Matrixx0070 data-analyze
When to use
Use this when someone asks a question a dataset can answer — from "what was revenue last month?" to "why did churn spike in Q3 and what do we do?" Right-size the effort to the ask.
Not for: profiling an unfamiliar dataset (use data-explore first), writing the SQL itself (use data-write-query), formal significance testing (use data-statistical-analysis), or QA of a finished analysis (use data-validate).
Method
- Classify the tier and say which you chose: (a) lookup — one metric; (b) comparison — segment/time deltas; (c) diagnostic — why X changed; (d) stakeholder — decision + recommendation. Decision point: if the ask implies an action, it is tier (d), not (a).
- Restate the question precisely: metric definition, grain, time window, filters, population. Surface hidden assumptions before computing.
- Locate and confirm the source. Verify the grain and reconcile one total against a known figure before trusting any number.
- Compute the smallest correct operation for the tier. Show the query or transformation.
- Interpret: magnitude, direction, and signal vs noise. For diagnostics, decompose the change by segment, mix, and rate to find the driver.
- Recommend — tier (d) only. Tie each insight to one concrete action and its expected impact.
Example
Ask: "Why did Q3 revenue drop 8%?" (tier c). Decompose: revenue = customers × ARPU. Customers fell 2%, ARPU fell 6%. ARPU decline concentrates in the SMB segment (−14%), where a March discount campaign expired and was replaced by higher churn among discounted accounts. Finding: the drop is an SMB retention problem, not a pricing or volume problem across the board.
Pitfalls
- Building a 5-section report for a one-number lookup, or answering a strategic ask with one bare figure.
- Reporting a delta without checking it is meaningful rather than normal fluctuation.
- Skipping reconciliation, then presenting a number from an incomplete extract.
- Claiming a cause from a single correlated segment without decomposing alternatives.
Output format
Answer: <number + units + time window>, first line.
How: <one-line method + source, reproducible>
Context: <comparison / trend / benchmark that gives it meaning>
Caveats: <data gaps, definitions, confidence limits>
Recommendation: <tier (d) only — action → expected impact>
Lead with the answer; never bury the number under methodology. If the data cannot answer the question, say so and state exactly what would be needed.