Three exercises in seeing what a table hides. Split a demand series into its parts, meet four datasets that share one statistics table, then name the relationship in your own data before you pick a chart.
SKU-4471, 60 monthly observations, 2021-01 to 2025-12. Synthetic series, STL components
from feasts::STL().
STL splits the series into trend, seasonal and remainder. Each panel below is one of those parts, drawn on its own scale.
Draws the planted trend and the counterfactual: what 2025 would have looked like had growth of 12 units a month continued.
STL estimates. It does not reveal ground truth. The largest gap is 22.6 units, and one monthly estimate carries a standard error of 40 / √5 = 17.9 units, so every gap sits inside about 1.3 standard errors.
Four datasets, eleven points each. One view tells you they are the same. The other tells you they are not.
Anscombe, F.J. (1973), “Graphs in Statistical Analysis”,
The American Statistician 27(1), 17–21.
Data is datasets::anscombe, shipped with R.
Mean x and sd x match to the last digit across all four sets. Everything else agrees only to two decimal places: mean y is 7.50091 for sets 1, 2 and 4 but 7.50000 for set 3, and r ranges from 0.81624 to 0.81652.
All four sets share mean x = 9.0, mean y = 7.50, r = 0.816, and the same fitted line y = 3.00 + 0.500x. They are identical to two decimal places on every summary a regression printout gives you.
The table view says nothing has changed. The plot view shows four different worlds: honest noise, a curve that a straight line has no business describing, a clean linear relationship wrecked by one outlier, and a set where a single point at x = 19 is doing all the work. Only the first one is a case where reporting r = 0.816 means what you think it means.
The working habit: plot the data before you quote a statistic about it. Every one of these four failures is invisible in the table and unmissable in the chart. Anscombe made this argument in 1973 and it has not aged.
Name the relationship first. The chart type follows from it. Nine categories, in the order the FT poster lists them.
Full poster and chart list: Financial Times Visual Vocabulary on GitHub. Category descriptions on this page are quoted from that README. The supply chain examples are ours, not the FT's.
Most bad charts are not badly drawn. They answer a different question than the one the reader has. A stacked column chart is excellent for part-to-whole and poor for ranking, and no amount of styling fixes that mismatch.
Working order: write the sentence you want the reader to leave with, name the relationship in that sentence, then pick from the shortlist. The chart choice is the last decision, not the first.
Two categories are worth a warning. Correlation charts invite a causal reading you did not intend, so say what you mean in the subtitle. Spatial is for when precise location matters more than anything else. A map of demand by region usually loses to a ranked bar chart, because the reader wants to know which region is worst, not where it sits.