Most candidates prep for a data-science consulting interview as if it were one long ML oral exam. It is not. BCG X, QuantumBlack, Bain's analytics track, and Big Four advanced-analytics roles run three different screens that reward three different things, and the sheet fails you when you bring the wrong skill to the room.
The thesis is simple: the business case scores judgment under ambiguity, the technical case scores whether you can turn a decision into a model without inventing data you do not have, and the coding screen scores whether you can ship working logic under time pressure. After this piece you should be able to name which screen you are in, what the interviewer is marking, and walk a business question through to a model choice and a go/no-go decision—exactly the translation BCG X describes as the heart of the role.
If you are targeting BCG X specifically, read the dedicated BCG X interview guide next. This article covers the shared format across the analytics consulting track.
What the firms say they hire for
Start with the job pages, not folklore.
BCG X frames the data-science role as translating real-world business problems into a quantitative context and building models that answer those questions. Its Stockholm process description is unusually explicit: a coding test, technical case interviews, and/or a live coding challenge. Older BCG GAMMA recruiting materials (the analytics practice that folded into BCG X) laid out a similar sequence: a multi-hour coding challenge, then a technical case with a data scientist, then a business case with a BCG consultant.
QuantumBlack, AI by McKinsey hires people who apply advanced analytics and modeling, then interpret results and recommend action—not people who only optimize a leaderboard metric. McKinsey's North America technical opportunities page lists, for data science, timed technical challenges, technical interviews, problem-solving interviews, and personal experience interviews. A McKinsey analytics recruiter has publicly described analytics hiring as typically three to five case interviews that test both technical work and business aptitude, because "successful data scientists are not only great with codes and data; they are also current or potential business leaders."
Bain Vector embeds data science, AI, and engineering inside industry teams rather than parking them in a pure research lab. Bain's interviewing guidance is role-specific: consulting tracks get cases; technical tracks may include portfolio walkthroughs, aptitude challenges, or coding. Big Four advanced-analytics postings (for example Deloitte Strategy & Transactions analytics roles) repeat the same demand: translate client issues into analytics problems and turn solutions into insight stories for executives.
None publish a confidential scoring rubric. They do publish process shape. Treat that as the prep map.
The three screens, side by side
| Screen | Typical format | What the sheet rewards | What kills the score |
|---|---|---|---|
| Business case | 25–40 min client-style problem, often with a partner or engagement manager | Structure, clarifying questions, prioritization, a recommendation with risks | Jumping to "I'd train an XGBoost" before the decision is clear |
| Technical / modeling case | 30–45 min with a data scientist; design an experiment or model end to end | Target definition, data feasibility, evaluation tied to cost of error, monitoring | Architecture theater: transformers and deep nets with no data story |
| Coding screen | Timed online test and/or live pair programming (Python / SQL / pandas) | Correctness, clean data handling, edge cases, readable code | Perfect algorithms that do not run, or brittle one-liners |
Variants exist. BCG X ANZ recruiting describes a coding assessment before virtual technical interviews with cases and live coding. QuantumBlack mixes timed challenges with problem-solving and personal experience interviews. Ask your recruiter which of the three you face in which round.
Business case
This is still a consulting case. The interviewer is listening for whether you can own a messy commercial question: margin, churn, capacity, fraud dollars, supply reliability. You clarify the objective, time horizon, and constraints. You structure. You do rough math. You land a recommendation a client could act on Monday.
The analytics twist is that your recommendation may be "instrument first, then model" or "do not build a model." Saying that out loud, with a reason, scores higher than naming a library. When I sit on these panels, I stop listening the moment a candidate treats feature engineering as the opening move. The opening move is the decision the business needs to make.
Technical / modeling case
Here a practicing data scientist wants to know whether you can frame the ML task correctly. Binary classification vs ranking vs forecasting. What is the unit of prediction (customer, SKU-day, claim). What labels exist today, and how noisy they are. Which features you would ask for first, and which you would refuse because they leak the future. How you would evaluate: AUC is rarely the whole story when a false positive costs $5 and a false negative costs $500.
They also score communication. QuantumBlack's own role copy asks for people who explain complex ideas to colleagues and clients. If you cannot explain why a simpler baseline beats a flashy model for this decision, you are not ready for client rooms.
Coding screen
This one is binary in spirit: does the code work, and can a teammate read it. Expect pandas joins, group-bys, null handling, simple SQL, maybe a small algorithm. Latency and elegance matter less than correctness under a clock. Firms that publish process details (BCG X coding test; McKinsey timed technical challenges) are telling you this screen is a gate, not a conversation.
Ned's rule. In a data-science consulting interview, name the decision first, the prediction second, and the model family third. Candidates who reverse that order look clever and still get dinged.
Worked mini-case: FreshLane produce stockouts
Prompt (original). FreshLane is a 120-store regional grocery chain. Fresh produce stockouts on high-velocity SKUs (berries, bagged salad, bananas) cost an estimated $2.4M per year in lost margin and substitution to cheaper private-label items. Waste on the same SKUs runs about $1.1M. Leadership asks: should we build a demand-forecasting model to change ordering, and if so, what kind?
Step 1 — Business decision (what the partner scores)
Clarify:
- Decision unit: store × SKU × day order quantity for the next 7 days.
- Objective: cut stockouts without raising waste above today's $1.1M.
- Constraint: store managers still place the final order; any model is a recommendation, not an auto-PO.
- Success metric for a pilot: stockout rate on a defined SKU list down 20% while waste stays within +5% of baseline.
A weak opening: "I'd use Prophet / LSTM / a transformer on sales history." A strong opening: "The decision is whether to change order quantities; the model only earns its keep if managers trust and act on it, so we need a pilot design before architecture."
Step 2 — Model framing (what the data scientist scores)
| Choice | Why it fits FreshLane | Why it might fail |
|---|---|---|
| Heuristic baseline: 7-day moving average × day-of-week factor | Ships in days; managers understand it | Misses promotions and weather |
| Gradient-boosted regression on store-SKU-day demand | Handles promotions, holidays, local events as features; tabular data is the natural shape | Needs clean promo calendars and inventory snapshots |
| Deep sequence model | Attractive on paper for long seasonality | Overkill for short pilots; hard to explain to store ops |
I would propose a two-stage path: ship the heuristic baseline in week one as the control-comparable recommender, then train a gradient-boosted regressor on 18 months of store-SKU-day sales with features for day-of-week, promo flags, local temperature, and on-hand at open. Prediction target: units sold next day (censored by stockouts—so use a simple imputation or a two-part model that first predicts stockout probability). Evaluation: pinball loss or MAE on sold units and a decision metric—pilot lift in stockout rate vs waste dollars.
Cost of error: over-forecast burns waste dollars; under-forecast burns the $2.4M stockout margin. Weight the loss asymmetrically toward under-forecast on the top 30 SKUs by contribution margin.
Step 3 — Coding screen (what the online test scores)
You will not build FreshLane's model in the interview. You might get a 45-minute task: given a CSV of daily sales and inventory, write pandas to compute stockout days (on_hand == 0 and demand_proxy > 0), join a promo calendar, and output top-10 SKUs by stockout frequency. Correct joins and null handling beat a clever unused API.
Step 4 — The decision you should say aloud
"Pilot 20 stores for six weeks. Control stores keep current ordering. Treatment stores get the boosted recommender as a suggested order, with the moving-average as fallback when confidence is low. Ship only if stockouts fall ≥20% and waste stays within +5%. If managers override more than 40% of suggestions, stop and fix trust and UX before more modeling."
That last sentence is what separates analytics consultants from model hobbyists. The sheet is listening for it.
How prep should split your hours
| Week focus | Share of practice time | Concrete work |
|---|---|---|
| Business cases | 40% | Profitability, ops, and "should we build this" cases with a recommendation |
| Modeling cases | 35% | Target, features, leakage, evaluation, monitoring—spoken out loud |
| Coding | 25% | Timed pandas/SQL drills; one live pair session per week |
PhD candidates often over-index on modeling novelty; industry ML engineers underweight the business case. Both patterns are visible from the interviewer chair. See also Is a PhD an advantage at MBB? and BCG X roles and interviews.
Practice this week
Run one full loop on an original prompt of your own: pick a business metric that moved, force yourself to state the decision, then design a model with a pilot go/no-go rule. Do it out loud in 25 minutes.
Then take a five-minute first rep on CoachNed at /start—three typed case turns with instant scores, no account required—and one structure or math drill if your business-case muscle is soft. Everything on the platform is open for seven days with no card; after that it is $120 for a recruiting season or $49 a month. Voice practice with Ned lives at /interview if you need the live case format.
Frequently asked questions
Do data science consulting interviews include a normal case interview?
Usually yes, at least once. BCG X and QuantumBlack both mix business problem-solving with technical assessment. Treat the business case as a consulting case first; technical vocabulary is a tool, not a substitute for a recommendation.
Will I have to code live for BCG X or QuantumBlack?
Often yes for data-scientist tracks. BCG X career pages describe coding tests and live coding; McKinsey's technical recruiting pages list timed technical challenges for data science. Confirm the format with your recruiter for your office and role.
Is Bain Vector the same as a product-company data science interview?
No. Bain embeds analytics inside consulting delivery (Vector). Expect more client-communication and business framing than a pure FAANG-style ML system design, though coding and modeling still appear for technical roles.
What model should I default to in a technical case?
Start with the simplest model that can answer the decision with available data—often a baseline heuristic plus a gradient-boosted tabular model. Escalate complexity only when you can name the failure mode the baseline cannot fix.
How is this different from Capital One or other quant banking interviews?
Product and lending analytics shops lean harder into product metrics and coding depth. Consulting analytics interviews keep a heavier business-case weight. For that banking variant, see the Capital One case interview guide.
Sources
- Explore roles — BCG Stockholm (BCG X process) — BCG X role framing and coding test / technical case / live coding sequence. Checked 2026-09-24.
- BCG X Associate Recruitment in ANZ — Coding assessment then technical and live coding interviews. Checked 2026-09-24.
- Data Scientist — QuantumBlack careers — Role expectations: modeling, interpretation, recommendations, communication. Checked 2026-09-24.
- North America technical opportunities — McKinsey — Data science interview components: timed challenges, technical, problem-solving, personal experience. Checked 2026-09-24.
- Inside tips from a McKinsey recruiter (analytics) — Three to five analytics case interviews; business-leader aptitude. Checked 2026-09-24.
- Bain Vector digital — How Bain embeds data science and AI in consulting teams. Checked 2026-09-24.
- Interviewing at Bain — Role-customized interviews, including technical aptitude formats. Checked 2026-09-24.
CoachNed is not affiliated with McKinsey, BCG, Bain, Deloitte, or any firm named above.