An exhibit (a chart or a table) is a sentence made of ink instead of words, and it obeys the same laws: it must make one point, state that point in its title, and strip everything that does not serve it. Most analyst exhibits fail because they are built as data containers ("here is everything I have on revenue") rather than arguments ("here is the one thing about revenue you need to believe"). The discipline, drawn from Cole Nussbaumer Knaflic's Storytelling with Data, Edward Tufte's The Visual Display of Quantitative Information, and Gene Zelazny's Say It With Charts, reduces to five rules.
How do you write a chart title that states the message?
Rule 1. One message per exhibit, and put it in the title. Before drawing anything, write the sentence the exhibit must prove. Then make that sentence the title, an action title, exactly as for a slide (see the executive slide deck). A chart titled "Revenue by segment, FY15âFY25" is a data container; the reader must find the point themselves. The same chart titled "Jewellery has gone from 78% to 88% of revenue â the watches business is now a rounding error to the thesis" tells the reader what to see before they look. The title is not a caption naming the axes; it is the "so what" rendered as a headline. If you cannot write the one-sentence message, you do not yet know what the exhibit is for. Do not draw it.
Advertisement
Which chart type fits which comparison?
Rule 2. Choose the right form for the comparison (Zelazny's chooser). The data does not pick the chart; the comparison you want the reader to make picks the chart. Zelazny's mapping, which you should internalize:
| The comparison you're making | The right form |
|---|---|
| A trend over time | Line (many points) or column (few points) |
| Ranking of items at one time | Horizontal bar, sorted by magnitude |
| Part-to-whole at one time | Single stacked bar, or a bar chart of the parts (a pie only for 2â3 slices, ideally never) |
| Correlation between two variables | Scatter |
| A single number or one comparison | Just state the number, big. No chart |
The most common exhibit crime is the mismatch: a pie chart with nine slices (the eye cannot rank angles, so use a sorted bar), a line chart across unordered categories (implies a trend that does not exist, so use a bar), a dual-axis chart juxtaposing two series on two scales (the "crossing lines" are an artifact of the arbitrary scaling and imply a relationship that may be invented, so it is almost always misleading; split into two charts or index both to 100). When in doubt, a sorted horizontal bar chart is the most-underused, hardest-to-misread exhibit in analysis.
How do you remove chartjunk and guide the eye?
Rule 3. Strip the chartjunk (Tufte's data-ink ratio). Every drop of ink that is not data is a tax on the reader's attention. Delete: heavy gridlines (light gray at most, or none), chart borders, background fills and gradients, 3-D effects (they distort the very quantities they depict; a 3-D pie is a lie factor generator), redundant legends when you can label directly, and decorative icons. Tufte's principle: maximize the share of ink that carries information, and erase non-data ink and redundant data-ink without mercy. A clean chart looks almost empty; that emptiness is the point â it is all signal.
Rule 4. Label directly and guide the eye. A legend the reader must decode ("which line is blue again?") is friction; label each line or bar at the series itself. Then use pre-attentive attributes (color, bold, size) to point at the one thing that matters and gray out everything else. If the message is "our company's margin diverged from peers in FY23," draw your company's line in a single strong color and every peer in light gray; the reader's eye goes to the message in milliseconds, before conscious reading. Color used to decorate (every series a different bright hue) creates a "where do I look?" scramble; color used to direct (one highlight, the rest muted) is doing analytical work. Use color to mean something, never to brighten.
How do you design a table that makes an argument?
Rule 5. Design tables like arguments too. Tables are exhibits, and most are built wrong. Right-align numbers (so magnitudes line up visually), use consistent decimal places, order rows by magnitude or logic (never alphabetically; alphabetical order hides the story), include a summary/total row, and use light emphasis (bold, a single shaded row) to point at the number that matters. A ten-column table with every cell competing for attention is a data dump; a table that sorts the segments by profit contribution and bolds the one that is the whole thesis is an argument. The same "one message" test applies: what is the single thing this table is here to show, and does its layout make that thing jump out?
The two illustrative exhibits worked in Worked example 2 (a bad Titan segment chart redesigned, and a bad IT-services margin table redesigned) show all five rules operating together. The through-line: an exhibit is not where you store your data; it is where you make one argument, and every design choice either serves that argument or is chartjunk to be deleted.