Data Science Interview Questions: What's Actually Asked (and How to Prepare)

About 70% of data science candidates who clear the resume screen fail the technical interview — not because they don't know the material, but because they've prepared for the wrong questions. Most prep guides recycle the same 50 conceptual definitions. Real interviews at companies like Meta, Amazon, and mid-size startups lean heavily on applied statistics, SQL under time pressure, and open-ended case problems where there's no single right answer.

This guide breaks down exactly what data science interview questions look like in practice, organized by the type of round you'll face, with specific examples and the preparation approaches that actually move the needle.

How Data Science Interview Loops Are Structured

Before drilling into individual data science interview questions, it helps to know what you're walking into. Most companies run 4-6 rounds:

• Recruiter screen — background, salary, timeline. Skip-worthy for prep purposes.
• Technical phone screen — usually 1-2 SQL or Python questions, sometimes a stats question. 30-45 minutes.
• Take-home or timed analytics challenge — given a dataset, asked to explore, model, and present findings. 2-8 hours.
• On-site / virtual on-site — 3-5 back-to-back rounds: SQL, ML theory, coding (Python/pandas), product/metrics, and behavioral.
• Hiring manager round — communication, past work, culture fit.

The rounds that eliminate the most candidates are SQL under time pressure and the product/metrics round. Most candidates over-prepare for ML theory and under-prepare for everything else.

Statistics and Probability: The Data Science Interview Questions That Trip People Up

These questions aren't hard in the academic sense — they're hard because candidates memorize definitions instead of building intuition. Interviewers notice immediately.

Common questions and what they're testing

• "Explain p-value to a non-technical stakeholder." — They're testing whether you can communicate. The correct answer has nothing to do with 0.05. A good answer: "It's the probability of seeing results this extreme if there were actually no real effect. A small p-value means your result would be very unlikely if chance alone explained it — but it doesn't tell you the effect is large or practically important."
• "You run an A/B test. Variant B wins with p=0.03. Should you ship it?" — No single right answer. They want you to ask: What's the sample size? Was the test run for the correct duration? Is there a multiple comparisons problem? What's the practical effect size? What are the risks of a false positive here?
• "What's the difference between Type I and Type II error? Which is worse?" — Context-dependent. Worse to have a false positive in a medical trial; worse to have a false negative in a fraud detection system. They want to hear you reason about costs, not recite definitions.
• "A coin is flipped 10 times and lands heads 8 times. Is it fair?" — Walk through a quick binomial calculation. P(8+ heads | p=0.5) ≈ 5.5%. Not conclusive evidence, but you'd want more data.

What to actually practice

Work through problems where you derive expected values, conditional probabilities, and confidence intervals from scratch — not just recognize their definitions. Bayesian vs frequentist framing comes up often at FAANG and quantitative hedge funds.

SQL and Python: The Data Science Interview Questions Most People Underestimate

SQL rounds are where strong ML engineers get knocked out. You'll be writing queries in real time on a shared screen with no autocomplete, usually against a schema you've never seen before.

SQL patterns that appear constantly

• Window functions — RANK(), ROW_NUMBER(), LAG(), LEAD(). Questions like "Find users who made a purchase in consecutive months" require window functions or self-joins. If you can't write PARTITION BY fluently, fix that first.
• Retention and cohort analysis — "What percentage of users who signed up in January were still active in March?" This is a self-join on user activity tables. It comes up constantly at product-analytics-heavy companies.
• Funnel analysis — Given an events table, calculate conversion rate at each step of a signup funnel. The trick is handling users who skip steps or re-enter the funnel.
• Deduplication — "This table has duplicate rows due to a bug. Write a query to return only the most recent record per user." Common answer uses ROW_NUMBER() in a CTE.

Python / pandas interview questions

Python rounds are less about syntax and more about thinking out loud. Common patterns:

• Merge two dataframes, handle NaN values, explain your choices
• Implement a rolling 7-day average on a time series
• Given a list of transactions, compute the top N customers by spend — they want to see if you use groupby + nlargest or write a loop
• Write a function to detect outliers (IQR method vs z-score — explain the tradeoff)

Machine Learning: What Data Science Interview Questions Actually Test

ML questions at most companies aren't asking you to derive backpropagation from scratch. They're testing whether you can choose the right approach for a business problem and explain your reasoning clearly.

The framing questions

• "How would you build a recommendation system for our platform?" — This is deliberately open. Start with clarifying questions: Is this cold-start or returning users? What data do we have? What's the latency requirement? They want collaborative filtering vs content-based vs hybrid reasoning, not a full implementation.
• "Your model has 95% accuracy on the test set. The business isn't happy. What's wrong?" — Class imbalance. A model that predicts the majority class every time achieves 95% accuracy if positives are 5% of the data. Ask about precision, recall, and what the false negative cost is.
• "When would you use a random forest vs gradient boosting?" — Random forests are faster to train, less prone to overfitting on small datasets, more interpretable. Gradient boosting generally achieves better accuracy on tabular data when tuned but requires more careful hyperparameter work. Use XGBoost or LightGBM when you have time to tune; random forest when you need something reliable out of the box.
• "Explain regularization to me." — L1 (Lasso) drives coefficients to exactly zero — useful for feature selection. L2 (Ridge) shrinks them toward zero but rarely to zero — useful when you want to keep all features but reduce variance. Elastic net is both.

The experiment design questions

Increasingly common, especially at product-led companies: "How would you measure whether feature X improved user engagement?" These require you to define your metric, identify confounders, choose between A/B test and observational methods, and reason about statistical power. Candidates who've only studied ML algorithms struggle here.

Product and Metrics Rounds: The Data Science Interview Questions Nobody Prepares For

At companies with mature data teams, the product/metrics round is often the hardest. A typical question: "Daily active users dropped 15% last Tuesday. Walk me through how you'd investigate."

The structure they want to see:

1. Sanity check the data — Is this a tracking bug? Did the logging pipeline fail? Check whether other metrics moved correspondingly.
2. Segment the drop — Is it platform-specific (iOS vs Android vs web)? Geography? User segment (new vs returning)? Feature area?
3. External factors — Holiday? Competitor outage driving traffic to us normally? Major news event?
4. Internal changes — Any deployments that day? A/B tests launched? Marketing campaigns paused?
5. Form a hypothesis, test it — Based on what you find, what's the likely cause and what data would confirm it?

This is a structured thinking exercise. Interviewers are watching your process, not waiting for you to name a specific cause.

Top Courses to Prepare for Data Science Interviews

The courses below are worth recommending because they build the applied skills — SQL, Python, statistics — that the interview rounds above actually test, rather than covering ML theory in isolation.

Introduction to Data Analytics (Coursera)

Builds the foundational analytical thinking that underpins the product/metrics interview round. Strong on framing business problems as data problems — the skill most self-taught candidates are missing.

Tools for Data Science (Coursera)

Covers the practical toolchain — Jupyter, GitHub, SQL, Python — that interviewers assume you know. Useful for candidates coming from non-CS backgrounds who need to close tooling gaps quickly.

Python for Data Science, AI & Development by IBM (Coursera)

IBM's Python track is unusually good at connecting syntax to real data tasks. The pandas and data wrangling sections map directly to what you'll write in a Python interview round.

Analyze Data to Answer Questions (Coursera)

Focused entirely on the analysis layer — taking a dataset and deriving defensible answers. The kind of structured thinking this teaches is exactly what case-study interview rounds reward.

Process Data from Dirty to Clean (Coursera)

Real-world data is almost always messy. This course covers data validation, cleaning, and transformation — skills that come up in take-home challenges and SQL rounds alike.

Python Data Science (edX)

edX's offering moves faster and goes deeper into statistical modeling than most Python courses. Worth it if you already have basic Python and want to close the statistics-in-code gap before a technical screen.

FAQ: Data Science Interview Questions

How long does it take to prepare for a data science interview?

Realistically, 4-8 weeks of focused preparation if you're already working in a data-adjacent role. If you're switching from a different field, plan for 3-6 months to close the SQL and statistics gaps. The mistake most people make is cramming definitions in the last two weeks instead of building applied fluency over months.

What's the most common reason candidates fail data science interviews?

SQL. Candidates who've primarily worked in Python or R often haven't written complex SQL under time pressure. Window functions, CTEs, and cohort queries are the specific gaps. Practice writing queries against a timer with no IDE support — that's the actual interview condition.

Do I need to know deep learning for a data science interview?

Depends entirely on the role. For product analytics or business intelligence roles: no. For ML engineering or applied research roles: yes, you should be comfortable with the architecture and training process even if you're not deriving math. Most general data scientist roles focus more on regression, tree-based models, and experimental design than deep learning.

Should I memorize common data science interview questions?

Don't memorize answers — build understanding. Interviewers probe follow-up questions. If you've memorized "p-value is the probability of observing results as extreme as..." but can't explain what happens to your p-value if you double the sample size, you'll get caught. Build intuition, not flashcard decks.

How important is GitHub for data science interviews?

Interviewers do look at GitHub portfolios before on-site rounds, but a GitHub profile rarely makes or breaks a decision. A strong profile with a well-documented project can serve as a conversation anchor — interviewers will ask you to walk through it. An empty GitHub profile is a mild negative signal for mid-to-senior roles but rarely a dealbreaker for entry-level positions.

What salary can I expect after passing data science interviews?

Entry-level data scientist roles in the US typically land between $90K-$130K total compensation. Mid-level (3-5 years) ranges from $130K-$180K at most companies, with FAANG and quant funds paying significantly above market. The salary gap between candidates who interview well and those who don't is substantial — strong interviewers often clear 20-30% higher offers than equally-skilled candidates who struggled through rounds.

Bottom Line

Data science interview questions test a specific combination of skills that no single course covers completely: applied statistics, SQL fluency, Python for data manipulation, ML model selection, and business problem framing. The candidates who perform best have practiced the applied work — not just read about it.

If you're starting from scratch, prioritize in this order: SQL (specifically window functions and cohort analysis), Python pandas, statistics intuition, then ML concepts. Save deep learning for last unless the job description explicitly requires it.

The courses above from Coursera and edX cover these foundations well. Pick the one that closes your biggest gap rather than trying to complete all of them — focused preparation beats breadth every time.