Data Scientist Interview Questions: 30 Questions With Sample Answers

Data scientist interview questions test four things: whether you reason correctly about uncertainty (statistics and experiments), whether you can build and judge a model, whether you can get data out of a database yourself, and whether you can turn a result into a business decision. The three questions you are most likely to get are "How would you design an A/B test for this feature?", "Explain the bias-variance tradeoff" and "Tell me about a model you built and how you knew it worked." This guide gives you 30 questions grouped by round, one line on what a strong answer shows for each, and five fuller sample answers we wrote.

How the data scientist interview usually runs

Stages vary by employer, team and level, so treat this as a typical shape rather than a rule. Product-analytics teams lean on experiments and metrics. Research and modeling teams lean on machine learning depth. Ask the recruiter which rounds you will get.

Some companies add a take-home assignment instead of, or in addition to, the technical screen. If you get one, the write-up matters as much as the code: state your assumptions, show what you tried and why you stopped.

Statistics and experiment questions

This is where most data scientist loops separate candidates. Interviewers want correct reasoning in plain language, not memorized formulas.

1. What is a p-value? A strong answer shows you know it is the probability of seeing data at least this extreme if the null hypothesis were true, and that it is not the probability the hypothesis is true.
2. How would you design an A/B test for a new checkout button? Shows you pick one primary metric, guardrail metrics, a randomization unit, a minimum detectable effect and a sample size before launch.
3. What goes wrong if you check results every day and stop when p is below 0.05? Shows you understand that repeated peeking inflates the false positive rate, and know fixes such as a fixed horizon or sequential testing methods.
4. What is statistical power, and what drives it? Shows you can name sample size, effect size, variance and significance level, and explain the trade-off.
5. Explain Type I and Type II errors with a product example. Shows you can tie abstract errors to business cost: shipping a useless feature versus killing a good one.
6. When would you not run an A/B test? Shows judgment: tiny traffic, network effects between users, ethical limits, or a change that is obviously required.
7. What is the central limit theorem and why does it matter for experiments? Shows you know why sample means are roughly normal for large samples, which is what makes common tests work.
8. You tested 20 metrics and one is significant. What do you conclude? Shows awareness of multiple comparisons and corrections such as Bonferroni or false discovery rate control.
9. Explain Simpson's paradox with an example. Shows you check segments before trusting an aggregate.
10. What is the probability of rolling at least one six in four rolls of a fair die? Shows you reach for the complement: 1 - (5/6)^4, about 0.518.

The American Statistical Association's statement on p-values (March 7, 2016) is the cleanest primary source for question 1. Two of its six principles are worth quoting in your own words: a p-value does not measure the probability that the hypothesis is true, and it does not measure the size or importance of an effect (ASA statement on statistical significance and p-values).

Machine learning questions

Expect depth on the models you list on your résumé. If a project says "XGBoost", be ready to explain how gradient boosting works and why you picked it over a simpler baseline.

11. Explain the bias-variance tradeoff. Shows you connect underfitting and overfitting to model complexity and can say what you would change in each case.
12. How do you detect and prevent overfitting? Shows held-out validation, cross-validation, regularization, simpler models, more data and early stopping, chosen for the situation.
13. What is the difference between L1 and L2 regularization? Shows you know L1 can push weights to exactly zero (feature selection) while L2 shrinks weights smoothly.
14. Your classifier has 98% accuracy on fraud data. Is it good? Shows you ask about class balance first, then move to precision, recall, PR-AUC and the cost of each error type.
15. When would you choose precision over recall? Shows you tie the threshold to business cost, for example spam filtering versus disease screening.
16. What is data leakage, and how have you caught it? Shows you know features that would not exist at prediction time, and splits that let the future leak into training.
17. How does a random forest differ from gradient boosting? Shows bagging of independent trees versus sequential trees that correct earlier errors, and the tuning and overfitting trade-offs.
18. How would you handle missing data? Shows you first ask why it is missing, then pick between dropping, imputing, or adding a missingness indicator.
19. How do you explain a model's prediction to a non-technical stakeholder? Shows you can use feature importance or example-based explanations without overstating causality.
20. How would you know a deployed model is degrading? Shows monitoring of input drift, prediction drift and delayed ground truth, plus a retraining plan. Production-heavy depth belongs to the machine learning engineer interview.

SQL, coding and product case questions

Most data scientist loops include at least one hands-on round. The SQL bar overlaps with analyst roles, so use our SQL interview questions guide and the LeetCode SQL 50 guide for drills.

21. Write a query for 7-day retention by signup cohort. Shows joins on dates, conditional aggregation and a clear definition of "retained" stated before you write code.
22. Find the second-highest salary per department. Shows window functions (DENSE_RANK()) and how you handle ties.
23. In Python, compute a rolling 7-day average per user. Shows pandas groupby plus rolling, and that you sort by date first. Our Python interview questions cover the language basics.
24. Weekly active users dropped 10%. How do you investigate? Shows you check data quality first, then segment by platform, region and user type, then look at what shipped.
25. How would you measure the success of a new recommendation feature? Shows one primary metric, guardrails, and a plan to separate novelty effects from lasting change.

Behavioral and situational questions

26. Tell me about a model or analysis that changed a decision. Shows impact, not just technique: what the decision was, what changed, how you measured it.
27. Tell me about a time your analysis was wrong. Shows ownership, how you found the error and what you changed in your process.
28. Describe a time a stakeholder disagreed with your result. Shows you can defend evidence calmly and still update when they have context you lacked.
29. A product manager wants to ship based on a test that is not significant. What do you do? Shows you lay out the risk honestly and offer options such as running longer.
30. You have two weeks and messy data. How do you scope the project? Shows prioritization: a simple baseline first, a clear stopping point and early check-ins.

Build these stories in advance with the free STAR story builder, and see STAR answer examples for structure.

Sample answers to common data scientist interview questions

These are examples we wrote to show structure and depth. Replace the details with your own work.

"Explain the bias-variance tradeoff."

Example (any level): "Prediction error has three parts: bias, variance and noise we can't remove. Bias is error from a model that is too simple to capture the real pattern, like fitting a straight line to a curve. Variance is error from a model that is too sensitive to the specific training sample, so it changes a lot if you retrain on new data. Making a model more flexible usually lowers bias and raises variance. In practice I look at training and validation error together: if both are high, I add capacity or better features; if training error is low and validation error is high, I regularize, simplify or get more data."

"How would you design an A/B test for a new checkout button?"

Example (product data scientist): "First I'd agree on the goal with the team. Say it's completed purchases per visitor, so that's the primary metric, with average order value and refund rate as guardrails. I'd randomize by user, not by session, so one person always sees the same version. Then I'd pick the smallest lift worth shipping, maybe one percent relative, and compute the sample size for 80% power at a 5% significance level. That tells me the run length, and I'd round up to whole weeks to cover weekday patterns. I would not stop early on a good-looking day. After the test I'd check that the split was actually 50/50 before reading results."

"Your fraud model has 98% accuracy. Is it good?"

Example (any level): "I can't tell yet. If two percent of transactions are fraud, a model that predicts 'not fraud' every time also scores 98%. I'd look at the confusion matrix, then precision and recall for the fraud class, and the precision-recall curve because the classes are imbalanced. Then I'd ask the business what each error costs. A missed fraud might cost hundreds of dollars, and a false alarm costs a support ticket and an annoyed customer. That trade-off sets the decision threshold, not accuracy."

"Tell me about a time your analysis was wrong."

Example (experienced hire): "I reported that a new onboarding flow lifted 30-day retention by four points. Two weeks later a teammate noticed the treatment group had more users from one paid campaign, because the assignment happened after a redirect that some ad links skipped. I told the product lead the same day, reran the analysis on the correctly randomized users, and the lift was under one point and not significant. Since then I add a sample-ratio check and a pre-period comparison to every test readout, and I've caught two more assignment bugs that way."

"Weekly active users dropped 10%. How do you investigate?"

Example (new grad): "I'd first confirm the drop is real: check whether logging, the metric definition or a data pipeline changed that week. If the data is sound, I'd segment by platform, country, new versus returning users and app version, because a drop in one segment points to a different cause than an even drop everywhere. Then I'd line the timing up against releases, marketing changes and holidays. I'd come back with the most likely cause, the evidence for it, and what data would confirm it."

What interviewers listen for, by skill

Questions to ask the interviewer

• "How do data scientists here split time between experiments, modeling and ad hoc analysis?"
• "How does a model or analysis go from a notebook to a decision or to production here?"
• "What does the experimentation platform handle for you, and what do you build yourselves?"
• "What would a strong first 90 days look like in this role?"

More options, sorted by interviewer, are in our list of questions to ask the interviewer.

How to practice data scientist interview questions

Reading answers is not the same as saying them. Statistics answers in particular fall apart when spoken for the first time, because the order of ideas matters. A practice routine that works:

1. Pick five questions from each group above and answer each one out loud in under two minutes.
2. Record yourself, then check: did you define terms before using them, and did you give a decision at the end?
3. Do two SQL problems and one pandas problem a day without autocomplete.
4. Rehearse one project walkthrough until you can explain the baseline, the metric, what failed and the result.

CoPilot Interview's Mock Interview builds questions from your résumé and the job description, reads each one aloud, and scores your spoken answer out of 10 with what to fix and a model answer. The free browser mock interview has a data science and ML track with no install. If you are also interviewing for analyst roles, compare this list with our data analyst interview questions, and see data science interview help for how each round is structured.

FAQ

What are the most common data scientist interview questions?

The most common are how to design an A/B test, what a p-value means, the bias-variance tradeoff, how to evaluate a classifier on imbalanced data, a SQL retention or ranking query, and a walkthrough of a model you built and how you knew it worked.

How much statistics do I need for a data scientist interview?

You need hypothesis testing, p-values, power and sample size, confidence intervals, the central limit theorem, multiple comparisons and basic probability. Interviewers care most that you can explain each one correctly in plain words and apply it to an experiment.

Do data scientist interviews include coding?

Most do. Expect SQL at minimum, often Python with pandas, and sometimes a take-home assignment. Algorithm-heavy coding is less common than in software engineering loops, but it varies by company, so ask the recruiter.

How is a data scientist interview different from a data analyst interview?

Data scientist loops usually go deeper on statistics, experiment design and machine learning, while analyst loops focus more on SQL, dashboards and business metrics. Both test product sense and communication.

How should I talk about a machine learning project in an interview?

Cover the business goal, the baseline you started with, the metric you optimized and why, what did not work, the final result in numbers, and what you would do differently. Interviewers often probe the failures more than the wins.

How can I practice data scientist interview questions for free?

Answer questions out loud and record yourself. CoPilot Interview has a free browser mock interview with a data science and ML track, and its desktop Mock Interview scores each spoken answer out of 10 on the free plan.