Deloitte · Primly Community

Deloitte data scientist interview (SQL, case, stats): what each round looks like

ds_dmitri · 4 replies

Just finished the Deloitte DS loop. This was for a senior data scientist role in their AI & Data practice. Sharing a breakdown because the prep resources for Deloitte DS specifically are thin.

The format I went through (4 rounds over 3.5 weeks)

Round 1: Recruiter screen Usual Deloitte recruiter questions. Travel, background, why the firm. They mentioned the role would involve client-facing deliverables, not internal product work. If you're coming from a pure product company DS background, that context matters.

Round 2: SQL + analytics challenge (async) Three SQL problems, about 45 minutes. One was a GROUP BY + window function problem, one was a multi-table JOIN with some filtering logic, one was closer to a business question where you had to write SQL to answer "what percentage of clients had a drop in activity month over month."

Not LeetCode hard but you need to be fluent. Window functions (LAG, LEAD, RANK, ROW_NUMBER) showed up. Practice those specifically.

Round 3: Live case + stats interview (60 min) Two-part interview. First half: a case scenario, something like "a client's ML model for credit scoring is underperforming. How do you diagnose and fix it." This is less about code and more about your mental model for model debugging. Talk through: data drift, feature importance, validation methodology, class imbalance.

Second half: statistics concepts. Bias-variance tradeoff, when to use which model, p-value interpretation, A/B testing design. Real questions, not trick questions. Know your stats cold.

Round 4: Behavioral + presentation Presented a slide deck summarizing a past project. 10 minutes, then Q&A. Questions were about methodology decisions, stakeholder communication, how I explained findings to non-technical clients.

What stood out vs typical DS loops The client communication piece is everywhere at Deloitte. Even in the model debugging case, they asked how I'd explain my findings to an executive. If you've only worked in internal-facing DS roles, practice articulating technical results in business terms. That's the gap I see most DS candidates miss here.

4 replies

analyst_ana

The window function callout is so true. I bombed an analytics screen because I'd been writing pandas in Jupyter for 18 months and completely neglected SQL. Don't let that happen to you. Brush up on LAG, LEAD, PARTITION BY, all of it.

ml_mike

The 'diagnose a failing model' question is my favorite interview question to get, honestly. Shows whether someone actually understands what can go wrong vs just knowing how to train a model. Data drift, label leakage, distribution shift, those are the real answers. Not 'try a different architecture.'

ds_dmitri

Exactly. And they went a level deeper: how would you build a monitoring system to catch drift before the client notices. That question separated the people who've operated models in production from the people who've just trained them.

de_derek

The presentation round is a wildcard people don't prep for. Picking the wrong project to present can sink you. Pick something where you had a clear impact, can explain the methodology simply, and have data to back the outcome. Don't pick the most technically impressive project if you can't tell a clean story about it.