What would your risk committee ask before it signed off on a synthetic population?

Thank you for downloading the white paper, “The Compliance Function Is a Model Risk Problem. Treat It Like One.”

You likely finished it with one question: If your detection models cannot be validated against a censored historical record, and the proposed answer is synthetic populations, what validates the synthetic populations? It is the right question, and it is the one this second paper answers.

Most write-ups of synthetic data start at generation. That is already too late. By the time a persona exists, the questions your risk committee will ask have either been answered upstream or never will be: what data it was derived from, who approved that data for this use, how fidelity was measured, and who can use the result and for what.

“It was generated, so it must be safe” is an assumption. The paper walks through the gate that tests it and the record it has produced so far: 74% to 92% alignment with real-world outcomes on feature and pricing priorities, onboarding attrition predicted within 87% of the observed result. All inside your own AWS account, where the evidence is yours.