Data Scientist Interview Questions (2026) and How to Answer
By FindAJobAI team Updated How we wrote this page
Data scientist interviews usually include a statistics and probability round, a SQL or Python exercise, a machine learning discussion, and a product or business case, such as how you'd measure a new feature. Many add a take-home analysis that you present.
Interviewers look for sound reasoning more than fancy models: clear problem framing, honest evaluation, and recommendations a business can act on.
Below: 30 questions with what each one really asks and an answer outline, full sample answers to the 10 most asked, red flags to avoid and a 7-day plan.
Answering with STAR
Tell each story in four steps: the situation in a sentence, your task, the actions you took (most of the answer) and the result, with a number if you have one. Aim for about two minutes; the outlines show what to include, the details must be yours.
Behavioural questions (1–10)
How you've handled real situations. Answer each with one true story, in STAR order.
1.Tell me about a model or analysis that changed a business decision.
What they're really asking: Impact: did anyone act on your work?
Answer outline (STAR)
- SituationA business question with money or users at stake.
- TaskGive decision-makers an answer they could trust.
- ActionFrame the question, choose the method, validate it, present the trade-offs.
- ResultThe decision made, and its measured result.
Full sample answer (fictional example)
Our sales team gave the same discount to every renewal. I analysed three years of renewals and found that high-usage customers renewed at about the same rate with or without a discount. I built a simple model that flagged the 30 percent of accounts at real risk and recommended discounts only for them. Leadership agreed to try it for a quarter with a holdout group. Renewal rates matched the holdout, and discount spend fell 22 percent, around $400,000 a year. The sales director now asks for the risk list before each quarter.
2.Tell me about an analysis where the data surprised you.
What they're really asking: Curiosity and rigour: do you check before you trust?
Answer outline (STAR)
- SituationA result that contradicted what everyone expected.
- TaskDecide whether it was real or an artefact.
- ActionCheck data quality, segment the data, test other explanations.
- ResultWhat turned out to be true, and what it changed.
Full sample answer (fictional example)
Our new onboarding flow looked like a success: setup completion rose 18 percent. But activation a week later was flat, which surprised me. Before telling anyone, I segmented the data and found the lift came entirely from a tracking change: a new step sent a 'setup complete' event even if users skipped it. I told the product manager the same day, we corrected the event, and the true lift was three percent. I wrote a short note on checking events before trusting a funnel, and we added a test for it.
3.Describe a time you explained a complex model to non-technical people.
What they're really asking: Communication, often the deciding skill in data science hiring.
Answer outline (STAR)
- SituationA model whose outputs drove decisions.
- TaskHelp stakeholders trust it and use it correctly.
- ActionExplain what it does with examples, show where it fails, avoid jargon.
- ResultHow it was adopted, and the questions people stopped asking.
Full sample answer (fictional example)
Our support team used a model that ranked tickets by urgency, but agents ignored it because they didn't understand it. I met five agents and showed examples: this ticket ranks high because it mentions a payment failure and the customer has written twice. I also showed where it fails, such as sarcastic messages, so they knew when to override it. I avoided talking about features and probabilities. Within a month agents followed the ranking on about 80 percent of tickets, and first-reply time on urgent tickets fell by a quarter.
4.Tell me about a project that failed, or one you decided to stop.
What they're really asking: Judgement, and honesty about what doesn't work.
Answer outline (STAR)
- SituationA model that couldn't beat the baseline, or a question the data couldn't answer.
- TaskDecide whether to keep going.
- ActionCompare against the baseline, estimate the value, recommend stopping or pivoting.
- ResultTime saved, and what you learned.
Full sample answer (fictional example)
I spent six weeks building a model to predict which trial users would convert, hoping to target sales calls. It never beat a simple rule based on whether users invited a teammate. I checked honestly: the rule had almost the same precision with no maintenance, and the extra revenue from the model was below the cost of running it. I recommended stopping and presented the numbers to the head of sales, who agreed. We used the rule instead, and I wrote up what I'd learned about the data so the next person wouldn't repeat the work.
5.Tell me about working with engineers to put a model into production.
What they're really asking: Can your work leave the notebook?
Answer outline (STAR)
- SituationA prototype model that needed to run inside a product.
- TaskGet it deployed and monitored.
- ActionAgree an interface, write tests, monitor data and performance drift.
- ResultTime to production, and how it performed live.
6.Tell me about working with messy or incomplete data.
What they're really asking: Can you get useful answers from imperfect data, honestly?
Answer outline (STAR)
- SituationA key table with missing values, duplicates or inconsistent IDs.
- TaskReach a defensible answer without waiting for perfect data.
- ActionProfile the gaps, choose and document a handling rule, test how sensitive the result is.
- ResultThe conclusion, how robust it was, and the data fix you requested.
7.Tell me about an experiment that went wrong.
What they're really asking: Do you learn from flawed tests instead of hiding them?
Answer outline (STAR)
- SituationA test with tracking bugs, uneven groups or an early readout.
- TaskSalvage what you could and avoid a bad decision.
- ActionSpot the problem, tell stakeholders, rerun or discard, add health checks.
- ResultThe decision that was protected, and the checks added.
8.Tell me about a metric you designed or changed.
What they're really asking: Can you define success in a way that guides good behaviour?
Answer outline (STAR)
- SituationA team steering by a metric that was easy to game or lagging.
- TaskCreate a measure that tracked real value.
- ActionInterview users, test candidates on past data, pilot, document the definition.
- ResultThe behaviour it changed, and what you'd refine.
9.Tell me about influencing a roadmap with data.
What they're really asking: Do you shape decisions or just answer questions?
Answer outline (STAR)
- SituationA team about to build a feature with no evidence of demand.
- TaskBring evidence early enough to matter.
- ActionSize the opportunity from usage data, run a quick test or survey, present options.
- ResultWhat was built or dropped, and the result.
10.Tell me about prioritising requests as the only data scientist.
What they're really asking: Can you protect your time for the work with most impact?
Answer outline (STAR)
- SituationMany teams asking for help at once.
- TaskFocus on what moved the business.
- ActionRank by impact and effort with leaders, say no with reasons, build self-serve tools for repeats.
- ResultRequests answered, and the large project you finished.
Technical and role-skill questions (11–22)
Can you do the work itself? Think out loud and name the trade-offs.
11.How would you design an A/B test for a new checkout flow?
What they're really asking: Experimentation, the most common product data science question.
Answer outline
- Choose the primary metric (conversion) and guardrails such as revenue per user and errors.
- Work out the sample size from the baseline rate, the minimum detectable effect and the power.
- Randomise by user, fix the duration in advance, and don't stop early because it looks good.
Full sample answer (fictional example)
First I'd define the goal. The primary metric would be checkout conversion, with guardrails such as revenue per user, refund rate and page errors. Then I'd calculate the sample size from the baseline conversion, the smallest lift worth detecting and the power I want, typically 80 percent, and turn it into a duration of at least one or two full weeks to cover weekly patterns. I'd randomise by user, not session, check the groups are balanced, and fix the end date in advance. I wouldn't stop early because it looks good.
12.Explain a p-value to a product manager.
What they're really asking: Do you really understand p-values, and can you explain them simply?
Answer outline
- If there were really no difference, how often would a result this extreme happen by chance?
- It isn't the probability that the change works.
- Always pair it with the effect size and a confidence interval.
Full sample answer (fictional example)
The p-value answers one question: if the change really did nothing, how often would we see a difference at least this large by chance? A p-value of 0.03 means about three times in a hundred. It is not the probability the change works, and it doesn't tell you how big the effect is. So I'd always show the effect size and a confidence interval alongside it, say 'conversion rose between 0.5 and 2.1 points', so the decision-maker can judge whether the effect is large enough to matter.
13.How do you handle an imbalanced classification problem?
What they're really asking: Do you know why accuracy misleads when one class is rare?
Answer outline
- Pick metrics that reflect it: precision, recall or PR-AUC, not accuracy.
- Resample or weight the classes, and tune the threshold to the business cost of each error.
- Keep the real class ratio in the validation data.
14.How do you know your model isn't overfitting?
What they're really asking: Can you trust your own results, and prove it?
Answer outline
- Compare training and validation performance, using cross-validation.
- Split by time when the data has an order, to avoid leakage.
- Regularise or simplify, and touch the held-out test set only once.
15.Which features would you build to predict customer churn?
What they're really asking: Can you turn knowledge of the business into predictive features?
Answer outline
- Usage trends such as falling logins, support tickets, billing events and contract dates.
- Make sure no feature leaks information from after the churn date.
- Start with a simple, explainable baseline model.
16.How would you tell whether a feature caused higher retention, or is only correlated with it?
What they're really asking: Causal reasoning, which separates strong data scientists.
Answer outline
- Correlation can reflect self-selection: engaged users are the ones who try features.
- The best answer is an experiment; failing that, a natural experiment or difference-in-differences.
- Be clear about the assumptions each method needs.
Full sample answer (fictional example)
Correlation alone can mislead because of self-selection: engaged users are the ones who try new features, and they'd retain better anyway. The cleanest answer is an experiment: randomly show the feature to some users and compare retention. If that isn't possible, I'd look for a natural experiment, such as a staggered rollout, or use difference-in-differences or matching on past behaviour. I'd be clear about the assumptions each method needs, and say how confident I am. Without an experiment, I'd call the result evidence, not proof.
17.Write a SQL query for 7-day retention by sign-up cohort.
What they're really asking: SQL fluency on a real product question.
Answer outline
- Group users into cohorts by sign-up week.
- Join to activity and count users active between day 1 and day 7.
- Divide by cohort size, and leave out cohorts too young to measure.
18.How does a random forest differ from gradient boosting?
What they're really asking: Can you compare common tree models and choose between them?
Answer outline
- A random forest averages many deep trees trained independently on random samples.
- Gradient boosting builds shallow trees in sequence, each fixing the last one's errors.
- Forests resist overfitting with little tuning; boosting is often more accurate but needs care.
19.What goes wrong when you run many tests or look at results repeatedly?
What they're really asking: Do you know why peeking and many comparisons give false wins?
Answer outline
- Each extra comparison raises the chance that something looks significant by luck.
- Stopping a test as soon as it looks good inflates false positives.
- Fix the sample size in advance, correct for multiple comparisons, or use sequential methods.
20.What is regularisation, and how do L1 and L2 differ?
What they're really asking: Do you understand how models are kept from overfitting?
Answer outline
- It adds a penalty on large coefficients to the loss, keeping the model simpler.
- L1 can shrink coefficients to exactly zero, selecting features; L2 shrinks them smoothly.
- Choose the strength by cross-validation, and scale features first.
21.How would you forecast next quarter's demand?
What they're really asking: Can you approach a time series sensibly before reaching for complex models?
Answer outline
- Plot the series and look for trend, seasonality, holidays and breaks.
- Start with a simple baseline such as last year's value or a seasonal average.
- Compare models by rolling-origin validation, and give a range, not one number.
22.How do you handle missing data?
What they're really asking: Do you treat missingness as information, not just a nuisance?
Answer outline
- Ask why it is missing: at random, related to other fields, or related to the value itself.
- Options: drop rows, impute with simple or model-based values, or add a missing indicator.
- Test how much the conclusion changes under each choice, and document the one you used.
Take-home or live exercise: tips
Data science loops often include a take-home analysis with a short presentation, a live SQL or Python exercise, and a product case on measuring a feature.
- Lead with the recommendation, then the evidence, then the caveats, in your write-up or slides.
- State your assumptions and the checks you ran on the data before modelling.
- Keep the model simple unless a complex one clearly wins on a fair comparison.
Skills employers test for
Technical rounds follow the job ad. These skills are named most often in the 483 newest data scientist job ads we analysed (out of 2,454 live postings), shown as the share of ads that mention each. Expect questions on the top few.
- 1Python81%
- 2Machine learning73%
- 3SQL59%
- 4Statistics50%
- 5LLMs37%
- 6Data analysis36%
- 7R24%
- 8AWS24%
Full list: Data Scientist resume keywords.
From live job ads · computed November 2, 2026 · refreshed daily
Situational questions (23–30)
Scenarios that test judgement: say what you'd do first, and why.
23.An A/B test shows a significant lift, but the PM's instinct says it's wrong. What do you do?
What they're really asking: Can you defend a result while checking whether it's really true?
Answer outline
- Check the test's health: sample ratio mismatch, novelty effects, tracking bugs.
- Look at segments and guardrail metrics.
- Share the evidence, and propose a follow-up test if real doubt remains.
Full sample answer (fictional example)
I'd treat the instinct as a prompt to check, not as an insult. First I'd test the experiment's health: is the split the planned ratio, were there tracking changes, did the lift appear only in the first days, which suggests a novelty effect? Then I'd look at segments and guardrail metrics to see whether something else moved the other way. I'd share the evidence with the PM, including what would change my mind. If real doubt remains, I'd propose a short follow-up test on fresh traffic rather than argue.
24.You have two weeks to answer a question that deserves two months. What do you do?
What they're really asking: Can you deliver a useful answer fast, and say how sure you are?
Answer outline
- Agree which decision the answer will inform.
- Deliver a directional answer with clear caveats.
- Say what a fuller study would add, and whether it's worth the time.
25.A model's performance drops a month after launch. How do you investigate?
What they're really asking: Do you know that models decay, and how to catch it?
Answer outline
- Look for pipeline changes and drift in the input features.
- Compare the new data with the training data, segment by segment.
- Retrain, recalibrate or roll back, and add drift alerts.
Full sample answer (fictional example)
I'd start by checking whether anything changed in the pipeline: a new data source, a changed feature definition, or a failed job that fills features with defaults. Then I'd compare live input data with the training data feature by feature to find drift, and break performance down by segment to see where it's worse. If the world changed, I'd retrain on newer data. If a pipeline broke, I'd fix it. If I can't tell quickly, I'd roll back to the previous model. Then I'd add drift alerts and a monitored performance metric so it's caught earlier.
26.Leadership wants a machine learning project, but a simple rule would do. What do you say?
What they're really asking: Do you optimise for value rather than for interesting work?
Answer outline
- Show how well the rule meets the problem's needs.
- Explain the cost of building and maintaining a model.
- Recommend the rule now, and a model when the data justifies it.
27.A stakeholder asks for a dashboard of everything about customers. What do you do?
What they're really asking: Can you turn a vague request into a decision-focused deliverable?
Answer outline
- Ask which decisions they make and which numbers they'd look at first.
- Propose a small first version with three to five metrics and clear definitions.
- Agree how to measure its use, and add to it only on request.
28.Your experiment can't reach significance with the traffic you have. What are your options?
What they're really asking: Do you have practical alternatives to just waiting?
Answer outline
- Choose a more sensitive metric, or cut variance with a pre-experiment covariate.
- Accept a larger minimum detectable effect, or run the test longer.
- If neither works, decide on other evidence and say how uncertain you are.
Full sample answer (fictional example)
I'd look at what I can change before accepting defeat. A more sensitive metric, say clicks on the feature instead of purchases, needs fewer users. Variance reduction with a pre-experiment covariate, such as past spend, can cut the sample size noticeably. I could accept a larger minimum detectable effect, if only a big win would change the decision, or run the test longer. If none of that is feasible, I'd tell the team honestly that a test can't answer this, use other evidence such as user research, and state how uncertain we are.
29.A colleague's analysis contradicts yours. How do you resolve it?
What they're really asking: Can you settle a technical disagreement without ego?
Answer outline
- Compare data sources, filters, definitions and date ranges side by side.
- Walk through each step together until the numbers diverge.
- Agree which is right, or why both are, and write down the shared definition.
30.Your model works, but business users don't trust it. What next?
What they're really asking: Do you treat trust as part of the product?
Answer outline
- Ask what makes them doubt it: past errors, no explanation, or a clash with intuition.
- Show examples of where it is right and wrong, and explain its main drivers.
- Run it beside their current process, so they see results before relying on it.
Red flags to avoid in your answers
Each one is quick to fix before the day.
Leading with an algorithm instead of the business question.
Start with the decision the work informs, then explain the simplest method that answers it.
Quoting accuracy on an imbalanced problem.
Name the metric that fits the cost of each error, and show the baseline.
Saying 'the data shows' about a correlation.
Say what the evidence supports, what it can't prove and which test would settle it.
Hiding limitations to sound confident.
Give a range or a caveat. Interviewers trust people who know where their results are weak.
How to prepare for a data scientist interview in 7 days
An hour or two a day is enough; the order matters more than the length.
Day 1: Parse the role
Check whether the ad leans towards experimentation, modelling or analytics, and rank your stories to match.
Day 2: Statistics refresh
Rework p-values, intervals, power and multiple testing on paper, explaining each in plain words.
Day 3: Product cases
Practise two cases aloud: measuring a new feature and diagnosing a falling metric, with metric, method and decision.
Day 4: SQL and Python
Solve four timed problems: retention, a funnel, a window function and a pandas group-by.
Day 5: Model questions
Prepare answers on overfitting, imbalance, leakage and metrics, each with an example from your own work.
Day 6: Project stories
Write two projects with impact, one failure and one disagreement in STAR form, with numbers.
Day 7: Take-home rehearsal
Present one past analysis in ten minutes, recommendation first, to a friend. Then rest.
Questions to ask the interviewer
Pick two or three; the answers tell you whether you want the job.
- How do data scientists here split their time between analysis, experiments and models?
- How are models deployed and monitored, and who owns them afterwards?
- How mature is your experimentation platform?
- Which recent decision was driven by data science work?
- How do you measure a data scientist's impact here?
Preparation tips
- Revise statistics: distributions, hypothesis tests, confidence intervals and power.
- Practise product cases out loud: metric, hypothesis, method, decision.
- Prepare two projects with business impact, including what you'd do differently.
- In a take-home, lead with the recommendation, then the evidence, then the caveats.
Prepare the general questions too: tell me about yourself, why do you want to work here? and why are you leaving? Afterwards, send a thank-you email; when the offer comes, negotiate the salary.
How we wrote this page
The FindAJobAI team wrote the questions, outlines and sample answers for data scientist interviews. The people, employers and numbers in the samples are invented: copy the structure, and use details from your own work.
The live skills section is based on 483 data scientist job ads analysed, the newest we track, with the same deterministic skill matcher as our in-demand skills report: no AI model reads the postings. It was last computed on November 2, 2026 and refreshes daily.
Last reviewed on November 1, 2026.