Data Scientist case study round: framing problems end to end

You are given a scenario such as 'design a system to reduce customer churn' and asked to walk through your approach. There is rarely a single correct answer; the interviewer probes your process. Most Data Analyst / Data Science Jobs at product companies now include a round like this before extending an offer.

• Problem framing: restating the business goal and the decision the model supports.
• Data and features: what you would collect, and what leakage or bias risks exist.
• Method and metric: the model class, the evaluation metric, and why they fit.
• Validation and rollout: how you would test and monitor in production.

Strong candidates keep the business objective in view at every step; weak ones jump to a model before defining success.

Scenario: 'Build an approach to reduce churn on a subscription product.'

Strong walkthrough: 'First, is the goal predicting churn or reducing it? Those differ. I would define churn precisely, then frame it as predicting churn probability so we can target retention offers. Features: usage trend, tenure, support tickets, payment failures, watching for leakage like cancellation-flow events. I would evaluate on precision-recall at the intervention budget, validate on a temporal holdout, and A/B test the retention action rather than assuming the model alone reduces churn.' This links model to decision.

Weak walkthrough: 'I would train a random forest on all the columns and get high accuracy.' No definition, no leakage check, no link to an action. It sounds technical but answers the wrong question.

The case is a proxy for how you would scope real ambiguous work.

• Judgement: do you choose the simplest method that meets the goal?
• Rigour: do you catch leakage, bias, and the difference between prediction and causation?
• Product sense: do you connect the model to a decision and a measurable outcome?

Naming trade-offs unprompted, such as interpretability versus accuracy, signals seniority. So does saying what you would do differently with more data or time.

Build a reusable scoping structure so open prompts stop feeling intimidating.

• Practise a loop: clarify goal, define target, list data and risks, choose method and metric, plan validation and rollout.
• Take five common scenarios (churn, recommendation, fraud, forecasting, ranking) and walk each end to end aloud.
• Force yourself to state at least two trade-offs per case.

Because the round is spoken, rehearse narrating a full design without notes. A free AI voice mock interview on InterviewPrep can generate a case-style prompt from your CV and a real job description, then score your structure, pace and filler words, so your reasoning sounds deliberate rather than rushed.