Diverging palette

A diverging palette is two ramps glued together. Each arm runs from a pale neutral out to a saturated dark end, one arm in each of two hues, and the join is placed on the value that separates one kind of thing from the other: profit from loss, warmer than normal from cooler, ahead of target from behind. Colourfulness and darkness both grow with distance from that join, so the eye reads two things at once, which side a value is on and how far out it went.

Everything depends on the midpoint being real. Zero is the obvious one, but an average, a target, a forecast, or last year’s number can all be genuine dividing lines. What breaks the encoding is a midpoint chosen for the picture rather than for the data: slide the neutral away from zero and every value between the old middle and the new one flips its apparent sign, so a month that was above average is painted in the colour that says below. Nothing about the chart announces this, which is why an off-centre diverging scale is one of the quieter ways a chart can mislead. If the domain is lopsided, keep the neutral where it belongs and let one arm be shorter, or clamp both arms to the same distance and say so.

Hue choice carries a load here too. The red and blue pairing is common because it maps onto a cultural warm and cool split, but red and green is the pairing to avoid: it is the exact axis most colour vision deficiency sits on, so the two arms collapse into one and the midpoint becomes invisible. Purple with orange, or teal with brown, survive that and print in greyscale as a light middle with dark ends, which is the shape the palette is supposed to have.

When there is no meaningful middle, this is the wrong tool. Data that only runs one way belongs on a sequential palette, where lightness carries the order and the darkest end is simply the most. Groups with no order at all belong on a categorical palette, where the whole point is that no colour reads as more than another.