Model Development

Our development process starts before a single parameter is estimated. We begin our engagements with meetings with the line-of-business managers who own the portfolios being modeled — the people who understand how the loans were underwritten, how the portfolio has behaved, and where the data limitations are. We have been told no other outside firm makes these meetings standard practice. We do.

From there we follow a transparent, fully documented process through developmental database construction, model design, estimation, testing, and implementation. We document successes and dead-ends with equal care. Our code and models are built to be controlled, reproducible, and explainable — not black boxes that require us to operate them for you.

Every development engagement includes an independent parallel verification process: a second set of calculations to confirm model outputs before anything goes to a board, an examiner, or a regulator. This is not an add-on. It is standard practice on every engagement.

Our team uses AI tools — including generative AI for data exploration, preliminary code development, and documentation drafting — as part of our standard workflow. Every AI-generated output is produced in short, reviewable segments, independently tested, and documented alongside the human-authored components. We document exactly where and how AI was used so your validators and examiners can see the seam between human judgment and algorithmic output. This is how AI should be used in a regulated environment — and it is how we work.

Two-thirds of any development engagement is data preparation. That is not the natural cost of modeling — it is the cost of poor data governance. Client data almost never arrives in the form needed to build reliable models. Merger histories create fragmented data warehouses. Production systems were never designed to support model development. Key variables were never tracked or were tracked inconsistently.

We treat data preparation as a modeling problem, not an administrative task. The developmental data warehouse we build is itself a deliverable — structured, documented, and designed so your team can maintain it and add to it after we leave.

Our approach is robust in the statistical sense: we look for corroborating evidence rather than stopping at the first model that fits. When the data allows two reasonable specifications, we build both and explain why we chose one. When it doesn’t allow a strong choice, we say so and document the constraints. That discipline is what allows our models to hold up in examiner review. That is what it means to elevate the industry standard — not to cite it as a justification, but to examine whether it actually applies to your portfolio.

AI tools accelerate the labor-intensive parts of this process — data exploration, pattern recognition, preliminary specification testing — without replacing the senior judgment that determines whether a model is right. A four-person team using AI intelligently can produce output that competes with firms ten times its size. Not because the AI replaces expertise — it does not — but because it frees the expertise to focus on the decisions that matter.