Data scientist interview questions: 23 questions and a scorecard for interviewers
The short answer
Good data scientist interview questions test how the candidate turns a business problem into something a model or experiment can answer, designs and reads A/B tests, builds and evaluates models honestly, checks data quality, explains results to people who are not technical and gets models into production responsibly. Ask them to walk through one project from the business question to what happened after it shipped, and listen for how they chose the approach, how they knew it worked and what they did about bias, privacy and model drift. If the client uses a take-home or live exercise, keep it short, give every candidate the same brief and time, and score it against criteria agreed in advance. Score every candidate against the same five criteria so the decision rests on evidence.
All 23 questions with why you ask each one, what a strong answer shows and follow-ups, plus the data scientist scorecard and rating guide. Free to download and adapt; no sign-up needed.
When to use these questions
Data scientists use statistics, experiments and machine learning to help a business make better decisions or build better products, so an interview needs to show how someone frames the right problem, chooses a method that fits it and proves the result is real, not only which libraries and algorithms appear on their resume. The strongest evidence comes from one project described in depth: the business question, the data, the baseline, how the model or experiment was evaluated and what changed because of it.
The title covers different jobs, from analytics roles built around experiments and metrics to machine learning roles that ship models inside products, so agree with the hiring manager which side matters most, which languages and platforms the team uses, how much production engineering the role involves and what kind of data the person will handle. For agency recruiters, the screen before the client’s technical rounds is where to confirm those essentials and judge communication directly; leave detailed modeling and coding assessment to the client’s data scientists.
23 data scientist interview questions
Grouped by what they test. Pick the questions that match the role, ask every candidate the same ones in the same order, and score each answer against the scorecard below.
Problem framing and business impact
Start with one real project in depth. Look for a candidate who asks what decision the work supports before choosing a method.
Question 1: Walk me through a data science project from the original business question to what happened after it was delivered. What was your part?
- Why ask it:
- Establishes what the candidate owned and whether the work changed anything.
- A strong answer shows:
- A clear business question, a baseline to beat, their own decisions on data and method, and a measurable result or decision that followed.
- Follow-up:
- What was the simplest approach you considered, and why did you go beyond it?
Question 2: A manager asks you to “use machine learning to reduce customer churn.” What do you ask before you touch any data?
- Why ask it:
- Tests whether the candidate frames the problem before picking a technique.
- A strong answer shows:
- Questions about how churn is defined, what action the business can take for at-risk customers, how success will be measured and whether a simple rule might be enough.
Question 3: Tell me about a time you recommended against building a model, or replaced one with something simpler.
- Why ask it:
- Shows judgment and a focus on value over technique.
- A strong answer shows:
- A specific case where a rule, a report or a heuristic met the need, with the reasoning and how the stakeholder responded.
Question 4: Before you commit weeks to a project, how do you judge whether it is worth doing?
- Why ask it:
- Tests commercial judgment about where data science effort goes.
- A strong answer shows:
- Sizing the likely impact, checking that the data exists and is usable, confirming how the output will be used, and agreeing a small first milestone.
Statistics and experimentation
Ask about tests the candidate designed or analyzed themselves. A recruiter can score these against the guidance without being a statistician.
Question 5: How would you design an A/B test for a change to a sign-up page? Talk me through the decisions you would make before it launches.
- Why ask it:
- Tests experiment design from start to finish.
- A strong answer shows:
- A clear hypothesis, one primary metric plus guardrail metrics, a sample size or test length worked out in advance, random assignment and a decision rule agreed before results come in.
- Follow-up:
- What would you recommend if the test showed no meaningful difference?
Question 6: A product team wants to stop a test early because the new version is ahead after three days. What do you tell them?
- Why ask it:
- Tests understanding of a common and costly experimentation mistake.
- A strong answer shows:
- Explaining that repeatedly checking and stopping at the first good result raises the chance of a false win, that early days may not reflect a normal week, and offering a pre-agreed stopping rule or a method designed for early stopping.
Question 7: Tell me about an experiment whose result surprised you or the team. How did you check it before you reported it?
- Why ask it:
- Shows rigor when results look too good or too strange to be true.
- A strong answer shows:
- Checks such as whether the groups were split as planned, tracking or logging errors, novelty effects and results by segment, done before any conclusion was shared.
Question 8: When a randomized test isn’t possible, how do you estimate the effect of a change?
- Why ask it:
- Many business questions can’t be answered with a clean experiment.
- A strong answer shows:
- Approaches such as comparison groups, matching or difference-in-differences, explained along with their assumptions and limits rather than presented as proof.
Question 9: How do you explain a confidence interval or a p-value to a stakeholder who only wants to know whether the change worked?
- Why ask it:
- Tests whether the candidate can turn statistics into a decision.
- A strong answer shows:
- Plain language about uncertainty and the likely size of the effect, a clear recommendation, and no overstatement of what the number shows.
Modeling, evaluation and data quality
Listen for honest evaluation and careful data habits. Strong candidates explain why a model works, not only which algorithm they used.
Question 10: How did you choose and evaluate the model in a recent project? Which metric did you optimize for, and why that one?
- Why ask it:
- Tests whether evaluation reflects the real cost of the model’s mistakes.
- A strong answer shows:
- A baseline, a metric chosen for the business cost of false positives and false negatives, a proper holdout or time-based validation, and a comparison of simple and complex models.
Question 11: Your model scored very well in testing but performs poorly on live data. What could have gone wrong?
- Why ask it:
- Tests understanding of leakage, overfitting and differences between training and live data.
- A strong answer shows:
- Features that wouldn’t be available at prediction time, a validation split that didn’t match real use, overfitting and shifts in the live data, with how they would check each one.
Question 12: Tell me about a feature you engineered that made a real difference to a model. Where did the idea come from?
- Why ask it:
- Feature work often matters more than the choice of algorithm.
- A strong answer shows:
- A feature grounded in understanding of the business or in conversations with domain experts, tested properly and checked for leakage.
Question 13: What do you check in a new dataset before you build anything on it?
- Why ask it:
- Tests the data quality habits that prevent wrong conclusions.
- A strong answer shows:
- How the data was collected, missing values, duplicates, outliers, label quality, time coverage and whether it represents the people or cases the model will be used on.
Question 14: How do you choose between a model that is more accurate and one that is easier to explain?
- Why ask it:
- Tests judgment about trade-offs the business will have to live with.
- A strong answer shows:
- Weighing the use case, how decisions may need to be explained to customers or reviewers, how large the accuracy gap really is, and a real example of the choice.
Communication, production and responsible use
A model only matters once people trust and use it safely. Any interviewer can judge the communication question; listen to the rest for ownership after launch.
Question 15: Explain a model you built to me as if I were the sales director who will use its output.
- Why ask it:
- Tests communication directly, and a non-technical recruiter can judge it.
- A strong answer shows:
- What the model predicts, how reliable it is in business terms, how to act on it and when not to trust it, without jargon.
Question 16: Tell me about a model you helped put into production. What happened between your notebook and the live system?
- Why ask it:
- Shows whether the candidate’s work reaches users and what part they played.
- A strong answer shows:
- Working with engineers on packaging, data pipelines, testing and deployment, agreeing batch or real-time requirements, and their own role in each step.
Question 17: How do you monitor a model once it is live, and how do you know when it needs retraining?
- Why ask it:
- Models degrade as data and customer behavior change.
- A strong answer shows:
- Tracking input data drift, the spread of predictions and real outcomes as they arrive, alerts with a named owner and a plan to retrain or roll back.
- Follow-up:
- Tell me about a time monitoring caught a problem. What did you do?
Question 18: How do you check whether a model treats different groups of people unfairly?
- Why ask it:
- Biased models can harm people and expose the business to serious risk.
- A strong answer shows:
- Comparing error rates and outcomes across groups where the data allows, looking for features that act as proxies, involving legal or ethics colleagues, and changing the model or process when a gap appears.
Question 19: What steps do you take to protect personal data when you work with it?
- Why ask it:
- Data scientists often handle sensitive information about customers or employees.
- A strong answer shows:
- Using only the data needed, removing or masking identifiers where possible, following the company’s access and retention rules, and raising concerns before using data for a new purpose.
Must-haves and logistics
Ask these of every candidate before the client’s technical rounds, and check the answers against the resume.
Question 20: Which languages, libraries and platforms have you used for data science work in the last two years, such as Python, R, SQL, a cloud machine learning platform or an experimentation tool?
- Why ask it:
- Confirms hands-on experience with the client’s stack.
- A strong answer shows:
- Specific tools tied to real projects and recent dates, with honesty about anything used only in courses or side projects.
Question 21: Has your work been mostly experimentation and analytics, machine learning that runs in products, or both, and in which industries?
- Why ask it:
- The title covers different jobs, and clients usually want one side more than the other.
- A strong answer shows:
- A clear account of where their experience sits, with examples that overlap with the client’s needs and candor about gaps.
Question 22: Can you work [remote, hybrid or on-site, and location]? What notice period are you on, and what pay range would you expect for this role?
- Why ask it:
- Rules out arrangement, timing and budget mismatches early, without asking about pay history.
- A strong answer shows:
- A clear yes or the specific constraint, a firm start date and a range you can compare with the client’s budget.
Question 23: Are you legally authorized to work in [country] for this employer, and will you need visa sponsorship now or in the future?
- Why ask it:
- Confirms eligibility; ask every candidate the same question in the same way.
- A strong answer shows:
- A direct answer, recorded the same way for every candidate.
Data scientist interview scorecard
Five criteria for this role, with what a score of 1, 3 and 5 looks like. Scores of 2 and 4 sit between them.
| Criterion | What it means | Score 1 looks like | Score 3 looks like | Score 5 looks like |
|---|---|---|---|---|
| Problem framing and impact | Starts from the business decision and delivers work that changes something. | Jumps straight to algorithms and cannot name a decision their work influenced. | Frames a project around a clear question and gives one example of measurable impact. | Clarifies the decision first, sizes value before starting, chooses simple solutions where they fit and shows repeated impact. |
| Statistics and experimentation | Designs sound experiments and reads results with care. | Cannot explain how a test was designed or treats any early lift as proof. | Designs a sensible A/B test with a primary metric and planned sample size. | Anticipates pitfalls such as early stopping and uneven splits, uses suitable methods when tests aren’t possible and explains uncertainty plainly. |
| Modeling and data rigor | Builds and evaluates models honestly on data they have checked. | Reports accuracy with no baseline or validation approach, and assumes the data is clean. | Uses a baseline, a suitable metric and a holdout set, and checks data quality before modeling. | Guards against leakage, ties metrics to business costs, engineers meaningful features and explains every trade-off. |
| Communication | Explains models and findings so non-technical people can act on them. | Relies on jargon or describes methods without reaching a recommendation. | Gives a clear explanation and recommendation a non-specialist can follow. | Adapts to each audience, is open about limits and uncertainty, and makes results easy to act on. |
| Production and responsible practice | Gets models into use and looks after them, fairly and safely. | Work ends in a notebook, with no thought for monitoring, bias or privacy. | Has helped deploy a model and describes sensible monitoring and data protection habits. | Owns models after launch, catches drift early, checks for unfair outcomes and raises privacy concerns before they become problems. |
The 1–5 rating scale
The same scale for every criterion and every candidate.
| Score | Level | What it means |
|---|---|---|
| 1 | Well below requirement | No relevant evidence, or an answer that contradicts the requirement. |
| 2 | Below requirement | Partial evidence with important gaps. |
| 3 | Meets requirement | Clear, relevant evidence at the level the role needs. |
| 4 | Above requirement | Strong, specific evidence beyond the expected level. |
| 5 | Exceptional | Repeated high-quality evidence with clear impact. |
How to run the interview with these data scientist interview questions
- 01
Step 01
Agree the must-haves first
Confirm the essential credentials, experience and availability with the hiring manager or client before any interviews. - 02
Step 02
Pick 8 to 12 questions
Take the must-have questions, then the questions that test what this role needs most. Use the same set, in the same order, for every candidate. - 03
Step 03
Ask for real examples
When you hear “we” or “I would”, ask what the candidate personally did, and what happened in the end. - 04
Step 04
Score before you discuss
Rate each criterion on the scorecard with the evidence behind it, then compare with other interviewers. - 05
Step 05
Verify before you submit
Check licenses, certifications and right to work against the original source before you put the candidate forward.
Red flags, and questions not to ask
- Proposes a complex model before asking what decision the result will support or what a simple baseline would achieve.
- Quotes impressive accuracy figures but can’t say how the model was validated or what it was compared against.
- Treats an A/B test as finished as soon as the numbers look good, or can’t explain how the sample size was set.
- Has never followed a model or analysis past delivery to see whether anyone used it.
- Dismisses bias, privacy or consent questions as someone else’s job.
- Age, marital or family status, pregnancy or plans for children, religion, ethnicity or national origin, sexual orientation or gender identity. These are protected characteristics under the UK Equality Act 2010 and US federal law, and they say nothing about whether someone can do the job.
- Health, sickness absence or disability before an offer. You can ask whether the candidate needs any adjustments for the interview, and whether they can do the essential tasks of the job.
Screen data scientist applicants before the first call
Add these questions to a Beatview AI interview and every applicant answers them on video or audio, with the same time limit. Beatview scores each answer against your criteria and shows the reasoning, and you can share the shortlist with your client through a password-protected link. AI interviews are on the Pro plan; the Free plan screens resumes for one active job.
Data scientist interview questions: frequently asked questions
Still deciding?
Bring a live vacancy and we’ll walk through where automation ends and recruiter review begins.
Ask the candidate to walk through one project from the business question to the result, then ask questions that test problem framing, statistics and A/B testing, modeling and evaluation, data quality, communication, putting models into production and responsible use of data. Add must-have questions on the client’s tools, type of experience, work arrangement, right to work and start date, and ask every candidate the same questions in the same order.
Ask the candidate to explain a model as if you were the person who will use its output, and to describe a decision their work changed. You can judge clarity, ownership and whether they talk about business results or only techniques. Use the listening notes for the experimentation and modeling questions, confirm which tools they used recently and on what, and leave deeper technical assessment to the client’s data scientists.
A short exercise can show skill better than a description, but keep it fair. Agree with the client who sets and scores it, state the expected time up front and keep it short, use made-up or public data rather than a real client problem so candidates aren’t doing unpaid work, and give every candidate the same brief, data and time. Score it against criteria agreed in advance, offer adjustments when a candidate asks, and use a follow-up discussion of the work to judge reasoning and communication, not only the final result.
A data analyst mainly answers business questions with existing data, using SQL, spreadsheets and dashboards. A data scientist goes further into statistics, experiments and predictive models, and may build models that run inside products. A data engineer builds and maintains the pipelines and warehouses that both rely on. The boundaries vary between companies, so check the client’s job description and weight the questions toward the work the role actually involves.
Get the data scientist interview questions template
All 23 questions with why you ask each one, what a strong answer shows and follow-ups, plus the data scientist scorecard and rating guide.
Opens in Excel, Google Sheets or Numbers. Version 2 October 2026.