Y

Data visualisation and storytelling · Advanced

From analysis to a decision

3 min read

Why it matters

Stakeholder management and communication appeared in about three quarters of the Data Analyst job descriptions we read. At senior levels the job is to change a decision, and that depends on how you present the analysis as much as on the analysis itself. Many analysts stall at mid level because they present everything they did instead of what the reader should do.

The concept

Lead with the answer. A common structure (from Barbara Minto's pyramid principle) is:

  1. Situation: what the reader already agrees with.

  2. Complication: what changed or what is at risk.

  3. Resolution: your recommendation, with the one or two numbers that support it.

Then the evidence, one chart per point, each titled with its takeaway. Methods and caveats go last or in an appendix, but they must be there.

Write for someone who reads only the first three lines. If they stop there, do they know what to do?

Worked example: a one-page memo on customer retention

Using the RFM segmentation from the pandas shelf on the UCI Online Retail II data:

retention_memo.md
**Recommendation: run a win-back offer for 650 "at risk" customers before the next gift season.**

Situation: A third of our customers (1,519 "champions") bring 74% of revenue, and they bought within the last month.

Complication: Another 650 customers who used to order often have been quiet for a median of 96 days.
They account for 12% of last year's revenue, about 1.05 million. If they do not return by September,
we lose them for the peak season, when revenue doubles (September to November 2011 averaged
1.05 million a month against 0.58 million from December to August).

Resolution: Email the 650 with an offer on the products they bought most (wicker picnic baskets,
hanging heart T-light holders). Hold out a random 20% as a control group, so we can measure the effect.

Evidence
1. Chart: revenue share by segment (champions 74%, at risk 12%, hibernating 9%, recent 6%)
2. Chart: monthly revenue showing the September to November peak
3. Table: top products for the at-risk segment

Caveats
- One year of data, December 2010 to December 2011, from one UK online retailer.
- 25% of order lines have no customer id and are excluded.
- Segments use quartile scores; thresholds are a choice, not a law.

Notice the order: the reader knows the recommendation, its size and the ask in the first line. The control group turns the campaign into an experiment, which answers the obvious follow-up question: "did it work?".

Common mistakes

  • Chronological storytelling. "First I cleaned, then I joined..." belongs in the appendix.

  • Many findings, no recommendation. Pick the one that matters most. Others go in a list below.

  • Hiding caveats. Stakeholders who discover a caveat later stop trusting you. State them plainly.

  • Numbers without scale. "650 customers" is abstract; "12% of revenue, about 1.05 million" is a decision.

Interview angle

Case interviews and take-home debriefs end with "what would you recommend?". Answer in one sentence first, then support it. Practise saying a recommendation in under 20 seconds.

Practice task

  1. Rewrite one of your past analyses as a one-page memo with this structure.

  2. Present it to a friend in two minutes. Ask them what they would do next. If their answer is not your recommendation, rewrite.

Further reading