Data Scientist Interview Questions: The Complete Guide

Why the Data Scientist interview is different

A Data Scientist interview loop is unusually broad. In one loop you might be asked to write a SQL query, derive a probability from first principles, design an A/B test, debug a metric that dropped 10% overnight, and then explain all of it to a panel of non-technical stakeholders. Few other roles compress this many different skills into a single interview process.

Interviewers are evaluating whether you can:

• Pull and shape data yourself — SQL is not optional, even at companies that primarily use Python
• Reason about uncertainty correctly — the difference between “the numbers look different” and “the numbers are statistically significantly different”
• Design experiments that produce a trustworthy answer, not just a p-value
• Translate a business question into an analysis plan, and an analysis result back into a business recommendation
• Know enough applied ML to build and evaluate a model, without necessarily being a specialist in it

The role varies more by company than almost any other in tech — “Data Scientist” at a hedge fund, a growth-stage startup, and a Big Tech ads team can mean three different jobs. This guide covers the categories common to nearly all of them, and flags where company type changes the emphasis.

The typical Data Scientist interview loop

Analytics-heavy roles weight SQL, statistics, and case studies more; “full-stack” or ML-adjacent Data Scientist roles add a heavier applied-ML round closer to what an ML Engineer would face.

Category 1 — Motivation and background questions

Sample questions:

1. “Walk me through a project where your analysis directly changed a business decision.”
2. “Why data science here, specifically, versus a more analytics-heavy or more ML-heavy role elsewhere?”
3. “Tell me about an analysis you did that turned out to be wrong. What did you learn?”

What strong answers do: lead with the decision that changed, not the technique used. “I ran a regression” is weaker than “I found that discount depth had no effect on repeat purchase rate, which stopped the team from launching a promotion that would have cost $400K with no retention benefit.”

Category 2 — SQL and data manipulation

SQL is the single most consistently tested skill across Data Scientist loops, and it’s where strong analytical candidates most often lose points on avoidable syntax and edge-case mistakes.

Sample questions:

1. “Write a query to find the second-highest salary in each department.” (window functions: DENSE_RANK or ROW_NUMBER)
2. “Write a query to compute a 7-day rolling average of daily active users.” (window function with ROWS BETWEEN)
3. “Find customers who made a purchase in January but not in February.” (anti-join / NOT EXISTS)
4. “Given a table of user events, compute month-over-month retention.” (self-join or window lag on cohort tables)
5. “Write a query that returns duplicate rows in a table, including all the duplicates.” (GROUP BY + HAVING COUNT(*) > 1, or window COUNT() OVER (PARTITION BY ...))

Worked example (Q1):

SELECT department, employee_id, salary
FROM (
  SELECT
    department,
    employee_id,
    salary,
    DENSE_RANK() OVER (PARTITION BY department ORDER BY salary DESC) AS rnk
  FROM employees
) ranked
WHERE rnk = 2;

Say out loud why you chose DENSE_RANK over ROW_NUMBER: if two employees tie for the highest salary, ROW_NUMBER would arbitrarily pick one as “second,” while DENSE_RANK correctly treats the tie and moves to the next distinct value. Naming this trade-off unprompted is a strong signal — interviewers are grading whether you know why, not just whether the query runs.

Category 3 — Statistics and probability

Sample questions:

1. “Explain p-values to someone with no statistics background. What’s a common misinterpretation?”
2. “What’s the difference between a Type I and Type II error, and how do you trade off between them?”
3. “You flip a coin 100 times and get 60 heads. Is the coin biased? Walk me through your reasoning.”
4. “Explain the Central Limit Theorem and why it matters for A/B testing.”
5. “What’s the difference between correlation and causation? Give an example of a confound you’ve seen in real data.”
6. “Explain Bayes’ theorem with a concrete example — say, a medical test with a known false-positive rate.”
7. “When would you use a t-test vs. a chi-squared test vs. a Mann-Whitney U test?”

How these are graded: interviewers are listening for correct intuition explained simply, not textbook definitions recited fast. For Q1, a strong answer names the common misinterpretation explicitly: “a p-value of 0.03 does NOT mean there’s a 3% chance the null hypothesis is true — it means, if the null were true, we’d see data this extreme or more 3% of the time.” Candidates who skip that distinction are a common false-positive filter for interviewers.

Category 4 — Experimentation and A/B testing

Sample questions:

1. “Walk me through how you’d design an A/B test for a new checkout flow.”
2. “How do you decide on a sample size before running an experiment?”
3. “The experiment showed a statistically significant 2% lift, but the product team is skeptical. How do you investigate whether it’s real?”
4. “What is peeking, and why is it a problem? How do you avoid it?”
5. “You can’t randomize at the user level for this feature — what do you do instead?”
6. “How would you detect novelty effects or seasonality contaminating your results?”

A strong answer, worked example (Q1):

“First, I’d clarify the primary metric — what does ‘better’ mean here, conversion rate, revenue per session, or completion time? Then I’d check for a guardrail metric to make sure we’re not trading one for another, like refund rate. I’d compute the minimum detectable effect and required sample size given our current traffic and baseline conversion rate, so we know the test duration up front rather than stopping when it ‘looks significant.’ I’d randomize at the user level, not session level, to avoid contamination from users who see both variants. And I’d pre-register the analysis plan — the metric, the test, and the stopping rule — before looking at any data, specifically to avoid p-hacking by peeking.”

Notice the shape: metric first, guardrails named, sample size computed rather than guessed, randomization unit justified, and a pre-registered analysis plan — each element addresses a specific way A/B tests go wrong in practice.

Category 5 — Applied machine learning for data science

Sample questions:

1. “How would you build a model to predict customer churn? Walk through your approach end to end.”
2. “What evaluation metric would you use for a churn model, and why not accuracy?”
3. “How do you decide which features to include when you have 200 candidate features and limited time?”
4. “Explain overfitting to a product manager who doesn’t have a technical background.”
5. “You built a model with 95% accuracy on an imbalanced dataset. Why might that number be misleading?”

What strong answers include: business framing before modeling technique. For Q1, a strong answer starts with “what decision will this model drive, and what’s the cost of a false positive vs. a false negative” before naming an algorithm — a data scientist who jumps straight to “I’d use XGBoost” without framing the business decision is a common gap interviewers probe for.

Category 6 — Case study and product-sense questions

This category is somewhat unique to Data Scientist loops — open-ended business scenarios with no single correct answer, testing structured thinking under ambiguity.

Sample prompts:

1. “Daily active users dropped 15% yesterday. Walk me through how you’d investigate.”
2. “How would you measure the success of a new ‘save for later’ feature on an e-commerce site?”
3. “The company wants to know if raising prices by 5% will increase or decrease total revenue. How would you approach this?”
4. “Design a dashboard for a VP of Sales. What metrics would you include, and why?”

The framework to use, narrated out loud (worked on Q1):

1. Clarify scope first. Is the drop across all platforms, all geographies, all user segments? Sudden or gradual?
2. Rule out measurement issues before behavioral ones. Check for a tracking/logging bug, a timezone issue, or a pipeline failure before concluding users actually changed behavior — this is the single most commonly skipped step and the fastest real answer in practice.
3. Segment the drop. New vs. returning users, by channel, by device, by geography — find where the drop concentrates.
4. Correlate with known events. A deploy, a marketing campaign pause, a competitor action, a holiday.
5. Form a hypothesis and state what data would confirm or refute it.
6. Recommend a next action, not just a diagnosis — what would you tell the VP to do today?

Category 7 — Data visualization and communication

Sample questions:

1. “You have five minutes with a VP who is skeptical of your analysis. What do you say?”
2. “How would you visualize a result with high uncertainty without misleading the audience?”
3. “Critique this chart [interviewer shows a misleading visualization — e.g., a truncated y-axis].”

What panels are grading: whether you lead with the conclusion (similar to a BLUF format), whether you can explain uncertainty in plain language (“this could be anywhere from a 2% to 8% lift — we’re confident it’s positive, less confident about the exact size”), and whether you can spot a manipulated or misleading chart on sight.

Category 8 — Behavioral questions

Sample questions:

1. “Tell me about a time a stakeholder disagreed with your analysis or recommendation.”
2. “Describe a time your analysis was wrong, and how you found out.”
3. “Tell me about a time you had to say ‘we don’t have enough data to answer that.’”
4. “Describe a project where the data available didn’t match what you needed. How did you adapt?”

A strong STAR answer, worked example (Q3):

Situation: A product lead wanted a definitive answer on whether a new onboarding flow caused a specific downstream retention lift, based on three weeks of post-launch data with no control group. Task: Give an honest answer without either overstating confidence or being unhelpfully vague. Action: I explained that without a control group, we couldn’t isolate the onboarding change from seasonality or concurrent marketing changes — but I proposed a retroactive matched-cohort comparison against similar users from before the launch as a directional read, while flagging its limitations clearly in the write-up. Result: The directional analysis showed a plausible but not definitive lift; I recommended we run a proper randomized test for the next iteration rather than treat the retroactive number as ground truth, and the team agreed. What I’d do differently: I’d push earlier in the project for a held-out control group before launch, rather than trying to reconstruct one after the fact.

Category 9 — Company-specific patterns

Full question bank

Motivation / background

1. Walk me through a project where your analysis changed a business decision.
2. Why data science here, specifically?

SQL 3. Find the second-highest salary in each department. 4. Compute a 7-day rolling average of daily active users. 5. Find customers who purchased in January but not February. 6. Compute month-over-month retention from a raw events table. 7. Find all duplicate rows in a table.

Statistics & probability 8. Explain p-values, including a common misinterpretation. 9. Type I vs. Type II error — how do you trade off between them? 10. Is a coin that landed 60/100 heads biased? Show your reasoning. 11. Explain the Central Limit Theorem and its relevance to A/B testing. 12. Correlation vs. causation — give a real confound example. 13. Explain Bayes’ theorem with a concrete example.

A/B testing & experimentation 14. Design an A/B test for a new checkout flow. 15. How do you determine sample size before running a test? 16. A significant result looks fishy to the product team — how do you investigate? 17. What is peeking, and how do you avoid it?

Applied ML 18. Build a churn prediction model — walk through your approach. 19. What metric would you use for an imbalanced churn dataset, and why not accuracy? 20. Explain overfitting to a non-technical product manager.

Case study / product sense 21. DAU dropped 15% overnight — walk through your investigation. 22. How would you measure the success of a new product feature? 23. Design a dashboard for a VP of Sales.

Communication 24. You have five minutes with a skeptical VP — what do you say? 25. Critique a misleading chart.

Behavioral 26. Tell me about a time a stakeholder disagreed with your analysis. 27. Describe a time your analysis turned out to be wrong. 28. Tell me about a time you had to say “we don’t have enough data.”

Common mistakes that eliminate candidates

• Writing SQL that “mostly works” without handling ties, nulls, or duplicates. Naming the edge case unprompted is a stronger signal than getting to the answer fast.
• Reciting statistics definitions without the intuition. “A p-value is the probability of the null hypothesis” is a common wrong answer that eliminates candidates on the spot.
• Jumping to a model before framing the business decision it drives. Applied-ML answers that open with an algorithm name, not a decision, read as technique-first thinking.
• Treating case-study questions as a single-answer quiz. These are structured-thinking exercises; panels want to hear your process, not guess the “right” answer.
• Confusing statistical significance with business significance. A significant 0.1% lift may not be worth shipping — strong candidates say so.
• Weak behavioral answers with no specifics. “I communicated the limitations” without stating what was actually said reads as rehearsed.

A two-week prep plan

Week 1 — Technical foundations

• Days 1-2: Drill SQL — window functions, joins, GROUP BY/HAVING, and at least one query per day involving ties, duplicates, or nulls.
• Days 3-4: Refresh statistics — hypothesis testing, p-values, confidence intervals, common distributions. Write your own plain-language explanation for each of the top statistics questions above.
• Days 5-6: Drill A/B testing — sample size calculations, and 2-3 full experiment-design walkthroughs using the framework above.
• Day 7: Write 3 STAR stories covering a wrong analysis, a stakeholder disagreement, and a data-limitation situation.

Week 2 — Case studies and simulation

• Days 8-9: Practice 3-4 case-study/product-sense prompts out loud, timed at 30 minutes each, using the 6-step investigation framework.
• Days 10-11: Do a full timed take-home simulation on a public dataset — SQL extraction, cleaning, analysis, and a one-page write-up.
• Day 12: Run a mock interview covering one SQL round and one statistics/A-B-testing round back to back.
• Days 13-14: Research the specific company’s DS emphasis (see the company-specific table above), review your notes, and rest before the loop.

Go deeper

This tutorial covers the interview itself. If you want to build the underlying skills — from Python and SQL through statistics, A/B testing, and business intelligence — follow the structured learning path:

Data Scientist roadmap — the complete stage-by-stage path from foundations through certification, with free courses and tutorials at every stage.

Once you’ve built and practiced these skills, validate them with an industry-recognized credential:

Data Science Specialization certification — an adaptive exam covering the full data science pipeline, with a digital badge for your resume and LinkedIn profile.