About 60% of data science candidates fail the technical screen — not because they can't do the job, but because they prepare for the wrong questions. Hiring managers at mid-to-large companies run structured loops: a statistics/probability round, a coding round (usually SQL + Python), an ML concepts round, and a take-home or case study. If you've been grinding LeetCode hard problems thinking that's enough, you're in the wrong lane.
This guide covers the actual data science interview questions that come up repeatedly across FAANG, fintech, and mid-sized startups — grouped by round type, with the reasoning interviewers are looking for, not just the textbook answers.
How Data Science Interview Questions Are Structured
The format varies by company size, but most loops at companies with a dedicated data science function follow this pattern:
- Recruiter screen — background, compensation alignment, timeline
- Technical phone screen — SQL or Python, sometimes a probability question
- Take-home or live coding — a dataset, 3-5 hours, real analysis
- Onsite / virtual onsite — 3-5 rounds covering stats, ML, product sense, and sometimes a presentation of your take-home
Startups under 200 people often skip the take-home and go straight to a working session. Consulting-adjacent roles (McKinsey QuantumBlack, BCG GAMMA) weight the case study heavily and care less about Python syntax.
Statistics and Probability Data Science Interview Questions
This is where most candidates underestimate the depth required. You need to explain concepts without hiding behind formulas.
What is the central limit theorem and why does it matter in practice?
The CLT states that the sampling distribution of the mean approaches a normal distribution as sample size increases, regardless of the population distribution — provided samples are independent and identically distributed. In practice: it's why A/B test results can be analyzed with z-tests even when the underlying metric (revenue per user, session length) isn't normally distributed. Interviewers want you to connect the theorem to a concrete situation, not recite a definition.
Explain p-value to a product manager who has never studied statistics.
This is a communication test as much as a stats test. A good answer: "A p-value tells you how surprising your result would be if there were actually no real difference. A p-value of 0.03 means: if the feature truly had no effect, you'd see a result this extreme only 3% of the time by chance. It doesn't tell you the size of the effect or whether it's practically meaningful." The trap answer is "probability that the null hypothesis is true" — wrong, and interviewers know it.
What's the difference between Type I and Type II errors?
Type I: false positive — you conclude an effect exists when it doesn't (controlled by α, your significance threshold). Type II: false negative — you miss a real effect (controlled by β; power = 1-β). The follow-up question is almost always: "When would you tolerate a higher Type I error?" Good answers involve high-velocity testing environments where the cost of a false positive is low and speed matters, versus medical or financial decisions where false positives are expensive.
Other common statistics questions
- What's the difference between correlation and causation? Give me an example where confounding drove a spurious correlation.
- How do you handle multiple testing problems? (Bonferroni, FDR correction)
- When would you use a t-test vs. a Mann-Whitney U test?
- Explain Bayesian vs. frequentist inference — which would you use for a low-traffic A/B test?
- You run an experiment and get p=0.06. The business wants to ship. What do you say?
SQL and Python Data Science Interview Questions
SQL questions in data science interviews tend to involve aggregation, window functions, and self-joins. Python questions focus on pandas, data cleaning, and sometimes writing a function from scratch (e.g., implement k-means in 20 lines).
SQL: Find the second-highest salary in each department
Classic question. The clean answer uses a window function:
SELECT department_id, salary
FROM (
SELECT department_id, salary,
DENSE_RANK() OVER (PARTITION BY department_id ORDER BY salary DESC) AS rnk
FROM employees
) ranked
WHERE rnk = 2;Interviewers look for whether you reach for DENSE_RANK() vs. a subquery with MAX() — the former is cleaner and handles ties correctly.
Python: How do you handle missing data in a dataframe?
The answer they're looking for isn't just "df.dropna() or df.fillna()." It's a decision framework: understand why data is missing first (MCAR, MAR, MNAR), then choose the strategy. Dropping rows is only safe when missingness is random and the subset is large enough. Mean/median imputation distorts variance. Model-based imputation (MICE, KNN imputer) is better for MAR. Always check whether missingness itself is informative — sometimes adding a binary "was_missing" flag is more useful than the imputed value.
Other SQL/Python questions that come up frequently
- Write a query to calculate 7-day rolling average of daily active users.
- Given two tables (orders, users), find users who ordered more than 3 times in their first 30 days.
- In pandas, how would you identify and handle outliers without removing valid extreme values?
- Implement a function to calculate cosine similarity between two vectors without using sklearn.
Machine Learning Data Science Interview Questions
ML questions span conceptual understanding, implementation choices, and debugging. Senior-level interviews go deeper into regularization, bias-variance tradeoff, and production considerations.
Bias-variance tradeoff
High bias = underfitting (model too simple, misses signal). High variance = overfitting (model too sensitive to training data, doesn't generalize). The tradeoff: as you increase model complexity, bias decreases and variance increases. You want the sweet spot where total error (bias² + variance + irreducible noise) is minimized. Practical follow-up: "How do you detect each?" Bias: high training error. Variance: large gap between training and validation error.
When would you use a random forest vs. gradient boosting?
Random forests are faster to train, more robust to hyperparameter choices, and harder to overfit. Gradient boosting (XGBoost, LightGBM) typically achieves lower error on structured tabular data when tuned, but is more sensitive to learning rate and tree depth. In practice: start with random forest as a baseline, switch to gradient boosting if you need the extra 2-5% accuracy and have time to tune. For production with strict latency requirements, random forest inference is more predictable.
How do you handle class imbalance?
Multiple levers: resampling (SMOTE, undersampling majority), adjusting class weights in the model, choosing the right evaluation metric (precision-recall curve or F1 over accuracy), and at inference time, adjusting the classification threshold. The wrong answer is "I use SMOTE" without explaining what problem you're solving — interviewers want to see that you think about the business cost of false positives vs. false negatives first.
Other ML questions
- Explain regularization (L1 vs. L2) and when you'd use each.
- How do you prevent data leakage in a time-series forecasting problem?
- You deploy a model and its performance degrades over 3 months. Walk me through your investigation.
- What's the difference between precision and recall? Build a confusion matrix for a fraud detection model and explain the tradeoffs.
- How does cross-validation work, and why is k-fold CV inappropriate for time-series data?
Product Sense and Case Study Data Science Interview Questions
At companies where data scientists work closely with product teams (most of FAANG, growth-stage startups), you'll face questions that test whether you can connect analysis to business decisions — not just run models.
A metric drops 15% on Tuesday. Walk me through how you diagnose it.
Structure matters here. Good answer: (1) verify the data isn't a logging/tracking error before panicking; (2) segment by platform, geography, user cohort, feature area; (3) check whether there was a code release, external event, or campaign change on Monday; (4) look at upstream funnel metrics to isolate where the drop begins; (5) quantify impact and form a hypothesis before diving into root-cause analysis. Interviewers are watching whether you jump to explanations before checking the data pipeline.
How would you measure the success of a new recommendation algorithm?
Short-term metrics (CTR, conversion rate on recommendations) are easy to game and don't capture quality. Long-term metrics (retention, revenue per user, diversity of consumed content) matter more. A good answer acknowledges the tradeoff between optimizing for clicks (which can lead to clickbait) and optimizing for satisfaction (harder to measure, often requires user surveys or implicit signals like completion rate). Run an A/B test, but define a guardrail metric — you want to improve recommendation CTR without degrading overall session length.
Top Courses to Close Skill Gaps Before Your Interview
These courses map directly to the interview rounds described above. Don't take all of them — identify your weakest area and start there.
Introduction to Data Analytics
Covers the analytical mindset and core statistical reasoning that underpins most interview questions. Good starting point if your stats foundation is shaky or you're transitioning from a non-quantitative background.
Tools for Data Science
Practical grounding in the Python/R/SQL stack interviewers actually test. Particularly useful for the coding screen — covers pandas, Jupyter workflows, and data manipulation patterns that come up in take-home exercises.
Python for Data Science, AI & Development by IBM
If Python is your weak spot, this course covers the specific libraries (NumPy, pandas, Matplotlib) you'll need to write clean, readable analysis code under interview conditions — not just "run this notebook" familiarity.
Analyze Data to Answer Questions
Focuses on translating business questions into analytical queries — exactly the skill tested in product sense rounds. Covers SQL aggregation, joins, and the reasoning process for structuring an analysis.
Process Data from Dirty to Clean
Data cleaning is underrated as an interview topic. Interviewers frequently give messy datasets in take-homes specifically to see how you handle nulls, duplicates, and type inconsistencies. This course covers it systematically.
Python Data Science (EDX)
A strong alternative for candidates who prefer a university-style curriculum. Covers statistical analysis and machine learning fundamentals at a level of depth that maps well to the ML concepts round.
FAQ: Data Science Interview Questions
How many rounds does a typical data science interview have?
Most mid-to-large companies run 4-6 rounds: a recruiter screen, 1-2 technical screens (SQL/Python, stats), a take-home assignment, and an onsite with 3-4 interviewers covering ML, product sense, and behavioral. Smaller companies often condense this to 2-3 rounds with no take-home.
Is SQL tested in data science interviews?
Yes, almost universally. Even at companies with full data engineering teams, data scientists are expected to write production-quality SQL for analysis. Window functions (RANK, LEAD/LAG, running totals), CTEs, and self-joins come up most often. Practice on real datasets, not toy examples.
How much machine learning theory do I need to know?
You need to understand the intuition behind algorithms — why they work, when they fail, and how to tune them — not memorize proofs. The exception is roles explicitly titled "ML Engineer" or "Research Scientist," which go deeper into mathematical foundations. Applied data scientist roles care more about your ability to select the right model and interpret results.
What's the best way to prepare for the take-home assignment?
Practice on public datasets (Kaggle, UCI repository) with a time limit. The common failure mode isn't statistical — it's presentation. Interviewers penalize notebooks with no narrative, charts with no axis labels, and analyses that don't end with a clear recommendation. Write your findings as if you're presenting to a product manager who won't read your code.
Do I need a degree in data science or statistics to pass these interviews?
No, but you need the knowledge. Many successful candidates come from physics, engineering, economics, or self-taught backgrounds. What matters is whether you can answer the questions accurately and demonstrate that you've applied the concepts on real projects. A portfolio of 2-3 well-documented analyses carries more weight than a degree from an unrecognized program.
How should I prepare for behavioral questions in a data science interview?
Use the STAR format (Situation, Task, Action, Result), but focus on quantifying the result. "I improved the model" is weak. "I improved prediction accuracy from 72% to 81%, which reduced churn-related refunds by ~$40K/quarter" is strong. Have 3-4 stories ready about handling ambiguous problems, influencing decisions without authority, and recovering from a wrong analysis call.
What to Prioritize
If you have two weeks: spend the first week on SQL (window functions, aggregation, common interview patterns) and probability fundamentals (CLT, p-values, confidence intervals). Spend the second week on a realistic take-home practice session and one mock interview. The take-home is where most candidates lose the offer — not because they can't build a model, but because they deliver an unreadable notebook with no business narrative.
If you have a month: add a focused pass through ML concepts (bias-variance, regularization, class imbalance, model selection) and practice explaining your reasoning out loud. The onsite rounds reward candidates who think clearly under pressure, not candidates who memorized the most formulas.
The courses listed above cover all of these areas. Pick the one that maps to your weakest round and go deep on it — surface-level familiarity across five courses is worse than real competence in two.