Capital One · Primly Community

Capital One data scientist interview (SQL + case + stats): what each round actually tests

ds_dmitri · 7 replies

just finished the capital one DS interview loop for a mid-level data scientist role (not the analyst track, the DS track). took detailed notes. sharing everything.

the structure (5 rounds total): recruiter screen (30 min): resume, background, comp range technical screen with a DS (45 min): SQL + some stats case study (60 min): live product/business case, see below coding (python) interview (45 min): not ML-heavy, more pandas/data manipulation behavioral round (45 min): STAR, two interviewers

SQL round: this is real. they're a SQL shop at the DS level. you will write queries. in my case: a join-heavy query across three tables (transaction, customer, account) a window function question (rank customers by spend within each product category) a question about finding anomalies in time-series spending data

they wanted clean readable SQL, not just a correct answer. CTEs are better than nested subqueries here.

stats round (same interviewer): conceptual but specific. questions: explain p-value to a non-technical PM if we see a 5% lift in an A/B test, how do we decide if it's real what's the difference between type I and type II error in the context of a fraud model (nice and on-brand) how do you handle class imbalance

the business case: given: credit card customer data, a sharp drop in average transaction size over the past 90 days in a specific merchant category. figure out what's going on and what you'd recommend.

no data was actually given. you have to structure your investigation, say what data you'd pull, what hypotheses you'd test, in what order. they interrupt with clarifying questions. this round is mostly about structured thinking and being able to communicate with business stakeholders.

python round: pandas-heavy. read a csv, clean it (missing values, wrong dtypes), compute some aggregations, produce a summary table. not scikit-learn or model building. they want to see whether you can work with messy data efficiently.

7 replies

analyst_ana

the class imbalance question for fraud models is so standard and so many people mess it up. what answer did you give?

ds_dmitri

talked through oversampling (SMOTE), undersampling, and class-weighted loss functions. mentioned that in production fraud, precision and recall trade-offs matter more than accuracy, and that you tune the threshold based on the business cost of false positives vs false negatives. they seemed happy with that.

analyst_ana

the business cost framing at the end is the thing that separates the answer. noting this.

ml_mike

curious if they got into any ML model questions at all or was it really just stats/SQL/case the whole way

ds_dmitri

minimal ML. one question about what algorithm i'd use for a churn prediction problem and why. but they didn't ask me to implement anything. the DS role at capital one at mid-level is more analytics-oriented than modeling-heavy. if you want modeling-heavy look at their ML engineer track.

numbers_only

comp check: what was the offer range for mid-level DS in 2026? seeing a lot of variance in what people report for this role

ds_dmitri

recruiter mentioned $130-155K base for mid-level DS in McLean. RSUs on top but nothing dramatic at this level. total comp probably $150-175K all-in depending on the RSU grant.