The best way to prepare for a Data Science internship interview is to drill three areas: SQL queries, core statistics, and a structured way of talking through a business case. This article gives University of Mumbai students a question bank for all three, with model answer outlines, a scoring sheet and a one-week plan. Rules on internships and training differ between colleges and programmes, so a short list at the end tells you what to confirm with your own college.
How Data Science internship interviews are usually structured
Formats vary by employer, so treat this as a typical pattern rather than a rule. Many entry-level Data Science interviews combine a short introduction, a technical screen, a discussion of one project from your resume and a practical or case round. Smaller teams may fold everything into a single 30 to 45 minute conversation.
- Introduction (5 minutes): who you are, what you have built, why this role.
- SQL or data handling (10 to 15 minutes): reading a schema, writing joins, aggregations and window functions.
- Statistics and machine learning basics (10 to 15 minutes): distributions, hypothesis testing, bias and variance, evaluation metrics.
- Case or project discussion (10 to 15 minutes): framing a problem, choosing metrics, explaining trade-offs.
Interviewers for internships rarely expect deep production experience. They look for clean reasoning, honest limits on what you know, and the ability to explain a result to a non-technical person.
The preparation scorecard
Before you start the question bank, rate yourself honestly on each skill from 1 (cannot do) to 4 (can teach it). Anything scoring 1 or 2 gets your first study hours.
| Skill area | What a 3 or 4 looks like | My score (1-4) |
|---|---|---|
| SQL joins and filters | Write INNER, LEFT and self joins without looking anything up, and explain why a join duplicated rows | |
| SQL aggregation and windows | Use GROUP BY, HAVING, ROW_NUMBER, RANK and running totals correctly | |
| Descriptive statistics | Choose between mean and median, explain skew and outliers | |
| Inference | Explain p-value, confidence interval, Type I and II errors in plain words | |
| Machine learning basics | Explain overfitting, cross-validation, regularisation and when to use which metric | |
| Case framing | State objective, metric, data needed, method and risk in under two minutes | |
| Project storytelling | Describe one project in three minutes, including a mistake you fixed | |
| Communication | Explain a technical result to someone outside the field |
Free 2-Week CareerSwami Internship
Online, 14 days, with a free offer letter and a free completion certificate once you finish the work. Check with your college that it fits your requirement before you enrol.
Question bank 1: SQL
SQL is the most common hands-on screen because it is quick to test. Practise by writing queries on paper or in a local database, then checking the output yourself. The sample tables below are illustrative.
Assume these tables for the questions: orders(order_id, customer_id, order_date, amount), customers(customer_id, name, city, signup_date) and products(product_id, category, price) with order_items(order_id, product_id, quantity).
| # | Prompt | Model answer outline |
|---|---|---|
| S1 | Find the total order amount per city. | Join orders to customers on customer_id; GROUP BY city; SUM(amount). Mention that an INNER JOIN drops customers with no orders, which does not matter here. |
| S2 | List customers who have never placed an order. | LEFT JOIN customers to orders, filter where orders.order_id IS NULL. Alternative: NOT EXISTS. Note why NOT IN can misbehave with NULLs. |
| S3 | What is the difference between WHERE and HAVING? | WHERE filters rows before grouping; HAVING filters groups after aggregation. Give one example of each. |
| S4 | Find the second highest order amount. | Use DENSE_RANK() over amount descending and pick rank 2, or a subquery with MAX below the overall MAX. Discuss ties. |
| S5 | Show each customer’s most recent order. | ROW_NUMBER() OVER (PARTITION BY customer_id ORDER BY order_date DESC), keep row number 1. |
| S6 | Compute a running total of amount by date. | SUM(amount) OVER (ORDER BY order_date) after aggregating to daily totals. Explain the default window frame. |
| S7 | Explain INNER, LEFT, RIGHT and FULL joins. | Describe which unmatched rows each keeps. Use two tiny tables of three rows to show it. |
| S8 | Why did my join return more rows than expected? | One-to-many or many-to-many keys duplicate rows. Check key uniqueness, aggregate before joining, or use DISTINCT only after understanding the cause. |
| S9 | Find the top three product categories by revenue. | Join order_items to products; revenue = SUM(quantity * price); GROUP BY category; ORDER BY revenue DESC; LIMIT 3 (or a rank filter). |
| S10 | Calculate month-over-month growth in revenue. | Aggregate by month, then LAG(revenue) OVER (ORDER BY month); growth = (revenue – previous) / previous. Guard against division by zero. |
| S11 | How do you find duplicate rows? | GROUP BY the identifying columns with HAVING COUNT(*) > 1. To remove, use ROW_NUMBER and delete rows beyond 1. |
| S12 | What is an index and when can it hurt? | A structure that speeds lookups. It costs storage and slows writes, and offers little benefit on low-selectivity columns. |
SQL answering habits that interviewers notice
- Restate the table grain first: one row per order, or per order item?
- Say how you would treat NULLs and ties before you are asked.
- Write the simple version, then offer an optimisation.
- Check your result against a tiny example in your head.
Question bank 2: Statistics and probability
Statistics questions test whether you understand what a number means, not whether you can recite a formula. Keep answers short, then give an example.
| # | Prompt | Model answer outline |
|---|---|---|
| T1 | When would you report the median instead of the mean? | When the distribution is skewed or has outliers, such as incomes or delivery times. The median resists extreme values. |
| T2 | Explain a p-value to a non-technical manager. | If there were truly no effect, how surprising would data this extreme be? A small value is evidence against “no effect”, not proof of a big or useful one. |
| T3 | What is a confidence interval? | A range built by a procedure that would capture the true value in a stated share of repeated samples. Avoid saying the parameter has a 95 percent chance of lying inside this one interval. |
| T4 | Type I versus Type II error? | Type I: rejecting a true null (false alarm). Type II: failing to reject a false null (missed effect). Link to alpha and power. |
| T5 | State the central limit theorem and why it matters. | Sample means approach a normal distribution as sample size grows, given finite variance. It justifies many standard tests and intervals. |
| T6 | Correlation versus causation. | Correlation shows association. Confounders, reverse causality or chance may explain it. Randomised experiments are the cleanest route to causal claims. |
| T7 | How would you design an A/B test for a new app button? | Define the metric, randomise users, estimate sample size from the minimum detectable effect, run for full weekly cycles, avoid peeking early, then test and check guardrail metrics. |
| T8 | Two dice are rolled. What is the probability the sum is 8? | Five favourable outcomes (2-6, 3-5, 4-4, 5-3, 6-2) out of 36, so 5/36. |
| T9 | Explain Bayes’ theorem with an example. | P(A|B) = P(B|A)P(A) / P(B). Use a rare-disease test: even an accurate test gives many false positives when the base rate is low. |
| T10 | What is the bias-variance trade-off? | Simple models underfit (high bias); flexible models chase noise (high variance). Cross-validation and regularisation help find a balance. |
| T11 | Your classes are 98 percent negative. Is 98 percent accuracy good? | No, a model that predicts all negatives gets it. Use precision, recall, F1 or PR-AUC, and consider the cost of each error type. |
| T12 | How do you handle missing values? | First ask why they are missing. Options: drop, impute with median or a model, add a missing indicator. Fit imputers on training data only. |
A quick formula sheet worth memorising
- Precision = TP / (TP + FP); recall = TP / (TP + FN).
- F1 = 2 x precision x recall / (precision + recall).
- Standard error of a mean = standard deviation / square root of n.
- Conditional probability: P(A|B) = P(A and B) / P(B).
Question bank 3: Case discussion
Case questions are open-ended on purpose. A reliable structure keeps you calm: objective, metric, data, method, risks, next step.
| # | Prompt | Model answer outline |
|---|---|---|
| C1 | A food delivery app sees falling repeat orders. How do you investigate? | Define repeat rate and the time window. Segment by city, cohort, restaurant type and delivery time. Check if the drop is across the board or localised. Form hypotheses (delays, pricing, quality), then test. |
| C2 | How would you predict which customers will cancel a subscription? | Define churn clearly, build features from usage and billing history, avoid leakage from after the churn date, start with logistic regression, evaluate with recall at a chosen precision, plan how the business will act on scores. |
| C3 | Estimate how many chai stalls are in a large city. | Break into population, share of people who buy tea outside, purchases per stall per day. State assumptions out loud and sanity-check the result. |
| C4 | A model has 90 percent accuracy offline but performs poorly in production. Why? | Data drift, training-serving mismatch, leakage, different population, delayed labels. Propose monitoring and a shadow deployment. |
| C5 | How would you recommend products on a small e-commerce site with little data? | Start with popularity and category-based rules, then item co-occurrence. Move to collaborative filtering only as data grows. Mention the cold-start problem. |
| C6 | The sales dashboard shows a sudden 30 percent drop. What do you check first? | Data pipeline failures, duplicate or missing loads, definition changes, then genuine causes such as seasonality, outages or campaigns. |
| C7 | How would you explain a model’s prediction to a non-technical stakeholder? | Use top drivers (feature importance or SHAP), a concrete example, and plain language about uncertainty and limits. |
Worked illustrative example: a case answered end to end
This example is illustrative. The numbers, table names and query are invented to show structure and have not been run against real data.
Prompt (C1 variant): “An online grocery service says repeat purchases fell last month. What would you do?”
- Objective: “I would first confirm what ‘repeat purchase’ means here: a customer ordering again within 30 days of a prior order.”
- Metric: 30-day repeat rate = customers with a second order within 30 days / customers with a first order in that month.
- Data needed: order timestamps, customer IDs, city, delivery time, discount flags.
- Method: Compare this month with previous months, then split by city, new versus old customers, and delivery delay band to find where the drop is concentrated.
- Risks: Recent customers have had less time to reorder, so the latest cohort looks worse. Compare cohorts at equal age.
- Next step: If one city explains most of the drop, check operations there, then propose a controlled test of a fix.
A first-pass query for the metric might look like this (illustrative, PostgreSQL-style):
WITH first_orders AS (
SELECT customer_id, MIN(order_date) AS first_date
FROM orders
GROUP BY customer_id
)
SELECT DATE_TRUNC('month', f.first_date) AS cohort_month,
COUNT(DISTINCT f.customer_id) AS new_customers,
COUNT(DISTINCT o.customer_id) AS repeated_in_30_days
FROM first_orders f
LEFT JOIN orders o
ON o.customer_id = f.customer_id
AND o.order_date > f.first_date
AND o.order_date <= f.first_date + INTERVAL '30 days'
GROUP BY 1
ORDER BY 1;In the interview, narrate the query: the CTE finds each customer’s first order, the LEFT JOIN keeps customers who never returned, and dividing the last two columns gives the repeat rate. Then mention the cohort-age caveat from step 5, which is the detail that separates a careful answer from a rushed one.
How to score a practice answer
Practise with a friend or record yourself, then score each answer on the sheet below. Aim for steady gains over a week rather than a perfect first round.
| Criterion | 0 points | 1 point | 2 points |
|---|---|---|---|
| Correctness | Wrong or unsupported | Mostly right with gaps | Correct and justified |
| Structure | Rambling | Some order | Clear sequence stated up front |
| Assumptions | Hidden | Mentioned late | Stated early and checked |
| Edge cases | None | One mentioned | Relevant ones handled (NULLs, ties, leakage) |
| Plain-language summary | None | Technical only | One-sentence takeaway a manager could use |
A seven-day preparation plan
- Day 1: Complete the scorecard. Revise joins and GROUP BY; do S1 to S3.
- Day 2: Window functions; do S4 to S6 and S10.
- Day 3: Descriptive statistics, distributions and the central limit theorem; do T1, T5 and T8.
- Day 4: Hypothesis testing, A/B tests and Bayes; do T2 to T4, T7 and T9.
- Day 5: Machine learning evaluation, leakage and missing data; do T10 to T12.
- Day 6: Two case questions out loud with a timer; score them using the sheet above.
- Day 7: Rehearse your three-minute project story and a mock interview with a classmate.
Building a project story that survives follow-up questions
Interviewers often spend more time on one project than on all the textbook questions combined. Pick the project you understand best, not the most impressive title.
- Problem: what question were you answering and for whom?
- Data: where did it come from, how large was it, what was wrong with it?
- Approach: why that model or method over a simpler baseline?
- Result: which metric, compared with what baseline?
- Mistake and fix: one thing that went wrong and what you changed.
- Next step: what you would do with more time or data.
If a project came from a course assignment, say so. Honest framing is easier to defend than inflated claims, and follow-up questions will test the details.
Getting project experience if your resume is thin
Many students reach the interview stage with only classroom exercises. A short, structured internship can add a project you can discuss. The free two-week online internship from CareerSwami runs for 14 days, comes with a free offer letter and a free completion certificate after successful completion, and is offered across 26 tracks. It is an independent programme, so check that your tasks fit the Data Science topics above, and treat the work as material for your project story rather than a substitute for practice.
University of Mumbai and Data Science interviews: OJT rules, CTPC and syllabus links
Interview preparation is easier to plan when it is tied to what the university publishes. Several University of Mumbai documents bear directly on a Data Science internship, though the rules for your own programme depend on your college and syllabus year.
For postgraduate students, the University’s On-the-Job Training (OJT) guidelines are framed under NEP 2020 and apply from 2023-24. They state that PG OJT carries 4 credits and 120 clock hours, and is to be done in Semester II or the break after it, with any unfinished OJT cleared before the end of Semester III. In practice, interviews should be completed early in Semester II, so the SQL, statistics and case practice above should start before then. The file name includes “Draft”, so check that you are reading the current version.
The same guidelines split evaluation 50% external (host organisation) and 50% internal (department mentor), out of 100 marks. The suggested internal components include weekly reporting, a written report and a viva or presentation. This is a reason to choose a project you can explain clearly: the viva and the interview reward the same skill. The appendices include an undertaking, outreach letter, resume template, student diary or log, attendance sheet, completion certificate and report format. The undertaking has a “Mode of Internship” field, but the document does not clearly say how online internships are treated, so ask before accepting a remote offer.
The Career Training and Placement Cell (CTPC) page lists mock interviews, career guidance and internship or apprenticeship coordination among its functions, and says the University’s Internship Cell is part of it. The page snapshot dates from 2024 and may have changed, so confirm with the cell directly.
For syllabus mapping, the B.Sc. Data Science programme syllabus (Academic Council, 2021) includes an Artificial Intelligence course with practical in Semester 3. A T.Y.B.Sc. Computer Science Semester VI Data Science syllabus from 2023 is also hosted. Both may have been superseded, so use them only as a starting point for matching units to the question banks.
Checks before you book interviews
- Ask your department whether the PG OJT guidelines apply to your programme and whether the final version differs from the draft.
- Confirm the OJT window for your semester and the deadline for submitting the log, report and completion certificate.
- Ask whether a remote or online internship is acceptable and how to fill the “Mode of Internship” field.
- Contact the CTPC or Internship Cell about mock interviews and current internship notices.
- Download your current syllabus from the University syllabus page and mark the units that cover SQL, statistics and machine learning.
Data Science interview mistakes that cost University of Mumbai candidates offers
- Memorising answers word for word. Interviewers rephrase questions, and rote answers collapse.
- Using DISTINCT to hide a join problem instead of finding its cause.
- Quoting accuracy on imbalanced data without discussing recall and precision.
- Listing tools you cannot explain. Every skill on your resume is fair game.
- Skipping questions for the interviewer. Ask about the team’s data, the tools they use and what a good first month looks like.
Frequently Asked Questions
How much SQL do I need for a Data Science internship interview?
Be comfortable with joins, GROUP BY with HAVING, subqueries, CASE expressions and basic window functions such as ROW_NUMBER, RANK and LAG. Questions S1 to S11 above cover the usual range for entry-level screens.
Do I need machine learning projects, or are SQL and statistics enough?
Requirements differ by role. Analyst-leaning internships weigh SQL, statistics and communication heavily, while modelling roles expect at least one end-to-end project. Read the posting and match your preparation to it.
How should I answer if I do not know a question?
Say what you do know, state the assumption you would make, and reason aloud toward an answer. Admitting a gap and showing how you would find the answer is better than guessing confidently.
Which Python topics should I revise alongside this?
Pandas operations that mirror SQL (merge, groupby, pivot), handling missing values, basic plotting, and train-test splits with scikit-learn. Be able to explain what each line of your own notebook does.
How long should I spend preparing?
A focused week using the plan above is a reasonable start if you already know the basics. If your scorecard shows several 1s, allow two to four weeks and spread the sessions out.
Can I use a free online internship as project experience in an interview?
Yes, if you can explain what you did, what data or tools you used and what you learned. Describe it accurately, including its length and scope, and do not present it as something larger than it was.
Will an internship certificate count for academic credit at my college?
That depends entirely on your college and programme. Ask your coordinator in writing using the list above before assuming any credit or mark allocation.
Which questions about the team’s data, tools and mentoring should I ask a Data Science interviewer at the end?
Good options: what the intern’s first project would be, what data and tools the team uses, how interns are mentored, and how success is judged by the end of the internship.
Related Data Science guides
- Data Science Internship Subfields Compared for JNTUH Students: Analytics, ML and Data Engineering
- Turning an Academic Project into a Data Science Internship Demo: A Guide for JNTUK Students
- Lovely Professional University Students: Data Science Internship Portfolio and Project Guide
- All Data Science internship guides
Ready to put this into practice?
Join the free two-week CareerSwami online internship in Data Science. The internship, the offer letter and the completion certificate are free.