Data Science Interview Questions
Data science interviews filter for people who reason about data honestly: what could bias this, what would falsify it, does this metric mean what we think. Modelling technique matters less than statistical judgement at every company worth joining.
Q1Explain the bias-variance trade-off with a concrete example of each failure mode.
What they're checking: Whether the core ML concept is intuition or memorisation.
A strong answer: High bias: linear model on non-linear data, underfits both sets. High variance: deep tree memorising noise, train-test gap. Strong answers tie to model capacity, data size and regularisation.
Q2Your model has 95% accuracy. Why might that be terrible?
What they're checking: Metric literacy — the most common trap in applied ML.
A strong answer: Class imbalance: 95% accuracy on 95% negatives is a coin flip on the class you care about. Pivot to precision/recall, PR-AUC, and choosing metrics from the cost of each error type.
Q3Design an A/B test for a checkout change. What could silently invalidate your result?
What they're checking: Experimental rigour beyond "split traffic 50/50".
A strong answer: Unit of randomisation vs interference, sample-size before peeking, novelty effects, and Simpson-paradox segments. Naming a pre-registered metric and stopping rule marks real experience.
Q4What is data leakage? Give an example subtle enough to pass code review.
What they're checking: The failure that ships confident nonsense to production.
A strong answer: Future information in training: scaling fit on the full dataset before the split, or a feature like "account_closed_date" that encodes the label. Strong answers include how to catch it — time-based splits, suspiciously good scores.
Q5Write a SQL query for each user's second-highest purchase. Now explain its cost on a billion rows.
What they're checking: SQL beyond SELECT — window functions plus scale awareness.
A strong answer: ROW_NUMBER() over user partitions ordered by amount, filter rank 2 — then discuss the sort behind the window, partition pruning, and pre-aggregation if this runs daily.
Q6How do you detect and handle drift once a model is in production?
What they're checking: Whether your ML experience ends at the notebook.
A strong answer: Monitor input distributions and prediction distributions against training baselines, plus delayed ground-truth metrics. Distinguish data drift from concept drift; retraining cadence tied to detection, not calendar.
Q7Explain p-values to a product manager without saying "null hypothesis".
What they're checking: Communication — half the job is translation.
A strong answer: "If the change truly did nothing, how surprising would this result be?" — plus what it does not mean: not the probability the change works, and significance is not business impact.
Q8When would you choose a simple model over a state-of-the-art one?
What they're checking: Engineering maturity versus leaderboard chasing.
A strong answer: Interpretability requirements, small data, latency budgets, maintenance cost, and baseline-first discipline: ship logistic regression, measure, then justify every point of added complexity.
Now try answering these under interview pressure
InterviewGPT runs a mock Data Science interview — asks questions like these tailored to your situation, listens to your spoken answers, and scores every one with written feedback. First session free, no credit card.
Start a free Data Science interview