UPI fraud-pattern detection with a cost-based alert threshold

Sahyadri Co-operative Bank processes about 9 lakh UPI transactions a day. Around 1 in 2,000 is later reported as fraud: mostly "collect request" scams, SIM-swap takeovers and mule accounts that receive many small payments and empty out within an hour. Last quarter customers reported ₹3.1 crore of fraud, and the bank refunded about a third of it.

Today the bank uses three hand-written rules, such as "block any transfer over ₹50,000 to a payee added in the last 24 hours". The rules catch about 30% of fraud and raise 4,000 false alerts a week. The fraud desk has six analysts. Each can review about 120 alerts a shift, so anything above roughly 700 alerts a day is simply not looked at.

The risk head wants to know whether a simple machine-learning model can do better, and she wants the answer in rupees. Missing a fraud costs the bank the average refunded loss (about ₹18,000). A false alert costs about ₹40 of analyst time plus an annoyed customer.

Real transaction data cannot leave the bank, so you will generate a realistic synthetic dataset with the fraud patterns built in, or use the public PaySim dataset as a stand-in. You will build a baseline from the current rules, train two models, and choose a threshold that minimises total cost within the 700-alerts-a-day limit.

Done looks like a notebook anyone can re-run, a cost table comparing rules and models, and a one-page note for the risk head with the threshold you recommend.