Three rounds in, a Meta interviewer asks you to "explain p-value to a non-technical stakeholder." You freeze. Not because you don't know statistics — you passed the exam — but because you've never had to compress it into two clear sentences under pressure. That's the gap most data science interview prep misses: the difference between knowing something and being able to use it on the spot.
This guide covers the data science interview questions that actually come up at tech companies, analytics teams, and startups — organized by category, with the kinds of answers interviewers find credible. Skip the memorized definitions. Focus on reasoning.
How Data Science Interview Questions Are Structured
Most data science interviews run four to six rounds, and each round targets a different skill layer. Understanding the structure helps you allocate prep time correctly instead of over-indexing on one area.
- SQL / data manipulation — Almost universal. Expect window functions, CTEs, joins on messy keys, and "write a query to find X" prompts.
- Statistics and probability — A/B testing, distributions, Bayes, confidence intervals. Companies that run experiments (most of them) weight this heavily.
- Machine learning concepts — Bias-variance tradeoff, regularization, evaluation metrics, when to use which model. More conceptual than implementation at the interview stage.
- Python / pandas — Usually a take-home or live coding component. Manipulation, aggregation, handling nulls.
- Product / case study — "How would you measure success for Feature X?" Common at FAANG and growth-stage companies.
- Behavioral — Often underestimated. Stories about stakeholder conflict, ambiguous data, and shipping under pressure.
Entry-level roles lean on SQL and statistics. Senior and staff roles shift weight toward ML system design, experimentation methodology, and the product/case questions.
Common Data Science Interview Questions by Category
SQL Interview Questions
SQL is the most consistent filter. Candidates who stumble here rarely advance regardless of their ML knowledge.
- Window functions: "Write a query to find the top 3 revenue-generating products per region." Expect
RANK(),ROW_NUMBER(),LAG(). - Self-joins: "Find all users who logged in on two consecutive days." Tests whether you can think relationally.
- Aggregation with conditions: "Find the percentage of orders that were returned, broken down by product category." Simple concept, easy to get the join wrong.
- NULL handling: "Why does COUNT(*) differ from COUNT(column)?" Interviewers use this to distinguish people who understand SQL semantics from people who just write it.
The most common failure mode is writing a technically valid query that returns wrong results on edge cases — duplicate rows from bad joins, NULL propagation errors, or off-by-one in date range filters. Always think through edge cases out loud.
Statistics and Probability Questions
These questions reveal whether you've actually run experiments or just read about them.
- "Explain p-value without using the words 'null hypothesis' or 'probability.'" The intended answer: "If there were no real effect, a p-value of 0.03 means we'd see a result this extreme in only 3% of experiments run under identical conditions." Most candidates botch this.
- "When would you use a t-test vs. a Mann-Whitney U test?" Interviewers want you to say: when the normality assumption is questionable and sample sizes are small, go non-parametric.
- "How would you design an A/B test for a checkout flow change?" Expected elements: randomization unit (user vs. session), metric choice (conversion rate + guardrail metrics), minimum detectable effect, sample size calculation, and runtime commitment.
- "You run an A/B test and get p=0.04. The PM wants to ship. What do you do?" This is about multiple comparisons, peeking, and practical vs. statistical significance — not just "ship it."
- Bayes basics: "If a disease test is 99% accurate and 1% of the population has the disease, what's the probability someone who tests positive actually has it?" (~50%. Most people get this wrong intuitively.)
Machine Learning Interview Questions
Interviewers rarely ask you to implement an algorithm from scratch. They're checking whether you understand trade-offs.
- Bias-variance tradeoff: "Your model performs well on training data but poorly on holdout. What are your next three steps?" Regularization, more data, simpler model — in that diagnostic order.
- Feature selection: "You have 500 features. How do you decide which to keep?" Variance thresholds, correlation analysis, LASSO, permutation importance — with reasons.
- Imbalanced classes: "Your churn model has 2% positive labels. What metric do you use, and why not accuracy?" Precision-recall AUC or F1, because accuracy is trivially optimized by predicting the majority class.
- Overfitting in practice: "Your model works perfectly in staging but degrades in production within two weeks. What's your diagnostic process?" Feature drift, label leakage, training-serving skew — real candidates have seen these.
- Model choice justification: "Why would you use a gradient boosted tree over logistic regression for this problem?" Interviewers want interpretability trade-offs, not "XGBoost is better."
Product and Case Study Questions
These are increasingly common at mid-to-large companies and weed out candidates who've only done Kaggle competitions.
- "How would you measure the success of Instagram Stories?" Define the north star metric, secondary metrics, and guardrails. The framing matters as much as the answer.
- "Daily active users dropped 10% last Tuesday. Walk me through your investigation." Segment by platform, geography, user cohort, feature. Rule out instrumentation errors first. Work top-down.
- "We want to add a new recommendation algorithm. How do you evaluate it before shipping?" Offline evaluation (precision, recall), then A/B test with guardrail metrics on engagement, retention, and revenue.
What Interviewers Actually Want to See
Across categories, strong candidates share three behaviors that weak candidates skip:
- Clarify before solving. "Is the goal to maximize precision or recall here?" or "Is this a cold-start scenario?" shows you think about problem framing, not just execution.
- Think out loud about trade-offs. Don't just state the answer — state what you're giving up by choosing it. "I'd use SMOTE for class imbalance but it can introduce synthetic noise, so I'd validate on real holdout data only."
- Acknowledge what you don't know. "I'd want to check whether the feature is normally distributed before running that test" lands better than a confident wrong answer.
Interviewers at top companies are not trying to catch you out. They're trying to see if they'd trust you with production data and ambiguous business questions. That comes through in how you reason, not whether you remember every formula.
How to Prepare: A Realistic Approach
Spray-and-pray prep (doing 200 LeetCode SQL problems without a theme) is less effective than targeted skill-gap closure. Run this diagnostic first:
- Do a mock SQL round. If window functions take you more than 3 minutes to set up, that's your gap.
- Explain the central limit theorem out loud. If you stumble, stats is the gap.
- Describe the last model you built and why you chose it. If you can't defend the choice, ML fundamentals need work.
- Solve one product metrics question with a timer. If you don't have a structured framework, that's the gap.
Four to six weeks of deliberate prep is enough for most people who already have a working data background. The bottleneck is usually statistics and product case questions — not SQL, which most candidates have used — because these require building new mental models, not just practice repetition.
Top Courses to Close Interview Prep Gaps
These are courses worth the time specifically for interview preparation, not just general skill-building.
Introduction to Data Analytics (Coursera)
Covers the full analyst workflow from data collection through interpretation — useful for grounding your answers to product/case questions in a structured process. Rated 9.8/10 across thousands of learners.
Tools for Data Science (Coursera)
Builds the Python, SQL, and notebook fluency that shows up in live coding rounds. Better than isolated tool tutorials because it situates each tool in a real workflow context.
Python for Data Science, AI & Development by IBM (Coursera)
IBM's curriculum is unusually good on pandas manipulation and data wrangling — exactly what take-home components test. Rated 9.8/10 and covers the libraries that show up most in interview problems.
Analyze Data to Answer Questions (Coursera)
Specifically designed around translating business questions into data operations, which is the core skill tested in case-study rounds. Strong on SQL aggregation patterns.
Process Data from Dirty to Clean (Coursera)
NULL handling, deduplication, and transformation logic — the unglamorous work that SQL interviewers specifically test because it's what you'll actually do on the job.
Python Data Science (edX)
More rigorous on the statistical foundations than most Python courses. If stats is your weak area, this bridges the gap between knowing Python and being able to answer A/B test design questions under pressure.
Data Science Interview Questions: FAQ
How many rounds does a data science interview typically have?
Most companies run four to six rounds: a recruiter screen, a SQL or Python technical screen, one or two case study rounds, a statistics/ML round, and a final team fit conversation. FAANG companies often add a system design round for senior roles. Total time from application to offer is typically four to eight weeks.
Is SQL tested in every data science interview?
At companies with structured data pipelines (which is most of them), yes. Even ML-heavy roles at companies like Airbnb or Lyft include a SQL round because data scientists pull their own data. The only exception is research-track roles at companies like DeepMind where the expectation is pure modeling work with a data engineering team handling extraction.
What statistics topics come up most often?
A/B testing design and interpretation is the most common. After that: confidence intervals, hypothesis testing assumptions (normality, independence), Type I vs. Type II error, and Bayesian vs. frequentist framing. Probability distributions (normal, Poisson, binomial) and conditional probability appear in probability-specific roles. You won't be asked to derive maximum likelihood estimates from scratch.
Do I need to know deep learning for data science interviews?
For most industry data science roles (as opposed to ML engineering or research), no. Interviewers at non-AI-product companies rarely go deep on neural network architectures. You should be able to describe when you'd choose deep learning over gradient boosted trees and name the situations where each fails. For AI-product companies (OpenAI, Cohere, etc.), the bar is much higher.
How do I answer "tell me about a project you've worked on"?
Use the STAR format but make the "result" quantitative. "I improved churn prediction accuracy" is weak. "I reduced false negatives in our churn model by 18%, which the retention team estimated prevented $400K in annual churn" is strong. If you don't have work experience, frame a personal project or Kaggle competition the same way — hypothesis, approach, result, what you'd do differently.
What's the single biggest mistake candidates make in data science interviews?
Jumping to a solution without clarifying the problem. Interviewers at technical companies specifically probe whether you'll ask the right question before writing a line of code or pulling a dataset. A candidate who says "Before I write that query, can you tell me whether we're counting by user or by session?" immediately signals senior-level judgment, regardless of how the rest of the question goes.
Bottom Line
Data science interview questions fall into predictable patterns, and most candidates fail not from lack of knowledge but from lack of structured preparation. Close your specific gaps — SQL edge cases, A/B test design, or product framing — rather than grinding randomly. Four to six weeks of targeted prep, including at least one mock interview in each category, gets most candidates with a real data background to a competitive position.
Start with your weakest area. If you're unsure what that is, the SQL and statistics diagnostic questions above will tell you quickly.