CareerGuide 13 of 36
Career guide · Guide 13 of 36
Top 50 Data Science Interview Questions and Answers
Top 50 data science interview questions with clear answers: statistics, probability, A/B testing, pandas and case studies, for freshers and experienced.
On this page
These 50 questions cover what data science interviews ask, from freshers to experienced candidates: statistics and probability, A/B testing and experiments, pandas, causal questions and case studies that test how you think. Each answer is written in plain words and gets straight to the point. Questions about specific algorithms (trees, boosting, k-means and so on) are in the Top 50 machine learning interview questions.
Question list
Show all 50 questions
- What does a data scientist do?
- What is the data science lifecycle?
- What is the difference between a population and a sample?
- What is the central limit theorem, and why does it matter?
- What is the difference between descriptive and inferential statistics?
- What is a normal distribution?
- What is the difference between variance and standard deviation?
- What is a p-value, in simple words?
- What are Type I and Type II errors?
- What is the difference between correlation and causation?
- What is exploratory data analysis (EDA)?
- How do you handle outliers in data science work?
- Which Python libraries does a data scientist use?
- What is the difference between a data scientist, a data analyst and an ML engineer?
- How do you explain a technical result to a business audience?
- How do you design an A/B test?
- How do you calculate the sample size for an A/B test?
- What is statistical power?
- What is the difference between a t-test, a z-test and a chi-square test?
- What is a confidence interval, and how do you interpret it?
- What is the multiple testing problem?
- What is Bayes’ theorem? Give an example.
- What is conditional probability?
- How would you analyse whether a new feature increased engagement without an A/B test?
- How do you use pandas to find the top 3 products by revenue in each category?
- How do you join and aggregate data in pandas compared with SQL?
- What is sampling bias, and how do you avoid it?
- What is regression analysis used for in data science?
- What is the difference between bias and variance in a model, with an example?
- How do you evaluate a regression model?
- What is cohort analysis, and how is it used in data science?
- How do you detect and handle multicollinearity?
- What is the difference between a parameter and a statistic?
- What are common pitfalls when analysing an A/B test?
- How do you handle imbalanced classes in a data science project?
- Case: how would you build a model to predict customer churn?
- Case: our sign-up conversion dropped 15% this week. How do you investigate?
- Case: how would you estimate the impact of a price increase?
- How would you measure the success of a recommendation system?
- What is CUPED, and why is it used in experiments?
- How do you decide whether a model is ready for production?
- How do you prevent data leakage in a real project?
- How would you build a demand forecast for an online shop?
- What would you do if your model performs well offline but poorly in production?
- How would you build a model when you have very few labelled examples?
- How do you prioritise data science projects?
- Tell me about a data science project you are proud of.
- How do you communicate uncertainty in your results?
- How do you handle a stakeholder who wants a specific answer from the data?
- How are LLMs changing data science work?
Beginner data science interview questions (freshers)
1. What does a data scientist do?
A data scientist uses data, statistics and machine learning to help a business make better decisions. Typical work includes exploring data to understand a problem, designing and analysing experiments such as A/B tests, building models that predict outcomes like churn or demand, and explaining the results to people who will act on them.
Day to day, much of the time goes into getting and cleaning data with SQL and Python, checking whether the data can answer the question at all, and turning results into a clear recommendation. The best data scientists combine solid statistics, practical coding and good judgement about what the business actually needs.
2. What is the data science lifecycle?
A common way to describe it:
- Understand the problem: what decision will this support, and how will success be measured?
- Get the data: find sources, query them and check access and quality.
- Explore and clean: look at distributions, missing values and odd records; fix or document problems.
- Model or analyse: start with a simple baseline, then try better methods.
- Evaluate: use honest validation and a metric that matches the business goal.
- Communicate and deploy: share findings, or put the model into a product or process.
- Monitor and improve: track performance and update as data changes.
In interviews, stress that the steps loop back: evaluation often sends you back to the data.
3. What is the difference between a population and a sample?
The population is the whole group you want to learn about, such as all customers of a bank. A sample is the part you actually measure, such as 2,000 customers surveyed. Because measuring everyone is usually impossible or too expensive, you use the sample to estimate things about the population, such as the average balance.
The key question is whether the sample represents the population. A random sample usually does; a sample of only customers who answer emails may not, because they can differ from the rest. Statistics such as confidence intervals tell you how much a sample estimate might differ from the true population value.
Learn more: Probability Distributions and the Central Limit Theorem
4. What is the central limit theorem, and why does it matter?
The central limit theorem says that if you take many random samples of a reasonable size from almost any population and calculate each sample’s mean, those means will be approximately normally distributed around the true mean, even when the original data is not normal. The spread of those sample means gets smaller as the sample size grows.
It matters because it lets us use normal-distribution tools, such as confidence intervals and z-tests or t-tests, to make statements about averages from skewed real-world data, like revenue per user. It is the reason A/B tests on averages work. It needs independent observations and a large enough sample; very heavy-tailed data needs larger samples or other methods.
Learn more: Probability Distributions and the Central Limit Theorem
5. What is the difference between descriptive and inferential statistics?
Descriptive statistics summarise the data you have: averages, medians, percentages, standard deviations and charts. For example, “the average order value last month was 52”. They describe; they do not generalise.
Inferential statistics use a sample to draw conclusions about a larger population or about cause and effect, with a measure of uncertainty. For example, “the new checkout increased conversion by 1.2 percentage points, with a 95% confidence interval of 0.4 to 2.0”. Hypothesis tests, confidence intervals and regression models are inferential tools. Most business questions need both: describe first, then infer carefully.
Learn more: Descriptive Statistics: Mean, Median, Spread and Outliers
6. What is a normal distribution?
A normal distribution is the symmetric, bell-shaped distribution defined by its mean and standard deviation. Values near the mean are most common, and values become rarer further away. About 68% of values fall within one standard deviation of the mean, about 95% within two and about 99.7% within three.
Many measurements, such as heights or measurement errors, are roughly normal, and averages of many observations tend to be normal because of the central limit theorem. Much real business data is not normal, though: incomes, order values and session lengths are usually skewed with long right tails. Always plot the data before assuming normality.
Learn more: Probability Distributions and the Central Limit Theorem
7. What is the difference between variance and standard deviation?
Both measure how spread out values are. Variance is the average of the squared differences from the mean. Standard deviation is the square root of the variance, which brings it back to the same units as the data. If order values are in rupees, the variance is in rupees squared, which is hard to interpret, while the standard deviation is in rupees.
In practice you report the standard deviation, while variance appears inside formulas and models, such as when combining independent sources of variation, since variances add. For samples, divide by n − 1 instead of n to get an unbiased estimate of the population variance.
Learn more: Descriptive Statistics: Mean, Median, Spread and Outliers
8. What is a p-value, in simple words?
A p-value is the probability of seeing a result at least as extreme as the one you observed, assuming there is really no effect (the null hypothesis is true). A small p-value, such as 0.01, means your result would be rare if nothing were going on, which is evidence against the null hypothesis.
What it is not: it is not the probability that the null hypothesis is true, and it says nothing about how big or important the effect is. A tiny effect can be significant with a huge sample. Always report the effect size and a confidence interval alongside it, and decide the significance level before looking at the data.
Learn more: Hypothesis Testing and P-Values Explained Simply
9. What are Type I and Type II errors?
A Type I error is a false positive: concluding there is an effect when there is none, such as launching a feature that does not really help. Its probability is the significance level, alpha, often 5%. A Type II error is a false negative: missing a real effect, such as dropping a feature that would have helped. Its probability is beta, and power equals 1 − beta.
There is a trade-off: lowering alpha makes false positives rarer but false negatives more common, unless you increase the sample size. Choose based on which mistake is more costly for the decision.
Learn more: Hypothesis Testing and P-Values Explained Simply
10. What is the difference between correlation and causation?
Correlation means two things move together; causation means changing one actually changes the other. Customers who use a feature may spend more, but that might be because engaged customers both use more features and spend more, not because the feature causes spending.
Correlation can come from a hidden common cause (confounding), from reverse causation, or from chance. To establish causation, the gold standard is a randomised experiment such as an A/B test. When experiments are impossible, methods such as difference-in-differences, matching or instrumental variables can help, but they rely on assumptions you must state clearly.
11. What is exploratory data analysis (EDA)?
Exploratory data analysis is the first look at a dataset to understand it before modelling. You check the size and structure, data types, missing values, duplicates, the distribution of each variable, outliers and relationships between variables, using summary statistics and charts such as histograms, box plots and scatter plots.
The goals are to find data quality problems, understand what each column really means, spot patterns worth testing, and form hypotheses. Good EDA often changes the plan: you may discover the target is defined differently than expected or that a key field is mostly empty. Write down what you find, because it shapes every later decision.
12. How do you handle outliers in data science work?
First decide whether an outlier is an error or a real but unusual value. Errors, such as an age of 250 or a price with extra zeros, should be fixed or removed. Real extreme values, such as a very large customer, carry information and usually should stay.
For real outliers, choose methods that are robust to them: report medians as well as means, cap extreme values at a percentile (winsorising), transform skewed variables with a log, or use models like tree-based methods that are less sensitive. Always check whether your conclusions change with and without the outliers, and say so in your results.
Learn more: Descriptive Statistics: Mean, Median, Spread and Outliers
13. Which Python libraries does a data scientist use?
The core set: pandas for tables, NumPy for numerical arrays, matplotlib and seaborn (or plotly) for charts, scikit-learn for classic machine learning, SciPy and statsmodels for statistics and statistical models, and XGBoost or LightGBM for gradient boosting. For deep learning, PyTorch is the most common choice. Notebooks such as Jupyter are used for exploration.
For larger data, Polars, DuckDB or PySpark help. Interviewers care less about the list and more about whether you can use pandas and scikit-learn fluently and explain what the code does.
14. What is the difference between a data scientist, a data analyst and an ML engineer?
A data analyst focuses on describing and explaining what happened, using SQL, spreadsheets and dashboards. A data scientist goes further into statistics, experiments and predictive models, often with Python. An ML engineer focuses on building, deploying and maintaining machine learning systems in production, with more software engineering.
In small companies one person may do all three. In larger ones, a data scientist might build a churn model, an ML engineer turns it into a reliable service, and an analyst tracks its business impact in a dashboard. See data engineer vs data scientist vs data analyst.
15. How do you explain a technical result to a business audience?
Start with the decision and the answer, not the method: “We should send the retention offer to the top 10% riskiest customers; it will likely save about 300 customers a month.” Then give the evidence in one or two clear charts. Explain uncertainty in plain words (“we’re fairly confident it’s between 200 and 400”), and mention the main limitation.
Avoid jargon such as AUC or p-values unless the audience uses them; translate them into what they mean for the decision. Keep the technical details ready for questions or an appendix. Finish by agreeing on the next step.
Statistics, experiments and coding questions
16. How do you design an A/B test?
- Define the goal and metric: one primary metric, such as conversion rate, plus guardrail metrics that must not get worse, such as refunds.
- Write the hypothesis: the new design increases conversion.
- Choose the unit of randomisation, usually the user, so one person always sees the same version.
- Calculate the sample size from the baseline rate, the smallest effect worth detecting, the significance level and the power.
- Run the test for the planned time, ideally whole weeks to cover weekly patterns, without stopping early.
- Check the test worked: balanced groups and no sample ratio mismatch.
- Analyse with the planned test, report the effect with a confidence interval, and decide.
Learn more: A/B Testing: Design, Sample Size and Analysis
17. How do you calculate the sample size for an A/B test?
You need four inputs: the baseline value of the metric (for example 5% conversion), the minimum detectable effect you care about (for example an increase to 5.5%), the significance level (usually 5%) and the power (usually 80%). A power calculation, using a formula or a library such as statsmodels, then gives the number of users needed per group.
The main intuition: smaller effects and noisier metrics need much larger samples, since required size grows roughly with one over the effect size squared. If the required sample is more traffic than you have, test a bolder change, use a more sensitive metric, run longer, or use variance-reduction techniques such as CUPED.
Learn more: A/B Testing: Design, Sample Size and Analysis
18. What is statistical power?
Power is the probability that a test detects a real effect of a given size, if it exists. A power of 80% means that if the true effect is as large as you assumed, you will get a significant result 80% of the time, and miss it 20% of the time (a Type II error).
Power increases with larger samples, larger true effects, lower variance and a higher significance level. Underpowered tests are a common problem: they often miss real improvements, and when they do find a significant result, the estimated effect tends to be exaggerated. Plan power before running a test, not after.
Learn more: Hypothesis Testing and P-Values Explained Simply
19. What is the difference between a t-test, a z-test and a chi-square test?
A z-test compares means or proportions when the population standard deviation is known or the sample is large; it is often used for conversion rates in big A/B tests. A t-test compares means when the standard deviation is estimated from the sample; it is the usual choice for averages such as revenue per user, and becomes almost identical to a z-test for large samples. A chi-square test checks whether categorical variables are related, such as whether the device type and the purchase outcome are independent, by comparing observed counts with expected counts.
Choose based on the data type (means or counts) and sample size, and check the assumptions, such as independence.
Learn more: Hypothesis Testing and P-Values Explained Simply
20. What is a confidence interval, and how do you interpret it?
A confidence interval is a range of plausible values for something you estimate. For a 95% confidence interval, if you repeated the study many times and built an interval each time, about 95% of those intervals would contain the true value.
In practice: “the conversion lift is 1.2 points, 95% CI 0.4 to 2.0” means the data are consistent with a real lift somewhere between 0.4 and 2.0 points. It is not strictly “a 95% probability the true value is in this interval”, although that informal reading is common. Narrow intervals mean precise estimates; if an interval for a difference includes zero, the result is not significant at that level.
Learn more: Hypothesis Testing and P-Values Explained Simply
21. What is the multiple testing problem?
If you test many hypotheses at a 5% significance level, some will look significant by chance alone. Test 20 metrics with no real effects, and on average one will show a “significant” difference. Reporting only the winners leads to false discoveries.
Solutions: choose one primary metric in advance; treat the others as supporting or exploratory; adjust significance levels with methods such as Bonferroni (simple, strict) or Benjamini-Hochberg (controls the false discovery rate); and confirm surprising results with a follow-up test. The same problem appears when you slice results into many segments after the test.
Learn more: Hypothesis Testing and P-Values Explained Simply
22. What is Bayes’ theorem? Give an example.
Bayes’ theorem updates a probability when you get new evidence: P(A | B) = P(B | A) × P(A) / P(B).
Classic example: a disease affects 1% of people. A test detects it 99% of the time, and gives a false positive 5% of the time. If someone tests positive, what is the chance they have the disease? Out of 10,000 people, 100 have it and 99 test positive; of the 9,900 without it, 495 test positive. So the chance is 99 / (99 + 495), about 17%, far lower than most people guess. The lesson: when the condition is rare, even accurate tests produce many false positives.
23. What is conditional probability?
Conditional probability is the probability of an event given that another event has happened, written P(A | B). It equals P(A and B) / P(B). For example, the probability a customer buys, given that they added an item to the cart, is much higher than the overall probability of buying.
It matters because most business questions are conditional: churn probability for customers with a support complaint, or fraud probability for transactions from a new device. A common mistake is confusing P(A | B) with P(B | A): the probability that a fraudulent transaction is from a new device is not the same as the probability that a new-device transaction is fraudulent.
24. How would you analyse whether a new feature increased engagement without an A/B test?
Without randomisation, you must compare carefully and state assumptions. Options:
- Before-and-after with a control: compare the change in users who got the feature with the change in similar users who did not, over the same period (difference-in-differences). This removes general trends and seasonality, assuming both groups would have moved in parallel.
- Matching: compare users who adopted the feature with similar non-adopters, matched on past behaviour.
- Staggered rollouts by region or time can act as natural experiments.
Watch for self-selection: engaged users adopt features first, which inflates naive comparisons. Report results as estimates with caveats, and recommend a proper experiment for big decisions.
Learn more: A/B Testing: Design, Sample Size and Analysis
25. How do you use pandas to find the top 3 products by revenue in each category?
Calculate revenue per product, then rank within each category:
import pandas as pd
sales = pd.DataFrame({
"category": ["A", "A", "A", "A", "B", "B"],
"product": ["p1", "p2", "p3", "p4", "p5", "p6"],
"revenue": [100, 300, 200, 50, 80, 120],
})
top3 = (sales.groupby(["category", "product"], as_index=False)["revenue"].sum()
.sort_values(["category", "revenue"], ascending=[True, False])
.groupby("category")
.head(3))
print(top3) category product revenue
1 A p2 300
2 A p3 200
0 A p1 100
5 B p6 120
4 B p5 80groupby(...).head(3) keeps the first three rows of each category after sorting. If ties matter, use rank(method="dense") within each group and filter on the rank instead, and say how ties should be handled.
In an interview, also say how you would check the result: compare the summed revenue with the total in the raw data, and look at one category by hand to confirm the ranking.
26. How do you join and aggregate data in pandas compared with SQL?
pandas has direct equivalents: merge is a SQL JOIN (how="left", "inner" and so on), groupby(...).agg(...) is GROUP BY, boolean filtering is WHERE, filtering after aggregating is HAVING, sort_values is ORDER BY, and drop_duplicates is DISTINCT. Window functions map to groupby(...).transform(...), rank and shift.
Two practical tips: always check the row count after a merge, because duplicate keys multiply rows, just as in SQL, and use validate="many_to_one" in merge to catch that early. For large data, run the heavy joins in the database or a tool like DuckDB, and bring only the result into pandas.
27. What is sampling bias, and how do you avoid it?
Sampling bias happens when the data you analyse is not representative of the population you want to learn about, so conclusions do not generalise. Examples: a survey answered only by very happy or very unhappy customers; analysing only customers who completed sign-up, which hides those who dropped out; or survivorship bias, such as studying only companies that are still in business.
Avoid it by defining the target population first, sampling randomly where possible, checking the sample against known totals such as age or region mix, and weighting results when some groups are under-represented. When bias cannot be removed, state it clearly and explain how it might change the conclusion.
28. What is regression analysis used for in data science?
Regression models the relationship between an outcome and one or more explanatory variables. Data scientists use it to predict values, such as sales from price and promotion, and to estimate effects, such as how much a 1% price increase changes demand, holding other factors constant.
For explanation, the coefficients, their confidence intervals and the model’s assumptions matter: linear relationships, independent errors, no strong multicollinearity and no important missing variables. A coefficient shows association, not causation, unless the design supports causal claims. For prediction, what matters most is performance on new data, measured with validation.
29. What is the difference between bias and variance in a model, with an example?
Bias is error from wrong assumptions: a straight-line model fitted to a curved relationship will be wrong in a consistent way, however much data you give it. Variance is error from sensitivity to the particular training sample: a very deep decision tree fits each training set closely, so its predictions change a lot with small changes in the data.
Example: predicting house prices with only the number of rooms has high bias; a model that memorises every training house has high variance. The best model balances the two, which you find with validation, regularisation and enough data.
Learn more: Overfitting, Cross-Validation and Data Leakage
30. How do you evaluate a regression model?
Common metrics: MAE (mean absolute error) is the average size of the errors, easy to explain and robust to outliers. RMSE (root mean squared error) penalises large errors more, useful when big misses are especially costly. R-squared shows the share of variance explained, useful for comparison but not for judging business usefulness alone. MAPE gives percentage errors, but breaks down when actual values are near zero.
Always compare against a simple baseline, such as the average or last year’s value, check errors across segments, and plot predicted against actual values to spot patterns the metrics hide.
31. What is cohort analysis, and how is it used in data science?
Cohort analysis groups users by when they started, such as the month of sign-up, and tracks each group’s behaviour over time, such as the percentage still active after one, two and three months. It shows whether newer customers behave better or worse than older ones, which an overall average can hide when the customer mix changes.
Data scientists use cohorts to measure retention, the long-term effect of product changes, and customer lifetime value, and to build features for churn models. See retention, cohort and funnel analysis in SQL for how to build one.
32. How do you detect and handle multicollinearity?
Multicollinearity is when explanatory variables are strongly correlated with each other, such as house size in square feet and number of rooms. It does not hurt prediction much, but it makes individual coefficients unstable and hard to interpret, with large standard errors and sometimes surprising signs.
Detect it with a correlation matrix or the variance inflation factor (VIF; values above about 5 to 10 are a warning). Handle it by removing or combining correlated variables, using domain knowledge to choose the most meaningful one, applying regularisation such as ridge regression, or using dimensionality reduction. If you only need predictions, you may be able to leave it.
33. What is the difference between a parameter and a statistic?
A parameter is a number that describes the whole population, such as the true average spend of all customers. It is usually unknown. A statistic is a number calculated from a sample, such as the average spend of 1,000 sampled customers, and is used to estimate the parameter.
The difference matters because statistics vary from sample to sample. Statistical inference, through confidence intervals and hypothesis tests, is about quantifying how far a sample statistic might be from the population parameter. Greek letters such as μ and σ usually denote parameters; Latin letters such as x̄ and s denote statistics.
34. What are common pitfalls when analysing an A/B test?
- Peeking: checking results daily and stopping when significant inflates false positives.
- Sample ratio mismatch: if a 50/50 split shows 52/48, assignment or logging is broken, and results cannot be trusted.
- Novelty and learning effects: early behaviour may differ from long-term behaviour.
- Wrong unit of analysis: randomising by user but analysing by session understates uncertainty.
- Too many metrics and segments without correction.
- Network effects: in marketplaces or social products, treatment can affect control users.
- Ignoring practical significance: a significant 0.1% lift may not be worth the cost.
Plan the analysis before the test starts, and check the data before reading the result.
Learn more: A/B Testing: Design, Sample Size and Analysis
35. How do you handle imbalanced classes in a data science project?
First, measure with the right metrics: precision, recall, F1, the precision-recall curve, not accuracy. Then improve the model with class weights, oversampling the rare class (for example SMOTE) or undersampling the common one, applied only to the training data. Often the most important step is choosing the decision threshold from business costs, such as the cost of missing a fraud versus reviewing a false alarm.
Also ask whether more rare-class examples can be collected, and remember that resampling changes predicted probabilities, so recalibrate them if they are used directly. More detail is in the machine learning interview questions.
Case study and experienced data science interview questions
36. Case: how would you build a model to predict customer churn?
- Define churn with the business, such as no purchase in 60 days, and the prediction point: “on the first of each month, predict churn in the next 60 days”.
- Build the dataset: one row per customer per month, with features computed only from data before the prediction date, such as recency, frequency, spend trend, support tickets and plan type.
- Split by time: train on earlier months, test on later ones.
- Model: start with logistic regression, then gradient boosting.
- Evaluate with precision and recall among the highest-risk customers, since the business can only contact some.
- Act and measure: test the retention offer with a control group to measure real saved customers, not just model accuracy.
37. Case: our sign-up conversion dropped 15% this week. How do you investigate?
First, check it is real: data pipeline delays, tracking changes, a new definition, or a holiday can all fake a drop. Then segment: by traffic source, device, browser, country, new versus returning, and funnel step, to find where the drop is concentrated. A drop limited to one browser suggests a bug; a drop in one marketing channel suggests a traffic-quality change.
Next, list what changed that week: releases, experiments, campaigns, pricing, competitor moves, outages. Form a hypothesis, test it with data, and quantify the impact. Report clearly: what happened, the evidence, the estimated cost and the recommended fix, with an owner.
38. Case: how would you estimate the impact of a price increase?
The best method is an experiment: randomly show the new price to some customers or regions and compare conversion and revenue. Where that is not possible, use past price changes, comparing affected and unaffected products or regions over the same period (difference-in-differences), or estimate price elasticity with a regression that controls for seasonality, promotions and competitor prices.
Look beyond immediate revenue: churn, customer mix, support contacts and long-term lifetime value can change too. Present a range of outcomes with assumptions clearly stated, and recommend a staged rollout so the estimate can be checked with real data.
39. How would you measure the success of a recommendation system?
Offline, before launch: ranking metrics on historical data, such as precision at k, recall at k, NDCG and coverage of the catalogue. These help compare models but do not prove business value.
Online, the real test: an A/B test measuring click-through rate, add-to-cart, purchases, revenue per user and long-term engagement, with guardrails such as returns and diversity of recommendations. Watch for feedback loops, where the system only learns from what it already shows, and popularity bias, where it keeps recommending the same top items. Track performance for new users and new items separately, because they are the hardest cases.
40. What is CUPED, and why is it used in experiments?
CUPED (controlled-experiment using pre-experiment data) is a variance-reduction technique for A/B tests. It adjusts each user’s metric using their behaviour before the experiment, such as their spend in the previous month, which is correlated with the outcome but not affected by the treatment. Removing that predictable part of the variation makes the treatment effect easier to detect.
The benefit is the same statistical power with fewer users or a shorter test, often a substantial reduction for metrics with strong user-level consistency. It requires reliable pre-period data for users, and the covariate must be measured before randomisation so the treatment cannot influence it.
Learn more: A/B Testing: Design, Sample Size and Analysis
41. How do you decide whether a model is ready for production?
Check five things. Performance: it beats the current approach or a simple baseline on honest, time-based validation, and the gain matters to the business. Robustness: it performs consistently across important segments and over time. Fairness and compliance: no unacceptable gaps between groups, and explanations where needed. Operational fit: latency, cost, data availability at prediction time, and monitoring are in place. Decision process: everyone knows how the predictions will be used and what happens when the model is wrong.
Then launch gradually, ideally as an experiment, so the real impact is measured rather than assumed.
42. How do you prevent data leakage in a real project?
Think about time for every feature: “would I have known this at the moment of prediction?” Build features as of a prediction date and use time-based splits. Watch for target leakage, such as a cancellation reason field when predicting cancellations, and for features updated after the event. Fit scalers, encoders and imputers inside a pipeline on training data only. Remove duplicate or related records that cross the train-test boundary, such as the same customer in both.
A warning sign is a model that is suspiciously good. Investigate the top features: leakage usually shows up as one feature with surprisingly high importance.
Learn more: Overfitting, Cross-Validation and Data Leakage
43. How would you build a demand forecast for an online shop?
Start with questions: what level (product, category, warehouse), what horizon (days or weeks ahead), and what decision it supports (ordering stock). Build simple baselines first: last year’s value and a seasonal average. Then add models: classical time-series methods per series, or a gradient boosting model across many products with features such as lags, rolling averages, calendar effects, promotions and price.
Validate with rolling backtests over past periods, measure error in units the business understands, and provide ranges rather than single numbers so the ordering decision can account for uncertainty. Monitor errors after launch, especially around promotions and holidays.
44. What would you do if your model performs well offline but poorly in production?
Look for differences between the training setup and the real world. Common causes: data leakage that inflated offline results; training-serving skew, where features are calculated differently in production; data drift, where the population has changed; missing or delayed data at prediction time; and a feedback loop where the model’s own decisions change the data it sees.
Investigate by comparing feature distributions and predictions between training data and live data, and by recomputing production features with the training code for a sample. Fix the root cause, add monitoring for it, and update your validation so the problem would be caught next time.
45. How would you build a model when you have very few labelled examples?
Several options, often combined. Start simple: rules or heuristics from domain experts can be a strong baseline. Use pre-trained models and transfer learning, which need far fewer labels, or ask a large language model to label or classify examples, then check a sample by hand. Spend labelling effort wisely with active learning, where the model picks the most uncertain examples for people to label. Semi-supervised learning can also use unlabelled data.
Keep a small, carefully labelled test set that is never used for training, so you can measure honestly whether each approach works. Sometimes the best answer is to invest in collecting labels as part of the product.
46. How do you prioritise data science projects?
Estimate each project’s value (how much a better decision is worth), its likelihood of success (is the data available and good enough?), the effort, and how quickly results can be used. Favour projects tied to a clear decision with an owner who will act. Quick wins, such as a simple analysis that changes a pricing decision, often beat ambitious models.
Share the ranking and reasoning with stakeholders and your manager, revisit it as you learn more, and stop projects early when the data shows they will not pay off.
47. Tell me about a data science project you are proud of.
Use a clear structure:
- The business problem and why it mattered.
- The data and its challenges.
- Your approach, including the baseline and why you chose the final method.
- How you validated it, with numbers.
- How it was used, and the measured impact, such as revenue gained or hours saved. Mention one thing that went wrong and how you handled it.
Interviewers look for judgement (choosing simple methods when they work), rigour (honest evaluation), and impact (what changed). Prepare a two-minute version and be ready to go deeper into any step.
48. How do you communicate uncertainty in your results?
Give ranges, not just single numbers: “we expect 300 more sign-ups, likely between 200 and 400”. Explain what drives the uncertainty, such as limited data or assumptions about seasonality, and what would reduce it. Use visuals such as error bars or forecast bands.
Translate statistics into decision terms: “even in the pessimistic case, the change pays for itself” is more useful than a p-value. Be honest when the data cannot answer the question, and suggest how to get better evidence, such as running an experiment.
49. How do you handle a stakeholder who wants a specific answer from the data?
Stay neutral and transparent. Agree the question, the metric and the method before looking at the results, so the analysis is not shaped to fit a conclusion. Share the data, assumptions and limitations openly, and invite the stakeholder to challenge the method rather than the result.
If the result disagrees with their view, present it calmly with evidence, acknowledge what the data cannot rule out, and suggest a test that could settle it. Your credibility depends on being trusted to report what the data shows, even when it is unwelcome.
50. How are LLMs changing data science work?
Large language models help with coding (writing and explaining pandas or SQL), exploring data faster, summarising text data such as reviews and tickets, and turning unstructured text into structured features. Many text classification tasks that once needed custom models can now start with a prompted model.
The fundamentals still matter: defining the problem, getting good data, honest evaluation and communicating results. Data scientists who can evaluate LLM outputs rigorously and combine them with classic methods are in demand. See the LLM interview questions for the basics interviewers now ask.
How to prepare
- Statistics first: be able to explain p-values, confidence intervals, power, Type I and II errors and the central limit theorem in plain words.
- Practise two experiment questions: designing an A/B test, and analysing one that looks suspicious.
- Practise case questions out loud: a metric that dropped, a churn model, a price change.
- Be fluent in pandas and SQL: joins, group-bys and window functions come up in coding rounds.
- Follow the data science roadmap if you are still building the foundations.