Mock Interview Copilot Interview questions by role
Data scientist interview questions
Data scientist interviews test whether you can choose and validate the right method for a business problem, not just whether you know many algorithms. Expect questions on statistics and experiment design, how you would build and evaluate a model, a deep dive into a past project, and how your work changed a product or decision.
How the interview usually runs
Processes usually include a recruiter screen, a coding or SQL exercise, a statistics or machine learning round, a product or case round, and a project deep dive. The balance depends on the role: product data science leans on experiments and metrics, machine learning roles lean on modelling and deployment.
The project deep dive is where interviewers learn the most. They will ask why you chose that model over a simpler one, how you knew it worked, what went wrong, and whether anyone used it. A candidate who can defend a logistic regression that shipped is usually stronger than one who describes a complex model that never left a notebook.
Common questions and what the interviewer is checking
Walk me through a model you built, from problem to result.
What they are checking: The full chain: framing the business problem, data, features, baseline, model choice, evaluation, and what happened after it was handed over.
How would you design an A/B test for a new checkout button?
What they are checking: Choosing a primary metric and guardrails, randomisation unit, sample size and duration in principle, and what you would do about peeking or multiple metrics.
Your model does well offline but badly in production. Why might that be?
What they are checking: Awareness of data leakage, training-serving skew, distribution shift, feedback loops and wrong evaluation metrics.
How do you choose an evaluation metric for a classifier?
What they are checking: That you connect precision, recall and thresholds to the cost of each kind of error in the real problem, rather than reporting accuracy by default.
Explain a p-value to a product manager.
What they are checking: Correct understanding, and the ability to say it plainly without overstating what it proves.
When would you not use machine learning?
What they are checking: Judgement. Rules, a simple heuristic or a dashboard are often the right answer, and interviewers like candidates who say so.
How do you handle missing data or class imbalance?
What they are checking: Practical options with their trade-offs, and checking why data is missing before filling it in.
Tell me about a time your analysis was wrong or a project failed.
What they are checking: Honesty and learning: what you missed, how it was caught, and what you now check by default.
How would you explain your model's predictions to the people using them?
What they are checking: Interpretability methods at a sensible level, and a focus on what the users need to trust and act on the output.
How to structure a strong answer
For project questions, lead with the business problem and the baseline you had to beat. Interviewers want to see that you started simple, measured honestly and only added complexity when it earned its place.
- The problem and why it mattered, in business terms.
- The baseline and how you evaluated against it.
- The key modelling or design decision and its trade-off.
- The outcome in use, and what you would change.
Example answer
Walk me through a model you built, from problem to result.
The support team wanted to know which incoming tickets were likely to escalate, so urgent ones could be picked up first. The baseline was the existing rule of flagging tickets with certain keywords, which caught some escalations but flagged far too many.
I built a gradient-boosted model on ticket text features, customer history and product area, and evaluated it on later months than it was trained on, to mimic real use. We tuned the threshold with the team leads, because missing an escalation cost much more than a false alarm. It beat the keyword rule on both precision and recall, and the team used it to sort the queue. The thing I would do differently is involve the agents earlier, because their feedback on false alarms improved the features more than any tuning I did.
Mistakes to avoid
- Describing a sophisticated model without mentioning a simple baseline.
- Reporting accuracy on an imbalanced problem as if it means the model works.
- Designing an experiment without saying what result would make you stop or ship.
- Explaining statistics in jargon that a product manager could not repeat.
- Talking about notebooks and never about whether anyone used the result.
Frequently asked questions
What questions are asked in a data scientist interview?
Typically statistics and probability, experiment design, machine learning concepts and evaluation, SQL or coding, a product case, and a deep dive into a project you led, including what impact it had.
How do I answer A/B testing questions in a data science interview?
Define the hypothesis and primary metric, name guardrail metrics, choose the randomisation unit, explain how you would size and time the test, and say how you would decide. Mention pitfalls such as peeking and novelty effects.
How technical is a data scientist interview compared with a data analyst one?
Data science interviews usually go further into statistics, modelling and evaluation, while analyst interviews focus more on SQL, metrics and communication. Both expect you to tie your work to a business decision.
What should I say if my projects never reached production?
Be honest, and show you understand what production would require: monitoring, retraining, the gap between offline and online performance, and who would use the output.