Can you walk me through the complete data analytics lifecycle from a business question to a decision?
The data analytics lifecycle is the path from a vague business request to a decision someone acts on. Interviewers ask this to check whether you see analysis as a business process rather than a sequence of tool operations, so name each stage and say what done looks like before you move on.
Frame the question — turn the request into one measurable question with an agreed metric definition, population and time window; confirm what decision the answer drives and what result would change that decision.
Source and profile — identify the system of record, then check grain, row counts, date coverage, key uniqueness and null rates, and reconcile a known total against the source before trusting anything.
Clean and prepare — resolve missing values, duplicates, type errors and outliers, recording each decision so the same transformation can be re-run on the next refresh.
Analyse and validate — explore, segment and test the hypothesis, then deliberately attack the result: re-check the grain, look for confounders and segment reversals, and sanity-check against a number the business already trusts.
Communicate and follow up — deliver a recommendation with assumptions and limitations stated up front, then track whether the decision was taken and whether the metric moved.
Key terms: business question, metric definition, grain, data profiling, reconciliation, confounder, assumptions, recommendation, impact measurement
Answer with a real project as the spine rather than reciting stage names — say "the request was 'why is churn up', I reframed it as 'which cohort and which product line drives the increase', and here is what each stage produced". Interviewers listen for two things most candidates skip: that framing happens before any data work, and that validation is a deliberate step rather than something you assume.
Expect the follow-up "where do most analyses go wrong?" — the strong answer is unclear framing and unverified grain, not tooling. Avoid making it sound linear; mention that you loop back when profiling shows the data cannot support the original question.