Card fraud detector for rare frauds
Inspired by publicly reported work at Stripe
Catch fraudulent card payments that are fewer than 1 in 500, without blocking good customers.
The brief
Stripe has written about how Radar uses machine learning to score payments for fraud, and why choosing a threshold is a trade-off between blocking fraud and blocking good customers. Build a fraud detector on public card data, handle the extreme imbalance, and choose a threshold from the cost of fraud against the cost of a blocked customer. Good looks like: precision-recall curves rather than accuracy, a cost-based threshold, and a short note for a merchant on what the score means.
Read the case first
What Stripe has said in public about this kind of work, on its own pages:
- How we built it: Stripe Radar (Stripe) ↗
- A primer on machine learning for fraud detection (Stripe guide) ↗
Make it your own
Set it in an Indian payment setting: UPI collect requests, cash-on-delivery orders or college fee payments. Write the costs for that merchant and show how the threshold changes.
AITeky has no link with Stripe, which has not reviewed or endorsed this brief. The brief only draws on what the company has published, uses open data rather than its data, and is a learning exercise: don't present your project as Stripe's work or system.
What you will build
- Explore the imbalance and amounts
- Train logistic regression and gradient boosting
- Evaluate with precision-recall and AUPRC
- Choose a threshold from fraud and false-decline costs
- Explain the score for a merchant
Free, open data to build it with
Sign up free to get the data and links
It takes a minute with your email. AiTeky Pass (₹999/month or ₹9,999/year) is free for your first 30 days.