Financial Crime & Compliance · Module 2 of 4
The Peer Group
A money-service business on Keystone's book moves $1.45m in an entirely ordinary month. In one peer group that month is invisible and the system would find laundering inside it about nine times in a hundred; in another the same month, unchanged in every particular, is detected once in every eight hundred. Nothing about the customer moved. What moved is who they were being compared to, and telling the system who those people are is a field on a customer record that, in almost every bank, was populated once from the risk-rating field by somebody who thought of it as plumbing. It is most of the decision. Seven customer types have to be arranged into peer groups, which is 877 possible arrangements, and three forces decide what a band ends up being: how alike the group is, how much customers vary anyway, and how confident you can be about where the centre even sits, a term that is inversely proportional to the number of customers, and the reason a group of one has no normal. Two people in the room are both right and want opposite things: Model Risk wants fewer groups because every group is a model with an owner, a document and an annual review, and the head of the alert team wants narrower bands because every alert is a real person spending a real hour. The trap underneath is that a bank holds two orderings of its customers, by how risky they are rated and by how much money they actually move, and only one of them says what a monitoring band should be built on. Every number a monitoring function reports about itself improves as its bands widen, so which dial can you trust? Then a commercial win in an unrelated segment moves one customer type's behaviour and size together: does the obvious repair help, or hurt?