Brief #20 · Machine Learning
Mobile Wallet Fraud Detection
Flag fraudulent transfers in a bKash-style ledger where only 1 in 500 is fraud: features from behaviour, a time-based split, and a threshold chosen by what mistakes cost.
Outlines are free · building with tracking needs Pro
Fraudsters are emptying accounts through fake 'you won a prize' calls. Our rule ('block transfers over 25,000 taka at night') blocks honest people and misses most fraud. Build a model that catches more fraud than the rule while bothering fewer honest customers, and explain every alert it raises.
What you'll build
- 1Generate or use a public transaction dataset where fraud is about 0.2%, with users, devices, times and amounts.
- 2Split by time (train on earlier months, test on the last one) and explain why a random split would cheat.
- 3Build behavioural features: amount compared with the user's usual, new device, new recipient, transfers in the last hour.
- 4Train logistic regression and gradient boosting, and compare them on precision-recall curves, not accuracy.
- 5Choose the alert threshold from the cost of a missed fraud versus a blocked honest transfer, and explain each alert with its top features.
Step-by-step walkthrough
Walkthrough coming soon
This project's step-by-step walkthrough, with real outputs at every step, is still being written. You can read the brief and try it on your own meanwhile; tracking opens when the walkthrough is ready.
Done when
- On the held-out month, the model catches more fraud than the rule at the same or lower false-alarm rate.
- Precision-recall curves for both models and the rule, with your chosen threshold marked.
- No feature uses information from after the transaction happened.
- A one-page model card: what it's for, data, metrics, known weaknesses.
Stretch goals
- Score transactions in real time from a stream, under 50 ms each.
- Simulate fraudsters changing tactics next month and measure how fast the model decays.