Build a narrative whose claims remain bounded by the evidence

Data storytelling questions show up in senior and staff interviews as case prompts: "Here's a dataset and a finding — walk me through how you'd present it to the VP of Product." Interviewers are not testing whether you can make a chart. They're testing whether you know who decides what, whether your claim matches the strength of your evidence, and whether you'll hold the line on uncertainty when someone senior pushes you to make the number sound better. A weak answer recites chart best practices; a strong answer starts with the decision and works backward to the evidence.

Start from the decision, not the data

Data storytelling is the deliberate connection of an audience's decision, a clear claim, trustworthy evidence, explanatory context, and a next action. It is not decorating charts with a dramatic headline. Begin with who must decide what, what they already know, what consequence matters, what uncertainty they can tolerate, and what action is available. The same analysis may require an executive brief, operational walkthrough, public explanation, or technical appendix.

What changes when the room is an executive versus an analyst team is not just vocabulary — it's where the recommendation goes and how much of the method survives the first slide. An executive audience gets the finding, the size of the effect, the confidence in it, and the recommended action, in that order; the methodology belongs in an appendix that answers challenge, not in an opening that establishes attention. An analyst team gets the method up front because they will audit it, and the recommendation is the conclusion of the argument rather than its opening. Analysts often present in the order they discovered the analysis, and the result is a presentation where the point arrives after most of the audience has stopped following.

When an interviewer asks "how would you present this," the first thing they listen for is a question back: who's in the room, and what do they need to decide? An answer that starts with the chart type has already failed the framing probe.

Lead with the claim

State the main message early in plain language. A useful structure is context, change or question, evidence, explanation, alternatives, implication, and action — but the sequence should reflect reasoning rather than suspense. Put the strongest decision-relevant evidence near each claim, distinguish observed facts from interpretation and recommendation, and make assumptions visible. One message per chart and per slide: if a chart needs three sentences of caveats to say what it shows, it's two charts.

The takeaway is the headline, and the evidence supports it rather than leads to it. A useful interview exercise: state the finding in one sentence a non-analyst could repeat accurately. If you can't, the analysis isn't finished or the claim isn't clear. A common weak answer builds the narrative like a mystery — data, method, more method, and the conclusion on the last slide. Interviewers probe this directly: "What's the one thing the CEO should remember?" Fumbling it means the story was organized for the analyst, not the audience.

Make the number mean something

A figure without a baseline, denominator, comparison target, and time window is not a finding; it's a raw observation. "Churn is 4.2%" answers nothing. "Churn rose from 3.1% to 4.2% among monthly-plan users in Q3, while annual-plan churn held flat at 1.4%" is a claim someone can act on.

Define the baseline, denominator, time window, population, unit, and whether the period is complete. Absolute and relative change answer different questions; percent changes from tiny bases sound dramatic — a 50% increase on 40 users is 20 users. A 12 percent uplift on 340 users and a 12 percent uplift on 34,000 users are different claims, and presenting them identically means the audience cannot tell which decisions the number can carry.

Interviewers probe here with a deliberately ambiguous figure: "Revenue is up 20% — good news?" The weak answer evaluates the number; the strong answer asks up from what, over what period, for which segment, and compared to what — and notes that a 20% rise off a one-month trough after a pricing change is a different story than steady 20% year-over-year growth. Where an effect is not distinguishable from noise, say so rather than presenting the point estimate alone; a finding reported with its interval and later confirmed builds more credibility than three confident findings of which one reverses.

Choose the chart that makes the argument

Chart choice is an argument about what comparison matters, so make it deliberately. Position and length encode comparison — bar charts compare magnitudes and need a zero baseline, because truncating the axis multiplies the apparent difference. Lines imply continuity and belong on ordered time. Distributions show spread where the average hides it. Part-to-whole is where pies get proposed and where they fail: readers compare angles badly beyond three or four slices, so a sorted bar usually carries the same information more legibly. Maps only when geography matters.

The obvious alternative often fails for a specific reason, and naming that reason is what interviewers want. Dual axes invite a causal reading of two series that may be unrelated — use them rarely and label them carefully. The test is whether someone could reach the wrong conclusion by reading the chart correctly — if they can, the chart is making an argument you did not intend.

Honesty in the visual is a recurring interview probe because it's where pressure gets applied. Truncated axes, dual axes, cherry-picked windows, and misleading color or scale are the standard four, and the fix for each is known: zero baseline for bars, one axis or clearly justified duals, the longer series shown or the window defended as the decision horizon, and color that doesn't encode magnitude redundantly with a scale chosen to exaggerate it. The fix should not weaken the message — if the honest version of the chart kills the story, the story was the chart.

Put uncertainty in the narrative, not the footnote

Uncertainty and quality belong in the main narrative when they could change the decision. Explain sample size, interval, model or forecast assumptions, provisional data, missing coverage, revision, and sensitivity in language the audience can use. Avoid unsupported decimal precision. Show scenarios or ranges where a point estimate falsely implies certainty. Say what is known, uncertain, and not measured.

Honesty about uncertainty is what makes the rest of the story credible, and it's usually the first thing cut for brevity. State the sample size, the comparison period, and what was excluded, because those determine whether the finding survives. Interviewers often role-play the pressure: "Can we tell the board it's a 15% lift?" The weak answer is yes or no; the strong answer states what the interval supports, what would change with more data, and what claim survives the worst case — then lets the executive choose with eyes open.

Correlation, causation, and the claim you can defend

Separate descriptive, diagnostic, predictive, and causal claims. Time order, plausible mechanism, controls, randomization, identification assumptions, and robustness affect causal credibility. Do not turn a coincident trend or model importance score into "X caused Y." Present alternative explanations and what additional evidence would discriminate them.

This is the most common trap in a storytelling case: the analysis is observational, and the candidate's headline quietly upgrades it to causal. Interviewers probe with "what else could explain this" — a weak answer defends the finding; a strong answer names two or three rivals (seasonality, a concurrent launch, a composition shift) and says what data would rule them out.

People, access, and auditability

Represent people without exploiting them. Human examples make aggregate effects understandable but must not reveal identity, invite re-identification, or substitute one anecdote for population evidence. Minimize data, avoid stigmatizing labels, report subgroup heterogeneity and small numbers safely, and consider who is absent from collection. Privacy-preserving aggregation can still harm a group through framing.

Accessibility is part of the narrative, not a compliance afterthought: descriptive titles, logical reading order, plain language, sufficient contrast, redundant cues — never color or hover as the only carrier of meaning — plus chart summaries and accessible tables. Caption audio and video, explain acronyms, and test with assistive technology.

Make sources and transformations inspectable: dataset owner, metric definition, time and refresh, methodology, and links to the analysis. Preserve versioned queries, notebooks, and chart data where policy permits. Label estimates, simulations, and synthetic examples. A screenshot without definition, filters, source, and date is weak evidence even if visually polished.

Design for the objection and the action

Anticipate the strongest objection, the affected segment, the operational constraint, and the downside. Provide a concise executive layer and optional methodological depth without hiding inconvenient evidence. State the decision owner, the requested action, timing, success measure, monitoring, and reversal trigger. Record what was decided and update the story when data, method, or recommendation changes.

Review both accuracy and interpretation: subject-matter, analytical, privacy, and audience reviewers checking claims, calculations, charts, and exclusions. Test whether people can restate the main message, identify its uncertainty, and choose the intended next step without coaching. Measure comprehension, decision quality, and whether outcomes change — not applause or slide count.

Likely follow-ups after you present a structure: "What would you cut if you had two minutes instead of twenty?" (the claim, the effect size, the confidence, the ask), "What's the strongest counter-argument?" (you should have it prepared, not improvised), and "What would make you withdraw the recommendation?" (a stated reversal trigger is the mark of a claim bounded by its evidence). If you can answer those three, the narrative will survive the room.