You're making a lot of assumptions there. The guy probably had a training dataset with real cycle data and shopping data and went off of that.
I am, but that's because this is 1998. Where would you ever get such a training dataset with real cycle data tied to shopping data? Menstrual tracking apps weren't a thing then. And any anonymized medical studies that actually did such tracking certainly couldn't have been correlated with identifiable shopping data, I would think.
I think the idea is that the cycle is inferred by analyzing purchasing data.
The person I responded to literally said "training dataset with real cycle data".