I won’t say why or how, but I have insider info from RAT on what’s behind all this. This shutdown is part of a series of upcoming shutdowns. This week’s was a beta test for a new machine learning model specifically designed to target churning activity through unsupervised learning. Amex data scientists are improving an evolving neural network model that identifies churning behavior as it happens, and then assigns risk scores that factor in to anti bust-out or fraud scores. They’re treated the same, those all have different scores but their values affect your overall risk/reward score for Amex. If that overall score reaches a certain threshold (which they’re still experimenting with what exactly it will be), you will get shut down. What they’re doing is based on the same ML models that are used at Amex for AML and fraud prevention. The shutdowns will happen in small waves because one giant hit would be a massive drain on resources and disruption to their overall business, including drawing negative public attention outside of the churning community. That would scare people who are profitable but think they’re winning. They are also incrementing shutdown waves because of the resource usage it takes to identify accounts and assign a score based on the new indicated risk factors. When shutdowns are triggered, they begin combing through the new dataset. Right now, the model only looks back two years.

The clawbacks on self-referrals were a pilot program to this week’s events, not an entirety separate event. The same model that identified self-referrers was used for these shutdowns.

I’ll report back when I hear more info.

Also, I should add that if you had self-referrals clawed back, you’re already in the main data set. Those who were hit by the clawback were flagged as the highest priority for the model to comb through because of the strong correlation to proven churning/gaming behavior. Self-referrals were like a litmus test for who will be the low-hanging fruit for them to cut. As they tune the model to identify more positive activity correlations with churning, it will get more specific. Combing through level 3 data will become more prevalent in identifying churners, especially true if you have transactions from GCM and other obvious sites in your transaction history. If you were hit in the clawback but weren’t shut down this time, be cautious that you’re in their crosshairs. From what I’ve learned about the model, the inferred course of action here to avoid a shutdown is to not do anything with Amex except organic spend on your existing cards. Don’t carry a balance, no more MS, and no new cards. Keep in mind that the anti-churning model is fundamentally based on the neural architecture of their anti-fraud model.

PS I’ve read many comments loosly speculating about what RAT is. RAT isn’t a bunch of guys in India who are manually combing through the accounts with high MR balances, or checking referral sources for validity, etc.. This isn’t true. “RAT” is two data science and machine learning engineers in Phoenix who transitioned from the Amex fraud prevention team. There are two other engineers who have shared responsibilities with RAT and other anti-fraud teams.

The Hilton Aspire no AF link had absolutely nothing to do with these shutdowns. It’s a pure coincidence. The portfolio of recommended accounts to be shut down was sent off for approval more than a week ago. RAT doesn’t shut down accounts. They provide tranches of customer accounts suspected of abuse – along with correlating data. Those tranches go to evaluation from non-engineering management; who are the decision makers on shutting down a tranch of accounts or enacting other penalties. This is new to them and it’s somewhat uncharted territory, the speculations on here and on Reddit that they’re winging it are accurate. Their plan is to cut losses and effectively disable Amex churning first, then deal with the fallout afterwards. They’re legally within their rights so shutdown approvals are not made on a legal basis.