Data Scientist Interview Questions, by Round

You have probably prepared SQL and some machine learning, and you are unsure what else is coming. This section maps the five rounds these loops reuse so your study matches the scoring. Each round tests a different thing, and preparing only the technical two is the most common way strong candidates lose offers.

• SQL and coding. Query writing plus Python data manipulation, usually live.
• Statistics and probability. Interpretation and reasoning, not proofs.
• Experiment design. A/B tests: what to measure, how long to run, when to trust the result.
• Modelling or case. An open problem you frame, then solve at a high level.
• Behavioral and stakeholder. How you handled ambiguity, disagreement, and results nobody wanted.

Titles vary. A company that calls the role data scientist may run an analyst loop, and vice versa. If you want the analyst version of this page, it sits in the interview questions by role hub.

Statistics and probability

• What does a p value actually mean, and what does it not mean?
• Explain a confidence interval to a product manager in two sentences.
• When would you use a t test instead of a z test?
• What is the central limit theorem and why does it matter for your work?
• A metric moved and the result is significant. Why might it still be wrong?
• Explain Type I and Type II error using a decision your team actually makes.
• What is selection bias and how would you detect it in your data?
• When is a median a better summary than a mean?

Experiment design and A/B testing

• How would you design an experiment to test a new onboarding flow?
• How do you choose sample size, and what happens if you cannot reach it?
• Your test is significant after three days. Do you ship it?
• What is a novelty effect, and how would you separate it from a real lift?
• How do you handle multiple metrics moving in different directions?
• What would you do if randomisation was broken partway through the test?
• How would you measure something you cannot randomise, like a pricing change across a whole market?

The pattern to notice: almost none of these have a single correct answer. They are scored on whether you name the assumption, state the tradeoff, and say what you would check.

Modelling

• How would you build a model to predict customer churn? Start with how you would define churn.
• Your model has 95 percent accuracy on an imbalanced dataset. Is it good?
• Explain the bias variance tradeoff using a model you have shipped.
• How would you decide between a simple model and a complex one for this problem?
• How do you know a model has stopped working after deployment?
• Explain regularisation to someone who has not used it.

SQL and Python

• Write a query returning each user's first and most recent purchase date.
• Compute a seven day rolling average of daily active users.
• Find the conversion rate by cohort month.
• In pandas, reshape a long table to wide and explain when you would not.
• How would you find and handle duplicate rows in a transactions table?

Product sense and metrics

• How would you measure whether a new feature is successful?
• Daily active users are up but revenue is flat. Investigate.
• Which single metric would you put on the leadership dashboard, and what would you leave off?
• How would you tell a good retention curve from a bad one?

Reading a question list builds recognition. Interviews test production under time pressure with someone interrupting you, and those are different skills. The gap shows up most in the experiment and case rounds, where the answer is a structured argument rather than a fact.

• Answer out loud, on a timer. Six minutes per case question, no notes, no restarting.
• Say your assumptions first. Naming what you are assuming is scored, and it also buys you thinking time.
• Practice being interrupted. Real interviewers cut in at minute two. Rehearsing an uninterrupted monologue does not prepare you for that.
• Write SQL by hand. Live rounds often use a plain editor with no autocomplete and no way to run the query.
• Explain every result twice. Once technically, once to a stakeholder who does not care about your method.

A data scientist with four years of experience prepared for a marketplace loop by reviewing a hundred questions in a document, then froze on "your test is significant after three days, do you ship it" because she had never said an answer to it out loud. The content was in her head; the delivery was not. If you want these drilled with follow-ups, the mock interview mode runs the round and pushes back on your answers.