Visa · Primly Community

Visa data scientist interview: SQL, case, and stats breakdown from my recent loop

analyst_ana · 5 replies

Went through the Visa DS interview loop in Q1 2026 for a data scientist role on the risk analytics team. Three rounds total after the OA. Here's the real breakdown.

SQL round. This was more involved than I expected. Not just window functions and CTEs, but actual multi-step analytical queries. The scenario given: you have a transactions table, a merchants table, and a cards table. Write a query to find merchants with an unusual spike in declined transactions over the last 30 days relative to their 90-day baseline. That's a real, applied problem. You need window functions, date arithmetic, probably a CTE or two. Know your aggregation and partitioning cold. Practice against fintech-flavored datasets if you can.

Stats and probability. They asked a Bayesian question about fraud detection. Given a prior probability of fraud for a transaction type, a model sensitivity of X and specificity of Y, what's the posterior probability this flagged transaction is actually fraud? Classic confusion matrix + Bayes theorem. Then they asked how you'd choose a decision threshold for the model. Be ready to talk about precision/recall tradeoffs and business cost asymmetry (false negatives in fraud are more costly than false positives in most contexts).

Case/business problem. Open-ended: "Visa sees that authorization rates dropped 2% for a specific issuer in Mexico over the last 60 days. Walk me through how you'd investigate." Not a coding problem, pure analytical reasoning. I asked clarifying questions first, then segmented by card type, transaction type, time of day, merchant category. They wanted to see a structured diagnostic approach.

Behavioral. One session. Mostly around how I'd communicate findings to non-technical stakeholders and a time I had to change my analysis based on feedback.

Difficulty overall: genuinely applied, not LeetCode-DS. If you've worked on real data problems in fintech or done serious SQL work on transactional data, this plays to your strengths.

5 replies

analyst_ana

The Bayesian fraud detection question is one I've seen at like four fintech companies now. It's a standard. I've started keeping a "fintech interview question bank" and this is a repeat entry.

sec_sasha

The multi-table SQL scenario they described is basically a real Visa analytics problem. Which... makes sense. At least they're testing on something they actually care about and not binary tree reversal.

ml_mike

"False negatives in fraud are more costly" is a good frame but it flips in certain contexts (like credit line extension, where false positives have high cost). Did they probe on the nuance or was it a one-answer question?

sdr_sky

They probed. The interviewer asked me to consider a scenario where the false positive rate was high enough to cause merchant friction. So yeah, it wasn't just "always minimize false negatives." Context matters and they tested for that.

content_cole

Any signal on DS comp at Visa for mid-level? I've seen wide ranges across fintech and Visa is harder to pin down than the big names.