Data Scientist Mock Interview Guide
How to prepare for data science interviews: framing ML cases, model evaluation, data leakage, experiments, and talking honestly about what went wrong.
HotSeat team
Updated November 2026 3 min read
Asked how you'd predict customer churn, it's tempting to say ‘XGBoost’ within fifteen seconds. A stronger first move is a question: what will the team do with the prediction?
That difference is most of what a data science interview is testing. Anyone can name models. The harder skill is working out whether a model should exist, how you'd know it was working, and what happens when it meets reality.
In this article
An order for ML case questions
- Start with the decision. Who uses the prediction, what will they do differently and what does a wrong answer cost? Sometimes this is where you discover a simple rule would do.
- Define the target and data. How are the labels created? Will each feature actually exist at prediction time? Who's missing from the data?
- Set a baseline. Begin with something simple, so any improvement has something to beat. It also gives you an honest answer when asked whether the complex model was worth it.
- Choose how you'll evaluate. Pick metrics that match the decision, then cover offline testing, an experiment and monitoring after launch.
- Plan for failure. Talk about drift, feedback loops, unusual cases and where a person should stay involved. It's easy to stop at deployment, so bringing this up helps you stand out.
Questions that come up a lot
When would you choose a simpler model?
Why they ask: They're checking that complexity has to earn its place.
- Compare performance against the baseline
- Include latency, explainability, maintenance and data size
- Say what gain would justify more complexity
Likely follow-up: How much performance would you trade for explainability?
How would you detect data leakage?
Why they ask: This shows whether your evaluation reflects what's known at prediction time.
- Check when each feature gets created
- Keep training and test data properly separate
- Be suspicious of results that look too good
Likely follow-up: Can preprocessing cause leakage?
Which metric would you use for an imbalanced classification problem?
Why they ask: They want the metric to come from real error costs, and accuracy won't tell you much here.
- Ask what false positives and false negatives each cost
- Discuss precision, recall, thresholds and calibration as needed
- Don't pick a metric before you know the decision
Likely follow-up: Who chooses the threshold?
Tell them about something that failed
A failed model can make a better story than a successful one, because it shows you can diagnose a problem. Here's an invented example:
“Our churn model did well offline, but it produced too many alerts for the customer team to use. I'd optimised for recall without checking how many customers they could call each week. We worked backwards from their capacity, adjusted the threshold and added account value to the ranking. The shorter list led to more completed calls and a higher save rate.”
The mistake was theirs, and they say so. They also fixed it by talking to the people using the model.
Get ready for questions about your own project
Expect someone to push on each stage. Why did this need machine learning at all? How were the labels created and checked? What did you compare it against? Why that metric? What changed after launch? Where shouldn't the model be trusted?
The weak answers are easy to spot: ‘we had a lot of data’, ‘the dataset was already clean’, ‘it's state of the art’, ‘it gave the highest score’. The worst might be ‘we didn't find any major limitations’, which usually means nobody looked.
If your project never reached production, say so. A class project is fine as long as you can talk about what you'd have to solve to deploy it.
Statistics, in plain language
You'll probably be asked to explain something to a non-specialist, so practise on a friend. Can you explain why correlation isn't causation using a real example? Can you say what an estimate does and doesn't support? Who's missing from a sample, and why does that matter?
Experiments come up often too. Know randomisation, power and guardrail metrics, and what can break a test. If you can't explain a concept simply, you may not know it as well as you think, and an interviewer will find that out quickly.
Practise defending your data science choices
Work through model, metric, project and business questions in a mock interview.
Start a data science interviewSources and further reading
- Model selection and evaluation (scikit-learn)
- Common pitfalls and recommended practices (scikit-learn)
- Structured interviews (U.S. Office of Personnel Management)