Visualising retail numbers without misleading anybody
A chart is an argument. The choices in it — the axis, the aggregation, the comparison, what is left out — determine what a reader concludes, and most of those choices are made without much thought.
In commercial analytics, where the audience is busy and the stakes are real, chart conventions are an honesty question more than an aesthetic one.
What are the recurring problems?
Truncated axes that exaggerate movement, mixed bases on one chart, aggregations that hide variation, and comparisons chosen after the result was seen. Each of those is common and each changes the conclusion.
None of them are usually deliberate. They are defaults, and the defence is a short checklist applied before a chart goes into a deck.
When is a truncated axis acceptable?
When the variation of interest is small relative to the level and the truncation is clearly marked. A share series between narrow bounds is unreadable on a full axis, and truncating it is reasonable if the reader can see that it has been done.
The unacceptable version is a truncated axis on a volume chart where the visual impression of a large change is doing the persuading.
How should uncertainty be shown?
As a band or an interval where an estimate carries one, which for modelled figures is always. A line with no indication of uncertainty implies precision the estimate does not have, and readers reasonably take the implication.
This is where most commercial charting falls down, because uncertainty is visually inconvenient and commercially uncomfortable.
What about aggregation?
Every aggregation hides variation, and the question is whether the hidden variation matters. A national line concealing two regions moving in opposite directions is technically accurate and practically misleading.
Showing the distribution alongside the average — or at least the range — costs little and prevents the most common misreading of an aggregate.
What should be labelled on every chart?
The measure, the basis, the period, the universe and the source. Five items, and a chart missing them cannot be checked by anyone and will be quoted without them.
The basis matters most in retail, because a total-market line and a like-for-like line look identical and mean different things.
How many charts should a finding need?
Usually one. A finding that requires four charts to establish is usually not yet a finding, and the discipline of reducing it to one clarifies the analysis as well as the presentation.
The corollary is that supporting detail belongs in an appendix rather than in the argument. A deck that shows everything establishes nothing.
What about colour and the rest?
Consistency matters more than palette. Using one colour for one brand throughout a document lets a reader follow it without a legend; changing colours between charts makes every chart a fresh puzzle.
Beyond that, the usual accessibility considerations apply, and they are cheap to respect.
Frequently asked questions
Should I ever use a dual axis?
Rarely, because the relationship between two series on separate scales can be made to look like almost anything depending on the scaling. Where two measures must be compared, indexing both to a common base is usually clearer and harder to abuse.
Is a pie chart ever right?
For a small number of parts of a clearly defined whole, it is acceptable and rarely the best option. Shares over time, which is what retail usually wants, are better served by a line or a stacked area where the whole is genuinely the whole.
How do I show a small share change fairly?
With a truncated and clearly marked axis, or by charting the change itself rather than the level. Both are honest; the dishonest version is a truncated axis presented as though it were full.
What about indexed series?
They are useful for comparing growth across series of different sizes, and they hide the levels entirely. Pair an index chart with a levels table so that a reader can see both the relative movement and the absolute size.
Should charts show data labels?
Where the exact value matters and where there are few enough that the chart stays readable. A chart with a label on every point is a table with decoration, and a table is better at being a table.
How do I handle a broken series?
Break the line visually and annotate the reason. Connecting across a methodology change or a data gap asserts a continuity that does not exist.
What is the most common honest mistake?
Comparing periods of different lengths or different completeness. It happens constantly with current-period data and it always flatters or damages in a predictable direction.
Does the audience need the methodology?
They need enough to know what they are looking at, which is the five labels. The full methodology belongs where it can be found rather than on the chart.
What single habit improves charts most?
Writing the one-sentence conclusion first, then building the chart that supports it. Charts built before the argument tend to show what was easy to plot rather than what matters.
### What should a chart-review checklist contain?
Axis start, basis, period comparability, aggregation level, uncertainty shown, and the five labels. Six checks, applied before a chart leaves the analyst, and they catch nearly every honest error.
Apply it to other people's charts as well. A habit of asking what basis a chart uses spreads quickly and improves the standard of everything circulating in a business. Public series published by the Census Bureau are a reasonable model for careful labelling.
Working with retail data? See NRS Insights data solutions or contact NRS Insights.