Data Viz Design

data-viz-design

a good chart answers a question. a bad chart shows data.

start with the question

before picking a chart type, name the question the viewer should be able to answer at a glance:

• composition: what makes up the whole? (pie, stacked bar, treemap)
• comparison: which is bigger? (bar, column)
• trend: how is it changing over time? (line, area)
• distribution: how is it spread? (histogram, box plot, density)
• relationship: how do two variables move together? (scatter, bubble)
• ranking: in what order? (bar, ordered list)
• flow: what becomes what? (sankey, funnel)

if you can't name the question, you don't need a chart. you might need a single number.

defaults that work

bar charts

• horizontal bars when labels are long
• start the axis at zero unless there's a strong reason not to
• order by value, not alphabetically, unless category order has meaning
• one color for one series; add color only when comparing series

line charts

• thin lines, single color when one series; differentiated when multiple
• highlight the latest value or the value being discussed
• show the trend, not the noise (consider smoothing for very dense data)

tables

• numbers right-aligned, words left-aligned
• monospace numbers (tabular figures) so columns align
• color or icons for delta from previous period
• sortable, with a default sort that matches the question

color use

• one accent color for the focused series, neutral grays for context
• diverging palettes (red-to-green or similar) only when zero is meaningful
• sequential palettes (light to dark) for ordinal data
• categorical palettes for unordered categories; cap at 5-7 colors
• always provide a non-color cue (label, icon) so colorblind users aren't excluded

what to avoid

• 3d charts
• pie charts with more than 4-5 slices
• dual y-axes (they invent correlations)
• starting bar axes above zero to exaggerate differences
• chart junk: gridlines that don't help, borders, drop shadows
• legends placed far from the data they label

annotations

a good chart often has one piece of text on it that names the insight:

"users who completed onboarding had 3x higher 30-day retention"

the chart is the proof. the annotation is the point.

dashboards

• one job per dashboard. monitoring is not the same as exploration.
• the most important number, biggest, top-left.
• support context: time range, segment, comparison period.
• empty states for every chart (no data, loading, error).