake integration is named. Distribution and deployment run through a named public cloud marketplace as either on demand analytics or a managed service. Loan level non qualified mortgage performance data is sourced from a named provider, and the platform ingests and validates client data alongside public agency data, with stated access to third party data vendors.

Nothing on the institution's own side is identified. No servicing platform, loan origination system, portfolio accounting system or order management system is named, and there is no public API documentation or integration count. The named plumbing is how data reaches the platform, not how the platform reaches the client's book.

Deployment is described in specific terms, and residency is absent. The platform is cloud native and offered in two distinct commercial forms: on demand analytics that the client runs itself, or a managed service the vendor operates. Both are published on a public cloud marketplace. A modular structure lets a client start with core data management and analytics and extend to custom workflows and user defined analytics.

Snowflake integration supports clients keeping data in their own estate. No hosting region, residency commitment, tenancy model or single tenant option is stated anywhere. For a platform holding client portfolio positions alongside borrower level records, that is a question an institutional buyer will ask before the first meeting ends.

Pricing is not public: no price, tier, minimum or billing basis is published. The structure is described clearly enough to show what would drive cost. The platform is modular and can be taken as on demand analytics or as a managed service with consulting attached. Nothing indicates whether charging is by asset under analysis, by loan count or by user, or instead by compute consumed or by subscription. The public cloud marketplace listing may carry pricing terms, though none are confirmed.

Breadth runs along two independent dimensions, and both are documented. By buyer, the platform serves investors, traders, portfolio managers and risk managers. It also serves servicing rights owners, private credit managers and the model validation functions inside institutions, with consulting delivered to asset managers replacing dealer supplied risk systems.

By asset, it covers more than seventy classes spanning residential whole loans, mortgage servicing rights, agency mortgage backed securities, private label securities and private credit. Reach is quantified: clients managing over $30 trillion in assets and more than 45 million active mortgage loans. Use cases extend beyond trading into CECL allowance calculation and climate risk exposure, which are different institutional functions rather than the same one described twice.