Data analyst interview questions, with worked answers

How to use these questions

Read the short answer and say it out loud. That is your first thirty seconds. Then read the full answer and pick two or three points to add from your own work. Interviewers want to hear how you think and how you check, not a list of tools.

Things marked depends on the provider or depends on the release depend on a provider or a release. Check the current documentation before you quote a detail.

What a data analyst is expected to know

• SQLAsked in 57% of data analyst ads (as of 5 Nov 2026).The core tool: joins, grouping, window functions and checking your own results, on real company data.: joins, grouping, window functions, and checking your own results.
• ExcelAsked in 16% of data analyst ads (as of 5 Nov 2026).Pivot tables, lookups, clean formulas and charts; still where many business users read and check numbers.: pivot tables, lookups and clear formulas others can follow.
• A dashboard tool: Power BIAsked in 36% of data analyst ads (as of 5 Nov 2026).Build reports and dashboards: data models, measures and refresh schedules; common where firms run Microsoft. depends on the provider or TableauAsked in 11% of data analyst ads (as of 5 Nov 2026).Build interactive dashboards people can filter themselves, from a clean data source. depends on the provider, from a clean data model to a page people use every week.
• PythonAsked in 40% of data analyst ads (as of 5 Nov 2026).Python for analysis: pandas for tables, charts, and scripts that automate a monthly report. or RR for statistics and reports; common in research, health and public-sector teams.: for cleaning, automation and analysis that is too big or too repetitive for a spreadsheet, often with pandasLoad, clean, join and reshape tables in Python. (a library for tables in Python).
• Data analysisAsked in 33% of data analyst ads (as of 5 Nov 2026).Turn a vague question into a clear one, answer it with data, and say what to do next. and data visualisationAsked in 10% of data analyst ads (as of 5 Nov 2026).Choose the chart that answers the question; label it so it cannot be misread.: turning a vague question into a clear one, and a chart that answers it at a glance.
• StatisticsAsked in 12% of data analyst ads (as of 5 Nov 2026).Averages and spread, sampling, significance and confidence intervals; enough to say how sure a finding is.: averages and spread, samples, significance, and reading an A/B test.
• Metrics: KPIs (key performance indicators, the few numbers a team is judged on), funnels and cohorts.
• The modern warehouse: a cloud warehouse such as SnowflakeAsked in 13% of data analyst ads (as of 5 Nov 2026).A cloud data warehouse: large tables, fast SQL, and costs that depend on how you query. depends on the provider or BigQuery, and dbtAsked in 9% of data analyst ads (as of 5 Nov 2026).Build tested, documented data models in SQL inside the warehouse; common in modern data teams. depends on the release for tested, documented data models in SQL.
• Communication: turning a finding into a recommendation a stakeholder (the person who asked, or who acts on it) understands.

SQL and data

What is the difference between an inner join and a left join, and how can a join go wrong?

What they check: that you know joins well and check for the classic mistakes.

Short answer: "An inner join keeps only rows that match in both tables; a left join keeps every row of the first table and fills in empty values where there is no match. Joins go wrong when the key is not unique, so rows multiply and totals are too high, so I always compare row counts before and after."

The full answer: Say I join customers to orders. An inner join drops customers with no orders; a left join keeps them, which matters when I want to count customers who never bought. The most common error is joining on a key that appears more than once on both sides, for example a customer with two addresses: every order now appears twice and revenue doubles. My checks: the number of rows before and after, whether the key is really unique, and a total I can compare with a trusted report. A filter on the second table placed in the wrong part of a left join quietly turns it into an inner join, so I check that too.

A weak answer: "I always use an inner join." That silently drops rows you may need.

Follow-up: "Your revenue is higher than finance's figure." Look for duplicated rows from a join first, then for refunds, currencies and dates that are counted differently.

How would you show each month's sales next to a running total for the year?

What they check: that you know window functions, which come up in almost every SQL round.

Short answer: "I would first total sales by month, then add a window function that sums the months from the start of the year up to each month. A window function calculates across related rows without collapsing them, so each month keeps its own row."

The full answer: Grouping gives one row per month. A window function then looks at a "window" of rows around each one, here all earlier months in the same year, ordered by month, and adds them up. The same tool ranks rows (the top three products per region), compares with the previous row (this month against last month), or numbers rows (each customer's first purchase). I would check the last month's running total against the year's total. Months with no sales need a row too, or the chart will skip them.

A weak answer: "I would do it in Excel by hand each month."

Follow-up: "Now show the change against the same month last year." Use the window function that reads an earlier row, twelve rows back, or join the table to itself by month.

How do you check data quality before you trust a data set?

What they check: that you test the data before you analyse it.

Short answer: "I check that it is complete, unique, valid and fresh: the row counts I expect, no duplicate keys, sensible values and dates, and data up to yesterday. And I compare one or two totals with a source people already trust."

The full answer: Data quality is whether data is fit for the question. My first look: how many rows, over which dates, how many empty values per column, how many distinct keys. Then sense checks: no negative quantities, no orders before the customer existed, prices in the expected range. Then a reconciliation: does total revenue match the finance report for one month? Where the same checks are needed every day, I automate them, for example as tests in dbtAsked in 9% of data analyst ads (as of 5 Nov 2026).Build tested, documented data models in SQL inside the warehouse; common in modern data teams. depends on the release, so a broken load is caught before anyone reads the dashboard. Problems go back to the owner of the source, not only into my workaround.

A weak answer: "The data comes from the system, so it is correct."

Follow-up: "A tenth of the orders have no customer id." Find out why, from the people who run the system; meanwhile, report them as a separate group so totals stay honest.

Two reports show different revenue for the same month. How do you find out why?

What they check: a calm, step-by-step reconciliation, and knowing why definitions differ.

Short answer: "I compare their definitions first: which date, which currency, whether refunds, tax and cancelled orders are in. Then I narrow it down by day, product and region until I find the rows that differ. Usually it is a definition, not a bug."

The full answer: Revenue can be counted when the order is placed, shipped or paid; with or without tax; before or after refunds; in the local currency or converted at different rates. I write both definitions side by side. Then I split the difference: by day to see if it is one date, by source system, by region. When I find the rows, I explain the cause and suggest one agreed definition, written down where everyone can find it, so the question does not come back next month.

A weak answer: "I would tell them to use my report."

Follow-up: "Both definitions are reasonable." Keep both, name them clearly ("booked revenue", "paid revenue"), and show which one each report uses.

Metrics, dashboards and analysis

Weekly active users fell sharply last week. How do you investigate?

What they check: a structured approach that rules out data problems before business ones.

Short answer: "First I check the data is right: did tracking or a data load break? Then I check the calendar, such as a holiday. Then I split the drop by platform, country, new or returning users and product area, to find where it is concentrated, and look at what changed there."

The full answer: Step one, is it real? A broken tracking tag, a late data load or a changed definition causes many "drops". Step two, is it expected? Season, holidays, a marketing campaign that ended. Step three, where is it? Split by the main dimensions; a drop in one app version or one country points straight at a cause. Step four, why? Match the timing with releases, outages, price changes or a competitor's move. Then report what I know, what I do not, and what I recommend, early, rather than waiting for a perfect answer.

A weak answer: starting with a long list of theories without checking the data first.

Follow-up: "The drop is only on Android, starting the day of a release." Tell the app team at once with the evidence, and check whether the app or only its tracking broke.

What makes a good KPI, and how would you choose one for a customer support team?

What they check: that you link a metric to a goal and know how metrics can be gamed.

Short answer: "A good KPI is tied to a goal, clearly defined, measured the same way every time, and something the team can influence. For support, I would pair speed with quality, such as time to resolve a case together with customer satisfaction and the share of cases reopened, so one cannot be improved by hurting the other."

The full answer: A KPI, a key performance indicator, is one of the few numbers a team is judged on, so it changes behaviour. If support is measured only on speed, people close cases too early. That is why I pair metrics: one for speed, one for quality, and a guardrail that must not get worse. Each needs a written definition: when does the clock start, what counts as resolved, which cases are excluded. I would agree them with the team, show the trend over time rather than one week, and review them after a quarter.

A weak answer: "Number of tickets closed per day." Easy to game, and it says nothing about customers.

Follow-up: "Satisfaction went up but costs rose a lot." Add cost per case to the view and discuss the trade-off openly with the manager.

Design a dashboard for a sales manager. What goes on it?

What they check: that you design for the person and their decisions, not for every chart you can make.

Short answer: "I would ask what decisions they make each week and build the page around them: a few headline numbers against target at the top, the trend below, then a breakdown by region, product and salesperson that they can filter. Everything else stays off the page."

The full answer: A dashboard is a page of charts and numbers that people check regularly, built in a tool such as Power BIAsked in 36% of data analyst ads (as of 5 Nov 2026).Build reports and dashboards: data models, measures and refresh schedules; common where firms run Microsoft. depends on the provider or TableauAsked in 11% of data analyst ads (as of 5 Nov 2026).Build interactive dashboards people can filter themselves, from a clean data source. depends on the provider. Before building, I interview the manager: which questions, how often, what do they do next? At the top: three to five KPIs with target and last period. Below: the trend, then the breakdowns. Every chart answers one question and has a clear title. I keep the data model clean, so numbers match other reports, show when the data was last refreshed, and check after a month which parts are used and remove the rest.

A weak answer: twenty charts on one page, in every colour, because the data is there.

Follow-up: "The dashboard is slow." Pre-aggregate the data in the warehouse, reduce the visuals per page, and limit the default date range.

How would you analyse where users drop off in a sign-up funnel?

What they check: that you can build a funnel correctly and turn it into an action.

Short answer: "I define each step and count how many users reach each one in order within a time window, then find the step with the biggest drop. I split that step by device, channel and user type to see who drops, and look at what happens on that screen."

The full answer: A funnel is the series of steps users take towards a goal, such as landing page, form, email confirmation, first use. I count users, not page views, and only those who passed the earlier steps, within a sensible time, for example seven days. The biggest drop is usually where to start. Splits show whether it is everyone or one group, for example mobile users on one browser. Then I look for a cause: a slow page, a confusing field, an email that lands in spam. I suggest a fix and, where possible, an A/B test to prove it helps.

A weak answer: comparing page views of each step, which double-counts people who reload.

Follow-up: "The email confirmation step loses many users." Check delivery and spam rates, the time until the email arrives, and test letting people start before they confirm.

What is a cohort analysis, and when would you use one?

What they check: that you can separate growth from retention.

Short answer: "A cohort is a group of users who started in the same period, such as everyone who signed up in March. A cohort analysis follows each group over time, so I can see whether newer customers stay longer or shorter than older ones, which an overall total hides."

The full answer: Total active users can rise while retention gets worse, simply because many new users arrive. A cohort table puts the start month in rows and the months since start in columns, with the share of each group still active. Reading down a column compares cohorts at the same age; reading across a row shows how one group fades. It is the standard way to check whether a product change or a new marketing channel brings customers who stay.

A weak answer: "I look at the total active users each month."

Follow-up: "The newest cohorts retain worse." Split them by channel; often one new source of sign-ups brings people who never meant to stay.

Statistics and reasoning

Marketing shows you an A/B test where the new email won. What do you check before you agree?

What they check: that you read someone else's test critically and fairly.

Short answer: "I check that the groups were random and the right size, that the main metric was chosen in advance, that the test ran long enough and was not stopped as soon as it looked good, and that the difference is bigger than chance, with its range. Then whether the effect is big enough to matter."

The full answer: In an A/B test, two random groups see two versions and we compare a metric. The traps: groups that are not really random or not the planned size; picking the metric that happened to win among many; stopping early when it looks good; too few people to tell a real difference from noise. I would ask for the plan, the group sizes and the result with a confidence interval, the range the true effect probably lies in. If everything holds, the remaining question is business value: a real but tiny lift may not be worth changing anything.

A weak answer: "The new one had more clicks, so it won."

Follow-up: "It won on opens but lost on purchases." Purchases matter more; opens are only a step towards them.

Customers who use our mobile app spend more. Should we push everyone to the app?

What they check: that you do not confuse correlation with cause.

Short answer: "Not yet. The correlation may run the other way: loyal, high-spending customers are the ones who bother to install the app. To know whether the app itself raises spending, we need a test, or at least a comparison of the same customers before and after they started using it."

The full answer: Correlation means two things go together; it does not show which causes which, or whether a third thing causes both. Here, customer loyalty probably drives both app use and spending. Better evidence: invite a random group to the app and compare their spending with a group not invited. Without a test, compare customers' own spending before and after they installed it, against similar customers who did not. Then present the finding with its uncertainty.

A weak answer: "Yes, the data clearly shows the app makes people spend more."

Follow-up: "We cannot run a test." Use the before-and-after comparison with a matched group, and say clearly how sure you are.

Conversion went up in every region, but down overall. How is that possible?

What they check: that you understand how mixes change averages, a classic trap known as Simpson's paradox.

Short answer: "The mix changed: more of our traffic now comes from a region that converts less, so the overall rate falls even though each region improved. I would show the rates by region and the change in the mix side by side."

The full answer: An overall rate is a weighted average of the groups, weighted by their size. If a low-converting group grows, say a new market or a cheap traffic source, the total can fall while every group rises. So I report both: each group's rate and how the mix moved. Which one matters depends on the question. For "is the product getting better?", the group rates; for "will we hit the sales target?", the overall one.

A weak answer: "The data must be wrong."

Follow-up: "The executive only wants one number." Give the overall number, with one sentence on the mix change, so nobody draws the wrong conclusion.

Tools and ways of working

When do you use Excel, a BI tool or Python?

What they check: that you pick the tool for the job and the audience.

Short answer: "Excel for small, one-off analyses and when the business user wants to work with the numbers themselves. A BI tool for anything people check regularly. Python for large data, cleaning, repeated reports and statistics a spreadsheet does poorly."

The full answer: ExcelAsked in 16% of data analyst ads (as of 5 Nov 2026).Pivot tables, lookups, clean formulas and charts; still where many business users read and check numbers. is quick and everyone can open it, but it is easy to break and hard to check. A BI tool such as Power BIAsked in 36% of data analyst ads (as of 5 Nov 2026).Build reports and dashboards: data models, measures and refresh schedules; common where firms run Microsoft. depends on the provider gives one trusted version that refreshes on its own. PythonAsked in 40% of data analyst ads (as of 5 Nov 2026).Python for analysis: pandas for tables, charts, and scripts that automate a monthly report. with pandas, a table library, handles millions of rows, automates the boring parts and keeps every step written down so it can be rerun. Most real work mixes them: SQL to get the data, Python or the BI tool to shape it, Excel for the people who want to explore.

A weak answer: "Python for everything; Excel is outdated."

Follow-up: "Finance keeps a critical forecast in a huge spreadsheet." Help move the data part into the warehouse and keep Excel as the view they like.

What is a data model for analytics, and why do teams use a tool such as dbt?

What they check: that you know how modern data teams keep numbers consistent.

Short answer: "A data model turns raw tables into clean, well-named tables that answer business questions, such as one table of orders and one of customers with agreed definitions. dbt lets you write those transformations in SQL, test them, document them and run them in order, so every report uses the same logic."

The full answer: Raw data from source systems is messy and spread out. Analysts used to clean it again in every report, so numbers differed. A shared model fixes the definitions once. A common shape is a "star": fact tables of events, such as orders, surrounded by dimension tables that describe them, such as customers, products and dates. dbtAsked in 9% of data analyst ads (as of 5 Nov 2026).Build tested, documented data models in SQL inside the warehouse; common in modern data teams. depends on the release builds these models in the warehouse, such as SnowflakeAsked in 13% of data analyst ads (as of 5 Nov 2026).A cloud data warehouse: large tables, fast SQL, and costs that depend on how you query. depends on the provider, with tests (no duplicate order ids, no empty customer ids) and documentation, and keeps the code in GitAsked in 10% of data analyst ads (as of 5 Nov 2026).Version control for queries and models, so changes are reviewed and can be undone. so changes are reviewed. depends on the release

A weak answer: "Each analyst cleans the data in their own report."

Follow-up: "A test fails one morning." Stop the bad data reaching the dashboards, tell the users, fix the source or the model, then rerun.

A manager asks you for "all the data on our customers". What do you do?

What they check: that you turn a vague request into a clear question, and handle personal data carefully.

Short answer: "I ask what decision it is for, what they will do with it and by when. Usually the real question is narrower, such as which customers to call this month, and the answer is a short list or one chart, not a huge export. I also check who may see personal data."

The full answer: Big exports go stale, get emailed around and break data protection rules, and rarely answer the question. So I ask: what problem, what decision, who acts, how often? Then I propose the smallest output that answers it, agree it in writing, and deliver a first version quickly to check I understood. Personal data is shared only with people who need it, with the fewest fields necessary; in Europe that is a GDPRAsked in 7% of data analyst ads (as of 5 Nov 2026).Personal data in reports and exports: show only what is needed, and anonymise where you can. requirement.

A weak answer: sending a full export the same afternoon.

Follow-up: "They insist on the export." Explain the risk, remove fields they do not need, and agree how long they keep it.

Your analysis shows that a project the stakeholder loves is not working. How do you present it?

What they check: honesty, tact, and making the finding useful.

Short answer: "I tell them early and privately, before any big meeting, with the evidence and how sure I am. I show what is not working, what might, and what I recommend, so it becomes a decision about next steps rather than a verdict."

The full answer: First I double-check the analysis, ideally with a colleague, because the finding will be challenged. Then I meet the stakeholder alone and walk through it: the question, the data, the result, the limits. I separate facts from interpretation. I bring options: change the project, test a different version, or stop it and use the money elsewhere. I invite them to challenge the method; sometimes they know something the data does not.

A weak answer: presenting it for the first time in front of their manager.

Follow-up: "They ask you to leave that slide out." Offer to present it with more context, but do not hide a result that others will rely on.

Your CV for these questions

Every answer above is easier when your CV already says it, because interviewers ask about what is on it. Put one line under each job for each of these, in the words the ads use:

Then score your CV against a real job: you will see which of these the job needs and which your CV does not prove yet.