Roughly 60% of data science candidates who fail technical screens say they were blindsided by statistics questions — not Python or ML. They over-indexed on model tuning and got knocked out in the first round by a question about p-values or the central limit theorem. This guide covers the full range of data science interview questions you'll actually face: statistics, SQL, machine learning, coding, and case studies, with a framework for how to prepare each category systematically.
How Data Science Interview Questions Are Structured
Most data science interviews at mid-to-large companies follow a predictable pipeline. Understanding the format tells you where to spend your prep time.
- Recruiter screen — resume, background, timeline. No technical content.
- Technical phone screen — 45–60 minutes. Usually SQL or Python, occasionally one statistics question.
- Take-home assignment — a real dataset, 3–8 hours, submitted as a notebook. You'll need to clean data, run analysis, and communicate findings clearly.
- Virtual on-site — 4–6 rounds: statistics, ML concepts, SQL live coding, a product/business case, and behavioral. Sometimes a presentation of your take-home.
The on-site is where most candidates lose offers. Each round is with a different interviewer and tests a distinct skill — cramming "data science" broadly is less useful than mastering each module separately.
Statistics and Probability: The Most Common Data Science Interview Questions
Stats questions appear in nearly every data science interview, yet many bootcamp graduates arrive under-prepared for them. These aren't graduate-level — but they require genuine understanding, not surface-level definitions.
Central Limit Theorem
You'll be asked to explain it and then apply it. A common prompt: "We ran an A/B test for 5 days and got a significant result. Should we ship?" The right answer involves sample size, variance of the metric, and whether 5 days captures weekly seasonality — not just the p-value. Knowing the CLT means knowing why the p-value alone is insufficient.
A/B Testing and Hypothesis Testing
Expect questions like: "Walk me through how you'd design an A/B test for a new checkout flow." Interviewers want to hear you mention: null hypothesis, minimum detectable effect, statistical power (typically 80%), significance threshold (usually 0.05), sample size calculation, and how you'd handle multiple comparisons. If you can't calculate sample size from scratch, practice this.
Confidence Intervals vs. P-Values
Many interviewers ask candidates to explain a confidence interval to a non-technical stakeholder. A common trap: treating "95% CI" as "there's a 95% chance the true value is in this range." That's wrong. The correct interpretation trips up a surprisingly large proportion of candidates — including people with statistics degrees.
Bayes' Theorem
Classic prompt: "1% of the population has disease X. A test is 99% accurate. If someone tests positive, what's the probability they actually have the disease?" Many candidates guess ~99%. The actual answer is around 50% (using Bayes' theorem with the base rate). This tests intuition about false positives in imbalanced datasets — directly applicable to fraud detection, spam filtering, and medical diagnosis work.
SQL Interview Questions for Data Scientists
SQL is tested in almost every data science interview. The questions typically go beyond simple SELECTs into window functions, self-joins, and aggregation logic. You should be able to write these without reference material under time pressure.
Window Functions
Know ROW_NUMBER(), RANK(), DENSE_RANK(), LAG(), LEAD(), and NTILE() cold. A typical question: "Given a table of user sessions, find each user's second purchase." The intended solution uses ROW_NUMBER() OVER (PARTITION BY user_id ORDER BY purchase_date) filtered to row 2.
Retention Queries
"Calculate 30-day retention" is a staple. You need to self-join a user events table on user_id where the second event falls within 30 days of the first event date. Being able to write this fluently, with correct date arithmetic, separates candidates who've done real analytics work from those who've only done tutorials.
Funnel Analysis
Given a table of events (viewed_product, add_to_cart, purchased), calculate conversion rate at each step. This involves conditional aggregation — COUNT(DISTINCT CASE WHEN event = 'add_to_cart' THEN user_id END) — and interviewers watch for whether you handle distinct users correctly.
Machine Learning Interview Questions
ML questions in data science interviews are mostly conceptual and applied, not mathematical derivations. Interviewers want to see that you understand trade-offs, not that you've memorized textbook definitions.
Bias-Variance Trade-off
This comes up in almost every ML round. Describe a model that is overfitting, and what you'd do to fix it. The answer isn't just "add regularization" — it's knowing when to collect more data, reduce model complexity, add dropout, or use cross-validation to evaluate. Interviewers listen for whether you think empirically or just recite theory.
Handling Imbalanced Data
"Your fraud detection model shows 99% accuracy. Your client is thrilled. What's the problem?" This tests whether you understand class imbalance. The right follow-up: look at precision, recall, and F1, because a model that predicts "not fraud" 100% of the time also gets 99% accuracy on a 1% fraud rate dataset. Solutions include SMOTE, class weights, threshold adjustment, and evaluation via precision-recall curves rather than ROC-AUC alone.
Feature Selection and Engineering
A real prompt: "You have 500 features. How do you decide which to keep?" Good answers cover: correlation with target, mutual information, variance inflation factor for multicollinearity, tree-based feature importance, and the cost of maintaining features in production. Mentioning production cost signals seniority.
Model Deployment and Monitoring
Senior roles ask: "You've deployed a model. Six months later, accuracy drops. What happened and what do you do?" This is about data drift vs. concept drift — the input distribution changed vs. the relationship between inputs and output changed. Detection methods include population stability index (PSI), monitoring KL divergence on feature distributions, and model performance dashboards. This question is frequently asked and frequently fumbled.
Business Case and Product Questions
At product-focused companies (Meta, Airbnb, Lyft, DoorDash), one or two rounds will be business cases — also called "metrics" rounds. These are not technical but require a structured thinking framework.
A typical prompt: "Daily active users dropped 15% last Tuesday. Walk me through how you'd diagnose it." A strong answer segments the problem: platform (iOS/Android/web?), region, user cohort (new vs. retained?), feature area (core loop vs. notification?), external factors (holiday, competitor launch, app store update?). The interviewer is checking whether you think systematically before jumping to conclusions.
Another common one: "How would you measure the success of a new onboarding flow?" This tests whether you can define a north-star metric, identify leading indicators, and anticipate confounds — not just say "look at conversion rate."
Top Courses to Prepare for Data Science Interview Questions
The most useful preparation combines structured coursework with deliberate practice on real questions. These courses build the technical foundations that interviewers actually test.
Introduction to Data Analytics (Coursera)
Rated 9.8/10. Covers the full analytics workflow — data collection, cleaning, visualization, and communicating insights — which directly maps to what interviewers look for in take-home assignments and business case rounds.
Tools for Data Science (Coursera)
Rated 9.8/10. Gets you fluent in the practical toolkit — Jupyter, Git, Python, R — before you're in front of an interviewer with a live coding environment. Less about theory, more about being functional under pressure.
Python for Data Science, AI & Development by IBM (Coursera)
Rated 9.8/10. Strong foundation for the Python coding round. Pandas, NumPy, and data manipulation are covered at the depth that technical screens actually require — not just toy examples.
Analyze Data to Answer Questions (Coursera)
Rated 9.8/10. Focused specifically on taking a question, finding relevant data, and structuring an answer — which is exactly the skill being evaluated in case-study rounds and take-home assignments.
Process Data from Dirty to Clean (Coursera)
Rated 9.8/10. Data cleaning is consistently underestimated by candidates but tested heavily in take-homes. This course addresses common messy-data scenarios that show up in real assignments.
Python Data Science (edX)
Rated 9.7/10. Good complement if you want a second perspective on the core Python data stack. edX's pacing works well for people who need to fit prep around a current job.
FAQ: Data Science Interview Questions
How many rounds does a typical data science interview have?
Most mid-size and large tech companies run 4–6 rounds: a recruiter call, a technical phone screen, a take-home, and 3–4 virtual on-site rounds. Startups often compress this to 2–3 rounds. Consulting firms may add a separate presentation round.
Is SQL or Python more important for data science interviews?
Both are tested, but SQL appears more consistently across company types. Even companies with heavy Python workflows test SQL because it's the baseline for data access and ad hoc analysis. If you're short on time, prioritize SQL first — window functions and aggregation are the most-tested areas.
Do data science interviews test machine learning math?
Rarely in applied roles. Most interviews test conceptual understanding (bias-variance trade-off, regularization, model evaluation metrics) rather than derivations. Research scientist roles at ML-focused companies are the exception — those may ask you to derive gradient descent or backpropagation.
What statistics do I need to know for data science interviews?
Focus on: hypothesis testing (t-test, chi-squared, Mann-Whitney U), confidence intervals, p-values and their limitations, A/B test design (power, sample size), Bayes' theorem, and probability distributions (normal, Poisson, binomial). You don't need graduate-level statistics — you need to apply these correctly in realistic scenarios.
How long should I spend preparing for data science interviews?
For someone with a working foundation in Python and some statistics background: 6–10 weeks of focused prep is realistic for reaching interview-ready level at competitive tech companies. For someone switching from a non-technical field, budget 4–6 months. The take-home project round is often the deciding factor — practice completing full analyses from raw data to written summary under time pressure.
What's the most common reason people fail data science interviews?
Communication. Candidates who can code but can't explain their reasoning clearly tend to get passed over — especially in business case rounds and when presenting take-home work. Interviewers are evaluating whether they'd trust you in a stakeholder meeting. Practice talking through your thought process out loud while solving problems, not just getting to the right answer silently.
Bottom Line
Data science interviews test five distinct areas — statistics, SQL, ML concepts, Python, and business case reasoning — and most candidates are weak in at least two of them. The fastest way to identify your gaps is to do a mock technical screen and a take-home assignment under timed conditions, then trace the failures back to root causes.
Statistics is the highest-leverage area to strengthen if you haven't already. It shows up in every on-site, it's the round that separates candidates most clearly, and it's the one people most reliably underestimate. Build your SQL window function fluency in parallel — that skill transfers directly to day-one work and interviewers know it.
The courses above will build the technical foundations. Beyond that, the preparation that actually moves outcomes is deliberate practice on realistic prompts — not re-reading documentation.