In this article
What this page covers
Each entry gives a short answer, one line on why it is asked, and a link for more depth. The questions are reworded from candidate reports; the page does not claim any company asks them. For the wider family of case rounds, start from the product and analytics case study guide.
How to structure any domain case
What is a domain case round, and how does it differ from a product case?
A domain case round puts a business problem from the company's own field in front of you, for example fraud at a bank or tax rules at a hiring platform. A product case stays with generic product reasoning. In a domain case you are also judged on whether you ask what the field's constraints are, such as money lost, regulation or customer trust, before you propose anything.
Why it is asked: it checks whether you can reason in an unfamiliar field. Read the full explanation: product and analytics cases.
Why do companies outside big tech run domain rounds?
Companies outside big tech run domain rounds because they need people who can learn their field quickly, and a problem from their own business tests that more directly than a generic puzzle. No company states this publicly; treat it as a reasonable reading. Prepare by practising the method, not the industry facts.
Why it is asked: the round samples how you learn a field on the spot. Read the full explanation: product and analytics case study guide.
What structure works for an unfamiliar domain case?
Use the same six moves each time: clarify the goal and who is affected; name one success metric and one guardrail metric; split the problem into segments; state hypotheses; recommend an action with the data you would check first; and name the risks. Say your assumptions aloud and label any number you invent as an assumption.
Why it is asked: structure is visible even when the field is new. Read the full explanation: Meta product architecture interview.
Knowing it is half the job. Saying it is the other half.
Interview Clozo is an AI interview copilot for real interview rounds: it hears the question and puts a strong answer on your screen, and you say it in your own words. See the features or how it works.
Start Free Trial →Pricing and promotions
How can promo results guide pricing in every market?
Treat the promo as an experiment on price sensitivity. Compare buyers who saw the discount with a comparable group who did not, estimate how demand changed, and then ask whether that sensitivity holds in another market. Local income, competitors and habits differ, so use the estimate as a prior, test it in each region with a small experiment, and keep a guardrail such as margin per order. To compare results across currencies, express a discount as a share of the local price, or adjust for purchasing power.
Why it is asked: it tests causal thinking applied to a pricing decision. Read the full explanation: reading an A/B test like a product owner.
Is a percentage discount or a fixed-amount discount better?
It depends on the price spread. A percentage discount scales with the price, so it gives expensive items the largest cash discount. A fixed discount is the same cash on every item, so it is a larger share of cheap items. In an illustrative basket of items priced 20, 50 and 100 units, 20% off is 4, 10 and 20 units off, while a flat 10 off is 50%, 20% and 10% of the price.1 Pick by the margin you can afford.
Why it is asked: it checks arithmetic and the link to margin. Read the full explanation: ads and marketplace domain questions.
How do you tell whether a promotion actually caused extra orders?
Incrementality is the extra outcome a promotion causes beyond what would have happened without it. Average uplift in a promo week is not incrementality, because some buyers would have bought anyway and some pulled purchases forward. Use a randomised holdout (randomised controlled experiments are the standard for causal effect, per Kohavi et al., 2009) or a matched comparison group, count only the difference, and subtract the discount given to people who would have bought at full price.
Why it is asked: confusing lift with causal effect is a common error. Read the full explanation: reading an A/B test like a product owner.
What is price elasticity, and how would you use it?
Price elasticity is the percentage change in quantity demanded divided by the percentage change in price. If it is below minus one, demand is elastic, so a price cut raises revenue; above minus one, a cut loses revenue. In an illustrative case with elasticity of minus two, a 10% cut lifts units from 100 to about 120, so revenue rises from 100 to about 108.1 Estimate it from experiments, not from one promotion.
Why it is asked: it tests whether you can turn a definition into a decision. Read the full explanation: data science interview questions.
Fintech and fraud
Which bank transactions are most likely to be fraud, and what features would you use?
Start with behaviour that breaks a customer's own pattern: a new device, an unusual hour, a far-away location, a sudden jump in amount, many small tests in a row, or a new payee. Add history features such as spend over the last day against the customer's average. Check each feature for leakage, meaning information only known after the fraud was confirmed.
Why it is asked: it tests feature thinking and leakage awareness. Read the full explanation: data science interview questions.
Why is accuracy a poor metric for fraud detection?
Fraud is rare, so a model that flags nothing is almost always right. In an illustrative set of 100,000 transactions with 500 frauds, flagging nothing is 99.5% accurate and catches no fraud.1 The scikit-learn documentation says precision-recall "is a useful measure of success of prediction when the classes are very imbalanced" (scikit-learn, Precision-Recall example, accessed October 2026).
Why it is asked: it separates metric habit from metric judgement. Read the full explanation: data science interview questions.
What are precision and recall, in plain words?
Precision is the share of flagged transactions that really are fraud. Recall is the share of all real fraud that gets flagged. Raising the threshold usually lifts precision and lowers recall, and lowering it does the reverse. Say which error costs more: a missed fraud costs the loss, while a false flag costs a review and an annoyed customer. The formal definitions are in the scikit-learn user guide.
Why it is asked: you need both terms to discuss any threshold. Read the full explanation: data science interview questions.
How do you judge whether a fraud model is profitable?
Convert each outcome into money: prevented loss for a caught fraud, review cost for every flag, and a cost for annoyed customers if you can estimate it. Then compare thresholds by net value. In the illustrative example, a missed fraud loses 120 units and a review costs 30; the middle threshold nets 21,000 units against 16,500 for the high bar, while the low bar loses 13,500.1
Why it is asked: it moves from model scores to a business decision. The table below shows the worked example.
Text description of the diagram
- The balanced threshold has the highest net value in this illustrative example.
- The low bar catches 450 of the 500 frauds but flags 2,250 transactions.
The balanced threshold has the highest net value in this illustrative example. The low bar catches 450 of the 500 frauds but flags 2,250 transactions, and review cost outweighs the loss it prevents.1 Change the review cost and the winner moves, so state your assumptions and show the sensitivity.
What are the benefits and drawbacks of a virtual credit card?
Benefits: a number can be limited to one merchant or one purchase, so a leak does less damage, and it can be issued and cancelled quickly. Drawbacks: some merchants and hotels need a physical card at pickup, refunds and disputes can be harder to follow, and customers may lose track of many numbers. Weigh these against the fraud saving.
Why it is asked: it checks product reasoning before a coding or logic part. Read the full explanation: product and analytics cases.
Compliance and global hiring
How would you explain hiring in Germany for a customer based elsewhere?
Name the main choices without claiming legal detail you do not know. The customer can hire through its own local company, or work with an employer of record (EOR), which is a firm that legally employs the worker locally while the customer directs the daily work. State that employment law, tax and social contributions are local, ask about the role and duration, and note that in Germany an EOR arrangement can fall under temporary-agency-work rules. Say a lawyer must confirm the specifics.
Why it is asked: it tests structured reasoning under regulatory uncertainty. Read the full explanation: product and analytics case study guide.
Contractor or employer of record: what are the trade-offs?
They are not equivalent: a contractor is self-employed and engaged under a services contract, while an EOR is the worker's legal employer. A contractor is simpler and cheaper to start but risks misclassification if the work looks like employment. An EOR gives employee protections and local compliance at a higher fee. Say that the test varies by country.
Why it is asked: it tests whether you see a legal risk behind a cost choice. Read the full explanation: product and analytics case study guide.
A new country tax law affects every customer's staff. What do you do?
Triage first: which countries, customers and workers are affected, and from what date. Then get the rule interpreted by tax specialists, update payroll logic behind a flag, notify affected customers with what changes for them, and test on historical payrolls before release. Track errors found by customers and late corrections as guardrails.
Why it is asked: it tests calm prioritisation when the rules change. Read the full explanation: product and analytics case study guide.
Product diagnostics
Users say the product is slow on very large pages. How do you diagnose it?
Reproduce it, then measure where the time goes: network, server, rendering or the browser's own work. Segment by page size, device and browser, and check whether it began after a release. Likely causes are loading or drawing the whole page instead of the visible part, and heavy re-computation on each edit. Propose loading on demand and set a latency target per page size.
Why it is asked: it tests a diagnosis habit, not a single answer. Read the full explanation: how to investigate a drop in watch time.
How do you walk through a past production incident in a case round?
Tell it as a timeline: how it was detected, the impact, the cause, the fix, and what changed so it is less likely to repeat. Keep to what you know and mark guesses. The strong part is the follow-up: a better alert, a safer rollout, or a runbook. Do not blame individuals; describe the system gap.
Why it is asked: it shows how you reason about failure and learning. Read the full explanation: how to investigate a drop in watch time.
Follow-up questions
How do I prepare for a domain case round without industry experience?
Learn the field's money flow in plain words: who pays whom, where losses arise, and what a regulator cares about. Then practise the six-move structure on two or three problems out loud. A strong answer asks good questions about the field, which shows learning speed better than facts do.
Should I ask clarifying questions first?
Yes, but briefly: goal, who is affected, the constraint, and the time horizon. Restate what you heard, then commit to an assumption so the interviewer can steer you instead of waiting.
How do I show my numbers are assumptions?
Say "assume" before the number and give a range or a sensitivity. The worked fraud example above is labelled illustrative for exactly that reason, and a change in one input can flip the answer.
What if I do not know the regulation?
Say so, state the general principle, and name who would confirm the detail. Admitting the boundary while still giving a safe plan is stronger than a confident guess.
- Use one structure every time: clarify, metric and guardrail, segments, hypotheses, recommendation, risks.
- Judge a rare-event model with precision and recall and a money threshold, not accuracy.
- Discount average lift; ask what the promotion caused beyond what would have happened anyway.
- Choose between percentage and fixed discounts by price spread and margin.
- In compliance cases, state the options, the local-law caveat and who confirms the detail.
Sources
- Precision-Recall (scikit-learn documentation) — scikit-learn, 26 Oct 2026
- Metrics and scoring: quantifying the quality of predictions (scikit-learn user guide) — scikit-learn, 26 Oct 2026
- Controlled experiments on the web: survey and practical guide (Kohavi et al.) — Data Mining and Knowledge Discovery (Springer), 1 Feb 2009
