1. Compute month-over-month user retention from an events table.
Self-join cohort month vs activity month, count distinct users, divide by cohort size.
Data Scientist interview questions
Data Scientist loops are the broadest in data: SQL, Python, ML theory, statistics and experimentation, and product-sense cases, all scored on how you reason, not just the final answer. Real questions with answers below.
How a Data Scientist interview runs
- SQL
- Python / pandas
- ML theory & modeling
- Statistics & experimentation
- Case / product sense
- Behavioral
How to prepare for a Data Scientist interview
What the loop actually tests
The broadest loop in data. Six kinds of question, and the bar is not depth in all six but sound reasoning in each: SQL and Python you can write without hesitation, statistics you can explain to a product manager, ML theory you can connect to a decision, and a case where you turn a vague product question into an experiment with a metric.
The hidden test across every round is whether you know what you do not know. A candidate who says which assumption would break their answer scores higher than one who is confident and wrong.
The rounds, in order
Recruiter screen, then a technical screen that is usually SQL plus a statistics or probability question. The onsite is four to six sessions: SQL or Python coding, a statistics and experimentation round (p-values, power, A/B test design, the pitfalls), an ML theory round (bias and variance, regularization, evaluation, what you would do with this data), a product case, and a behavioral session. Some companies add a take-home model and a presentation.
The order matters less than the coverage. Every one of those rounds appears in the question bank below, grouped by round on the topic pages.
What interviewers score
In coding: a working solution with the complexity stated, then edge cases without being prompted. In statistics: the definition, then the assumption, then what you would do if it failed. In ML: which model and why not the other one, which metric and why, and what leakage would look like here. In the case: a metric you defend, an experiment you can size, and a decision rule you state before seeing results.
Communication is scored in every round, usually under a name like clarity or influence. Explaining a confidence interval to a non-technical stakeholder is a question in its own right at several large companies.
How to prepare, and what to do first
Start with a gap analysis against the actual job post, because the emphasis swings hard by company: a product-analytics team weights experimentation and cases, a modeling team weights ML theory and Python. Spend your time in proportion to what the post lists first.
Then rebuild statistics from the questions interviewers ask rather than from a textbook: p-value, power, Type I and II errors, confidence intervals, sample size, multiple comparisons, and the A/B test pitfalls. Those questions repeat across companies with small changes. Keep SQL warm with one timed problem a day.
Where candidates lose the offer
Explaining a p-value as the probability the null is true. Choosing accuracy as the metric for an imbalanced problem. A case answer that jumps to a model before defining what would be measured. And the quiet one: a strong coder who cannot explain a result to the product manager in the room. Practice the explanation as hard as the code.
12 Data Scientist interview questions
2. Given a pandas DataFrame, compute the 30-day moving average of a column.
df['ma'] = df['x'].rolling(30).mean().
3. Explain the difference between .apply(), .map(), and vectorized ops in pandas.
Vectorized ops are fastest, map is element-wise on a Series, and apply is flexible but slower (row or column).
4. Detect and remove outliers using the IQR method in pandas.
Compute Q1 and Q3 with quantile in pandas, take IQR as Q3 minus Q1, and keep rows between Q1 minus 1.5 IQR and Q3 plus 1.5 IQR. Report how many rows the fence removed.
5. Explain p-value in plain language.
Probability of observing data as extreme as this, assuming the null hypothesis is true.
6. How do you determine sample size for an A/B test?
Use power analysis: baseline rate, MDE, alpha, power (0.8) to compute n per arm.
7. What is the difference between Type I and Type II error?
Type I is a false positive (rejecting a true null), Type II is a false negative (failing to reject a false null).
8. Explain the Central Limit Theorem and why it matters.
Sample means approach normality as n grows, enabling inference regardless of population shape.
9. What is p-hacking and how do you avoid it?
Pre-register hypotheses, correct for multiple comparisons, fix sample size in advance.
10. Explain confidence intervals vs credible intervals.
A confidence interval is frequentist coverage over repeated samples, a credible interval is Bayesian probability given the data.
11. How do you detect and handle multicollinearity?
Check VIF and correlations, then drop or combine features, or use regularization or PCA.
12. What metrics would you use for an imbalanced fraud model?
Precision, recall, PR-AUC and F-beta. Avoid raw accuracy, and consider cost-weighted metrics.