Capital One Data Scientist Interview Guide: OA, Data Challenge, Mini Case and Power Day
Prepare for Capital One data scientist interviews with reported assessment patterns, interview questions, and preparation advice.
On this page
Last reviewed November 16, 2026.
For data scientists.
Built by the Codinzhub team from Capital One interview experiences that candidates shared. Each experience is one person's interview, and Capital One changes its process by team and level, so use this as a map, not a promise.
The 30-second version. Capital One data roles follow one of two paths. The analyst-style path (used by many DS teams): a 70-minute OA of 14 or 15 Excel and SQL questions, a take-home data challenge (usually the airline dataset), then a four-round Power Day of two business cases, behavioral and a presentation of your data challenge. The modelling path: a CodeSignal OA of four Python and ML tasks, a mini case, then a Power Day of a technical code round, a role play where you explain a model to a non-technical person, culture fit and a case. Business cases appear in every path: they are maths plus business judgment, and they decide most results. Offers often come within hours.
Related roles have their own guides: Senior data analyst · Data science intern · Quantitative analyst · Applied researcher
Round-by-round breakdown
Stage 1: Application, where candidates applied
How it works. Many candidates had a referral; others applied online or were contacted on LinkedIn by a recruiter, who then asked them to apply on the careers site. One referred candidate waited about a month before the OA arrived. One role was cancelled after the HR call because the headcount was removed; the same candidate got a new process weeks later through another recruiter.
Stage 2: Shortlisting, how candidates moved forward
How it works. A 15 to 20 minute recruiter call: salary, location, visa status and tools you know (one was asked about Tableau). No technical questions. Sponsorship candidates got an immigration questionnaire. Some teams added a hiring manager call that was mostly behavioral: working across teams, evaluating models, your end-to-end projects. Capital One also runs a case workshop before the mini case for some candidates.
Stage 3: Online assessment
There are two different OAs. Ask your recruiter which one you will get.
The Excel and SQL OA (analyst-style roles)
At a glance
- Length: 70 minutes, screen shared and recorded, camera and microphone on
- Questions per round: 14 or 15: about 14 multiple-choice data questions and 1 SQL question
- What was asked: filtering and summarising CSV files, merging tables, click-through rates, a SQL query with joins and grouping, t-tests
- Most common: an ad platform dataset with A/B test and watch-time data
How it works. Questions 1 to 7 use one CSV file. Question 8 is a SQL query with three requirements, and you only score if the whole query is right. Questions 9 to 15 use four more CSV files that you must merge yourself. Candidates solved the data questions in Excel with filters, pivot tables, SUMIFS, COUNTIF and VLOOKUP or INDEX and MATCH. The last two questions were t-tests, so know one-tailed versus two-tailed. Watch for dates: in one dataset only the first row had a full date. Time is very tight; one candidate finished only about half and still passed. Guess rather than leave blanks. Scores took two to three days, which suggested someone reviewed the recording.
Questions candidates reported:
| Question (our summary) | What it tests |
|---|---|
| Which week had the longest watch time while a given ad was running? | SUMIFS, date cleanup |
| Which ads were played most, and which had the highest click-through rate? | Merging four tables |
| A SQL query over exam scores: group those above a threshold, flag those below | JOIN, GROUP BY, WHERE, IS NULL |
| Is the difference between two groups significant? | One-tailed versus two-tailed t-tests |
The CodeSignal Python and ML OA (modelling roles)
At a glance
- Length: 70 to 90 minutes
- Questions per round: 4
- What was asked: merging CSVs in pandas, SQL-style logic in Python, feature engineering, training a model
- Most common: a rides dataset or a house prices dataset
How it works. Question 1 merges several CSV files and computes metrics; convert results to numbers or half the tests fail. Question 2 is a long SQL-style transformation in Python with many output columns and a fixed "today" date. Question 3 is feature engineering: fill missing values and ordinal-encode, with fit_transform on training data and transform on test data, writing two output CSVs. Question 4 trains a model on imbalanced data and checks precision and recall; a random forest was enough. One candidate advised finishing question 1 in ten minutes.
More questions: Browse Capital One questions.
Stage 4: Data challenge (take-home)
At a glance
- Length: officially 8 to 10 hours; candidates spent much longer, over about a week
- Questions per round: one dataset with several questions
- What was asked: clean the data, build metrics, answer business questions, recommend
- Most common: the airline dataset (routes, delays, profitability)
How it works. Candidates submitted a Jupyter notebook, sometimes an Excel file describing new fields, and a slide deck they later reused for the Power Day presentation. One candidate described it as complex at first; another said there is no single right answer as long as the reasoning holds. Treat it like a real business project: clean carefully, state assumptions, and tie the answers into one story. Results came back in two days to two weeks.
Stage 5: Mini case
At a glance
- Length: officially an hour; some finished in 15 to 20 minutes
- Questions per round: 1 case with 4 to 6 questions
- What was asked: a business with numbers: profit, break-even, expected value, a recommendation
- Most common: the energy company case; the TV show renewal case; the ride-share case
How it works. The interviewer reads a background, asks for factors to consider, then gives numbers to calculate, then asks for a recommendation and sometimes a 30-second summary. You may ask for missing data, and you are expected to. A candidate who failed said they treated it as a maths problem: open with a framework (financial, customer, strategic, risk), then calculate. Several heard back the same day; one got a Power Day invite 90 minutes later.
Questions candidates reported:
| Case (our summary) | What was calculated | What the interviewer pushed on |
|---|---|---|
| An energy company moving to renewables | Production needed for a 10% return; payback years for a solar and an ethanol test plant | Is the number reasonable against capacity? Solar has no variable cost |
| A streaming company deciding whether to renew a show | Two-year profit of two shows; expected value with a success probability | Ask for the probability; do not assume 50%. Cost structure; whether to sell the show |
| A ride-share service | Daily profit from rides, driver pay and fixed cost; break-even surge price | Which data you still need |
| A restaurant offered a coupon site deal | Daily profit before and after a 40% commission | Read the change in tables and coupons, then decide |
| A grocery loyalty card | Profit, break-even for a premium tier, losses from a three-month database outage | How to stop the outage happening again: metrics, dashboards, alerts |
| A new streaming service | Break-even subscriber count | Is it realistic, and how to reach it? |
| A cloud storage startup | Monthly profit; break-even if users double; the paid-user share needed to break even | Which pricing is better and the risks |
Stage 6: The Power Day
How it works. Four rounds of 45 to 60 minutes, usually one morning. The order varies, for example case, behavioral, case, data challenge. HR sends prep material, and calling HR to ask what each round covers helped one candidate a lot.
Business case rounds
At a glance
- Length: about 45 minutes each; usually two of them
- Questions per round: 1 case with 5 to 8 questions
- What was asked: profit, expected value, break-even, strategy, risks
- Most common: the amusement park; the "coder" overtime case; the shared workplace; weather insurance
Questions candidates reported:
| Case (our summary) | What was calculated | What the interviewer pushed on |
|---|---|---|
| An amusement park with day, five-day and annual passes | Revenue, average visits per annual pass holder, profit; profit with extra land and a higher fee; expected profit of a bid that wins 80% of the time | Whether to bid, and the risks in your numbers |
| A software firm deciding whether to take a client ("coder" case) | Cost of overtime versus hiring contractors; 1,000 versus 2,000 lines of code | Whether to take the client |
| A home internet provider | Net value per new customer over a 15-month contract | Does an 18-month contract change it? |
| Weather insurance | Profit per policy with servicing, regulatory and payout costs; which risk groups to insure | How to sell to the riskier groups |
| A co-branded credit card | Profit from three customer groups; new users to break even; two marketing options | Red flags in the assumptions |
| A credit product strategy, such as limit increases for a segment | Profit equals revenue minus losses minus operating cost | Define the goal before any model |
| 500,000 users in three segments | Simple totals, then rough graphs | Sketch the shape from business sense; do not compute it exactly |
Product case round
At a glance
- Length: about 45 minutes
- Questions per round: 1 open product discussion
- What was asked: your favourite app
- Most common: explain an app's business model to someone who knows nothing about it
How it works. Pick an app, explain it from zero, name its audience, competitors and revenue streams, propose a new revenue stream with three metrics, then list six improvements, pick one, say how you would measure it, and defend its cost.
Data challenge presentation
At a glance
- Length: about 45 minutes
- Questions per round: your slides, then questions on your code
- What was asked: your methods, your metrics, your code
- Most common: "why Python?", "how did you use delay time?", "which part did you like most?"
How it works. Present your slides; interviewers may stop you to go deeper on a metric you defined. Then they open your code and ask why you did each step.
Technical and role play rounds (modelling path)
At a glance
- Length: about an hour each
- Questions per round: a code task or a simulated conversation
- What was asked: reading and testing code; explaining a model to a non-technical partner
- Most common: pytest tasks; the flight delay model role play
How it works. The technical round had four steps: set up an environment, explain what given code does and why it is designed that way, point out readability problems in other code, then run pytest, explain the failure and write a test. Some got SQL with window functions like RANK() and data wrangling in Python, plus model questions such as handling missing values and overfitting in XGBoost. The role play asks you to walk a product manager through a model, such as the flight delay case, from EDA to model explanation, or why a model did well in validation but worse after launch.
Behavioral and culture fit round
At a glance
- Length: 30 to 45 minutes
- Questions per round: 3, each with one follow-up
- What was asked: an accomplishment, a mistake, helping others, ethics
- Most common: the "classic three": accomplishment, failure, helping someone
Questions candidates reported: an accomplishment; a failure or mistake; a time you helped someone; using an advanced technique to solve a problem; an ethics question; handling ambiguous requirements. Interviewers asked follow-ups like what your manager said and how long it took.
More questions: Browse Capital One questions.
Stage 7: Decision and offer
How it works. Very fast: one offer came under 50 minutes after the Power Day, others the same evening or within three days. Then comes team matching, where your result is shared with hiring managers; on a Friday this can take longer.
Topics to prioritise
Across these experiences, the most common topics are profit and break-even maths, expected value, structured business frameworks, Excel pivots and lookups, SQL joins and window functions, pandas merging, feature engineering and explaining models simply.
The Codinzhub Next-Round Playbook
What Capital One data candidates who moved on did. Put together by the Codinzhub team from the Capital One experiences in this guide.
To get past the OA
- Practise the Excel questions with pivot tables and lookups under a 70-minute timer.
- Do the Excel questions first and the SQL question after, to bank points.
- For the Python OA, practise
fit_transformversustransformand type conversions until they are automatic.
To get past the data challenge
- Treat it like a real project: careful cleaning, stated assumptions, one storyline across the questions.
- Know why you chose every method; you will be asked about your code.
To get past the cases
- Work through every commonly shared case, with the numbers, before the interview. One candidate failed the only one they skipped.
- Open with a framework, then calculate, then recommend, then name the risks.
- Ask for missing numbers, such as contract length or success probability.
- Check that each answer makes sense, for example against maximum capacity.
To get past behavioral
- Prepare true stories for accomplishment, failure and helping others, with details for follow-ups.
FAQ
Do I need coding-practice problems for data science at Capital One?
Mostly not. One candidate who received an offer said real project experience and good coding habits mattered more.
Is there one right answer to the cases?
No. Candidates who made different calls passed, as long as their reasoning held together.
About this guide
This guide is researched and written by the Codinzhub team. We read every Capital One interview experience we could find, kept only what candidates actually described, and rewrote it in our own words. Questions are our own short summaries; the practice links point to similar problems, not the exact interview questions. Found something out of date? Share your own experience on Codinzhub and we will update the guide.
Keep track of what you’ve learned.