Feature Store

1. The Incident

The fraud team ships a new model. Offline, it looks like a triumph. Evaluated against months of history, it catches 40% more fraud than the rules it replaces. In production, three weeks later, the numbers are in. It is catching less than the old rules. Same model, same weights. What changed between evaluation and production is everything else.

The investigation lands on one feature: card_velocity_1h, how many orders this card placed in the last hour