Data Analytics

Analysis that doesn't change a decision is decoration. Frame the question, respect the data, and land on an action someone owns.

When to use

• Business/product/operations questions with data behind them.
• Ad-hoc deep dives; recurring reports (make them earn their cadence).
• Reviewing someone's analysis before it drives a decision.

Method

1. Frame the question: decision first — who decides what, by when, and what data would change it? No decision → probably no analysis.
2. Know the data: provenance, coverage, definitions (a metric with two definitions is a bug), gaps and biases — before computing anything.
3. Descriptive, honestly: distributions before averages; segments before totals; trends on comparable windows (no cherry-picked baselines).
4. Then diagnostic: what moved the metric — decomposition, cohort views; correlation ≠ cause; triangulate mechanisms.
5. Statistical care when inferring: sample size/power, confidence, practical vs statistical significance, multiple-comparisons awareness.
6. Insight → action: so-what → recommendation with expected effect, owner, and measurement plan; state uncertainty plainly.
7. Report like an engineer: one page — question, data, method, finding, recommendation, caveats; charts per data-visualization; numbers reproducible.
8. Close the loop: after the action, predicted vs actual; update the model of the world.

Quality gates

• Decision + owner named before analysis.
• Data provenance/definitions documented; segments + distributions shown.
• Inference claims carry uncertainty; multiple-comparison risk checked.
• Recommendation has owner + measurement plan; results reproducible.

Red lines

• No analysis without a question; no causal claims from observational data alone; no charts that oversell.

Related

product-analytics · data-visualization · decision-making.