Quick answer
Marketing data science starts with a question, turns the available records into usable evidence and tests what the answer means. Decide what you’ll do differently before choosing the model. Otherwise you risk finishing with an interesting chart and no next step.
Write the decision in one sentence
“Would a clearer onboarding email reduce early cancellations?” gives the team something to investigate. “Find insights” leaves them guessing. Name the action, the audience and the result you care about before asking anyone to build a model.
Agree who will use the result, when they need it and what evidence would change their plan. If nobody can act on the finding, refine the question before collecting more data.
| Stage | Deliverable | Review question |
|---|---|---|
| Define | Audience, outcome and time window | Could another person measure the same thing? |
| Prepare | Documented dataset and exclusions | What is missing or duplicated? |
| Compare | A simple baseline | Does complexity add useful accuracy? |
| Evaluate | Results on separate data or a fair test | Will the finding hold beyond the sample? |
| Handover | Owner, monitoring and next action | Who responds when performance changes? |
Build the dataset around the timing
For retention analysis, establish when customers entered, how long they were observed and what cancellation means. New customers have had less opportunity to leave. Comparing them directly with older customers can produce a misleading result.
Document joins between people, accounts and transactions. Reconcile counts with source systems and record exclusions. The marketing integration guide covers the identifiers and ownership needed for dependable reporting.
Begin with a baseline
A simple cohort comparison or rule gives the team something to improve on. Inspect the errors, not just the average score. An approach that works for a large segment may fail for a smaller group with different behavior.
Reserve data for evaluation rather than repeatedly tuning against every available record. Google’s explanation of overfitting describes why strong performance on familiar examples can fail to carry over to new ones.
Test the action separately from the explanation
An association can suggest a useful hypothesis without proving its cause. Customers using a feature may retain longer because they were more engaged already. Promoting the feature needs its own evaluation.
Where a controlled test is practical, define eligibility, the comparison and the primary outcome before launch. Otherwise, state the limits of the observational evidence and avoid a stronger causal claim than the design supports.
Make the handover usable
Deliver the recommendation with assumptions, source definitions, unresolved issues and an owner for the next step. If a model enters production, agree monitoring and a fallback. If the result does not justify a change, record that decision too.
Separate reporting an observed outcome from estimating what would have happened without the campaign. See our guides to marketing incrementality testing and attribution, MMM and incrementality.
