Role interview guide
Data scientist interview questions and what to expect
Data scientist interviews usually test statistics, machine learning, SQL and Python, then how you'd use them on a business problem: choosing a metric, designing an experiment, or building a model and explaining its limits. Candidates commonly report a take-home or case study and a deep dive into a past project, and for product-focused roles, questions on A/B testing.
What candidates commonly report; it varies by employer, level and country. Last reviewed 18 October 2026.
The usual stages
- 1Recruiter or HR screen
- 2Online test or technical screen: SQL, Python and statistics
- 3Take-home or live case: analyse data or build a simple model, then present it (often)
- 4Machine learning and statistics interview, often built around your own projects
- 5Product or business case on metrics and experiments (for product-focused roles)
- 6Hiring manager and HR rounds
What interviewers check
- Statistics you can explain: distributions, hypothesis tests, confidence intervals
- Choosing a model and a metric that fit the problem, not the fashion
- Careful validation: proper splits, overfitting, no information from the future in training data
- SQL and Python for getting and cleaning data
- Turning a vague business question into something you can test
- Explaining results and uncertainty to people who aren't technical
Behavioral questions
Tell me about a model or analysis of yours that was used to make a decision.
A good answer covers
- The problem, the decision and who made it
- Your approach and why you chose it over the alternatives
- The impact, with a number, and how you checked it
Describe a project where the data turned out much messier than you expected.
A good answer covers
- What was wrong: missing values, a biased sample, definitions that changed over time
- How you found it and what you did about it
- How it changed your conclusions or your timeline
Tell me about a time a model did worse once it was in use than it had in testing.
A good answer covers
- The gap between offline and live results
- The cause: drift, mismatched data, or features that weren't available at prediction time
- What you changed in how you validate models
Tell me about a time you chose a simpler approach over a more sophisticated one.
A good answer covers
- Both options and what each would have cost
- Why simple was good enough: accuracy, explainability, time to deliver
- How it worked out
Questions about the job
Explain the bias-variance trade-off, and how you'd spot overfitting.
A good answer covers
- Too simple a model misses patterns; too complex a model learns noise
- Spotting it by comparing training and validation scores, or with learning curves
- Fixes: more data, regularisation, a simpler model, cross-validation
You're building a fraud detection model and only a tiny share of transactions are fraud. How do you evaluate it?
A good answer covers
- Why accuracy misleads when one class is rare
- Precision, recall and precision-recall curves, and choosing a threshold from the cost of each error
- Class weights or resampling, and testing on data with the real fraud rate
How would you design an A/B test for a new recommendation feature?
A good answer covers
- A hypothesis, one main metric and a few guardrail metrics
- What gets randomised (users, sessions), the sample size and how long to run it
- Reading the result: significance, practical size of the effect, novelty effects
Explain a confidence interval to a product manager.
A good answer covers
- A range of plausible values for the thing being estimated
- Why a wider range means less certainty, and what narrows it
- A plain example from a test result
What's the difference between correlation and causation, and how could you get closer to a causal answer without an experiment?
A good answer covers
- Confounders and reverse causation, with an example
- Natural experiments, matching, or a before-and-after comparison with a control group
- Being clear about what the analysis can't prove
How would you predict which customers are likely to cancel their subscription next month?
A good answer covers
- Defining cancellation and the prediction window precisely
- Features from usage, payments and support, built only from data before the prediction date
- A simple baseline first, the evaluation metric, and how the business would act on the scores
Compare logistic regression, random forests and gradient boosting. When would you pick each?
A good answer covers
- Interpretability versus accuracy on tabular data
- Training time, tuning effort and sensitivity to noisy data
- What you'd try first and why
How do you handle missing values, and when can that choice change your results?
A good answer covers
- Why the data is missing: at random, or for a reason that matters
- Dropping rows, simple filling, model-based filling, or adding a 'was missing' flag
- Checking how sensitive the results are to the choice
How would you use SQL to find each customer's first order, and whether they ordered again within 30 days? Talk me through it.
A good answer covers
- Finding the first order per customer with MIN or a window function
- Finding the next order with LEAD or a join back to the orders table
- Handling customers with one order, and two orders on the same day
How would you explain to a business team why the model made a particular prediction?
A good answer covers
- Overall feature importance versus the reasons for one prediction
- Tools such as SHAP values or partial dependence plots, in plain words
- The limit: a feature being important doesn't make it a cause
Your model's performance has been slowly dropping since launch. What could be happening, and what would you do?
A good answer covers
- Data drift, changes in what the target means, or changes in upstream data
- Monitoring input distributions and live performance once labels arrive
- When to retrain on fresh data and when to rethink the model
Motivation and fit
Why data science? How did you get into it?
A good answer covers
- What got you started, and a project that hooked you
- What you enjoy most: modelling, experiments, the business problem
- What you're working to learn next
What kinds of problems would you want to work on here as a data scientist?
A good answer covers
- A grasp of the business and where data can help it
- One or two specific problems you'd find interesting
- How you'd measure success on them
Tips
- Know two projects inside out: the data, the baseline, the metric, what failed and what you'd do next.
- Say what a simple baseline would score before you mention any complex model.
- Practise explaining one statistical idea (p-values, confidence intervals, overfitting) in two sentences without jargon.
- In case questions, ask how the result will be used; it decides the metric and the threshold.
- Expect to write SQL and Python while talking; many screens test them before any machine learning.
Practise it out loud
A mock interview asks what data scientist interviews usually test, follows up on your answers and gives you a report. The warm-up is free, needs no account and checks five of these questions as you answer.
About this guide
Written by Orbit from what candidates and interviewers commonly describe for this kind of job. The questions are Orbit's own practice questions, not questions taken from any employer's interviews. Processes change, so check the job post and what your recruiter tells you.