Data scientist interview questions in English

Data scientist interviews mix statistics, machine learning, coding and product sense, and in English the hard part is explaining a model or a test result simply. Prepare plain-English definitions of p-values, overfitting and bias-variance, practise designing an A/B test out loud, and have one project story where you can say the business impact in a single sentence.

What a data scientist interview looks like

1. Recruiter screen
Background, the kind of data science you do (product analytics, ML, research) and logistics.
English focus: Say which type of data scientist you are in the first sentence. The title means different things at different companies.
2. Statistics and experimentation
Probability, hypothesis tests, A/B test design and interpretation.
English focus: Define terms in plain English before you use the formula. Interviewers test understanding, not memory.
3. Machine learning
Model choice, features, evaluation metrics, overfitting, sometimes ML system design.
English focus: Explain trade-offs: 'I'd start with a simpler model because…'.
4. Coding (Python or SQL)
Data manipulation in pandas or SQL, sometimes a light algorithm question.
English focus: Think aloud. Say what each step does to the shape of the data.
5. Product sense and behavioural
Metrics for a product, past projects, working with engineers and PMs.
English focus: Link every model to a decision someone made because of it.

10 data scientist interview questions, with spoken answers

Each answer is written to be said out loud: short sentences, the main point first, and one concrete detail. Change the details to your own experience before you practise.

1. Explain a p-value to someone without a statistics background.

Statistics and experimentation

What they listen for: A correct, simple definition, and no claim that it is the probability the hypothesis is true.

“Imagine the change we made had no effect at all. The p-value is how likely we'd be to see a difference this big, or bigger, just by chance in that world. A small p-value means 'this would be surprising if nothing had changed.' It doesn't tell us how big or how important the effect is.”

2. How would you design an A/B test for a new checkout button?

Statistics and experimentation

What they listen for: Unit of randomisation, primary metric, sample size, duration and what could go wrong.

“I'd randomise at the user level, so one person always sees the same version. My primary metric would be checkout conversion. Before starting, I'd estimate the sample size from the current conversion rate and the smallest lift we care about, and run for full weeks to cover weekday effects. I'd also check the split is really 50/50 before reading the results.”

3. What is overfitting and how do you prevent it?

Machine learning

What they listen for: A clear definition and more than one practical fix.

“Overfitting is when the model learns the noise in the training data, so it looks great in training and worse on new data. I'd catch it by comparing training and validation scores. To reduce it: more data, a simpler model, regularisation, or fewer features. With trees, limiting depth helps a lot.”

4. Your model has 99% accuracy on a fraud dataset. Are you happy?

Machine learning

What they listen for: That you notice class imbalance immediately.

“Not yet. If only 1% of transactions are fraud, a model that says 'not fraud' every time also gets 99%. I'd look at precision and recall on the fraud class, and ask the business which mistake costs more: blocking a real customer, or missing a fraud.”

5. Explain the bias-variance trade-off.

Machine learning

What they listen for: That you connect it to model complexity and to a real choice you made.

“Bias is error from a model that's too simple to capture the pattern. Variance is error from a model that's too sensitive to the specific training data. As you add complexity, bias usually goes down and variance goes up, so you're looking for the point where total error on new data is lowest.”

6. Tell me about a model you built that was used in production.

Product sense and behavioural

What they listen for: The decision it supported, how you measured it, and what you would do differently.

“I built a model that ranked support tickets by how likely they were to escalate. The decision it supported was which tickets a senior agent picks up first. We tested it against the old first-in-first-out order, and escalations fell by about a fifth. If I did it again, I'd involve the support team earlier, because they knew signals I'd missed.”

7. How would you measure the success of a recommendation feature?

Product sense and behavioural

What they listen for: Short-term and long-term metrics, and awareness that clicks are not the same as value.

“Short term, I'd look at click-through and add-to-cart from recommendations. But clicks can be gamed by clickbait, so I'd also track a longer-term metric like 30-day retention or repeat purchase, and a guardrail like how varied the recommendations are.”

8. The test result is significant but the effect is tiny. What do you recommend?

Statistics and experimentation

What they listen for: That you separate statistical significance from business significance.

“Significant just means it's probably real, not that it's worth it. I'd compare the size of the lift with the cost of maintaining the change. If it's a one-line change with no downside, ship it. If it adds complexity for a 0.1% lift, I'd probably recommend not shipping it.”

9. Write a function that returns the top three categories by revenue.

Coding (Python or SQL)

What they listen for: Clean code, and that you ask about ties and empty input.

“I'll group by category, sum the revenue, sort descending and take the first three. Two quick questions: if two categories tie for third place, do you want both? And should I return an empty list if there's no data, or raise an error?”

10. Tell me about a time your analysis was wrong.

Product sense and behavioural

What they listen for: Ownership, how you found the error, and what you changed afterwards.

“I once reported a 12% lift from a pricing test, and two days later I found a bug: some users had been counted in both groups. I told my manager the same day, re-ran the analysis, and the real lift was about 4%. Since then I always run a sample ratio check before I share any test result.”

Phrases to have ready

Defining a term simply: “In plain terms, that means…”
Choosing a simpler model: “I'd start with a baseline, like logistic regression, so we know what 'good' looks like.”
Explaining a trade-off: “We'd gain recall, but we'd pay for it with more false positives.”
Talking about confidence: “The effect is real, but the confidence interval is wide, so I wouldn't promise a specific number.”
Asking about the business cost: “Which mistake is more expensive here: a false positive or a false negative?”
Admitting a limit: “That's outside what this data can tell us. We'd need an experiment to answer it.”

Data Scientist vocabulary in plain English

Baseline: The simplest reasonable model or rule, used to judge if a better one is worth it.
Precision: Of everything the model flagged, how much was really positive.
Recall: Of everything that was really positive, how much the model found.
Confidence interval: The range the true value probably sits in.
Data leakage: The model accidentally sees information during training that it won't have in real life.
Sample ratio mismatch: The test groups are not the sizes you planned, which usually means a bug.

English mistakes that cost points

Avoid: Starting with the formula when asked to explain a concept.
Instead: Give a one-sentence everyday explanation first. Add the formula only if they want it.
Avoid: Saying 'the model is good' without a metric.
Instead: Name the metric and the comparison: 'AUC went from 0.71 to 0.78 against the old rules.'
Avoid: Using 'significant' to mean 'big'.
Instead: Say 'statistically significant' for a real effect and 'large' or 'meaningful' for size. Interviewers notice the difference.
Avoid: Describing what 'we' did for the whole project.
Instead: Use 'I' for your work and 'the team' for the rest. 'I built the features; the platform team deployed it.'

FAQ

What questions are asked in a data scientist interview?

Most loops cover statistics and A/B testing, machine learning concepts and evaluation, Python or SQL coding, product metrics and behavioural questions. In English interviews you will often be asked to explain a concept to a non-technical person, so practise simple definitions.

How do I explain a machine learning project in English?

Use four sentences: the business problem, what you built, how you measured it, and the result. Then stop and let them ask. Most candidates lose the interviewer by starting with the algorithm.

Is the data scientist interview different from a data analyst interview?

Usually yes. Data scientist loops go deeper on statistics and machine learning, and may include model design. Data analyst loops focus more on SQL, dashboards and stakeholder work. Both test how clearly you explain results in English.

Ready to practise?

Turn interview English into a repeatable skill

Work through the Interview English Course, then book live coaching with engineers who give feedback on both your technical answers and how you deliver them in English.