Capital One Data Science Manager Interview Experience
Capital One · Mid level · Data Scientist
You can’t just be like, ‘Oh, I know math.’ You have to translate a business problem into algebra, keep track of the numbers, and still have the business intuition to know what the answer means.
Interview process
The process started with a 90 minute CodeSignal and then a remote Power Day with four back-to-back interviews: stats roleplay, quantitative business case, hiring manager fit, and a technical code review style round. I did two Power Days because the first one closed up on me. But my second Power Day was shorter since I had passed the first one.
Interview rounds · 5
- 1
Online assessment
CodingData AnalysisMachine LearningI started with a 90 minute CodeSignal in Jupyter notebooks, and it was mostly pandas-style data cleaning, wrangling, summary stats, and one basic modeling section, with the real pressure coming from the clock.
Q1. Complete four notebook-style tasks cleaning messy data, summarizing it, and building a basic model in 90 minutes.
How they answeredI moved quickly through messy data cleaning and manipulation, basic summary statistics, and a simple model with a train test split. I did really well on it, but the main challenge was pacing because 90 minutes goes fast and you cannot get stuck on any one notebook.
- 2
Other round
Statistics & ExperimentationMachine LearningCase StudyCross-FunctionalThe stats round felt like a consultant roleplay where I had to explain and critique a flawed model to a business stakeholder who was pretending not to be technical.
Q1. Do you want to evaluate a linear regression or a classification and regression tree, and can you walk me through what is wrong with this airline delay model?
How they answeredI chose linear regression because I felt more comfortable with it. I walked through the airline delay example and pointed out that time of day should be categorical, not treated like a straight numeric variable, because 11:55 p.m. and 12:05 a.m. are close in reality but look far apart numerically. I also said if the real output is delay probability, I would switch to logistic regression. They also showed scatterplots and a correlation matrix, so I talked through multicollinearity and why I would revisit the feature setup.
Follow-up questions- Which inputs should not be treated as linear variables?
- If the business really wants the probability a flight is delayed, what model would you use instead?
- What issues do highly correlated predictors create in a regression?
- 3
Other round
Case StudyAnalyticalEstimationThe business case was very interviewer-led and much more quantitative than a typical open-ended product case, with lots of numbers to track on paper.
Q1. How does this amusement park make money?
How they answeredI started with ticket sales, then added premium rides, food and drinks, and event rentals as other revenue streams. From there I translated their assumptions about visit days, season passholders, and per-visit customers into basic algebra to estimate total revenue. After that they had me work down an income statement with sales, cost of goods sold, fixed costs, and things like depreciation to get to net income. The math itself was not hard, but you had to stay organized and keep track of a lot of numbers.
Follow-up questions- Given total visit days, a mix of season passholders and per-visit customers, what is total revenue?
- Can you walk me through the income statement down to net income?
- 4
Final / onsite round
BehavioralArtificial IntelligenceThe hiring manager round was the loosest part of the day and felt like a mix of behavioral fit and whatever that specific manager cared about.
- Q1. What does your ideal team look like, and what type of people do you like working with?
Q2. What if we used AI to profile people coming into our brick and mortar bank locations? Would that be ethical?
How they answeredI was honestly pretty confused by this one because it begged a lot of questions. I pushed back that there is a huge difference between matching someone already known to have committed a robbery and saying a person is more likely to commit a crime because of physical attributes or what they are wearing, which I said felt very wrong ethically. I also questioned whether that is even a decision a company should be making in the first place versus just calling law enforcement.
Follow-up questions- Are we talking about someone we already know committed a crime, or are we inferring risk from physical traits or clothing?
- 5
Technical round
TechnicalCodingCode ReviewDebuggingThe technical round was more of a code reading and code review interview than a heavy live coding round.
Q1. Write a short shell script to download a package, unzip it, and install it into an Anaconda environment.
How they answeredI wrote a very short bash script using basic commands like curl or wget, unzip, and the install step into an Anaconda environment. It was only a few lines, so they were not looking for anything fancy, just whether I could handle straightforward command-line tasks.
Q2. Walk me through this Python script and class, explain what each function is doing, tell me what __init__ is, and say whether these pytest unit tests are sufficient.
How they answeredThey gave me a well-written Python script with a class and several methods, and I explained it more function by function than line by line. I answered the basic Python question that __init__ is the constructor method, then reviewed the pytest examples and said whether they covered enough. Where they did not, I proposed and wrote a simple extra unit test or two to cover the missing cases.
Follow-up questions- What additional unit tests would you add?
Q3. Review this poorly written script and point out the problems.
How they answeredI walked through the script and called out the kinds of issues they wanted to see: repeated code, hardcoded values, weak variable names, and other basic maintainability problems. I did not have to rewrite it, just identify the issues clearly and explain why they were problems.
Tips from the candidate
What worked for me was getting as much detail as I could on the four CodeSignal question types and then practicing against a clock. I used AI tools to generate messy pandas datasets, sample regression outputs, bash prompts, and both good and bad Python scripts so I could rehearse the exact style of questions they ask. For the stats roleplay, I had a friend act like the business stakeholder and push me on the model assumptions, because that round is really about explaining technical issues in plain English. For the case, I would practice interviewer-led quantitative cases, not just open-ended product cases, and for the fit round I would still have STAR answers ready even though the manager can throw in random questions.
Company culture
Capital One felt very standardized and process-heavy to me. The Power Day format seems baked in across job families, and for data science it is remote now, not in person. The interesting part is that passing does not necessarily mean you get that exact job. It means you cleared the bar for the job family, and then hiring managers can still decide whether your background is specific enough for what they need. For internal candidates at least, that pass stays valid for a year, so there is a real networking and team-matching layer after the interview itself.