Data Scientist Interview Questions and Answers

Questions and model answers

9 questions you are likely to hear.

1. 1

   Explain the bias-variance trade-off.

   Why they ask: This core concept underpins model selection and tuning.

   Sample answer: Bias is error from overly simple assumptions, so the model underfits and misses real patterns. Variance is error from sensitivity to the training data, so the model overfits and performs poorly on new data. Increasing model complexity usually lowers bias but raises variance. I manage the trade-off with cross-validation, regularisation, limiting tree depth, more training data or ensembles like random forests, which reduce variance. In a demand forecasting project, a deep gradient boosting model looked excellent on training data but poor on a holdout month; tuning depth and learning rate improved the holdout error considerably.

2. 2

   Your fraud model has 99 percent accuracy. Is it good?

   Why they ask: This tests understanding of imbalanced data and metric choice.

   Sample answer: Not necessarily. If only 1 percent of transactions are fraud, a model that predicts no fraud for everything also gets 99 percent accuracy and is useless. For imbalanced problems I look at precision and recall for the fraud class, the PR AUC and the confusion matrix at the chosen threshold. Then I tie the threshold to business costs: a missed fraud might cost far more than a manual review. In a payments project I chose a threshold that caught most fraud while keeping manual reviews within the team's daily capacity, and presented the trade-off as a simple table.

3. 3

   What is data leakage, and how have you prevented it?

   Why they ask: Leakage is one of the most common causes of models failing in production.

   Sample answer: Data leakage is when information that would not be available at prediction time gets into training, so offline metrics look much better than real performance. Common causes are features computed using future data, target-derived features, and preprocessing like scaling fitted on the full dataset before splitting. I prevent it with time-based splits for temporal data, scikit-learn pipelines so transformations are fitted only on training folds, and by reviewing each feature's timestamp. In a churn model, a feature called last_contact_reason turned out to be filled in after cancellation; removing it dropped AUC but made the model honest.

4. 4

   How would you build a churn prediction model for a subscription business?

   Why they ask: This case question tests end-to-end thinking from problem framing to action.

   Sample answer: I would first define churn precisely, for example no renewal within 30 days of expiry, and agree on how the business will act on predictions, such as retention offers. Features would include usage trends, support tickets, payment failures, tenure and plan changes, computed only from data before the prediction date. I would start with logistic regression as a baseline, then try gradient boosting, evaluated with time-based validation and PR AUC. SHAP values would explain drivers to the retention team. Finally I would run an experiment to confirm that targeting high-risk users actually reduces churn.

5. 5

   When would you choose logistic regression over XGBoost?

   Why they ask: Interviewers check practical model selection judgement.

   Sample answer: I choose logistic regression when interpretability and stability matter most, data is limited, or regulators need clear explanations, such as in credit scoring. It is fast, easy to monitor and its coefficients are understandable. XGBoost usually performs better on tabular data with non-linear relationships and interactions, and handles missing values well, but it needs more tuning and explanation effort. I often build both: in a lending project, XGBoost gave a modest lift in AUC, but the risk team chose logistic regression with carefully engineered features because it was easier to validate and explain.

6. 6

   How do you explain a complex model's predictions to business stakeholders?

   Why they ask: Communication is often what separates strong data scientists.

   Sample answer: I focus on what decision they need to make and translate the model into that language. I show global drivers with a SHAP summary, then a few individual examples, such as why this customer has a high churn risk. I avoid jargon and use simple comparisons like customers with three or more failed payments are much more likely to leave. I also show where the model is weak. For a sales team, I built a one-page view listing each account's risk score and top three reasons, which they actually used in weekly calls.

7. 7

   How do you decide the sample size for an A/B test?

   Why they ask: This checks statistical rigour in experimentation.

   Sample answer: Sample size depends on the baseline conversion rate, the minimum detectable effect the business cares about, the significance level, usually 5 percent, and the desired power, usually 80 percent. With a 4 percent baseline and a goal to detect a 10 percent relative lift, I would calculate the users needed per variant with a power calculator or statsmodels, then estimate how many days that traffic takes. I would run for full weeks to cover weekly patterns and avoid stopping early when results look good, since peeking inflates false positives.

8. 8

   Have you worked with LLMs? How would you evaluate an LLM-based solution?

   Why they ask: Many data science roles in 2026 include generative AI work.

   Sample answer: Yes. I built a retrieval-augmented support assistant that answered questions from internal documentation. For evaluation, I created a test set of around 200 real questions with reference answers, and measured retrieval quality with recall at k, plus answer correctness and faithfulness to sources using both human review and an LLM-as-judge with spot checks. I tracked latency and cost per query. We also monitored user feedback after launch. The key lesson was that improving chunking and retrieval improved answers more than changing the model.

9. 9

   Tell me about a model that did not perform well in production.

   Why they ask: Interviewers want honesty and understanding of real-world model behaviour.

   Sample answer: A demand forecasting model for a grocery client performed well in validation but its errors grew sharply three months after launch. I found that a new promotion strategy had changed buying patterns, so the training data no longer represented reality. We had no drift monitoring. I added monitoring of input distributions and forecast error by category, retrained with recent data and added promotion features. Error returned to acceptable levels. Since then I always set up monitoring and a retraining plan before a model goes live, not after.

Keep going.

• Machine Learning Engineer Interview Questions and Answers
• Data Analyst Interview Questions and Answers
• Data Engineer Interview Questions and Answers
• Data Science Job Support Online
• Technical Interview Tips That Actually Help
• AI mock interviews