Insights · Modelling and pricing

Inheriting a Prophet model nobody wants to sign

It arrives the same way every time. The person who built it has left, the documentation is a tab called Notes, and the results are close enough to last year that nobody has had to ask hard questions. Then something changes — a product, a standard, a reviewer — and suddenly somebody has to put their name to it.

Start with reproduction, not review

The first task is not to understand the model. It is to reproduce a known result exactly: take the last signed valuation, run the model as it stands, and match it to the last decimal. Until that holds, every subsequent finding is unattributable, because you cannot tell whether a difference is a bug you found or a difference you introduced.

This step is skipped constantly and it is the reason remediation projects overrun.

Then work outward in three passes

• Inputs. Data feeds, assumption tables, product parameters. Most errors that survive years of use live here, because they are plausible rather than wrong-looking.
• Structure. Product logic, decrements, the projection mechanics. Slow to review, and where the genuinely serious errors sit.
• Output and reporting. Aggregation, mapping, the spreadsheet layer between the model and the report. This layer is usually undocumented and frequently the least controlled part of the whole chain.

Document as you go, not at the end. The understanding you have in week two is the thing you are being paid for, and it evaporates.

When to stop repairing

Rebuilding is usually cheaper than repairing when three things are true together: the product structure has moved so far from what the model assumes that changes require workarounds; nobody currently in the business understands the structural layer; and a regulatory or reporting change is coming that will require substantive modification anyway.

Any one of those on its own argues for repair. All three at once and the repair path becomes a series of increasingly expensive patches on a foundation nobody can explain — which is exactly how the model reached this state to begin with.