Robinhood · Primly Community

Robinhood data scientist interview (SQL + case + stats), what to expect

de_derek · 4 replies

sat the robinhood DS loop a couple months ago, L4/L5 level. sharing my notes.

robinhood's DS interview is pretty thorough compared to some fintech companies. it's not just SQL and business case: they actually test statistics and ML concepts, which i appreciated.

what i saw across the rounds:

SQL (60 min): this was a real SQL interview. not 'write a basic join.' questions involved window functions, running totals, lag/lead for time series, and one question about identifying suspicious transaction patterns using self-joins. tables were clearly inspired by their actual data model (user trades, positions, balances). know window functions cold.

case / product analytics (45 min): given a user funnel for a new feature rollout, where would you look if conversion dropped 15% week over week? walked through my diagnostic framework: check if the data is right first, then segment. they pushed on A/B testing design: what's your minimum detectable effect, how long do you run the test, how do you handle novelty effect.

statistics / experimentation (45 min): this was the round that filters a lot of people. questions included: what is p-value and why is it often misunderstood, explain confidence intervals to a non-technical stakeholder, how do you handle multiple testing corrections, what would you do if your experiment shows different effects across user segments. also asked about power analysis.

ML concepts (30 min light): not a deep ML round but they asked me to walk through a model i built, explain feature selection choices, and how i'd handle class imbalance. this felt more like a conversation than a test.

behavioral (30 min): standard. influence without authority, prioritization, communicating ambiguous findings upward.

the overall bar: strong SQL, strong stats/experimentation, product curiosity. you don't need to be a deep ML engineer but you need the fundamentals.

4 replies

analyst_ana

the suspicious transaction pattern question using self-joins is so clever. i'm guessing they wanted something like: find users who have the same amount coming in and going out within a short window? is that close?

content_cole

close but more specific: they wanted to find circular transaction patterns. user A sends to B who sends to C who sends back to A within 24 hours, above a threshold amount. it's a graph pattern expressed in SQL, essentially. self-join on a transactions table by timestamp.

de_derek

the window functions thing: companies keep saying 'know SQL well' and candidates keep underpreparing. rank, denserank, rownumber, lag, lead, firstvalue, lastvalue. if you can write all of those from memory in a real schema you'll beat half the field.

finance_faye

the p-value misunderstanding question comes up constantly in fintech DS interviews. my best framing: p-value tells you the probability of your observed result given the null hypothesis is true, NOT the probability the null is true given your data. most people have it backwards and the consequences in financial analysis can be significant.