Data Scientist Interview Questions and Answers (With What Hiring Managers Actually Want)

Data science interviews test three things: statistics, code, and communication. Most candidates are strong on one and weak on another. The ones who get hired are the ones who can connect a model to a business outcome and explain it plainly.

Here are the questions that come up most often, with answers that work and what interviewers are really listening for.

The statistics questions

"Explain overfitting to a non-technical stakeholder"

This is a communication test dressed up as a technical one. The wrong answer is the textbook definition. The right answer uses a plain analogy.

"It's like a student who memorises the past papers but can't answer a new question. The model learned the training data so precisely, including the noise, that it does badly on anything it hasn't seen. I deal with it by holding out a validation set and checking whether performance holds up on data the model never touched."

That answer shows you understand it and can explain it to someone who signs the budget.

"How would you handle missing data?"

There is no single right answer, and saying so is part of the answer. "It depends why it's missing. If it's missing at random, I can impute. If it's missing because of a systematic issue, like users on a free plan never seeing a certain field, imputing hides the problem. I'd report the pattern first and decide with the stakeholder."

"When would you use a tree-based model over a linear one?"

Interviewers want to hear you think about the trade-off, not that you prefer one. "When the relationships are non-linear and interactions matter, and when interpretability is less important than predictive power. If the stakeholder needs to know why a decision was made, I lean linear."

The coding questions

"Write code to evaluate a classification model"

They want to see you reach for the right metric, not just accuracy. Mention precision, recall, and the confusion matrix, and explain why accuracy alone is misleading on imbalanced data. A short, clean function beats a clever one you can't explain.

"How would you query and aggregate a large dataset?"

This tests whether you think about scale. Talk about filtering early, aggregating in the database rather than in pandas, and avoiding pulling a full table into memory. If you mention window functions or partitioning, be ready to write one.

The business questions

"Tell me about a project where your model was used in production"

This separates people who have shipped from people who have not. Name the decision the model drove, how often it ran, and what happened when it was wrong. "It routed support tickets and saved the team about 15% of triage time. When it was wrong, the fallback was a human queue, so the cost of an error was low. That trade-off is why we shipped it."

"How do you know a model is worth building?"

They want to see you think about value before accuracy. "I ask what decision it changes and what that decision is worth. If a 1% improvement is worth a lot of money, a simple model is worth building. If the lift is tiny, the model isn't worth the maintenance."

Questions to ask them

• How mature is the data infrastructure here?
• Who consumes the model output, and how do they use it?
• What is the last model that went to production, and what happened to it?

What to avoid

• Don't name a technique you can't explain in one sentence.
• Don't say accuracy when you mean precision or recall.
• Don't describe a project without naming the business result.

Before you go in

Pick one project where your work reached a decision maker and rehearse the three sentence version: the problem, the model, the result. That is the answer that gets you the offer.

Want to practise out loud? Try the mock interview tool, and line up your CV with our data scientist CV examples first.