Storytelling with Data

2006, TED. Hans Rosling steps to the screen and shows three animated bubble charts. In 15 minutes he flips the world's view of developing countries: 'Sweden and Bangladesh today are Sweden in 1970 and Sweden in 1900 - not different worlds'. The same WHO numbers had been freely available for years. They changed decisions only when Rosling gave them a narrative.

• **Spotify Wrapped**: the annual personal listening recap went viral not for its data (which is dull) but for the storytelling wrapper: hero (you), conflict (unexpected top genre), resolution (your year in music)
• **FiveThirtyEight and The Pudding**: data journalists reinvented the genre by moving from 'tables and charts' to narrative data essays - 1 million views on a piece about Hollywood economics
• **Internal McKinsey & BCG decks**: pyramid principle and MECE slide structure are so codified that newcomers train for 6 months specifically on storytelling, not on analysis

Narrative Structure

Hans Rosling in 2006 at the TED conference used three animated bubble charts to flip the world audience's view of developing countries. The same WHO statistics had been publicly available for years and moved nobody. The difference is not in the data but in the narrative. Cole Nussbaumer Knaflic formalised this in 'Storytelling with Data': a fact without structure stays background noise. Any presentation of an ML model result or A/B test should follow a three-act structure: setup (context and problem), conflict (what was discovered, what is unexpected), resolution (decision and call to action).

1. **Setup**: 'Checkout conversion dropped 8% over the quarter' - what is observed and why it matters now
2. **Conflict**: 'The drop is concentrated in the desktop Chrome 119+ segment; mobile and Safari are stable' - an unexpected finding that breaks the obvious hypothesis
3. **Resolution**: 'Hypothesis: regression in the checkout.js update on Oct 12; recommendation: roll back the release and run an A/B test of the fix' - a concrete action with a measurable result

Anti-pattern: data dump. A 25-slide deck of dashboards has no narrative - the audience sees numbers but does not retain the conclusion. The McKinsey 'Pyramid principle': start with the answer, then 3 supporting points, then evidence. C-suite listeners decide in the first 60 seconds; the rest reinforces or undermines. Slides without a headline conclusion are communication failures.

Chart Selection

The chart type is determined by the question the chart must answer, not by the data. Andrew Abela's 'Chart Chooser' matrix identifies 4 fundamental comparison types: composition (part-of-whole), comparison (across categories or time), distribution (how values spread), relationship (how variables connect). Most chart-selection errors are compositional data plotted as bar charts or time series in pie charts.

Pre-attentive processing principles: the brain takes ~250 ms to read position, length, colour, area - in decreasing precision. Position (scatter plot) yields the most accurate comparison; length (bar) is also reliable; angle/area (pie) is much worse (Cleveland and McGill, 1984). Therefore pie charts are usable only for 2-3 categories, and any share comparison is better served by a horizontal bar chart.

Dashboard Design

A dashboard is not a place for every possible metric. Stephen Few in 'Information Dashboard Design' defines it strictly: 'single screen, persistent display of the most important information needed to achieve goals, consolidated and arranged for at-a-glance monitoring'. Key elements: (1) one screen, no scrolling; (2) 4-9 key metrics maximum (Miller's law on working memory); (3) F-pattern or Z-pattern layout - the most important content in the top-left corner.

Dashboard levels: Strategic (CEO, weekly) - 5-7 top KPIs, quarter/year trends; Operational (PM, daily) - product metrics, drill-down to segments; Analytical (data team, on demand) - hypothesis exploration. The main mistake is mixing levels on one screen: a KPI like MAU should not sit next to p99 latency distribution - they have different audiences and update cadences.

• **Signal vs noise**: drop gridlines, redundant legends, 3D effects, unnecessary decoration
• **Colour carries meaning**: grey = baseline, colour = highlight only for anomalies. Red/yellow/green only when thresholds exist
• **Context next to the number**: '142 orders' is useless; '142 orders (+18% WoW, target: 150)' is decision-grade
• **Targeted summary**: SLO target and current value on one line - no need for two separate widgets
• **Update timestamp**: every dashboard must show when data was last refreshed. Stale dashboards are a source of bad decisions

Presenting Results

Presenting ML results to stakeholders is a distinct skill. Technical leads make a typical mistake: they show the confusion matrix, ROC AUC, hyperparameter search trajectory. Business audiences see noise and lose interest in 90 seconds. Edward Tufte: 'Above all else show the data' - but 'data' means different things to a CEO and to an ML engineer. For the CEO data = dollar impact; for the PM = retention/conversion delta; for the DS = model metric.

Template for a business presentation of an ML project: (1) One slide of problem framing - which funnel gap is closed; (2) One slide of approach without model internals - 'trained a classifier on 6 months of transactions'; (3) One slide of dollar result - 'fraud loss reduced by $1.2M/year at the same false-positive rate'; (4) One slide of risks - 'requires quarterly retraining on new data, drift monitoring'; (5) One slide of roadmap - what next. Technical details belong in the appendix for the technical audience.

Key Ideas

• **Narrative** structures data through three acts: setup -> conflict -> resolution. Without narrative even strong data becomes a data dump
• **Chart type** is dictated by the question, not the data: composition, comparison, distribution, relationship - Abela's matrix prevents missteps
• **Dashboard** is a single screen for one audience, 4-9 key metrics, no mixing levels. Colour carries meaning; grey is baseline
• **ML presentation** requires translating metrics into business impact (dollars, retention, NPS) - AUC and F1 belong in the technical appendix, not the CEO slide

Related Topics

Hans Rosling in the opening did not surface new data - he redrew known statistics. That shows communication often matters more than analysis. Storytelling closes the data science loop from raw data to decisions and connects to several pipeline stages:

• EDA and Visualization — EDA charts for the analyst are drafts; storytelling charts for presentation are final edits of the same data for a different purpose
• A/B Testing — Test results are almost always pitched to management via storytelling: 'a 2.3% lift means +$X in revenue' rather than 'p-value=0.04'
• ML in Product — Models become real only when the team grasps their impact. Presenting ML results is the final step from R&D to production

Вопросы для размышления

• Rosling used animation (change over time) and it worked at TED. In corporate presentations animated charts are often annoying. What is the contextual difference?
• A product team asks for 'a dashboard with all the metrics'. Why is that request a red flag, and how should it be rephrased to produce a useful artefact?
• If a CEO gives 30 seconds for your ML model pitch, what three sentences do you say? What trade-off do you make between technical accuracy and clarity?

Связанные уроки

• ds-12 — Causal inference makes the stories honest - without it correlation lies
• ds-14 — Advanced visualization - the next step after narrative
• ml-05-evaluation — Metrics - the raw material for data storytelling
• prob-08-variance — Variance explains the uncertainty in data stories
• stat-08-correlation