What does a data scientist do?
A data scientist uses analytical tools and techniques to pull meaningful insights out of data, in the words of the Bureau of Labor Statistics. The BLS duty list includes deciding which data are useful for a project, building and testing models, presenting findings with visualizations, and making business recommendations to stakeholders.
The O*NET OnLine profile for data scientists adds detail. Its task list includes cleaning raw data with statistical software, testing and reformulating models so they predict accurately, and comparing models with metrics such as loss functions. It also lists identifying business problems that data analysis can address.
BLS reports a median annual wage of $120,230 for data scientists in May 2025. It projects employment to grow 35 percent from 2025 to 2035. In Canada, Job Bank's requirements page for data scientists (NOC 21211) says a bachelor's degree in statistics, mathematics, computer science or a related discipline, or a college program in computer science, is usually required, and that the occupation is not regulated.
Which questions come from a data scientist's duties?
Each main duty tends to produce one kind of question. The table maps the duties that BLS and O*NET publish to the question each one tends to prompt and what the interviewer is checking. The questions are drawn from the job's published duties. They do not come from a survey of real interviews or from any named company.
| Duty (BLS and O*NET) | Question it tends to produce | What the interviewer is checking |
|---|---|---|
| Decide which data are available and useful | How would you decide what data to use for this problem? | You start from the business question, then check data quality and coverage |
| Clean and manipulate raw data | Tell me about a messy dataset you worked with. | You find missing values, duplicates and leaks before you model |
| Create, validate and test models | How do you know your model will work on new data? | A held-out test set, cross-validation and a baseline to beat |
| Compare models with performance metrics | Which metric would you pick here, and why? | The metric matches the cost of each kind of error |
| Present findings with visualizations | Walk me through a chart you made for a non-technical audience. | One clear message and the decision it supports |
| Recommend solutions to stakeholders | Tell me about a recommendation someone pushed back on. | You defend a finding and change your mind on evidence |
What behavioral questions come up in data scientist interviews?
Behavioral questions ask about something you did, so the interviewer can judge how you will act next time. For a data scientist, they usually center on a project that went wrong, a stakeholder who disagreed, or a result you had to explain plainly. Answer with one real project, told as situation, task, action and result.
The STAR method guide covers that shape in detail. The list of behavioral interview questions has general ones to practice alongside the role questions below.
1. Tell me about a model that did not perform the way you expected.
What they listen for: you noticed the problem yourself, found the cause and said what you changed. Blaming the data alone sounds weak. Naming a check you now run every time sounds strong.
Example answer (made up for this page, in STAR order):
Situation: A churn model I built for a subscription product scored well in testing, then flagged the wrong customers in its first month of use. Task: I owned the model, so I had to find the cause before the retention team stopped trusting it. Action: I compared the training features with live data and found one field that was only filled in after a customer had cancelled, which leaked the answer into training. I removed it, retrained with a time-based split, and set up a weekly check of predictions against outcomes. Result: The new model scored lower in testing, and its live results matched its test results much more closely. I now check when every feature is recorded before I train anything.
2. Tell me about a time a stakeholder disagreed with your analysis.
What they listen for: you took the objection seriously, tested it with data and kept the working relationship. A good answer shows you changed the analysis when they were right, or explained it better when they were wrong.
3. Describe a time you explained a technical result to a non-technical audience.
What they listen for: you started with the decision, used one simple chart and kept the method for questions. O*NET lists presenting the results of modeling and analysis to management as a core task, so this is part of the job itself.
Example answer (made up for this page, in STAR order):
Situation: A regional sales director asked whether a discount campaign had worked. Task: I had compared stores with and without the discount, and she had ten minutes for the update. Action: I opened with the answer in one sentence and showed a single chart of sales in both groups before and after. I kept the method in an appendix, and when she asked how sure I was, I gave a range and said in plain words what it meant. Result: She chose to run the discount in one more region before a full rollout, which was the decision the data could support.
4. Tell me about a project where the data you needed did not exist.
What they listen for: you narrowed the question, found a proxy or set up collection, and you were open about the limits of what you could conclude.
What technical questions do data scientist interviews ask?
Technical questions check that you understand the methods you use well enough to choose between them. Expect questions on statistics, model validation, metrics and coding in SQL or Python. Give the definition in one sentence, then say when you would use the method and what can go wrong with it.
5. What is overfitting, and how do you guard against it?
What they listen for: a plain definition and at least two defences. The scikit-learn guide to cross-validation describes overfitting as a model that would score perfectly on the data it has seen and fail to predict anything useful on unseen data. Its standard defence is to hold out a test set.
Example answer (made up for this page):
"Overfitting is when a model learns the noise in its training data, so it scores well there and poorly on new data. I guard against it in three ways. I hold out a test set that I only touch at the end. While tuning, I use k-fold cross-validation, where the model trains on all folds except one, is checked on the one left out, and the scores are averaged. I also compare against a simple baseline, so I can see whether a complex model earns its extra complexity."
6. Explain a p-value to someone in sales.
What they listen for: accuracy without jargon. The NIST/SEMATECH e-Handbook explains that a significance level of 0.05 means the null hypothesis is rejected 5 percent of the time when it is in fact true. A strong answer also says what a p-value leaves out, such as the size of the effect or whether it matters to the business.
7. How would you choose between precision and recall for a fraud model?
What they listen for: you tie the metric to the cost of each mistake. The scikit-learn precision and recall example defines precision as true positives over true positives plus false positives, and recall as true positives over true positives plus false negatives. Missing a fraud case and blocking a good customer cost different amounts, so ask which is worse before you pick.
8. Write a SQL query that returns each customer's most recent order.
What they listen for: a working approach and clear talk while you write. One common approach uses a window function such as ROW_NUMBER, partitioned by customer and ordered by order date, then keeps the first row. Say how you would handle ties and customers with no orders.
What situational questions should you expect?
Situational questions describe a problem and ask what you would do. They test judgment more than recall, so think out loud and ask clarifying questions first. For a data scientist, the scenarios tend to involve an unclear goal, a surprising result, or pressure to deliver an answer the data cannot support.
9. Your A/B test shows a large lift after two days, and the product manager wants to ship. What do you do?
What they listen for: you check the test against the plan set before it started, and you can explain the wait without blocking the team.
Example answer (made up for this page):
"I'd first check the test against the plan we agreed before it started: the sample size, the run time and the main metric. Two days can't cover a full week of user behavior, so I'd want at least one full week. I'd tell the product manager what we know so far, what could still change, and the date we'll have a reliable answer. If there's a business reason to ship sooner, I'd write down the risk so the decision is made on purpose."
10. You are handed a dataset and asked to "find insights." How do you start?
What they listen for: you ask what decision the work should inform before you touch the data. Then you profile the data, check its quality and agree on one question with the requester.
11. A model you deployed starts to drift. How would you notice, and what would you do?
What they listen for: a monitoring plan, such as tracking input distributions and prediction accuracy over time. Add a clear threshold for retraining or rolling back.
12. A senior leader asks you to rerun an analysis until it supports their plan.
What they listen for: integrity with tact. Offer to test their idea properly, share the result either way, and write down what was done so anyone can check it.
What follow-up questions tend to come next?
Follow-ups dig into the part of your answer that sounded thinnest. If you were vague about your own role, expect to be asked what you did yourself. If you skipped validation, expect to be asked how you knew the result held. Prepare a second layer of detail for each story before the interview.
- What would you do differently now?
- How did you measure that, and against what baseline?
- Who else worked on it, and what was your part?
- What would happen if that assumption were wrong?
- How long did it take, and what slowed it down?
How should you prepare for a data scientist interview?
Prepare four or five project stories, a one-sentence explanation of each method on your resume, and a short plan for each round. Read the posting closely, because the tools and problem types it names are what the interviewer will probe. Practice your answers out loud, since explaining a model aloud is a different skill from building it.
The free interview questions tool takes the job posting and the round you pick, and lists likely questions, why each is asked and what the interviewer listens for. If your resume undersells your projects, the guide to resume bullet points shows how to describe the work plainly.
Related roles share many of these questions. See the pages on data analyst interview questions, data engineer interview questions and software engineer interview questions. To rehearse, the Mock room has you answer out loud on a clock, scores each answer out of 10 and checks it against your resume.