Company Interview Questions · Consulting & Big 4

QuantumBlack (McKinsey): Data Scientist

❓ 6 questions🏢 McKinsey QuantumBlack🆕 Recent reports

QuantumBlack (McKinsey): Data Scientist Interview Process and Questions. Rounds, real reported questions and model answers, compiled from public candidate reports (sources at the end). Practice mode hides every answer: say your answer first, press Show answer, then rate yourself. Use 🔁 Review again to build your revision list and 🎲 Random question for a quick drill.

★ Must-know 145 📝 Assignments

About this interview

Compiled from a public interview guide that aggregates seven QuantumBlack Data Scientist candidate reports, plus public candidate threads. The questions are paraphrased, and the model answers are written by EasyLearnSpace.

Role	Data Scientist / Senior Data Scientist (QuantumBlack, AI by McKinsey)
Stages	Online assessment → recruiter / HR screen → technical coding and data screen → ML, GenAI and project deep dives → collaborative case + final rounds with a director
Online assessment	Easy coding problem, a notebook task (clean data, build train/test helpers, fit a random forest), about 20 MCQs on git, ML, stats and GenAI
Timeline	About 2–2.5 months
Outcomes in reports	Mixed: 1 offer, several rejections, 1 withdrawal

01Rounds at a glance

Technical screen: easy–medium algorithms, pandas, SQL (joins, duplicates); once, "how would you troubleshoot an API returning a 500 error?"
Deep dives: classical ML (bagging vs boosting, bias-variance, multicollinearity, PCA + k-means), GenAI concepts, the impact of your projects.
Case: end-to-end and collaborative: frame the problem → EDA → model → interpretation (often churn), plus a product-metric question.

02Questions and model answers

What is the difference between k-fold and stratified k-fold cross-validation?

K-fold: split the data into k equal folds; train on k−1 and validate on the remaining one, rotating through all k.
Stratified k-fold: each fold keeps the same class proportions as the full dataset. This is essential for imbalanced classification, otherwise some folds may have very few positives.
Others: GroupKFold (no leakage across users or patients) and TimeSeriesSplit (forward-chaining) for temporal data.

Bagging vs boosting: how do they differ?

Bagging (random forest): trains many models in parallel on bootstrap samples (plus feature subsampling) and averages them. It reduces variance and is robust and hard to overfit.
Boosting (XGBoost, LightGBM, CatBoost): trains models sequentially, each fixing the previous errors (fitting gradients of the loss). It reduces bias, is often more accurate on tabular data, but needs tuning (learning rate, depth, early stopping) and can overfit noise.

Case: how would you predict customer churn from a dataset? Walk through end to end.

Frame it: what counts as churn (no purchase in 60 days? contract cancelled?), the prediction horizon, the intervention budget, and the success metric (retained revenue).
EDA: churn rate, cohorts, tenure curves, usage trends, complaints, price changes. Check for leakage (cancellation-request fields).
Features: recency/frequency/monetary (RFM), usage deltas, support tickets, plan changes, payment failures.
Model: logistic baseline → gradient boosting; time-based validation; PR-AUC and lift@top-10%; calibrate.
Interpret: SHAP drivers; segment personas.
Act: target by uplift (who will respond to an offer), not just risk; run a controlled experiment; monitor.

Average comments per user dropped 10% last month. How would you investigate?

Is it real? Check logging and instrumentation changes, bots or spam filters, and the definition (active users vs all users).
Decompose: comments per user = (comments / active commenters) × (commenters / users). Which part dropped?
Segment: platform (iOS/Android/web), region, new vs existing users, content type, app version.
Look for causes: a product change (comment-box UI, ranking), seasonality, an influx of new users (mix shift), an outage, competitor events.
Confirm with an experiment or a natural comparison, quantify the impact, and recommend a fix.

How would you troubleshoot an API that returns HTTP 500 errors?

A 500 is a server-side error. Read the server logs and stack traces for the failing requests and correlate them with timestamps and request IDs.
Reproduce with the same payload (curl/Postman), then check recent deploys, config or secret changes, and dependency health (DB, downstream APIs, model endpoint: timeouts, OOM).
Check inputs: unexpected nulls or types that crash the handler. Add validation that returns a 4xx instead.
Mitigate (roll back, or scale up if the cause is resource exhaustion), fix the root cause, and add alerts and tests.

How do you handle multicollinearity in regression?

Detect: a correlation matrix, VIF above 5–10, and unstable coefficients that flip sign.
Effects: inflated variances, so the coefficients become uninterpretable. Predictions may still be fine.
Fix:
Drop or combine correlated features.
Use domain-driven feature selection.
Ridge regularisation.
PCA (at the cost of interpretability).
Centre interaction terms.

03How to prepare for QuantumBlack

Practise a timed notebook: clean data → train/test helpers → random forest → evaluate, in about 60 minutes.
Rehearse a collaborative churn case out loud, structured like a consultant.
Revise ML theory (CV variants, bagging vs boosting, bias-variance, multicollinearity) and SQL (duplicates, top-N per group).

Sources

Created by Ajeetkumar