Categorical palette

Where a sequential palette spends lightness, a categorical one spends hue. Groups like traffic sources, product lines, or teams have no rank, so the palette’s job is to make each one findable while making none of them look bigger: the hues are pushed as far apart around the circle as they will go, and lightness and colourfulness are held roughly level across the set. Let one swatch get noticeably darker or more saturated than its neighbours and readers start treating that series as the important one, which is a claim the data never made.

The practical ceiling is low. Six to eight is where most people stop being able to match a patch in a legend back to a wedge in the chart, because past that the hues have to sit close enough together that two of them read as the same colour seen twice. When a chart genuinely has twenty categories, the answer is not twenty colours: group the small ones into an “other” bucket, colour only the few series the chart is about, or drop colour entirely and label directly. A palette that has run out of separation is worse than no palette, since it looks like it is encoding something.

Two constraints ride along with every set. Around one in twelve men has a colour vision deficiency, so any pair the palette relies on has to survive red and green collapsing into each other, which is why good sets vary lightness a little even while trying to keep it level. And hue on its own is never enough on its own terms: direct labels, patterns, or marker shapes have to repeat what the colour says, which is the whole content of use of colour. A legend is the weakest form of this, because it makes the reader hold six colour-to-name pairs in their head while looking somewhere else.

The other two data palettes answer questions this one refuses to. A sequential palette encodes an ordered quantity, so darker deliberately means more. A diverging palette adds a meaningful midpoint and lets the two directions mean opposite things. Reach for either of those the moment the groups turn out to have an order after all, because a categorical set applied to ordered data throws that order away.