Every chapter until now has been about producing a result. This one is about delivering it. The uncomfortable truth is that the quality of your statistics sets a ceiling on your impact, but your communication decides how close you get to it. A brilliant model explained badly loses to a mediocre one explained well, because people act on what they understand. The goal of communicating data is not to show what you did; it is to change what your audience thinks or does. That shift from “here is my analysis” to “here is what you should conclude” is the whole art.
This chapter is practical. Two toolkit notebooks give you report-ready chart and table recipes to paste into your own work, and three analysis notebooks each build the visuals and tables behind a real report. Then, downloadable beside each one, is the finished Word document a statistician actually wrote from those visuals, an executive one-pager, an A/B-test readout, and a full analytical report, in plain language for a non-technical reader.
Start with the Audience and the Message
Before a single chart, answer two questions: who is this for, and what is the one thing they should take away. An executive wants the decision and its cost; a fellow analyst wants the method and the caveats; a product manager wants the implication for the roadmap. The same finding is told three different ways, and guessing wrong wastes everyone's time.
The most important habit is to lead with the answer. Journalists call it burying the lede when you don't; the fix is BLUF, Bottom Line Up Front. State the conclusion in the first sentence, then support it with evidence, then keep the detail for those who want to dig. Busy readers get the message immediately; careful readers can still verify it. This is the inverted pyramid, and it is the opposite of how the analysis was actually done (data first, conclusion last).
One more distinction shapes everything: exploratory versus explanatory. While analyzing, you make dozens of rough charts to find what is true, that is exploratory, for you. What you present is explanatory, for the audience: a small number of deliberate visuals, each making one point. The great sin of data presentation is dumping the exploratory mess on the audience and making them do your thinking.
Effective Tables
Tables are underrated. When the audience needs to look up exact values, compare a few precise numbers, or see several metrics at once, a well-made table beats a chart. A chart is for the shape of the data; a table is for the numbers themselves. The failure mode is a raw data dump; the fix is a handful of deliberate choices.
| Do | Why |
|---|---|
| Right-align numbers, fixed decimals | Digits line up by place value so magnitudes are comparable at a glance. |
| Round to the decision's precision | Nobody decides on the seventh decimal. Show $536k, not 535878.94. |
| Sort by what matters | Order rows by the value the reader cares about, not alphabetically, so the ranking is the message. |
| Highlight the key cell | Bold or shade the one number that carries the point; let everything else recede. |
| Add a summary row | A total or average lets the reader see the whole and the parts together. |
| Format units in the cell | Currency, percent, and thousands separators in the number, not buried in a header. |
The Table Toolkit notebook turns each of these into a reusable pandas recipe, formatting numbers, shading the maximum, coloring by sign, adding a total row, and exporting the finished table straight to Excel and to a Word document. That last step, notebook to deliverable, is the whole point.
Designing Effective Graphs
A good chart answers one question and makes its answer obvious. The single most useful decision is matching the chart type to the question: are you comparing categories, showing a trend, revealing a distribution, a relationship, or a composition? Pick the wrong form and even honest data misleads.
The rest is restraint. Edward Tufte's phrase for the clutter, gridlines, 3-D effects, heavy borders, gratuitous color, is chartjunk, and the fix is to maximize the share of ink that actually encodes data. The Chart Toolkit notebook ships five styled, ready-to-paste recipes that already follow these rules: sorted labeled bars, an annotated trend line, before-and-after bars, small multiples, and a demonstration of an honest axis, which is the bridge to the next section.
How to Avoid Misleading Visuals
The same techniques that clarify can deceive, sometimes by accident, sometimes on purpose. Knowing the common traps lets you avoid them in your own work and spot them in everyone else's. The most notorious is the truncated axis: starting a bar chart's scale somewhere above zero so a small difference looks enormous.
The truncated axis has a whole family of relatives. Keep this catalog in mind, as an author and as a reader.
| The trick | Why it deceives, and the fix |
|---|---|
| Truncated axis | A non-zero baseline on bars inflates small differences. Fix: start bar axes at zero. (Line charts of an index may zoom, but say so.) |
| Dual y-axes | Two series on two different scales can be made to cross or diverge at will, implying a relationship that isn't there. Fix: avoid them; use two panels. |
| Cherry-picked range | Showing only the window that supports your story. Fix: show enough history for the trend to be honest. |
| Area and bubble scaling | Doubling a value but quadrupling the circle's area (scaling the radius, not the area) exaggerates. Fix: scale by area. |
| Pie and 3-D abuse | Angles and perspective are hard to compare; 3-D distorts sizes. Fix: a sorted bar almost always reads better. |
| Correlation as causation | A chart of two rising lines implies one caused the other. Fix: say “associated with,” and show a confounder if you can. |
A visual that creates a false impression is a failure even if every number in it is correct. The test is simple: would the takeaway survive an honest redraw? If fixing the axis, widening the range, or switching from a pie to a bar changes the story, the story was in the design, not the data. This standard connects directly to the Data Ethics, Bias & Fairness chapter that closes this part.
Reports and Stakeholder Dashboards
Charts and tables are the sentences; a report or a dashboard is the paragraph. Which one you build depends on whether the audience needs a one-time argument or an ongoing view.
A report is a static, self-contained document that makes a case at a point in time. Its anatomy is reliable: an executive summary (the BLUF), then key findings with their supporting visuals, then method and limitations, then an appendix for the detail. The genre flexes with the audience, and the three notebooks in this chapter build the three most common versions end to end:
| Report genre | Shape, and who it is for |
|---|---|
| Executive one-pager | Recommendation first, four or five KPIs, one or two charts, a next step. For leaders who decide, not analyze. |
| Experiment readout | The decision in bold, then method, a results table, a chart with its uncertainty, and honest caveats. For a team shipping a change. |
| Analytical report | Summary, data and method, full results tables and figures, interpretation, and limitations. For a technical reviewer who will check your work. |
A dashboard, by contrast, is a living view that updates as data arrives, the province of the BI tools from the last part. The design rules are stricter, not looser: lead with a few KPIs, arrange from most important (top-left) to least, let users filter and drill down, and resist the temptation to cram in every metric. A dashboard that shows everything communicates nothing. The same explanatory discipline applies: each panel should answer a question a stakeholder actually asks.
Draw the line in the right place. The evidence, the figures, tables, and numbers, should come from code, reproducibly, so it never drifts from the data and next quarter's version is one rerun away. That is the reproducibility habit applied to the analysis. But the argument, what the evidence means, why it matters, and what to do about it, is written by a person for a specific audience. That is exactly how the three reports in this chapter are made: the notebooks produce the charts and tables; a statistician then pastes those visuals into a Word document and writes the story around them, in plain language and inverted-pyramid order. A report auto-assembled from variables reads like a robot filled in a form; a report a human wrote persuades.
The Toolkit, the Analyses, and the Reports
Two notebooks to copy recipes from, and three that do the analysis and build the visuals behind a real report. All five run on one shared company dataset. Beside each analysis is the finished, downloadable Word document a statistician wrote from those visuals, so you can compare the raw output with the written argument.
Chart Recipes
Five report-ready, styled matplotlib charts to paste into your work: sorted labeled bars, an annotated trend line, before/after bars, small multiples, and an honest-vs-misleading axis.
Table Recipes
Format numbers for the decision, highlight the key cell, add a summary row, wrap it in a reusable function, and export the finished table to Excel and to Word.
Executive One-Pager
The analysis behind the executive one-pager: KPIs, a revenue trend, and a ranked regional chart. The finished report is the human-authored Word doc below.
A/B Test Readout
The analysis behind the A/B decision memo: rates, lift, a two-proportion test and confidence interval, and one honest chart. The written memo is the Word doc below.
Analytical Report
The analysis behind the analytical study: a regression of satisfaction drivers, a coefficient table, and two figures. The full written study is the Word doc below.
These three reports are human-authored write-ups, a statistician explaining each analysis to a non-technical reader in plain language and inverted-pyramid order (recommendation first). Each one leverages the exact charts its analysis notebook produces from the shared company dataset (three sheets: monthly sales, an A/B test, and a satisfaction survey). Open a report beside its notebook to see the difference between raw output and a written argument. Each opens in its own tab, and carries a link to download the Word original. The numbers are locked to the data: 2024 revenue of $2.15M (up 12% across the year), an A/B lift from 8.1% to 9.4% (significant, p = 0.012), and a satisfaction regression where ease of use is the top driver (R² = 0.84).
The five notebooks run on pandas,
matplotlib, and statsmodels, all preinstalled on Colab. The Table Toolkit also shows a
one-cell recipe for dropping a table into Word with python-docx; that package is not preinstalled on
Colab, so it self-installs on first run (or add a cell with !pip install python-docx, install name
python-docx, import name docx). The three downloadable reports were authored separately by a
person, not generated by the notebooks.
Communicating Model Results in ML & AI
Explaining a model to a non-technical stakeholder is its own challenge, because the honest answer always includes uncertainty, and audiences want certainty. The discipline is to translate without lying: give the decision-relevant number, the confidence around it, and the limits, in plain language.
| Practice | What it means for communication |
|---|---|
| Report uncertainty, always | A prediction without a confidence interval or error bar invites false certainty. Show the range, not just the point. |
| Translate the metric | “92% precision” means little to a manager; “of every 100 we flag, 92 really are fraud” means everything. |
| Model cards | A short standardized summary of what a model does, its data, its performance, and its known limitations, for the people who deploy or are affected by it. |
| Explainability for stakeholders | Feature-importance and example-based explanations (from the interpretability chapter) turn a black box into a story a human can trust or challenge. |
| Do not overclaim | State what the model cannot do as clearly as what it can. Overselling accuracy is the fastest way to lose a stakeholder's trust for good. |
The field has formalized this. Model cards (Mitchell and colleagues, 2019) proposed a one-page standard for reporting a model's intended use, performance across groups, and limitations, and they are now common at major AI labs. The motivation is precisely a communication failure: models were deployed with impressive headline accuracy and no honest account of where they broke, or for whom. Communicating the limits of a result, not just its strength, has become part of doing the work responsibly, which is exactly the bridge into the Data Ethics, Bias & Fairness chapter next.
🎓 Key Takeaways
- ✓Communication decides your impact: a correct analysis nobody understands or acts on has no value; the goal is to change what the audience thinks or does.
- ✓Start with the audience and lead with the answer (BLUF, the inverted pyramid), and show explanatory visuals, not your exploratory mess.
- ✓Tables are for exact numbers: format for the decision, round sensibly, sort by what matters, highlight the key cell, and add a summary row.
- ✓Match the chart to the question (comparison, trend, distribution, relationship, composition), then let the title state the takeaway and cut the chartjunk.
- ✓Avoid misleading visuals: zero-baseline bars, no deceptive dual axes or cherry-picked ranges; the takeaway must survive an honest redraw.
- ✓Reports argue once; dashboards show continuously. A report leads with an executive summary; a dashboard leads with a few KPIs and resists clutter.
- ✓Generate deliverables from code so they never drift from the analysis, and when explaining a model, report the uncertainty and the limits, never just the headline number.
Practice Challenges
Five exercises building on the toolkit and report notebooks. The three report notebooks are themselves full worked examples to adapt.
Rewrite the lede
Take a finding stated as “we ran a regression and R-squared was 0.84” and rewrite it BLUF-style for an executive: answer first, then the one supporting number.
Fix a misleading chart
Take the truncated-axis bar chart from the Chart Toolkit and redraw it honestly. Describe how the takeaway changes.
Build a report-ready table
Using the Table Toolkit, format a revenue-by-region table with currency, a shaded maximum, a total row, and export it to Word.
report_table(...) then the Word-export cell.Adapt an analysis, rewrite the lede
Change Analysis 1 to rank the regions by their number of orders instead of revenue, then write the one-sentence bottom line a statistician would put at the top.
Translate a metric
Take the A/B result and write two versions of the headline: one for the growth team, one for the CFO. Note what each audience needs.
Solutions notebook
All five challenges worked through: an example BLUF rewrite, the misleading chart redrawn honestly, a report-ready region table exported to Word, the regions re-ranked by order count with a fresh bottom line, and one A/B result translated for two audiences. The written solution report beside it works all five challenges, with the visuals and headlines they should produce.
Quiz: Test Yourself
Eight questions on audience, tables, charts, misleading visuals, and reports. Answer them, hit Check Answers, and keep refining until you score 100%. Your progress is saved.
Communicating clearly and communicating responsibly are two sides of one coin. Data Ethics, Bias & Fairness closes this part and the applied arc of the book: algorithmic bias, fairness metrics, privacy, and building analysis people can trust. Browse the full Contents for what is published and what is on the way.