Communicate Clearly: Storytelling, Annotation, and Decision Support

30 min

From Charts to Evidence Statements

Learning objective: write claims bounded by the data; separate observation, interpretation, and unknowns.

A chart does not speak for itself; the sentence written next to it does most of the persuading. The most common failure is a caption that claims more than the chart can support, quietly turning a correlation or a single period's snapshot into a causal or permanent conclusion.

before = df[(df["entity_id"] == "E01") & (df["period"] == "2023") & (df["metric"] == "emissions_intensity")]["value"].iloc[0]
after = df[(df["entity_id"] == "E01") & (df["period"] == "2025") & (df["metric"] == "emissions_intensity")]["value"].iloc[0]
change_pct = round((after - before) / before * 100, 1)
# The sentence must match this number exactly, and say nothing the chart cannot show

• Writing "caused" when the evidence only supports "correlated," with nothing else to support causation
• Turning a single period's snapshot into a claim that this is how things always look
• Leaving out what the chart cannot answer, so the reader assumes it covers every angle

Key takeaway: Write the sentence next to a chart as tightly as the chart itself: state the observation, label any interpretation, and never claim more than the evidence supports.

35 min

Annotation and Data Story Structure

Learning objective: use title, subtitle, annotations, source notes, and sequence to guide reading without overstating.

Good annotation directs attention to the part of the chart that matters, without putting words in the data's mouth. A title that states the finding, a subtitle that adds context, a callout at the relevant point, and a source note that lets the reader check the claim all work together to guide reading responsibly.

plt.plot(trend["period"], trend["value"])
plt.annotate("Efficiency program launched", xy=("2024", trend.loc[trend["period"] == "2024", "value"].iloc[0]),
             xytext=(10, 10), textcoords="offset points", arrowprops={"arrowstyle": "->"})
plt.title("Entity E01: emissions intensity fell 19% since the 2024 program")
plt.figtext(0.01, -0.05, "Source: DS103_Indicators_Clean.csv, synthetic teaching data", fontsize=8)

Key takeaway: Titles, subtitles, callouts, and source notes should guide the reader to the real finding without ever claiming more than the chart shows.

35 min

Dashboard Thinking: Overview → Detail

Learning objective: design a small information hierarchy; avoid KPI overload; support drill-down questions.

A dashboard with twenty KPIs on one screen usually communicates less than one with four, because the reader cannot tell which numbers matter. Good dashboard thinking picks a small overview, then gives the reader an explicit path to drill into detail when the overview raises a question.

candidate_metrics = df["metric"].unique()
# Overview: pick the few metrics the audience's primary question actually needs
overview_metrics = ["emissions_intensity", "water_intensity", "incident_rate", "energy_use"]
overview = df[(df["metric"].isin(overview_metrics)) & (df["period"] == df["period"].max())]

• After seeing the overview, a user's first reaction is "what does this mean?" rather than "this tells me enough"
• Not one KPI has an explicit target or comparison to judge it against
• The detail layer's breakdowns do not match the questions the overview actually raises

Key takeaway: A small, prioritized overview plus a clear drill-down path communicates more than a wall of equally sized KPIs.

35 min

Misleading Charts and Reviewing AI-Generated Visuals

Learning objective: identify truncated axes, dual-axis abuse, cherry-picking, aggregation masking, and AI chart errors.

Most misleading charts are not outright hallucinations; they use small, common technique choices that quietly favor one conclusion. Recognizing the pattern is the skill that transfers, whether the chart was made by a person or generated by an AI tool.

narrow = trend[trend["period"].isin(["2024", "2025"])]  # the window someone chose to show
full = trend  # the full available history
# Compare: does the narrow window's story survive when you look at the full range?

• Every time AI hands back a finished chart, ask first: does the filtering match the scope I stated?
• Confirm the aggregation (average, sum, median) matches the claimed grain
• Check the axis baseline and scale, not just whether the chart "looks right"

Key takeaway: Truncated axes, dual-axis abuse, cherry-picked windows, and aggregation masking are the same risks whether a person or an AI tool made the chart - always check the filtering, aggregation, and scale before trusting it.

40 min

Independent Lab: Evidence-Based Data Story

Learning objective: independently deliver 3-5 visuals, evidence statements, limitations, and an action-oriented narrative.

• Go to the DS 103 Practice page
• PreviewDownload the practice data (with chart-design challenges)
• Download the Independent Lab starter notebook
• PreviewDownload the data dictionary (Traditional Chinese)
• PreviewDownload the data dictionary (English)

1. 1Choose a question and audience
   State the decision or question your data story addresses, and who will read it.
2. 2Build 3-5 charts
   Use at least two different chart families (for example a comparison and a trend) from the practice dataset.
3. 3Check scale integrity
   Confirm each chart's axis baseline, scale, and sort order are honest.
4. 4Apply accessible encoding
   Make sure no chart relies on color alone, and every chart has a title stating its finding.
5. 5Write one evidence statement per chart
   Separate observation from interpretation, and note anything still unknown.
6. 6Sequence the story
   Order the charts so the narrative moves from the overview question to the supporting detail.
7. 7State one limitation
   Write one sentence about what this data story cannot yet support.
8. 8Log any AI assistance
   If AI helped draft a chart or the narrative, note what it helped with and what you verified yourself.

• 3-5 charts covering at least two chart families
• One evidence statement per chart, separating observation and interpretation
• No chart relying on color alone; every chart has a stated-finding title
• At least one limitation or unresolved question
• A record of AI assistance (if used)
• One stated next action the evidence supports

• Every chart must be able to answer the brief's question on its own, not stay in the deck just because it "looks nice"
• Sequence the story so the reader sees the overview-level finding before the supporting detail
• State limitations specifically, for example the period the data covers or a missing denominator, rather than a vague "the data might be incomplete"

Key takeaway: An Evidence-Based Data Story is not a set of pretty charts; it is a sequence of honestly scaled visuals, each paired with a bounded claim, that together support a stated next action.