Three different kinds of work to look for
Stripe's Radar documentation distinguishes automated evaluation from manual review. Its dispute documentation describes a separate process triggered when a cardholder questions a payment. A dispute is not automatically evidence of fraud.
| Workstream | Possible assignment | Ask the manager |
|---|---|---|
| Risk operations | Review cases and document evidence | How are uncertain cases escalated and reviewed? |
| Disputes | Assemble records and support responses | Which stages would I support? |
| Risk analytics | Analyze outcomes, data quality or model performance | What data and analytical skills would I use? |
Use the posting to identify the work below. A queue-based case-review internship and a statistical model-evaluation internship can share the same ‘risk’ title.
Example: detecting more fraud creates more false positives
Consider 10,000 fictional attempted payments: 100 fraudulent and 9,900 legitimate. Assume all outcome labels are complete and comparable. Each row is an attempted payment, not a unique customer.
| Outcome | Approach A | Approach B |
|---|---|---|
| Fraudulent attempts flagged | 80 | 70 |
| Fraudulent attempts missed | 20 | 30 |
| Legitimate attempts flagged | 220 | 110 |
| Total flags | 300 | 180 |
| False-positive rate among legitimate attempts | 220 / 9,900 = 2.22% | 110 / 9,900 = 1.11% |
A catches 80% of fraud and B catches 70%. But 80 ÷ 300 = 26.67% of A's flags are fraudulent, compared with 70 ÷ 180 = 38.89% of B's. Detection rate and accuracy within the flagged group have different denominators.
Assume every flag blocks the payment, every missed fraud costs $100 and every blocked legitimate payment loses $20 of contribution. A costs 20 × $100 + 220 × $20 = $6,400. B costs 30 × $100 + 110 × $20 = $5,200. B is cheaper under these assumptions despite detecting less fraud.
Find the cost that would reverse the choice
A prevents ten more fraudulent attempts but blocks 110 more legitimate ones. At $20 per legitimate block, that extra friction costs $2,200. The approaches tie when each missed fraudulent attempt costs $2,200 ÷ 10 = $220.
Above $220, A has lower modeled cost; below it, B does. That conclusion assumes the same population, outcomes and blocking process. The exercise omits manual-review costs, recoveries, different payment amounts and longer-term customer effects. It is a comparison to explain, not a production screening recommendation.
Do not score unknown outcomes as legitimate payments
Recent attempts may not have had enough time for an outcome to emerge. A low observed fraud rate can therefore reflect incomplete labels rather than better performance. Compare groups with the same observation window and report unresolved outcomes separately.
Likewise, repeated attempts and unique customers are different counts. Keep a stable synthetic payment ID in a practice dataset and explain how retries are treated. Those definitions determine whether your rates answer the intended question.
Prepare for the work the internship actually offers
For analytics, build a spreadsheet with the counts, denominators and cost comparison above, then test one different loss assumption. Add an incomplete-label example and explain why it cannot be evaluated as though every outcome is known.
For operations, write a fictional case note separating the event, available evidence, uncertainty and next permitted check. Use synthetic data for either sample. Keep actual customer records and employer procedures in approved systems, and describe your practice work honestly in the application.