Data Science Interview Questions for Freshers

Practise reasoning from evidence. These questions help you discuss a beginner investigation honestly and clearly; they are not a list of promised interview topics. They are grouped by theme. For each one we describe what a strong answer shows and what a weak answer sounds like.

Fresher interviews for analyst and data roles in India usually combine statistics basics, a SQL or spreadsheet exercise, a discussion of your project and a few questions about judgement. The project discussion is where most candidates separate themselves, so prepare it first.

Statistics and reasoning

Mean or median?

Strong answer: use a skewed example. For values 20, 22, 23, 25 and 90, the mean is 36 and the median is 23. You explain why the long tail matters and that a manager may need both a typical experience and information about serious delays. Weak answer: declaring one measure always better.

What is selection bias?

Strong answer: the inclusion process makes observed records differ from the population of interest. A satisfaction survey completed only by engaged users can miss frustrated people who left. You explain what you would inspect and how you would narrow a claim if the bias cannot be corrected. Weak answer: a definition with no example and no consequence for the conclusion.

Does correlation establish cause?

Strong answer: no, and you discuss confounding, selection or reverse causation using an example, then identify what additional design or evidence would help. You avoid turning the answer into "nothing can be learned"; an association can still guide a useful investigation when described accurately. Weak answer: a one-word "no" with nothing about what you would do next.

What does a p-value tell you, and what does it not?

Strong answer: how surprising the data would be if the null hypothesis were true; it does not give the probability that the hypothesis is true, and it says nothing about whether the effect matters. Weak answer: "below 0.05 means it is significant" with no sense of effect size or context.

Why does a larger sample not fix a biased one?

Strong answer: more data from the same distorted process gives a more precise wrong answer. You give an example. Weak answer: assuming size solves everything.

Working with data

How would you start an unfamiliar dataset?

Strong answer: clarify the question, source, permitted use and row meaning; inspect quality before modelling; look for a plausible-looking field that might be recorded after the outcome and therefore be unavailable for prediction. Weak answer: "I would run a model and see".

How do you handle missing values?

Strong answer: first ask why they are missing, because missingness itself can carry information; then choose between dropping, imputing or flagging depending on the question and the volume. Weak answer: "fill with the mean" as a reflex.

Write a query that finds the three slowest delivery zones last month.

Strong answer: you talk through the grouping, the aggregate, the date filter and the ordering before writing, and you ask what "slowest" should mean, mean or median. Weak answer: typing immediately and getting the grouping wrong.

How do you make analysis reproducible?

Strong answer: document sources, dependencies, transformations and randomisation where relevant; run the notebook from a clean session; ensure another person can obtain permitted inputs. Weak answer: "I saved the notebook".

Modelling and judgement

When would you build a model?

Strong answer: when prediction or another modelling objective helps answer the question and suitable data exists; a descriptive report may be enough for a question about past performance. You explain the baseline and evaluation you would use if modelling is justified. Weak answer: "whenever there is enough data".

How would you evaluate a classifier for a rare event?

Strong answer: accuracy is misleading when one class dominates; you discuss precision, recall, the cost of each kind of error and choosing a threshold for a business reason. Weak answer: quoting accuracy alone.

A stakeholder wants a conclusion the data does not support. What do you do?

Strong answer: show what the data can support, state clearly what it cannot, and propose what evidence would settle the question. Weak answer: giving them the conclusion anyway, or refusing without an alternative.

About your project

Explain a changed conclusion

Strong answer: choose a project where checking the data changed your interpretation. State the initial idea, the evidence that challenged it and the revised finding, and finish with the practical limitation. Weak answer: a project story in which nothing surprised you.

What is the weakest part of your analysis?

Strong answer: a specific limitation, its likely effect on the conclusion and what you would do to address it. Being able to explain what your analysis cannot establish is part of statistical competence, not a weakness to hide. Weak answer: "I cannot think of one".

How to talk about your project

Use a fixed shape and practise it until it takes about three minutes. Start with the question and who cared about the answer. Describe the data, how it was collected and what that means for what it can show. Say what you did yourself. Give the main finding with its uncertainty, one plot you would show, and the limitation that could change the recommendation. Then stop and invite questions.

Expect follow-ups on the data quality checks you ran, the alternative explanation you ruled out and the one you could not. If you do not know, say what you would check. Confident reasoning about limits is more credible than a clean story.

Questions to ask the interviewer

• Where does the team's data come from, and who owns its quality?
• What does a typical analysis request look like, and who asks for it?
• How are findings communicated: dashboards, memos, meetings?
• How much of the work is SQL and reporting versus modelling?
• What would a new joiner be expected to deliver in the first three months?

These questions show that you understand analysis is done for someone, not in isolation. Keep salary and leave questions for a later round unless the interviewer raises them.

Mistakes freshers make in these interviews

• Reciting algorithm names when the question was about the data.
• Giving a textbook definition with no example.
• Claiming certainty about a cause when the data shows an association.
• Having no story about something that went wrong.

How GullyMentor approaches it

A GullyMentor mentor runs mock interviews on the investigation you actually did, including SQL exercises and the uncomfortable follow-ups above. See the Data Science track at /career-tracks/data-science. We do not guarantee jobs, interviews or salaries.

Sources & further reading

• pandas getting started tutorials
• Khan Academy: Statistics and probability
• OpenIntro Statistics (free textbook)
• Kaggle Learn: Intro to SQL
• scikit-learn model selection and evaluation

Examples and practice recommendations are educational illustrations. Tool documentation and programme inclusions should be checked for your intended use.