How Should Stripe Capital Be Evaluated?

You are interviewing for a Data Scientist role at a fintech payments company similar to Stripe. The company is launching a product called **Capital**, which offers **pre-qualified loans** to existing merchants on the platform. Eligible merchants are selected using the company's internal merchant data. Repayment is collected automatically as **12% of the merchant's daily processed revenue** until the contractual repayment amount is completed. Assume you have access to merchant transaction history, loan offers, acceptance data, repayment outcomes, disputes and refunds, merchant attributes, and loan-level financial data. Answer the following: 1. **If you were building a dashboard for Capital, what metrics would you include?** - Be explicit about primary success metrics, guardrail metrics, and how you would segment the dashboard. - Consider tradeoffs between growth, repayment performance, portfolio risk, merchant health, and long-term unit economics. 2. **How would you decide that a merchant should not receive a pre-qualified loan offer?** - What early warning signals or predictive features would you gather? - How would you define a bad outcome in this setting, given that repayment is tied to daily revenue rather than a fixed installment schedule? - Discuss issues such as model calibration, selection bias, and policy rules versus predictive models. 3. **Should the company offer multiple loan options to each merchant, or a single loan amount?** - Discuss the pros and cons from the perspectives of conversion, merchant experience, self-selection, adverse selection, risk management, and operational complexity. - If you wanted to test this, how would you design the experiment and choose the evaluation metrics? 4. **Suppose Capital profit starts to decrease. How would you diagnose the problem?** - Provide a structured framework to decompose profit changes. - Consider changes in funnel conversion, merchant mix, credit quality, repayment duration, funding cost, pricing, collections, macro conditions, and data quality. - Explain how cohort analysis or segmentation could help avoid misleading aggregate conclusions.