Data science interviews test three things at once: statistical reasoning, coding fluency (Python/SQL), and the judgment to apply both to messy business problems. The strongest candidates explain their reasoning out loud and tie every technical answer back to a business decision, rather than just naming an algorithm. This guide covers the questions you'll actually get, from probability basics to system design and behavioral rounds.
Data science interviews rarely fail candidates because they don't know a technique — they fail because they can't explain their thinking under pressure, connect concepts to business impact, or handle the follow-up "why" questions that come after a correct answer. This article walks through the question categories you'll actually face, from statistics fundamentals to GenAI curveballs, with guidance on what each question is really testing and how to structure a response that holds up under scrutiny.
Statistics and Probability Questions
Statistics questions are rarely about reciting formulas. Interviewers use them to check whether you understand why a method works and when it breaks down in messy, real-world data.
- P-values: Expect "what does a p-value of 0.03 actually mean?" The trap is saying it's "the probability the null hypothesis is true" — it's the probability of seeing data this extreme if the null is true.
- Distributions: You'll be asked to identify when data looks binomial, Poisson, or normal, and why the Central Limit Theorem justifies using normal approximations even when underlying data isn't normal.
- Hypothesis testing: Questions often center on Type I vs Type II errors and how sample size, effect size, and significance level trade off against each other.
- Bayes' theorem: A classic is the medical test false-positive problem — interviewers want to see you reason through prior probability rather than trust the headline accuracy number.
- Confidence intervals: Be ready to explain that a 95% CI doesn't mean "95% chance the true value is in this range" — it's about the long-run behavior of the estimation method.
What interviewers are really checking: can you translate statistical output into a plain-English recommendation a non-technical stakeholder would trust?
Python and SQL Coding Questions
Coding rounds test fluency, not cleverness. You'll typically see prompts on window functions (running totals, ranking within groups), joins (especially self-joins and handling duplicates), pandas manipulation (groupby, pivot, merge), and list comprehensions versus loops.
- Window functions: Practice writing
ROW_NUMBER(),RANK(), andLAG()/LEAD()queries — these show up constantly for "top N per category" problems. - Joins: Know the difference between inner, left, and full outer joins cold, and be able to explain what happens to row counts when keys aren't unique.
- Pandas manipulation: Expect to reshape a messy dataframe — merging two tables, handling nulls, then aggregating — all in a few lines.
- List comprehensions: These test whether you write idiomatic, readable Python instead of verbose loops.
The skill interviewers actually grade is narration: talk through your approach before typing, name edge cases (nulls, duplicate keys, empty groups) as you go, and test your logic on a small example out loud. Silence during a coding round reads as uncertainty even when the code is correct.
If you're unsure when to reach for SQL versus Python for a given task, it's worth reviewing Python vs SQL for data work — interviewers often ask this directly as a judgment question, not just a coding one.
Master the right skills for your goal
Not sure which path fits? Get a free 1:1 consultation with our team.
Core Machine Learning Algorithm Questions
This section probes whether you understand algorithms as tools with trade-offs, not black boxes you call .fit() on. Interviewers push on assumptions, failure modes, and selection criteria.
- Regression: Be ready to explain the assumptions behind linear regression (linearity, homoscedasticity, independence of errors) and what happens when they're violated.
- Tree-based models: Expect questions comparing decision trees, random forests, and gradient boosting — specifically why ensembling reduces variance and how boosting differs from bagging.
- Clustering: A common follow-up is "how do you choose k in k-means?" — know the elbow method and silhouette score, and be honest about their limitations.
- Algorithm selection: The real test is scenario-based: "you have 500 rows and 200 features, what do you try first?" This checks judgment, not memorization.
A model that's 95% accurate can still be worthless — if the thing you're trying to predict only happens 2% of the time, a model that predicts "never" beats it on accuracy alone.
That tension between accuracy and actual usefulness is exactly why interviewers pair algorithm questions with scenario twists: imbalanced labels, correlated features, or a dataset that's too small for the model you'd default to. Explaining your reasoning for ruling out an approach is often worth more than naming the "right" one.
Model Evaluation and Validation Questions
Once you've built a model, interviewers want proof you can judge whether it's actually good — and whether that goodness will survive contact with production data.
- Precision and recall: Know how to explain the trade-off using a concrete scenario, like fraud detection or medical screening, where false negatives and false positives carry different costs.
- ROC-AUC: Be able to explain what the curve actually plots and why it can be misleading on heavily imbalanced datasets.
- RMSE and regression metrics: Expect a question on why RMSE penalizes large errors more than MAE, and when that's desirable versus problematic.
- Cross-validation: Know the difference between k-fold and time-series split, and why shuffling time-ordered data is a common, costly mistake.
- Overfitting: You should be able to describe symptoms (huge gap between train and validation performance) and fixes (regularization, more data, simpler models) without hesitation.
- Data leakage: This is a favorite senior-level trap question — interviewers want you to spot leakage from features that wouldn't be available at prediction time, like using a "cancellation date" field to predict churn.
The underlying question across all of these: do you trust your own metrics, or do you interrogate them before presenting results?
Feature Engineering and Data Preprocessing Questions
These questions are deliberately scenario-driven because real data is never clean, and interviewers want to see your default instincts before reaching for a library function.
- Missing data: Expect to justify a choice between deletion, mean/median imputation, or model-based imputation, and to explain when missingness itself is informative.
- Encoding categoricals: Know the difference between one-hot and target encoding, and why high-cardinality categorical features (like ZIP code) need special handling.
- Scaling: Be able to explain why tree-based models don't need feature scaling but distance-based models (k-NN, k-means, SVMs) absolutely do.
- Imbalanced classes: Interviewers expect more than "use SMOTE" — talk through class weighting, threshold tuning, and choosing the right evaluation metric as complementary strategies.
- Outlier treatment: Be ready to distinguish outliers that are data errors from outliers that are legitimate rare events, since the correct treatment differs completely.
For candidates working with large or distributed datasets, expect a tie-in question about processing feature pipelines at scale — this is where familiarity with What Is Apache Spark can separate a strong answer from a purely theoretical one.
Advanced and Modern Topics: Deep Learning, NLP, and GenAI
Senior interviews increasingly assume baseline fluency in deep learning and generative AI, even for roles that aren't pure ML engineering. The bar has shifted from "have you heard of transformers" to "can you reason about when to use one."
- Neural network basics: Expect questions on activation functions, backpropagation intuition, and why vanishing gradients make deep networks hard to train without careful design.
- Embeddings: Be ready to explain what an embedding actually represents — a dense vector capturing semantic similarity — and why they outperform one-hot encoding for text and categorical data with many levels.
- Transformers: Know the core idea of self-attention at a conceptual level: the model weighs the relevance of every other token when representing a given token, which is what allows it to capture long-range context.
- LLMs and RAG: Increasingly common is a question like "how would you reduce hallucination in an LLM-based product?" — a strong answer references retrieval-augmented generation, grounding responses in retrieved documents rather than relying purely on parametric knowledge.
If this area feels shaky, it's worth reviewing What Is an LLM and Generative AI vs traditional AI before your interview — both are frequently referenced as the conceptual foundation interviewers assume you already have. Structured, hands-on practice with these concepts is exactly what the Data Science with Python programme is built to cover, from classical ML through to modern GenAI workflows.
Case Study and Product-Sense Questions
Open-ended prompts like "design an A/B test to evaluate a new checkout flow" or "build a churn model for a subscription business" aren't really asking for a perfect answer — they're testing whether you can structure ambiguity.
- Clarify the objective first: Ask what success looks like and what decision the analysis will inform before touching data or metrics.
- Define the metric: State the primary metric and at least one guardrail metric, and explain why you chose them over alternatives.
- Design the approach: For an A/B test, cover randomization unit, sample size/power, and duration. For a churn model, cover the target definition, feature sources, and time-window construction.
- Anticipate pitfalls: Mention novelty effects, seasonality, or leakage risks specific to the scenario — this signals experience over textbook recall.
- Tie back to action: End with what decision the business would make based on different possible outcomes.
Interviewers are grading your framework more than your final number — walking through these steps out loud, even under time pressure, shows structured thinking that a rushed, jump-to-the-answer response can't.
Behavioral and Resume-Based Questions
These questions determine whether you'll be trusted to represent findings to stakeholders, not just produce them. The STAR method (Situation, Task, Action, Result) is the standard structure, but most candidates undersell the Result.
- Talking about a past project: Lead with the business problem, not the algorithm — interviewers want to hear why the work mattered before how it was built.
- Quantify impact: Whenever possible, state the result in numbers (revenue lift, time saved, error reduction) rather than describing the model's technical accuracy alone.
- Handling disagreement: A common prompt is "tell me about a time a stakeholder rejected your analysis." Strong answers show you sought to understand their concern rather than defending the model outright.
- Communicating uncertainty: Be ready to describe how you've presented a result you weren't fully confident in — honesty about limitations builds more trust than false certainty.
Organizations investing in enterprise data training solutions consistently find that technical skill gaps are easier to close than communication gaps — so treat the behavioral round with the same preparation rigor as the technical ones.
- Interviewers weight reasoning and tradeoffs over memorized definitions — always explain why, not just what.
- Coding rounds test SQL joins/window functions and Python/pandas manipulation as much as raw algorithm knowledge.
- Model evaluation questions (precision/recall, cross-validation, leakage) come up more often than exotic ML theory.
- Case study and product-sense rounds reward a structured framework: clarify the goal, define metrics, propose an approach, discuss risks.
- Behavioral answers should follow STAR and quantify impact — 'improved accuracy' is weaker than 'reduced false positives by 18%, saving 200 review-hours/month.'
Glossary
- P-value: Probability of observing a result as extreme as yours if the null hypothesis were true.
- Overfitting: A model fitting training data (including its noise) so closely that it fails to generalize to new data.
- Regularization: A technique (like L1/L2 penalties) that discourages overly complex models to reduce overfitting.
- Cross-validation: Splitting data into multiple train/test folds to get a more reliable estimate of model performance.
- Feature engineering: Creating, transforming, or selecting input variables to improve a model's predictive power.
- Confusion matrix: A table comparing predicted vs. actual classes, used to derive precision, recall, and accuracy.
Frequently asked questions
What is the difference between supervised and unsupervised learning?
Supervised learning trains on labeled data to predict a known target, like classification or regression. Unsupervised learning finds structure in unlabeled data, like clustering or dimensionality reduction. In an interview, give one concrete example of each from your own work rather than just the textbook definition.
Explain the bias-variance tradeoff.
Bias is error from overly simplistic assumptions that cause underfitting; variance is error from excessive sensitivity to training data that causes overfitting. A good model balances both, and you tune this with regularization, model complexity, or more data. Mention that cross-validation is how you actually measure this tradeoff in practice.
How do you handle missing data?
First figure out if data is missing at random or systematically, since that changes the fix. Options include deletion, mean/median imputation, model-based imputation, or adding a 'missingness' flag as a feature. Say which approach you'd pick and why, tied to how much data you'd lose and whether the missingness itself is predictive.
What is a p-value and how do you interpret it?
A p-value is the probability of observing your result, or something more extreme, if the null hypothesis is true. A small p-value (typically below 0.05) suggests the result is unlikely under the null, so you reject it. Be ready to add the caveat that a p-value is not the probability the null hypothesis is true — that's a common trap question.
What is overfitting and how do you prevent it?
Overfitting happens when a model learns noise in the training data instead of the underlying pattern, so it performs well on training data but poorly on new data. Prevent it with cross-validation, regularization (L1/L2), simpler models, more training data, or early stopping. Name a specific case where you diagnosed overfitting using a train/validation performance gap.
How would you explain a machine learning model to a non-technical stakeholder?
Skip the algorithm name and lead with the business outcome: what decision it improves and by how much. Use an analogy or a concrete example prediction, then briefly note the model's confidence and limitations. Interviewers use this question to check communication skill, not technical depth, so keep it under two minutes.
Ready to build this capability?
Browse our upcoming batches — live, instructor-led, delivered on Orbit.