You already know: the system design method (lesson 01), feature stores and point-in-time data (lesson 03), stream processing (D14), class imbalance and calibration (D3), and monitoring (D11). Fraud looks like binary classification and isn't, in four ways: mistakes have different prices, so the threshold is an economic choice; positives are rare, so accuracy is meaningless; labels arrive weeks later; and the opponent adapts. This case is the first in the bank to use the full fourteen-section format, and it sets patterns (cost-based tiers, early labels, rules beside models) that the spam, credit and ad cases reuse.
How to read this lesson. No code, on purpose. Numbers marked assumed are assumptions you state out loud; prices are illustrative.
1. The prompt, and the questions that change the design
A mid-level answer: "train a classifier on past fraud and block payments above 0.5". A staff candidate first asks what each mistake costs, which actions exist besides allow and block, how fast decisions and labels are, and who the adversaries are. Those answers produce cost-based thresholds, a challenge tier, real-time velocity and graph features with early labels, and rules that can change in minutes beside a model.
Assumptions you state: